KR102623390B1 - Method, apparatus and program for maintaining accuracy of equipment anomaly detection model - Google Patents

Abstract

A method for maintaining the accuracy of an equipment abnormality detection model according to various embodiments of the present invention is disclosed. The method includes: monitoring sensor data of each of one or more devices using a learned anomaly detection model to detect anomalies in the devices; When the abnormality detection model detects abnormal data in specific sensor data obtained from specific equipment, obtaining preventive maintenance information related to the specific equipment and performing a consistency check on the specific equipment; Based on the consistency check, recognizing whether the abnormal data detected by the abnormality detection model was detected due to preventive maintenance or due to a defect in the specific equipment; And when the abnormal data detected by the abnormality detection model is recognized as being detected by preventive maintenance, updating the abnormality detection model based on the abnormal data.

Description

Method, apparatus and program for maintaining the accuracy of equipment anomaly detection model {METHOD, APPARATUS AND PROGRAM FOR MAINTAINING ACCURACY OF EQUIPMENT ANOMALY DETECTION MODEL}

The present invention relates to a method, device, and program for maintaining the accuracy of an equipment anomaly detection model, and specifically relates to a method, device, and program for maintaining the accuracy of the detection model through the update and initialization process of the anomaly detection model. .

In general, factories that have established a mass production system have the manufacturing process divided into several stages in order to increase the efficiency of mass production of products, and automation equipment suitable for each process is operated and operated for each of the various stages of the division.

In the case of automated equipment, in the process of repeatedly performing the same process, situations may arise where the process is performed abnormally due to errors in the equipment itself or the influence of the surrounding environment. In a manufacturing process system where each process is continuous and organically linked, if a problem occurs in the equipment of some process, it may result in defects in the entire process and product. Accordingly, periodic detection of abnormalities in factory equipment operated for each process is very important in terms of maintenance of the factory operation system.

Detection of abnormalities in factory equipment can be accomplished by skilled workers frequently checking the operation status of the equipment through various sensor process data. However, no matter how skilled a worker is, there are limits to accurately detecting abnormal data (or abnormal data) by checking all the vast amounts of sensor process data flowing in in real time, and the abnormal data detection process can take a lot of time. In addition, as factory equipment becomes more complex, such as the introduction of FA systems, the knowledge and know-how required for workers increases, and it may be difficult for inexperienced workers to identify the factors that led to the abnormal state.

Meanwhile, as sensor process data that can be used temporarily or permanently by being stored in a database is accumulated, research is being conducted on automated processing of monitoring data from industrial equipment related to various fields. In particular, as the amount of information that can be processed increases with the development of computer technology, artificial intelligence is evolving at a rapid pace, and research and development is underway on technologies to detect abnormalities in process data using artificial intelligence.

Republic of Korea Patent Publication No. 10-2021-0128713 discloses a method of analyzing and predicting process quality by analyzing product process data collected in real time using deep learning.

The present invention was conceived in response to the above-described background technology and is intended to provide a method, device, and program for maintaining the accuracy of an equipment abnormality detection model.

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

According to an embodiment of the present invention for solving the problems described above, a method for maintaining the accuracy of an equipment abnormality detection model is disclosed. The method includes: monitoring sensor data of each of one or more devices using a learned anomaly detection model to detect anomalies in the devices; When the abnormality detection model detects abnormal data in specific sensor data obtained from specific equipment, obtaining preventive maintenance information related to the specific equipment and performing a consistency check on the specific equipment; Based on the consistency check, recognizing whether the abnormal data detected by the abnormality detection model was detected due to preventive maintenance or due to a defect in the specific equipment; And when the abnormal data detected by the abnormality detection model is recognized as being detected by preventive maintenance, updating the abnormality detection model based on the abnormal data.

In an alternative embodiment, when the anomaly detection model detects abnormal data in specific sensor data obtained from specific equipment, obtaining preventive maintenance information related to the specific equipment and performing a consistency check for the specific equipment. Recognizing whether preventive maintenance has been performed on the specific equipment based on the preventive maintenance information; and performing the consistency test when it is recognized that preventive maintenance has been performed on the specific equipment.

In an alternative embodiment, performing the consistency check may include acquiring sensor data for each of a specific device on which preventive maintenance has been performed and one or more devices operating with the same recipe as the specific device; sampling the sensor data and performing alignment on the sensor data acquired from each device; determining minimum, maximum, and median values for each device from the sampled sensor data; measuring similarity of sensor data corresponding to each device based on at least one of the minimum value, the maximum value, and the intermediate value for each device; and determining, based on the similarity, whether the specific equipment on which the preventive maintenance has been performed is normal.

In an alternative embodiment, based on the consistency check, the step of recognizing whether the abnormal data detected by the abnormality detection model is detected by preventive maintenance or is detected due to a defect in the specific equipment includes: When it is determined that the specific equipment is normal by performing the consistency check, and the abnormal data detected by the anomaly detection model is recognized as having been detected by preventive maintenance, and when it is determined that the specific equipment is not normal by performing the consistency check. It may include recognizing that abnormal data detected by the anomaly detection model was detected due to a defect in specific equipment.

In an alternative embodiment, updating the anomaly detection model includes: scaling the anomaly data to obtain normalized data; It may include the step of transfer learning a fully connected layer using the normalized data while maintaining the feature extraction layer of the anomaly detection model.

In an alternative embodiment, the step of transfer learning a fully connected layer using the normalized data while maintaining the feature extraction layer of the anomaly detection model includes detecting the anomaly. Comparing a first maximum value for detecting anomalies in a model and a second maximum value of the normalized data; and when the second maximum value is greater than the first maximum value, changing the first maximum value for detecting an anomaly in the anomaly detection model to the second maximum value.

In an alternative embodiment, the updated anomaly detection model is updated to increase the range of data for detecting anomaly, and initial training is performed on the anomaly detection model at preset intervals to maintain the accuracy of the anomaly detection model. It can be initialized to the model.

In an alternative embodiment, monitoring sensor data of each of one or more pieces of equipment using a learned anomaly detection model to detect anomalies in the equipment includes obtaining process sensor data and prior knowledge data corresponding to the process sensor data. steps; Generating reconstruction process sensor data corresponding to the process sensor data using a deep learning model; calculating a reconstruction rate error based on the process sensor data and the reconstruction process sensor data; and detecting abnormal operation based on comparison of the reconstruction rate error and a reference threshold, wherein the deep learning model includes: a first sub-model for extracting feature information corresponding to the process sensor data; a second sub-model that extracts interaction relationship information between each process sensor data based on the process sensor data and the prior knowledge data; an attention module that generates feature information by combining outputs of the first sub-model and the second sub-model; and a dimensional reconstruction model that restores the feature information to generate the reconstruction process sensor data.

According to one embodiment of the present invention for solving the above-described problems, a device is disclosed. The device includes: a memory storing one or more instructions; and a processor executing the one or more instructions stored in the memory, and the processor may perform the above-described methods by executing the one or more instructions.

According to an embodiment of the present invention for solving the above-described problem, a computer program is disclosed that is combined with a computer as hardware and stored in a computer-readable recording medium to perform the above-described methods.

Other specific details of the invention are included in the detailed description and drawings.

In the present invention, even if an abnormality is determined for the sensor data measured from the equipment, it is possible to determine whether the abnormality is due to preventive maintenance or whether an abnormality has occurred in the equipment through a consistency check with sensor data measured from equipment operating with the same recipe. You can judge.

In addition, the present invention can reduce the time and cost required for data collection and model learning by transfer learning the anomaly detection model that detects anomalies in equipment using data determined to be abnormal, and further maintains the accuracy of the anomaly detection model. You can do it.

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

1 is a diagram illustrating a system according to an embodiment of the present invention.
Figure 2 is a hardware configuration diagram of a computing device according to an embodiment of the present invention.
Figure 3 is a flowchart illustrating an example of a method for checking consistency between a plurality of manufacturing process equipment based on sensor data according to an embodiment of the present invention.
Figure 4 is a flowchart illustrating an example of a method for measuring similarity according to an embodiment of the present invention.
Figure 5 is a flowchart illustrating an example of a method for maintaining the accuracy of an equipment abnormality detection model according to an embodiment of the present invention.
Figure 6 is a flowchart illustrating an example of a method for determining whether to perform a consistency check for specific equipment according to an embodiment of the present invention.
7 to 9 are diagrams for explaining an example of a method for performing alignment on sensor data according to an embodiment of the present invention.
Figure 10 is a schematic diagram showing one or more network functions related to one embodiment of the present invention.

Various embodiments are now described with reference to the drawings. In this specification, various descriptions are presented to provide an understanding of the invention. However, it is clear that these embodiments may be practiced without these specific descriptions.

As used herein, the terms “component,” “module,” “system,” and the like refer to a computer-related entity, hardware, firmware, software, a combination of software and hardware, or an implementation of software. For example, a component may be, but is not limited to, a process running on a processor, a processor, an object, a thread of execution, a program, and/or a computer. For example, both an application running on a computing device and the computing device can be a component. One or more components may reside within a processor and/or thread of execution. A component may be localized within one computer. A component may be distributed between two or more computers. Additionally, these components can execute from various computer-readable media having various data structures stored thereon. Components can transmit signals, for example, with one or more data packets (e.g., data and/or signals from one component interacting with other components in a local system, a distributed system, to other systems and over a network such as the Internet). Depending on the data being transmitted, they may communicate through local and/or remote processes.

Additionally, the term “or” is intended to mean an inclusive “or” and not an exclusive “or.” That is, unless otherwise specified or clear from context, “X utilizes A or B” is intended to mean one of the natural implicit substitutions. That is, either X uses A; X uses B; Or, if X uses both A and B, “X uses A or B” can apply to either of these cases. Additionally, the term “and/or” as used herein should be understood to refer to and include all possible combinations of one or more of the related listed items.

Additionally, the terms “comprise” and/or “comprising” should be understood to mean that the corresponding feature and/or element is present. However, the terms “comprise” and/or “comprising” should be understood as not excluding the presence or addition of one or more other features, elements and/or groups thereof. Additionally, unless otherwise specified or the context is clear to indicate a singular form, the singular terms herein and in the claims should generally be construed to mean “one or more.”

Those skilled in the art will additionally recognize that the various illustrative logical blocks, components, modules, circuits, means, logic, and algorithm steps described in connection with the embodiments disclosed herein may be implemented using electronic hardware, computer software, or a combination of both. It must be recognized that it can be implemented with To clearly illustrate the interchangeability of hardware and software, various illustrative components, blocks, configurations, means, logics, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented in hardware or software will depend on the specific application and design constraints imposed on the overall system. A skilled technician can implement the described functionality in a variety of ways for each specific application. However, such implementation decisions should not be construed as departing from the scope of the present invention.

The description of the presented embodiments is provided to enable anyone skilled in the art to use or practice the present invention. Various modifications to these embodiments will be apparent to those skilled in the art. The general principles defined herein may be applied to other embodiments without departing from the scope of the invention. Therefore, the present invention is not limited to the embodiments presented herein. The present invention is to be interpreted in the broadest scope consistent with the principles and novel features presented herein.

In this specification, a computer refers to all types of hardware devices including at least one processor, and depending on the embodiment, it may be understood as encompassing software configurations that operate on the hardware device. For example, a computer can be understood to include, but is not limited to, a smartphone, tablet PC, desktop, laptop, and user clients and applications running on each device.

Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

Each step described in this specification is described as being performed by a computer, but the subject of each step is not limited thereto, and depending on the embodiment, at least part of each step may be performed in a different device.

1 is a diagram illustrating a system according to an embodiment of the present invention.

Referring to FIG. 1, a system according to an embodiment of the present invention may include a computing device 100, a user terminal 200, and an external server 300. The system shown in FIG. 1 is according to one embodiment, and its components are not limited to the embodiment shown in FIG. 1, and may be added, changed, or deleted as necessary.

In one embodiment, when measuring manufacturing process sensor data using artificial intelligence technology, if it exceeds the standard value, it is immediately determined as defective. However, when repairing or replacing equipment parts, there are cases where the sensor data is significantly different from before and sensor data that should be judged normal is judged to be defective. When this situation occurs, it is impossible for an engineer to compare sensor data determined as defective with other equipment or chambers. Therefore, when an abnormality occurs, the sensor data between devices (or chambers) is automatically compared to determine good or bad. There was a need to do it.

The computing device 100 according to an embodiment of the present invention may perform a consistency check between a plurality of manufacturing process equipment based on sensor data.

Specifically, the computing device 100 may acquire sensor data for each of the specific equipment on which preventive maintenance has been performed and one or more equipment operated with the same recipe as the specific equipment. Here, preventive maintenance may include, but is not limited to, at least one of replacement of parts constituting the equipment, repair of parts constituting the equipment, and repair of the equipment itself.

When sensor data is acquired, the computing device 100 may sample the sensor data and perform alignment on the sensor data acquired from each device. Here, alignment is a process of processing sensor data and may include a process of adjusting each sensor data acquired from one or more pieces of equipment to the same standard (eg, time, length, etc.).

After performing early-in, the computing device 100 may determine the minimum, maximum, and median values for each device from the sampled sensor data. Additionally, the computing device 100 may measure the similarity of sensor data corresponding to each device based on at least one of the minimum value, maximum value, and median value for each device. Additionally, the computing device 100 may determine whether the specific equipment on which preventive maintenance has been performed is normal based on the similarity.

Therefore, even if an abnormality is determined for sensor data, the computing device 100 of the present invention determines whether the abnormality is due to preventive maintenance or an abnormality in the equipment through a consistency check with sensor data measured from equipment operating with the same recipe. By determining whether this has occurred, you can increase the convenience of equipment management.

Hereinafter, a detailed description of how the computing device 100 performs a consistency check between a plurality of manufacturing process equipment based on sensor data will be described with reference to FIGS. 3 and 4 .

In one embodiment, an anomaly detection solution utilizing artificial intelligence technology in the manufacturing process determined an anomaly when data different from learned data was collected. However, since abnormal detection that occurs due to maintenance for equipment operation, such as repair and replacement of equipment parts, is not an actual defect, there is a problem that model accuracy decreases if new data is not reflected and learned. To solve this problem, it is necessary to determine whether the prediction result of the artificial intelligence model is actually defective or not, and if it is not, a process of learning by reflecting new data is necessary. However, the field engineer must repeat the series of processes for each equipment and process. It is difficult to introduce it in the field due to the time it takes to collect a sufficient amount of data needed for learning.

According to one embodiment of the present invention, the computing device 100 can maintain the accuracy of the equipment abnormality detection model.

Specifically, the computing device 100 may monitor sensor data of each of one or more devices using a learned anomaly detection model to detect anomalies in the equipment. Additionally, when the anomaly detection model detects abnormal data in specific sensor data obtained from specific equipment, the computing device 100 may obtain preventive maintenance information related to the specific equipment and perform a consistency check on the specific equipment. .

Based on the consistency check, the computing device 100 may recognize whether the abnormal data detected by the abnormality detection model was detected by preventive maintenance or was detected due to a defect in specific equipment. Additionally, when the computing device 100 recognizes that the abnormal data detected by the abnormality detection model is detected through preventive maintenance, the computing device 100 may update the abnormality detection model based on the abnormal data. Here, the update of the anomaly detection model may be processed using a transfer learning method, but is not limited to this.

Therefore, the computing device 100 of the present invention can reduce the time and cost required for data collection and model learning through transfer learning, and can maintain the accuracy of the anomaly detection model through continuous updates.

In addition, the present invention can minimize unnecessary intervention of field engineers in determining the predicted results of an abnormality detection model by checking the consistency of sensor data between preventive maintenance information (e.g., part replacement) and equipment.

In addition, compared to learning an initial model, transfer learning has the advantage of reducing the time and cost required for data collection and model learning, so it can automatically maintain the accuracy of the anomaly detection model without additional work by engineers, making it easy to introduce in the field. You can expect it.

Hereinafter, a detailed description of how the computing device 100 maintains the accuracy of the equipment abnormality detection model will be described with reference to FIGS. 5 and 6.

In various embodiments, the computing device 100 may provide web- or application-based services. However, it is not limited to this.

Computing device 100 may include any type of computer system or computer device, such as, for example, microprocessors, mainframe computers, digital processors, portable devices, and device controllers. However, it is not limited to this.

Hereinafter, the hardware configuration of the computing device 100 will be described with reference to FIG. 2 .

Meanwhile, the user terminal 200 may be connected to the computing device 100 through the network 400, and may check consistency between a plurality of manufacturing process equipment based on sensor data performed by the computing device 100 and manage the manufacturing process equipment. It may be the administrator's terminal. Additionally, the user terminal 200 may be an administrator's terminal that manages processes performed on the computing device 100 to maintain the accuracy of the equipment abnormality detection model.

Here, the user terminal 200 may include, for example, various types of computer devices. For example, the user terminal 200 may refer to various terminal devices such as a smartphone, tablet PC, desktop, or laptop.

The user terminal 200 includes a display in at least a portion of the terminal, and may include an operating system for running an application or extension program-based service provided by the computing device 100. For example, the user terminal 200 may be a smart phone, but is not limited to this. The user terminal 200 is a wireless communication device that guarantees portability and mobility, and may be used for navigation, personal communication (PCS), etc. System), GSM (Global System for Mobile communications), PDC (Personal Digital Cellular), PHS (Personal Handyphone System), PDA (Personal Digital Assistant), IMT (International Mobile Telecommunication)-2000, CDMA (Code Division Multiple Access)- 2000, all types of handheld-based wireless communication devices such as W-CDMA (W-Code Division Multiple Access), Wibro (Wireless Broadband Internet) terminals, smartpads, tablet PCs, etc. It can be included.

The external server 300 may be connected to the computing device 100 through the network 400, and the computing device 100 maintains the accuracy of the consistency inspection method and equipment abnormality detection model between a plurality of manufacturing process equipment based on sensor data. It is possible to transmit and receive various information/data necessary to perform each of the methods, and the computing device 100 provides a method for checking consistency between a plurality of manufacturing process equipment based on sensor data and a method for maintaining the accuracy of the equipment abnormality detection model. Various information/data generated as each operation is performed can be stored and managed.

For example, the external server 300 may be a database server that stores information used in a method for checking consistency between a plurality of manufacturing process equipment based on sensor data and a method for maintaining the accuracy of an equipment abnormality detection model. For another example, the external server 300 may be a server that provides information used in a method for checking consistency between a plurality of manufacturing process equipment based on sensor data and a method for maintaining the accuracy of an equipment abnormality detection model.

The network 400 may refer to a connection structure that allows information exchange between nodes such as a computing device, a plurality of terminals, and servers. For example, the network 400 includes a local area network (LAN), a wide area network (WAN), the World Wide Web (WWW), a wired and wireless data communication network, a telephone network, and a wired and wireless television communication network. do.

Wireless data communication networks include 3G, 4G, 5G, 3GPP (3rd Generation Partnership Project), 5GPP (5th Generation Partnership Project), LTE (Long Term Evolution), WIMAX (World Interoperability for Microwave Access), Wi-Fi, and Internet. (Internet), LAN (Local Area Network), Wireless LAN (Wireless Local Area Network), WAN (Wide Area Network), PAN (Personal Area Network), RF (Radio Frequency), Bluetooth (Bluetooth) network, NFC (Near- Field Communication) network, satellite broadcasting network, analog broadcasting network, DMB (Digital Multimedia Broadcasting) network, etc., but is not limited thereto.

Figure 2 is a hardware configuration diagram of a computing device according to an embodiment of the present invention.

Referring to FIG. 2, the computing device 100 according to an embodiment of the present invention includes one or more processors 110 and a memory 120 that loads a computer program 151 executed by the processor 110. , it may include a bus 130, a communication interface 140, and a storage 150 that stores a computer program 151. Here, only components related to the embodiment of the present invention are shown in Figure 2. Accordingly, anyone skilled in the art to which the present invention pertains will know that other general-purpose components may be included in addition to the components shown in FIG. 2.

The processor 110 controls the overall operation of each component of the computing device 100. The processor 110 may be composed of one or more cores, and may include a central processing unit (CPU), a general purpose graphics processing unit (GPGPU), and a tensor processing unit (TPU) of the computing device. unit) may include a processor for data analysis and deep learning. Alternatively, it may be configured to include any type of processor well known in the art of the present invention.

Additionally, the processor 110 may perform operations on at least one application or program for executing methods according to embodiments of the present invention, and the computing device 100 may include one or more processors.

In various embodiments, the processor 110 includes random access memory (RAM) (not shown) and read memory (ROM) that temporarily and/or permanently store signals (or data) processed within the processor 110. -Only Memory, not shown) may be further included. Additionally, the processor 110 may be implemented in the form of a system on chip (SoC) that includes at least one of a graphics processing unit, RAM, and ROM.

Memory 120 stores various data, commands and/or information. Memory 120 may load a computer program 151 from storage 150 to execute methods/operations according to various embodiments of the present invention. When the computer program 151 is loaded into the memory 120, the processor 110 can perform the method/operation by executing one or more instructions constituting the computer program 151. The memory 120 may be implemented as a volatile memory such as RAM, but the technical scope of the present invention is not limited thereto.

Bus 130 provides communication functionality between components of computing device 100. The bus 130 may be implemented as various types of buses, such as an address bus, a data bus, and a control bus.

The communication interface 140 supports wired and wireless Internet communication of the computing device 100. Additionally, the communication interface 140 may support various communication methods other than Internet communication. To this end, the communication interface 140 may be configured to include a communication module well known in the technical field of the present invention. In some embodiments, communication interface 140 may be omitted.

Storage 150 may store the computer program 151 non-temporarily. When performing a process according to an embodiment of the present invention through the computing device 100, the storage 150 may store various information necessary to perform analysis according to the disclosed embodiment.

The storage 150 is a non-volatile memory such as Read Only Memory (ROM), Erasable Programmable ROM (EPROM), Electrically Erasable Programmable ROM (EEPROM), flash memory, a hard disk, a removable disk, or a device well known in the technical field to which the present invention pertains. It may be configured to include any known type of computer-readable recording medium.

The computer program 151, when loaded into the memory 120, may include one or more instructions that cause the processor 110 to perform methods/operations according to various embodiments of the present invention. That is, the processor 110 can perform the method/operation according to various embodiments of the present invention by executing the one or more instructions.

In one embodiment, computer program 151 may include one or more instructions to perform various methods related to various tasks related to training a neural network model.

The steps of the method or algorithm described in connection with embodiments of the present invention may be implemented directly in hardware, implemented as a software module executed by hardware, or a combination thereof. The software module may be RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), Flash Memory, hard disk, removable disk, CD-ROM, or It may reside on any type of computer-readable recording medium well known in the art to which the present invention pertains.

The components of the present invention may be implemented as a program (or application) and stored in a medium in order to be executed in conjunction with a hardware computer. Components of the invention may be implemented as software programming or software elements, and similarly, embodiments may include various algorithms implemented as combinations of data structures, processes, routines or other programming constructs, such as C, C++, , may be implemented in a programming or scripting language such as Java, assembler, etc. Functional aspects may be implemented as algorithms running on one or more processors.

Figure 3 is a flowchart illustrating an example of a method for checking consistency between a plurality of manufacturing process equipment based on sensor data according to an embodiment of the present invention. Figure 4 is a flowchart illustrating an example of a method for measuring similarity according to an embodiment of the present invention.

According to an embodiment of the present invention, the computing device 100 may perform a consistency check between a plurality of manufacturing process equipment based on sensor data.

Referring to FIG. 3, the computing device 100 may acquire sensor data for each of the specific equipment on which preventive maintenance has been performed and one or more equipment operated with the same recipe as the specific equipment (S110). Here, a recipe may refer to instructions including work steps and conditions necessary to perform a manufacturing process.

Specifically, the computing device 100 may acquire first sensor data measured from specific equipment on which preventive maintenance has been performed. Additionally, the computing device 100 may acquire second sensor data measured from specific equipment before preventive maintenance is performed. Additionally, the computing device 100 may acquire third sensor data measured from each of one or more devices that operate with the same recipe as the specific device.

In one embodiment, each of the first sensor data, second sensor data, and third sensor data may be data measured by each equipment while producing a preset number of wafers using a recipe for producing wafers.

For example, each of the first sensor data and the third sensor data may be data measured while producing the last 20 wafers from the currently produced wafer. Additionally, the second sensor data may be data measured during production of the last 20 wafers starting from the wafer produced before preventive maintenance was performed. However, it is not limited to this.

The computing device 100 may sample sensor data and perform alignment on sensor data acquired from each device (S120).

Specifically, the computing device 100 may extract data from sensor data at preset time intervals to correct sampling errors of sensor data. Additionally, the computing device 100 may guess missing sensor data that has not been sampled and fill in the data values of the missing portion. Additionally, the computing device 100 may calculate the average processing time of sensor data. Additionally, the computing device 100 may change the sensor data to correspond to the average process time.

That is, in the present invention, alignment performed on sensor data may be a type of data preprocessing. This will be explained later with reference to FIGS. 7 to 9.

The computing device 100 may determine the minimum, maximum, and median values for each device from the sampled sensor data (S130).

Specifically, the computing device 100 may extract the first multiplication value obtained by multiplying the sensor data value for each device and the first value as the minimum value for each device. Additionally, the computing device 100 may extract a second multiplication value obtained by multiplying the sensor data value for each device and a second value greater than the first value as the maximum value for each device.

For example, the computing device 100 may extract a value corresponding to 20% of the sensor data values for each device as the minimum value, and extract a value corresponding to 80% of the sensor data values as the maximum value. However, it is not limited to this.

In various embodiments, the computing device 100 deletes values corresponding to the top 20% of sensor data values for each device, deletes values corresponding to the bottom 20%, and then selects the minimum value for each device from the sampled sensor data, Maximum and intermediate values can be determined.

That is, the computing device 100 may perform scaling to prevent errors in the minimum and maximum values due to unstable values.

When the minimum value for each device and the maximum value for each device are extracted, the computing device 100 may calculate the intermediate value for each device based on the minimum and maximum values.

For example, the computing device 100 may calculate the intermediate value by dividing the sum of the minimum and maximum values for each device by 2. That is, the middle value may mean the average value of the minimum and maximum values, but is not limited thereto.

When the minimum, maximum, and intermediate values for each device are determined, the computing device 100 may measure the similarity of sensor data corresponding to each device based on at least one of the minimum, maximum, and intermediate values (S140 ).

Specifically, referring to FIG. 4, the computing device 100 may determine the intermediate value determined from the third sensor data as the reference value (S141).

Then, when the computing device 100 determines the reference value, the computing device 100 may obtain a first similarity value by measuring the similarity between the first intermediate value determined from the first sensor data and the reference value (S142). Additionally, the computing device 100 may obtain a second similarity value by measuring the similarity between the second intermediate value determined from the second sensor data and the reference value (S143). Additionally, the computing device 100 may obtain a third similarity value by measuring the similarity between the minimum value extracted from the first sensor data and the reference value (S144). Additionally, the computing device 100 may obtain a fourth similarity value by measuring the similarity between the maximum value extracted from the first sensor data and the reference value (S145).

In other words, the first similarity value is the 'median value (reference value)' of the third sensor data measured in each of one or more devices operating with the same recipe as the specific equipment and the first sensor data measured in the specific equipment on which preventive maintenance was performed. It may be a similarity value between 'first intermediate values'.

In addition, the second similarity value is the 'median value (reference value)' of the third sensor data measured from each of one or more devices operating with the same recipe as the specific equipment and the second sensor data obtained before preventive maintenance on the specific equipment. It may be a similarity value between 'second intermediate values'.

In addition, the third similarity value is the 'median value (reference value)' of the third sensor data measured from each of one or more devices operating with the same recipe as the specific equipment and the first sensor data obtained after preventive maintenance on the specific equipment. It may be a similarity value between 'minimum values'.

In addition, the fourth similarity value is the 'median value (reference value)' of the third sensor data measured from each of one or more devices operating with the same recipe as the specific equipment and the first sensor data obtained after preventive maintenance on the specific equipment. It may be a similarity value between 'maximum values'.

In the present invention, the similarity measured between sensor data may mean the similarity between each amplitude change type (eg, each wavelength). Additionally, the similarity measured between sensor data can be determined through a dynamic time warping (DTW) algorithm. Dynamic time warping may be an algorithm that measures the similarity between two time series data with different movements depending on speed or length. Dynamic time warping may be characterized by finding a warping path that minimizes the cumulative distance by matching in the direction where the distance is minimized.

In general, when determining the similarity between two time series data, the Euclidean distance algorithm can be used. The Euclidean distance algorithm can determine the similarity between both data by calculating the distance on the same time line. For example, when using the Euclidean distance, similarity is determined by matching each time point, such as 0 second to 0 second, 1 second to 1 second, etc. However, this Euclidean distance has the disadvantage that it is difficult to find similarity in data that appears to have similar overall patterns as tremor and movement become more severe, and that it is difficult to evaluate similarity between short-length time series data. The sensor process data used in the present invention may each have different operating times. In other words, in the case of the Euclidean distance algorithm, the accuracy of calculating similarity can be very low in situations where the time axis is distorted. In other words, if the alignment between the two data is not correct, it may be difficult to calculate the distance and thus difficult to evaluate the similarity.

In various embodiments, when the computing device 100 measures similarity between sensor data, it may vectorize both data using cosine similarity comparison, calculate a distance through the angle between vectors, and evaluate similarity based on this. However, cosine similarity comparison may not reflect the sequence of time series data.

Accordingly, the present invention can measure the similarity between at least two pieces of data by utilizing a dynamic time warping algorithm. The dynamic time warping algorithm may be an algorithm that evaluates the similarity of the wavelengths of the time axis at two different speeds. The dynamic time warping algorithm can match not only data on the same timeline but also surrounding time points to more similar elements. Using this dynamic time warping algorithm has the advantage of enabling similarity evaluation between time series data of different lengths.

In various embodiments, when the computing device 100 obtains the first similarity value, the second similarity value, the third similarity value, and the fourth similarity value, the first similarity value, the second similarity value, the third similarity value, and A plurality of normalization values can be calculated by dividing the fourth similarity value by the largest one (S146).

Specifically, the computing device 100 calculates a first normalization value by dividing the first similarity value by the largest value, and divides the second similarity value by the largest value to calculate the second normalization value. , the third normalization value can be calculated by dividing the third similarity value by the largest value, and the fourth normalization value can be calculated by dividing the fourth similarity value by the largest value.

The plurality of normalization values of the present invention can be used by the computing device 100 to determine whether a specific device is normal in step S150, which will be described below.

Referring again to FIG. 3, the computing device 100 may determine whether the specific equipment on which preventive maintenance has been performed is normal based on the similarity (S150).

Specifically, the computing device 100 may determine whether the specific equipment on which preventive maintenance has been performed is normal based on a plurality of normalization values calculated based on the similarity in step S146.

If the first normalization value is not 1, the computing device 100 may determine that the specific equipment on which preventive maintenance has been performed is normal. And, if the first normalization value is 1, the computing device 100 may determine that a defect will occur in a specific piece of equipment on which preventive maintenance has been performed.

That is, if the first similarity value is the largest among the first similarity value, the second similarity value, the third similarity value, and the fourth similarity value, the first similarity value is divided by the first similarity value so that the first normalization value is 1. You can. In this case, the computing device 100 may determine that a defect will occur in a specific piece of equipment on which preventive maintenance has been performed.

For example, the 'second intermediate value' of the second sensor data obtained before preventive maintenance on specific equipment and the 'intermediate value (reference) of the third sensor data measured on each of one or more devices operating with the same recipe as the specific equipment. The second similarity value between 'values' may be the highest because it is a similarity value between sensor data in a state in which a specific rose and one or more devices are all operating normally. On the other hand, as preventive maintenance is performed on specific equipment, changes in sensor data occur. When the similarity between the median value of the data in which the change occurred (i.e., the first sensor data) and the median value of the third sensor data is the highest, It may be determined that a defect will occur in a specific piece of equipment.

Meanwhile, when the remaining similarity values excluding the first similarity value among the first similarity value, the second similarity value, the third similarity value, and the fourth similarity value are the largest, the first normalization value may not be 1. In this case, computing device 100 may determine that the specific equipment on which preventive maintenance was performed is healthy.

Therefore, the computing device 100 of the present invention, even if an abnormality is determined for the sensor data measured from the equipment through the method described above with reference to FIGS. 5 and 6, the sensor data measured from the equipment operated with the same recipe and Through the consistency test, it is possible to determine whether an abnormality was determined due to preventive maintenance or whether a problem occurred in the equipment.

In one embodiment, when sensor data of each of one or more devices is monitored using a learned anomaly detection model to detect anomalies in the equipment, anomalies may be detected through preventive maintenance, and the computing device (100) of the present invention ) determines whether the abnormality detection is an abnormality due to preventive maintenance or whether an abnormality has occurred in the equipment, thereby omitting the manager's confirmation process and further increasing the efficiency of equipment management.

In various embodiments, computing device 100 may classify equipment on which preventive maintenance has been performed as either a normal type or a defective type.

Specifically, the computing device 100 may classify it as a preset first normal type when the first normalization value is smaller than the third normalization value and the fourth normalization value. Here, the first normal type may mean a type in which the matching characteristics (tool matching characteristics) between the replaced parts and the existing parts are complete.

Additionally, the computing device 100 may classify the first normalization value as a preset second normal type when the first normalization value is greater than the third normalization value and the fourth normalization value and is less than the second normalization value. Here, the second normal type may mean a type in which tool matching characteristics have improved due to parts replacement, but are not perfect.

Additionally, the computing device 100 may classify the defect as a preset first defect type when the second normalization value is smaller than the third normalization value and the fourth normalization value. Here, the first defect type may refer to a type in which the previously well-matched tool matching characteristics are broken due to replacement of parts.

Additionally, the computing device 100 may classify the defect as a preset second defect type when the second normalization value is greater than the third normalization value and the fourth normalization value. Here, the second defect type may mean a type in which the previously unsuitable tool matching characteristics have become worse.

The computing device 100 of the present invention can increase the convenience of observing the prognosis of preventive maintenance and managing equipment by providing a type classification method according to various embodiments.

Figure 5 is a flowchart illustrating an example of a method for maintaining the accuracy of an equipment abnormality detection model according to an embodiment of the present invention. Figure 6 is a flowchart illustrating an example of a method for determining whether to perform a consistency check for specific equipment according to an embodiment of the present invention.

According to one embodiment of the present invention, the computing device 100 can maintain the accuracy of the equipment abnormality detection model.

Referring to FIG. 5, the computing device 100 may monitor sensor data of each of one or more devices using a learned anomaly detection model to detect anomalies in the equipment (S210).

Specifically, the computing device 100 may acquire process sensor data and prior knowledge data corresponding to the process sensor data. Additionally, the computing device 100 may utilize a deep learning model (i.e., an anomaly detection model) to generate reconstructed process sensor data corresponding to the process sensor data. Additionally, the computing device 100 may calculate a reconstruction rate error based on process sensor data and reconstruction process sensor data. Additionally, the computing device 100 may detect abnormal operation based on comparison of the reconstruction rate error and the reference threshold.

Here, the deep learning model is a first sub-model that extracts feature information corresponding to process sensor data, and a second sub-model that extracts interaction information between each process sensor data based on process sensor data and prior knowledge data. , It may include an attention module that generates feature information by combining the outputs of the first sub-model and the second sub-model, and a dimensional restoration model that generates reconstruction process sensor data by restoring feature information.

In one embodiment, process sensor data for learning used to learn a deep learning model may include only sensor data related to normality. In other words, process sensor data for learning may not include sensor data related to abnormal operations.

In other words, as the deep learning model is learned based on data accumulated over many years, when process sensor data similar to existing accumulated process sensor data is input, it can output reconstructed process sensor data similar to the process sensor data related to the input. In addition, when process sensor data that is not similar to existing accumulated process sensor data is input, reconstructed process sensor data that is not similar to the process sensor data related to the input can be output.

Additionally, the computing device 100 may calculate a reconstruction rate error between process sensor data and reconstruction process sensor data output in response to the process sensor data. Specifically, the computing device 100 calculates a larger reconstruction rate error as the difference between the process sensor data related to the input and the reconstruction process sensor data related to the output is larger, and the process sensor data related to the input and the reconstruction process sensor data related to the output The larger the difference between the two, the smaller the reconstruction rate error can be calculated. That is, the reconstruction rate error can be calculated based on the difference between the input (ie, process sensor data) and output (ie, reconstruction process sensor data) of the deep learning model.

Additionally, the computing device 100 may detect abnormal operation based on comparison of the calculated reconstruction rate error and the reference threshold. In one embodiment, the reference threshold is obtained during the learning process of a deep learning model, and is characterized in that it is determined based on a reconstruction rate error related to the maximum value among a plurality of reconstruction rate errors related to each of a plurality of learning data. You can. For example, the reference threshold is based on the maximum reconstruction rate error among the reconstruction rate errors corresponding to each of the 100,000 process sensor data acquired over the past three years and the reconstruction process sensor data corresponding to each process sensor data. can be decided. In other words, the reference threshold may be determined based on the reconstruction rate error of the least reconstructed process sensor data (i.e., the process sensor data with the largest reconstruction rate error) among the process sensor data over the past several years. This reference threshold can be a standard for detecting abnormal behavior. In a specific embodiment, the computing device 100 may determine normal operation when the reconstruction rate error is below a reference threshold, and may determine abnormal operation when the reconstruction rate error exceeds the reference threshold.

In other words, the computing device 100 may generate reconstruction process sensor data based on process sensor data and calculate a reconstruction rate error based on comparison of the process sensor data and reconstruction process sensor data. The computing device 100 may compare the reconstruction rate error with a reference threshold and, if the reconstruction rate error exceeds the reference threshold, determine that an abnormal operation has occurred during the facility process. For example, if the process sensor data is similar to the learning data used to learn the deep learning model (i.e., process sensor data accumulated over many years), the reconstruction rate error may be calculated to be small. Conversely, if the process sensor data is not similar to the training data used to learn the deep learning model, the reconstruction rate error may be calculated to be large. That is, when the reconstruction rate error is greater than the reference threshold calculated based on data accumulated over many years, the computing device 100 generates a data type that is different from the data type obtained in an existing normal situation (i.e., has never been used in the past). It can be determined that sensor data of a type that has not been experienced has been detected and thus determined to be an abnormal situation.

In the present invention, prior knowledge data may include information about the correlation between a plurality of sensors. For example, the prior knowledge data may include information that, in the semiconductor production process, a first sensor that measures temperature in response to a specific process step and a second sensor that measures pressure in the same process step are correlated. In other words, prior knowledge data may include information that specific sensor data in a specific process step can affect other sensor data, that is, information related to interrelationship. In embodiments, prior knowledge data may be structured in the form of a graph structure.

These process sensor data and prior knowledge data can be obtained during the production facility process and can be used to detect abnormal behavior. Abnormal operation may include various process situations that impede yield, or situations in which abnormal operation or defective conditions related to equipment failure are detected.

According to one embodiment, the semiconductor production process consists of hundreds of production steps, and the yield of produced wafers can be measured through final inspection at the last step. The wafer yield may refer to the ratio of good chips among all chips in the wafer. For example, if all chips are normal, the yield of the wafer may be 100%. In general, when wafer yield is low in a semiconductor process, process sensor data related to production can be analyzed. For example, through process sensor data, it is possible to analyze whether many problems occur in wafers that have passed through a specific facility, or whether many problems occur under specific production conditions. In the hundreds of steps that produce semiconductors, there are specific production conditions for each step, and if these are not met, the wafer yield can be affected. For example, in certain facilities, there may be a condition that requires the wafer to reach a set temperature within a certain time after being input, but the condition may not be met during the actual process, and this can be identified through process sensor data acquired at that stage. You can. In this way, analyzing hundreds of process sensor data acquired in real time during the process can be very helpful in solving yield problems.

Additionally, in one embodiment, in semiconductor processing, preventive maintenance can be very important. Preventive maintenance can refer to an analysis process that seeks to prevent the entire facility from being shut down by identifying and resolving abnormal signals or defective conditions before the facility completely breaks down. For example, in the case of semiconductor processing, defects smaller than 20 nm that occur due to equipment problems must be found on a 300 mm wafer, so finding defects can be very difficult. As technology advances, circuit line widths have become increasingly thinner, which can make it more difficult to detect defects. For a specific example, a semiconductor process can typically include about 500 processes and 1,000 measurement steps. If even a small problem occurs in the equipment during this process, it can result in a large loss, such as having to discard the entire wafer. To prevent such problems, it can be very important to collect and analyze data generated from process sensors in semiconductor facilities. The computing device 100 of the present invention utilizes artificial intelligence to acquire and analyze process sensor data generated during the semiconductor process, thereby automating the detection of abnormal operations during the process. When using artificial intelligence to detect whether an abnormal operation has occurred, the time required to detect abnormal data can be minimized, and the cause of the abnormal state can be identified even with respect to process sensor data from complex equipment, providing convenience to workers. can be provided.

When the anomaly detection model detects abnormal data in specific sensor data obtained from specific equipment, the computing device 100 may obtain preventive maintenance information related to the specific equipment and perform a consistency check for the specific equipment (S220 ).

Specifically, referring to FIG. 6, the computing device 100 may recognize whether preventive maintenance has been performed on specific equipment based on preventive maintenance information (S211).

Specifically, the computing device 100 may extract preventive maintenance information from log data related to specific equipment and, based on this, recognize whether preventive maintenance has been performed on the specific equipment. However, the present invention is not limited to this, and the computing device 100 may receive information about whether to perform preventive maintenance on specific equipment from an administrator terminal.

Meanwhile, when the computing device 100 recognizes that preventive maintenance has been performed on specific equipment, it may perform a consistency check (S222).

Specifically, the computing device 100 may acquire sensor data for each of the specific equipment on which preventive maintenance has been performed and one or more equipment operated with the same recipe as the specific equipment. Additionally, the computing device 100 may sample sensor data and perform alignment on sensor data acquired from each device. Additionally, the computing device 100 may determine the minimum, maximum, and median values for each device from the sampled sensor data. Additionally, the computing device 100 may measure the similarity of sensor data corresponding to each device based on at least one of the minimum value, maximum value, and median value for each device. Additionally, the computing device 100 may determine whether the specific equipment on which the preventive maintenance has been performed is normal based on the similarity.

The method by which the computing device 100 performs the consistency check in step S222 may be performed in the same manner as the methods described above with reference to FIGS. 5 and 6, and thus redundant description will be omitted.

On the other hand, if the computing device 100 recognizes that preventive maintenance has not been performed on the specific equipment in step S211, it may recognize that a problem has occurred in the specific equipment (S223).

In this case, the computing device 100 may record anomalies occurring in a specific device in log data related to the specific device and transmit the corresponding information to the administrator terminal.

Referring again to FIG. 5, the computing device 100 can recognize, based on the consistency check, whether the abnormal data detected by the abnormality detection model was detected by preventive maintenance or was detected due to a defect in specific equipment. There is (S230).

Specifically, when the computing device 100 determines that a specific piece of equipment is normal as it performs a consistency check, it may recognize that the abnormal data detected by the anomaly detection model was detected through preventive maintenance. Additionally, when the computing device 100 determines that a specific device is not normal while performing a consistency check, the computing device 100 may recognize that the abnormal data detected by the anomaly detection model was detected due to a defect in the specific device.

When the computing device 100 recognizes that the abnormal data detected by the abnormality detection model is detected by preventive maintenance, the computing device 100 may update the abnormality detection model based on the abnormal data (240).

Specifically, the computing device 100 may obtain normalized data by scaling abnormal data. Additionally, the computing device 100 can transfer learn a fully connected layer using normalized data while maintaining the feature extraction layer of the anomaly detection model.

In one embodiment, even if sensor data changes due to preventive maintenance, the features that the anomaly detection model extracts from the data may be the same. That is, the computing device 100 maintains a feature extraction layer that finds feature points in sensor data, omits unnecessary learning processes, and can only perform transfer learning on the fully connected layer related to the critical point of anomaly detection.

For example, when transferring a fully connected layer, the computing device 100 may compare a first maximum value for detecting an anomaly in an anomaly detection model with a second maximum value of normalized data. Additionally, when the second maximum value is greater than the first maximum value, the computing device 100 may change the first maximum value for detecting an anomaly in the anomaly detection model to the second maximum value.

As described above, the anomaly detection model of the present invention can be updated to increase the range of data for detecting anomalies. In this case, as updates continue, the range for detecting anomalies continues to increase, which may lower the accuracy of detecting anomalies. To prevent this, the computing device 100 may initialize the anomaly detection model to the initially learned model at each preset cycle.

Therefore, the computing device 100 of the present invention performs continuous updates and periodic initialization (restore to the initial model) through transfer learning, thereby reducing the time and cost required to learn the initial model and improving the accuracy of the anomaly detection model. It can be maintained.

7 to 9 are diagrams for explaining an example of a method for performing alignment on sensor data according to an embodiment of the present invention.

According to an embodiment of the present invention, the computing device 100 is an algorithm for sensor data used to check consistency between a plurality of manufacturing process equipment based on sensor data, learn an anomaly detection model, and maintain the accuracy of the anomaly detection model. phosphorus (or pretreatment) can be performed.

In the following description, an example of how computing device 100 performs alignment on data to train an anomaly detection model (i.e., generate training data for an anomaly detection model) is described, but is not limited thereto. , the same method can be applied to alignment of sensor data used to check consistency between multiple manufacturing process equipment based on sensor data or to maintain the accuracy of an anomaly detection model.

According to one embodiment, the computing device 100 may perform preprocessing of learning data to train an artificial intelligence model (i.e., an anomaly detection model) that detects an abnormal process.

Specifically, the data preprocessing method performed by the computing device 100 may include acquiring raw data including a plurality of sensor process data.

In one embodiment, process sensor data may include various types of data obtained at industrial sites. Process sensor data may include sensor data generated in seconds during the process.

For example, a production facility may generate hundreds of sensor data per second. In semiconductor production facilities, sensors are installed to detect temperature, pressure, and various chemical input amounts, and process sensor data can be obtained in real time through the sensors. In other words, process sensor data may refer to sensor data acquired in real time through a plurality of sensors based on the operation of semiconductor processing equipment. Process sensor data may include operating parameters of various devices for wafer manufacturing in a semiconductor fab and various sensor data acquired by device operation. For example, process sensor data may include lot equipment history data from a management execution system (MES), data from equipment interface data sources, processing tool recipes, processing tool test data, probe test data, and electrical test data. , combined measurement data, diagnostic data, remote diagnostic data, post-processing data, etc., but the present invention is not limited thereto.

In one embodiment, the plurality of sensor process data may include information about sensor values that change over time.

In one embodiment, acquiring raw data may involve receiving or loading data stored in memory 120. Acquisition of raw data may include receiving or loading raw data into another storage medium, another computing device, or a separate processing module within the same computing device based on wired/wireless communication means. In specific embodiments, raw data may be received either to the production device or through an external server.

The data preprocessing method performed by the computing device 100 may include performing preprocessing on raw data. In one embodiment, preprocessing of raw data may be performed to express a plurality of sensor process data as data to be used in a neural network. For example, since machine learning systems generally use tensors with multidimensional arrays as their basic data structure, preprocessing is performed to convert sensor process data into tensor-type data so that the data can be used in the artificial intelligence model of the present invention. It can be done. In one embodiment, the present invention can generate three-dimensional tensor data corresponding to the sensor process data through preprocessing of the sensor process data.

In one embodiment, the computing device 100 may perform resampling on raw data when constructing a learning data set.

Resampling performed by the computing device 100 is intended to match sampling times between sensor process data included in raw data, and may be characterized by extracting sensor process data based on regular time intervals. The computing device 100 may extract the sensor process data again at regular time intervals to correct sampling errors of the sensor process data.

Specifically, referring to (a) of FIG. 7, the corresponding operation time of each wafer may be different. Wafer 1 and wafer 2 may each have different operating times. Specifically, looking at the operation times of wafer 1 and wafer 2, it can be seen that the operation times do not have a constant interval. Accordingly, the computing device 100 may perform resampling so that the operation time has constant time intervals.

For a specific example, referring to (b) of FIG. 7, the computing device 100 may extract sensor process data again at regular time intervals (1 second) in response to wafer 1. As shown in (b) of FIG. 7, the sensor process data corresponding to wafer 1 has a constant time interval (1 second), that is, 21:20:31, 21:20:23, and 21:20. It may be resampled to include an operating time from minutes 35 seconds to 21 hours 20 minutes 37 seconds.

In one embodiment, when constructing a learning data set, the computing device 100 may perform interpolation on resampling raw data.

Interpolation performed by the computing device 100 may be characterized as interpolating missing values generated in response to each sensor process data as resampling is performed. That is, as shown in (b) of FIG. 7, as resampling occurs, missing values (N/A) may occur for each sensor at each time point. In this case, the computing device 100 may perform interpolation by inputting the sensor value into the missing value. In one embodiment, the computing device 100 may be characterized in that the sensor value corresponding to the missing value is estimated using the previous value. That is, as shown in (c) of Figure 7, it can be supplemented through 0, 0 based on the previous values 0, 0, corresponding to 21:20:31 and 32 seconds, and 21:20:35. It can be supplemented through 1, 1 based on the previous value 1, 1, corresponding to seconds to 37 seconds. As described above, the computing device 100 may perform resampling so that all sensor process data has the same time interval, and may perform interpolation for missing values related to the resampling result.

In one embodiment, when constructing a learning data set, the computing device 100 may perform padding on interpolated raw data.

Padding performed by the computing device 100 may calculate average process time information based on process time information of sensor process data included in raw data.

Specifically, each sensor process data may operate at different times. Referring to (a) of Figure 7, for wafer 1, it operates for 10 seconds from 21:20:30 to 21:20:40, but for wafer 2, it operates from 21:20:42 to 21:21. It can be operated for 26 seconds up to 08 seconds. In other words, it can be confirmed that the process time between each sensor process data is different. In this case, the computing device 100 may calculate average process time information through the average value of the process times of wafer 1 and wafer 2. In a specific example, the average process time information related to wafer 1 and wafer 2 can be calculated as 18 seconds ((10 +26)/2).

Additionally, in the step of performing padding, the computing device 100 may perform a step of correcting the process time of each sensor process data in response to the average process time information.

In one embodiment, the computing device 100 may adjust the operation time of each sensor process data in response to the average process time information. For example, for wafer 1, which has an operation time of 10 seconds, which is shorter than the average process time, the operation time can be increased to correspond to the average process time, and for wafer 2, which has an operation time of 26 seconds, which is longer than the average process time, the operation time can be increased to correspond to the average process time. The time can be reduced to correspond to the average process time.

In one embodiment, in the case of sensor process data whose operating time has increased in response to the average process time, the computing device 100 may supplement missing values that occur as time increases based on previous values.

For example, the process sensor data related to wafer 1, as shown in (a) of Figure 8, has an operating time of 10 seconds, and is calculated by averaging with other process sensor data (e.g., process sensor data related to wafer 2). It may be shorter than the average process time (18 seconds), which may increase the operating time by 8 seconds. That is, as shown in (b) of FIG. 8, the operating time related to 21:20:41 to 21:20:48 can be added.

Additionally, missing values that occur as operating time increases according to the average process time can be supplemented based on previous values. That is, based on 1, 0 corresponding to the previous value (i.e., each sensor value corresponding to 21:20:40), 1, 0 corresponding to 21:20:41 to 21:20:48 Sensor values can be supplemented. The detailed numerical description of the above-mentioned operating time and average process time is only an example and the present invention is not limited thereto.

In other words, the process time in all processes can be unified through padding. Through this preprocessing process, the sampling rate of all sensor process data can be unified regardless of the process, and sampling errors can be prevented from occurring.

In one embodiment, building a learning data set may include acquiring tensor data based on raw data on which padding has been performed. According to an embodiment, the tensor data generated through the above-described preprocessing process may be 3D tensor data based on a 3D axis related to wafer id, sensor number, and time.

As described above, preprocessing can be performed to convert sensor process data into tensor data in tensor form so that it can be used in an artificial intelligence model.

Specifically, the data preprocessing method performed by the computing device 100 may include selecting valid data from preprocessed raw data and constructing a learning data set for training an artificial intelligence model.

Valid data is learning data related to the learning of an artificial intelligence model that detects abnormal processes, and may be characterized as process sensor data related to normal processes.

In one embodiment, the artificial intelligence model may be a neural network model that detects abnormal situations based on process sensor data acquired in real time during the process. For example, abnormal operations may include various process situations that impede yield, or situations in which abnormal operations or defective conditions related to equipment failure are detected.

An artificial intelligence model may be a neural network model trained to generate output data similar to input data. The artificial intelligence model may include, for example, an autoencoder that outputs output data similar to input data. An autoencoder may include a dimensionality reduction network function (eg, encoder) and a dimensionality restoration network function (eg, decoder).

According to one embodiment, the auto encoder may include at least one hidden layer, and an odd number of hidden layers may be disposed between input and output layers. The number of nodes in each layer may be reduced from the number of nodes in the input layer to an intermediate layer called the bottleneck layer (encoding), and then expanded symmetrically and reduced from the bottleneck layer to the output layer (symmetrical to the input layer). The nodes of the dimensionality reduction layer and dimensionality restoration layer may or may not be symmetric. Autoencoders can perform nonlinear dimensionality reduction. The number of input layers and output layers may correspond to the number of sensors remaining after preprocessing of the input data. In an auto-encoder structure, the number of nodes in the hidden layer included in the encoder may have a structure that decreases as the distance from the input layer increases. If the number of nodes in the bottleneck layer (the layer with the fewest nodes located between the encoder and decoder) is too small, not enough information may be conveyed, so if it is higher than a certain number (e.g., more than half of the input layers, etc.) ) may be maintained.

The artificial intelligence model can be learned through process sensor data related to normality accumulated over many years. Accordingly, it can output output data corresponding to the input by using process sensor data acquired in real time as input.

Specifically, as the artificial intelligence model is learned through normal process sensor data accumulated over many years, when process sensor data similar to existing accumulated normal process sensor data is input, a reconstructed process sensor similar to the process sensor data related to the input is used. Data can be output. In this case, the reconstruction rate error, which is the difference between input data and output data, may be small.

Conversely, when process sensor data that is not similar to existing accumulated normal process sensor data is input, the artificial intelligence model may output reconstructed process sensor data that is not similar to the process sensor data related to the input. In this case, the reconstruction rate error, which is the difference between input data and output data, may be large.

In one embodiment, the artificial intelligence model may determine normal operation if the reconstruction rate error between input data and output data is below a certain standard value, and if the reconstruction rate error between input data and output data exceeds a certain standard value, It can be determined that abnormal operation has occurred. In other words, when the reconstruction rate error exceeds a certain standard value, the artificial intelligence model acquires a data type that is different from the data type obtained in existing normal situations (i.e., a type of sensor data that has never been experienced in the past is detected). This can be determined as an abnormal situation.

As described above, the artificial intelligence model can determine whether an abnormal operation has occurred by using process sensor data obtained in a process situation as input. In order to learn such an artificial intelligence model, a large amount of process sensor data accumulated over many years is required, as described above. In particular, among the large amount of process sensor data, process sensor data related to normal operation must be provided.

According to one embodiment, the computing device 100 may construct valid data by selecting only process sensor data related to normal operation from a large amount of process sensor data (e.g., process sensor data related to normal and abnormal) accumulated over many years. there is. In other words, the computing device 100 can select and obtain valid data suitable for learning an artificial intelligence model from all raw data for which good or bad has not been determined obtained during the process.

In a specific embodiment, the computing device 100 may obtain valid data, which is sensor process data related to normal, by performing first screening and secondary screening in response to raw data including a plurality of process sensor data. there is.

More specifically, when constructing a learning data set, the computing device 100 may perform first selection in response to a plurality of sensor process data. Here, the first selection may be performed by comparing the amplitude similarity of each of the plurality of process sensor data based on the reference value.

In one embodiment, when the computing device 100 performs first screening, a step of obtaining a reference value based on a plurality of sensor process data may be performed. Here, the reference value may be a standard value for calculating amplitude similarity with each of a plurality of sensor process data. For example, the reference value may be obtained through the median of a plurality of sensor process data.

In detail, the plurality of sensor process data may include information about sensor values that change over time. A plurality of sensor process data may be expressed in the form of a graph as shown in (a) of FIG. 9. In Figure 9, the x-axis may correspond to time, and the y-axis may correspond to sensor values obtained during the process. The computing device 100 may identify the median value of sensor values of a plurality of sensor process data at each time point, and a combination of the median values may be a reference value. For a specific example, a large amount of sensor process data includes first sensor process data to fifth sensor process data, and corresponding to the first point in time, the sensor value of the first sensor process data is 3, and the sensor value of the second sensor process data is 3. When the sensor value is 8, the sensor value of the third sensor process data is 7, the sensor value of the fourth sensor process data is 10, and the sensor value of the fifth sensor process data is 90, the computing device 100 Among the values 3, 7, 8, 10, and 90, the median value of 8 can be identified as the median value corresponding to the first time point. In this way, the computing device 100 can obtain a reference value based on the median values for each time point.

In one embodiment, when the reference value is calculated using the median, the influence of outlier values may be minimized.

For example, a plurality of sensor process data included in the raw data may include both sensor process data related to normal and abnormal conditions that have not been judged good or bad. Because it contains sensor process data related to abnormal operation, outlier values that are different from general sensor values may exist. In the above example, when the first to fifth sensor values corresponding to the first time point are 3, 7, 8, 10, and 90, the sensor value related to 90 may be related to an outlier. In other words, there may be an outlier where a specific sensor value is significantly different from other sensor values.

If a reference value is obtained using an average in a situation where sensor process data related to normal and abnormal are mixed, the influence of the outlier may increase, so the reference value is not an appropriate standard. For a detailed example, the average value of the sensor values in the above-described example may be 23.6 ((3+7+8+10+90)/5), while the median value may be 8.

In other words, when using the average value, it may be greatly influenced by one outlier, making it difficult to use it as a reference value. In the present invention, considering that raw data includes sensor process data related to normal and abnormal conditions, a reference value can be obtained using the median value to minimize the influence of outliers. Since the reference value obtained in this way has minimal influence on outliers, it can serve as a more appropriate standard for selecting normal data.

In one embodiment, in performing the first selection, the computing device 100 performs the first selection based on the similarity between the amplitude change form of each of the plurality of sensor process data and the amplitude change form of the reference value. can be performed. The computing device 100 may select sensor process data having an amplitude change form similar to that of the reference value from among a plurality of sensor process data. For example, sensor process data that does not have a certain level of similarity to the amplitude of the reference value may not be selected in the first place.

In a specific embodiment, the similarity between each amplitude change type (eg, each wavelength) may be determined through a Dynamic Time Wraping (DTW) algorithm. Dynamic time warping may be an algorithm that measures the similarity (or degree of similarity) between two time series data with different movements depending on speed or length. Dynamic time warping may be characterized by finding a warping path that minimizes the cumulative distance by matching in the direction where the distance is minimized.

In general, when determining similarity between two time series data, the Euclidean distance algorithm can be used. The Euclidean distance algorithm can determine the similarity between both data by calculating the distance on the same time line. For example, when using Euclidean distance, similarity is determined by matching for each time point, such as 0 second - 0 second, 1 second - 1 second, etc. However, this Euclidean distance has the disadvantage that it is difficult to find similarities in data that appears to have similar overall patterns as tremor and movement become more severe, and that it is difficult to evaluate the similarity between short-length time series data. The sensor process data used in the present invention may each have different operating times. In other words, in the case of the Euclidean distance algorithm, the accuracy of calculating similarity can be very low in situations where the time axis is distorted. In other words, if the alignment between the two data is not correct, it may be difficult to calculate the distance and thus difficult to evaluate the similarity.

In another embodiment, cosine similarity comparison can be used to vectorize both data, calculate the distance through the angle between vectors, and evaluate similarity based on this. However, the cosine similarity comparison has the disadvantage of not reflecting the sequence of time series data. there is.

Accordingly, the present invention can measure the similarity between both data (i.e., reference values and each sensor process data) by utilizing a dynamic time warping algorithm. A dynamic time warping algorithm may be an algorithm that evaluates the similarity of wavelengths of time bases at two different speeds. The dynamic time warping algorithm can match not only data on the same timeline but also surrounding time points to more similar elements. Using this dynamic time warping algorithm has the advantage of enabling similarity evaluation between time series data of different lengths.

In one embodiment, the computing device 100 may remove sensor process data whose similarity to a reference value is less than a certain threshold and select (i.e., first select) sensor process data whose similarity to a reference value is greater than or equal to a certain threshold. Figure 9(b) may be a graph related to the results of the first screening. Referring to Figures 9(a) and 9(b), it can be identified that sensor process data that is not similar to the reference value is removed. That is, the computing device 100 may perform the first selection based on determining the similarity of the amplitude change pattern (eg, wavelength) based on the reference value.

In the step of constructing learning data, the computing device 100 may perform a second selection in response to the first selection of sensor process data. Here, the secondary selection may be performed based on timing similarity between sensor process data.

In a specific embodiment, performing the second selection may include identifying one or more abnormal data through timing similarity comparison between a reference value and each first selected sensor process data.

In a specific embodiment, timing similarity comparison between each data may be performed using an Affine Dynamic Time Warping (ADTW) algorithm. The affine operation time warping algorithm can evaluate the similarity related to the time difference between data. In the case of the affine operation time warping algorithm, time variation information can be extracted from the DTR path and the ADTW distance can be calculated based on this. In other words, when using the affine operation time warping algorithm, there is an advantage of being able to evaluate the similarity related to the time difference (i.e. timing) between data based on time variation information. As an example, each sensor process data may have different sensor values depending on changes in time. That is, each sensor process data can be expressed on a graph to have a specific wavelength. For example, sensor process data having the same wavelength as the reference value may be selected in the first selection process and may be subject to the second selection. In this case, the secondary selection may be performed based on whether the timing of changes, such as the rise or fall of the wavelength of the reference value, is similar.

In other words, secondary screening may be performed depending on whether a change in wavelength occurs at a time similar to the reference value. In an embodiment, the computing device 100 may identify sensor process data with timing that is not similar to a reference value as one or more abnormal data. One or more abnormal data may be sensor process data related to a timing that is different from the reference value.

Additionally, performing the second selection may include obtaining valid data by removing one or more abnormal data from the first selection of sensor process data.

That is, the computing device 100 removes sensor process data having waveform change timing that is not similar to the waveform change timing of the reference value and selects sensor process data having waveform change timing similar to the reference value and the waveform change timing (i.e., the second tea selection) can be done. Figure 9(c) may be a graph related to the results of secondary screening. Referring to Figure 9(a), Figure 9(b), and Figure 9(c), it can be seen that as a result of the second selection, sensor process data related to timing that is not similar to the reference value is removed. That is, the computing device 100 may perform the second selection based on determining the similarity of the waveform change timing based on the reference value.

As described above, the computing device 100 can obtain a reference value through a combination of median values at each time point of a plurality of sensor process data corresponding to raw data, and determine the similarity of the amplitude change form based on the reference value. By performing the first selection and determining the similarity of the timing of waveform changes, the second selection can be performed to obtain valid data suitable for learning an artificial intelligence model. In this case, since the reference value corresponding to the plurality of sensor process data is obtained based on the median value that is not affected by outliers, it can provide an appropriate standard for excluding abnormal data.

Additionally, when the first selection and the second selection are performed in response to raw data, process sensor data having amplitude change magnitude and change timing similar to the reference value may be selected as valid data. For example, process sensor data having an amplitude change magnitude and timing that are not similar to a reference value may be process sensor data related to abnormal operation.

That is, valid data can be obtained by determining only process sensor data with similar amplitude magnitude and similar timing as the standard reference value as normal sensor process data.

In a further embodiment, the computing device 100 may be characterized in that it obtains a correction reference value based on the second selected sensor process data and performs the first selection and the second selection based on the correction reference value. You can. When a correction reference value is obtained and the first screening and the second screening are performed again based on this, the accuracy of screening may be improved.

In more detail, the step of acquiring valid data includes obtaining a correction reference value based on the second selected sensor process data, performing first selection using the correction reference value, and first selection It may include performing secondary selection in response to the processed sensor data.

As described above, the second selected sensor process data may be process sensor data having amplitude change magnitude and change timing similar to the reference value. In other words, the second selected sensor process data may be determined to be normal sensor process data. If a reference value (i.e., a correction reference value) is generated again based on sensor process data from which sensor process data related to abnormalities have been filtered out, this can be a more accurate standard for selecting process sensor data related to normal compared to the initial reference value. For example, in the process of generating a reference value (i.e., a correction reference value) again, sensor process data in which data with abnormal amplitude or timing have been primarily filtered out is used, so it is possible to have high accuracy in selecting normal data. Accordingly, when the first screening and the second screening related to the magnitude and timing of the amplitude are performed again using the corresponding correction reference value, sensor process data related to abnormalities can be more appropriately removed.

In an embodiment, training data for training an artificial intelligence model should include only training data related to normality. For example, even if sensor process data related to a small number of abnormalities is included in the learning data, the output accuracy of the trained artificial intelligence model can be greatly reduced. In addition, as the artificial intelligence model is learned through process sensor data related to abnormalities, if it cannot detect abnormal signals or defective conditions in real time, it may result in large losses, such as having to discard the entire wafer. In other words, it can be very important to distinguish sensor process data that may be abnormal from being included as learning data. In other words, it may be important to construct learning data by selecting only sensor process data related to normal conditions through clear criteria. By using a correction reference value, the present invention can select only sensor process data related to normal using a more accurate standard.

In the above description, it was explained that the first screening and the second screening are performed again to obtain a correction reference value that serves as a more accurate classification standard. However, the first screening and the second screening are performed in the third and fourth rounds or It will be apparent to those skilled in the art that more than that can be performed, and thus a secondary correction reference value and a tertiary correction reference value may be obtained. In other words, as the first selection and second selection processes are repeated a plurality of times, the nth correction reference value has a clearer data classification standard, and the accuracy of filtering sensor process data related to abnormalities can be improved. This can prevent learning data from including sensor process data related to abnormalities, ultimately resulting in improved output accuracy of the artificial intelligence model.

Figure 10 is a schematic diagram showing one or more network functions related to one embodiment of the present invention.

Throughout this specification, artificial intelligence model, neural network, network function, and neural network may be used with the same meaning. A neural network can generally consist of a set of interconnected computational units, which can be referred to as “nodes”. These “nodes” may also be referred to as “neurons.”

A deep neural network (DNN) may refer to a neural network that includes multiple hidden layers in addition to the input layer and output layer. Deep neural networks allow you to identify latent structures in data. In other words, it is possible to identify the potential structure of a photo, text, video, voice, or music (e.g., what object is in the photo, what the content and emotion of the text are, what the content and emotion of the voice are, etc.) . Deep neural networks include convolutional neural networks (CNN), recurrent neural networks (RNN), auto encoders, generative adversarial networks (GAN), and restricted Boltzmann machines (RBM). machine), deep belief network (DBN), Q network, U network, Siamese network, etc. The description of the deep neural network described above is only an example and the present invention is not limited thereto.

A neural network may be trained in at least one of supervised learning, unsupervised learning, and semi-supervised learning. Learning of a neural network is intended to minimize errors in output. In neural network learning, learning data is repeatedly input into the neural network, the output of the neural network and the error of the target for the learning data are calculated, and the error of the neural network is transferred from the output layer of the neural network to the input layer in the direction of reducing the error. This is the process of updating the weight of each node in the neural network through backpropagation. In the case of teacher learning, learning data in which the correct answer is labeled in each learning data is used (i.e., labeled learning data), and in the case of non-teacher learning, the correct answer may not be labeled in each learning data. That is, for example, in the case of teacher learning regarding data classification, the learning data may be data in which each learning data is labeled with a category. Labeled training data is input to the neural network, and the error can be calculated by comparing the output (category) of the neural network and the label of the training data. As another example, in the case of non-teachable learning for data classification, the error can be calculated by comparing the input training data with the neural network output. The calculated error is backpropagated in the reverse direction (i.e., from the output layer to the input layer) in the neural network, and the connection weight of each node in each layer of the neural network can be updated according to backpropagation. The amount of change in the connection weight of each updated node may be determined according to the learning rate. The neural network's calculation of input data and backpropagation of errors can constitute a learning cycle (epoch). The learning rate may be applied differently depending on the number of repetitions of the learning cycle of the neural network. For example, in the early stages of neural network training, a high learning rate can be used to increase efficiency by allowing the neural network to quickly achieve a certain level of performance, and in the later stages of training, a low learning rate can be used to increase accuracy.

In the learning of neural networks, the training data can generally be a subset of real data (i.e., the data to be processed using the learned neural network), and thus the error for the training data is reduced, but the error for the real data is reduced. There may be an incremental learning cycle. Overfitting is a phenomenon in which errors in actual data increase due to excessive learning on training data. For example, a phenomenon in which a neural network that learned a cat by showing a yellow cat fails to recognize that it is a cat when it sees a non-yellow cat may be a type of overfitting. Overfitting can cause errors in machine learning algorithms to increase. To prevent such overfitting, various optimization methods can be used. To prevent overfitting, methods such as increasing the learning data, regularization, or dropout, which omits some of the network nodes during the learning process, can be applied.

The steps of the method or algorithm described in connection with embodiments of the present invention may be implemented directly in hardware, implemented as a software module executed by hardware, or a combination thereof. The software module may be RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), Flash Memory, hard disk, removable disk, CD-ROM, or It may reside on any type of computer-readable recording medium well known in the art to which the present invention pertains.

The components of the present invention may be implemented as a program (or application) and stored in a medium in order to be executed in conjunction with a hardware computer. Components of the invention may be implemented as software programming or software elements, and similarly, embodiments may include various algorithms implemented as combinations of data structures, processes, routines or other programming constructs, such as C, C++, , may be implemented in a programming or scripting language such as Java, assembler, etc. Functional aspects may be implemented as algorithms running on one or more processors.

Those skilled in the art will understand that various illustrative logical blocks, modules, processors, means, circuits and algorithm steps described in connection with the embodiments disclosed herein can be used in electronic hardware, (for convenience) It will be understood that the implementation may be implemented by various forms of program or design code (referred to herein as “software”) or a combination of both. To clearly illustrate this interoperability of hardware and software, various illustrative components, blocks, modules, circuits and steps have been described above generally with respect to their functionality. Whether this functionality is implemented as hardware or software depends on the specific application and design constraints imposed on the overall system. A person skilled in the art may implement the described functionality in various ways for each specific application, but such implementation decisions should not be construed as departing from the scope of the present invention.

The various embodiments presented herein may be implemented as a method, apparatus, or article of manufacture using standard programming and/or engineering techniques. The term “article of manufacture” includes a computer program, carrier, or media accessible from any computer-readable device. For example, computer-readable media include magnetic storage devices (e.g., hard disks, floppy disks, magnetic strips, etc.), optical disks (e.g., CDs, DVDs, etc.), smart cards, and flash memory. Includes, but is not limited to, devices (e.g., EEPROM, cards, sticks, key drives, etc.). Additionally, various storage media presented herein include one or more devices and/or other machine-readable media for storing information. The term “machine-readable media” includes, but is not limited to, wireless channels and various other media capable of storing, retaining, and/or transmitting instruction(s) and/or data.

It is to be understood that the specific order or hierarchy of steps in the processes presented is an example of illustrative approaches. It is to be understood that the specific order or hierarchy of steps in processes may be rearranged within the scope of the present invention, based on design priorities. The appended method claims present elements of the various steps in a sample order but are not meant to be limited to the particular order or hierarchy presented.

Above, embodiments of the present invention have been described with reference to the attached drawings, but those skilled in the art will understand that the present invention can be implemented in other specific forms without changing its technical idea or essential features. You will be able to understand it. Therefore, the embodiments described above should be understood in all respects as illustrative and not restrictive.

Claims (10)

A method performed by a computing device comprising at least one processor, comprising:
Monitoring sensor data of each of one or more devices using a learned anomaly detection model to detect abnormalities in the equipment;
When the abnormality detection model detects abnormal data in specific sensor data acquired from specific equipment, obtaining preventive maintenance information related to the specific equipment;
Recognizing whether preventive maintenance has been performed on the specific equipment based on the preventive maintenance information;
When it is recognized that preventive maintenance has been performed on the specific equipment, performing a consistency test on the specific equipment;
Based on the consistency check, recognizing whether the abnormal data detected by the abnormality detection model was detected due to preventive maintenance or due to a defect in the specific equipment; and
When recognizing that the abnormal data detected by the anomaly detection model is detected by preventive maintenance, updating the anomaly detection model based on the abnormal data;
Including,
The step of performing the consistency check is,
Obtaining first sensor data measured in the specific equipment on which preventive maintenance was performed, obtaining second sensor data measured in the specific equipment before the preventive maintenance was performed, and operating with the same recipe as the specific equipment Obtaining third sensor data measured from each of the above devices;
sampling the sensor data and performing alignment on the sensor data acquired from each device;
determining minimum, maximum, and median values for each device from the sampled sensor data;
A plurality of similarity values are generated based on the first sensor data, the second sensor data, and the third sensor data, which are data measured by each equipment while producing the preset number of wafers with the recipe for producing wafers. calculating step; and
calculating a plurality of normalization values by dividing each of the plurality of similarity values by the largest value among the plurality of similarity values; and
determining whether the specific equipment on which the preventive maintenance was performed is normal based on the plurality of normalization values calculated based on the similarity;
Including,
A method for maintaining the accuracy of equipment anomaly detection models.
delete delete ◈Claim 4 was abandoned upon payment of the setup registration fee.◈ According to claim 1,
Based on the consistency check, the step of recognizing whether the abnormal data detected by the abnormality detection model was detected due to preventive maintenance or due to a defect in the specific equipment is,
If it is determined that the specific equipment is normal by performing the consistency check, it is recognized that the abnormal data detected by the anomaly detection model was detected by preventive maintenance, and it is determined that the specific equipment is not normal by performing the consistency check. In this case, recognizing that the abnormal data detected by the abnormality detection model is detected due to a defect in a specific equipment;
Including,
A method for maintaining the accuracy of equipment anomaly detection models.
According to claim 1,
The step of updating the anomaly detection model is,
Obtaining normalized data by scaling the abnormal data;
Transfer learning a fully connected layer using the normalized data while maintaining the feature extraction layer of the anomaly detection model;
Including,
A method for maintaining the accuracy of equipment anomaly detection models.
According to clause 5,
The step of transfer learning a fully connected layer using the normalized data while maintaining the feature extraction layer of the anomaly detection model includes:
Comparing a first maximum value for detecting an anomaly in the anomaly detection model and a second maximum value of the normalized data; and
If the second maximum value is greater than the first maximum value, changing the first maximum value for detecting an anomaly in the anomaly detection model to the second maximum value;
Including,
A method for maintaining the accuracy of equipment anomaly detection models.
◈Claim 7 was abandoned upon payment of the setup registration fee.◈ According to claim 1,
The updated anomaly detection model is,
Updated to increase the range of data for detecting abnormalities,
In order to maintain the accuracy of the anomaly detection model, initializing the anomaly detection model to an initially learned model at a preset period,
A method for maintaining the accuracy of equipment anomaly detection models.
According to claim 1,
The step of monitoring sensor data of each of one or more devices using a learned anomaly detection model to detect anomalies in the device is:
Obtaining process sensor data and prior knowledge data corresponding to the process sensor data;
Generating reconstruction process sensor data corresponding to the process sensor data using a deep learning model;
calculating a reconstruction rate error based on the process sensor data and the reconstruction process sensor data; and
detecting abnormal operation based on comparison of the reconstruction rate error with a reference threshold;
Includes,
The deep learning model is,
a first sub-model for extracting feature information corresponding to the process sensor data;
a second sub-model that extracts interaction relationship information between each process sensor data based on the process sensor data and the prior knowledge data;
an attention module that generates feature information by combining outputs of the first sub-model and the second sub-model; and
a dimensional reconstruction model that restores the feature information to generate the reconstruction process sensor data;
Including,
A method for maintaining the accuracy of equipment anomaly detection models.
A memory that stores one or more instructions; and
A processor that executes the one or more instructions stored in the memory
Contains,
The processor executes the one or more instructions,
An apparatus for performing the method of claim 1.
A computer program combined with a computer as hardware and stored on a computer-readable recording medium so as to perform the method of claim 1.