Method and system for predicting process defects

Abstract

The present invention relates to a method and a system for predicting process defects. More specifically, the present invention relates to a method and a system for predicting process defects capable of determining whether a product is defective by using time series data extraction.

Description

Process defect prediction method and system

[0001] The present invention relates to a method and system for predicting process defects. More specifically, the present invention relates to a method and system for predicting process defects that can determine whether a product is abnormal by utilizing time-series data extraction.

[0002] Product production proceeds sequentially through the Component Assembly Line, which is a component-level process, the Sub Assembly Line, which combines components, and the Total Assembly Line, which assembles the final product.

[0003] In this regard, the display manufacturing process consists of a series of continuous steps requiring numerous precision processes and high-level quality control, and minute environmental changes or operational deviations occurring at each stage can directly affect product quality. In particular, due to the process characteristics of display products that demand high precision, problems such as bubble formation and curing defects may occur depending on changes in minute variables.

[0004] Traditionally, defects occurring during the product manufacturing process were primarily identified retrospectively through inspection equipment or operator judgment. Consequently, this frequently led to issues such as increased operating costs for inspection equipment, production delays, and resource waste resulting from defect detection late in the process. Furthermore, some defects were difficult to detect visually or only became apparent after a period of time, making early response challenging.

[0005] To address these issues, active attempts have recently been made to utilize AI-based anomaly detection technology to predict the likelihood of defects in real-time during the process and to respond proactively. However, conventional anomaly detection technologies are prone to degraded performance due to class imbalance, where the proportion of defective data is extremely small compared to the total data. Furthermore, prediction accuracy can be low because models often fail to sufficiently learn key features, such as time intervals between processes and interaction variables like temperature and pressure. Moreover, conventional models are generally trained with fixed parameters, making it difficult to flexibly respond to real-time changing process conditions or equipment status.

[0006] Therefore, there is still a need for technology capable of determining product defects by analyzing various production data collected during the manufacturing process.

[0007] The present invention aims to provide a process defect prediction method and system capable of determining whether a product is defective by analyzing various production data collected during the product manufacturing process.

[0008] More specifically, the present invention aims to solve the class imbalance problem and provide a process defect prediction model capable of predicting whether a product is normal or defective.

[0009] In particular, the present invention is intended to provide a process defect prediction model capable of detecting and predicting defects occurring in the sub-assembly line during the product manufacturing process in advance.

[0010] Furthermore, the present invention aims to provide a method and system for predicting process defects that can reduce the cost of the display manufacturing process and improve the efficiency of quality control.

[0011] To solve the problem described above, a computerized process defect prediction method according to the present invention, comprising the following steps, may include: specifying process data configured to include data corresponding to a produced product; configuring a training data set and a verification data set using the process data; training an artificial intelligence model to be trained using the training data set; acquiring an artificial intelligence model trained with the training data set based on the training; inputting the verification data set into the trained artificial intelligence model; obtaining a prediction value for the verification data set from the trained artificial intelligence model; and determining whether at least one product included in the verification data set is defective using the prediction value for the verification data set.

[0012] In an embodiment, the process data is configured to include a plurality of records having values ​​for a plurality of different columns, and in the step of configuring the training data set and the verification data set, the training data set and the verification data set can be configured based on a pre-set configuration criterion.

[0013] In an embodiment, the process data is configured to include at least one of a process name, time information, environmental variables, and defect status, and the pre-set configuration criteria may be related to configuring the training data set and the verification data set based on a value corresponding to at least one column among the plurality of columns.

[0014] In an embodiment, the at least one column includes at least one of a first specific column and a second specific column, wherein the first specific column is a column indicating whether a produced product is defective, and the second specific column may be a column indicating work order information of a produced product.

[0015] In an embodiment, the value corresponding to the first specific column is configured to have a first value or a second value depending on whether the produced product is defective, and the artificial intelligence model can be trained to predict the value for the first specific column.

[0016] In the embodiment, the value corresponding to the second specific column may be configured to have different identification values ​​depending on the work order information of the produced product.

[0017] In an embodiment, normal data corresponding to a normal product among the produced products may be configured to have the first value, and defective data corresponding to a defective product among the produced products may be configured to have the second value.

[0018] In an embodiment, the method may further include the step of specifying handwritten data configured to include data corresponding to a produced product, and the step of mapping the handwritten data to the process data based on the production time of the produced product.

[0019] In an embodiment, the method may further include the step of extracting a plurality of inter-process time interval data and investigation time data using at least one of the process data and the handwritten data.

[0020] In the embodiment, the process data may further include the plurality of time interval data between processes and the investigation time data.

[0021] In an embodiment, in the step of configuring the training data set and the verification data set, at least a portion of the plurality of records may be included in each of the training data set and the verification data set according to the preset configuration criteria, such that the configuration ratio of a first record containing the first value for the first specific column and a second record containing the second value for the first specific column among the plurality of records satisfies the preset configuration ratio criteria.

[0022] In an embodiment, the pre-set configuration criteria may be related to configuring the training data set and the validation data set such that, based on the value corresponding to the first specific column and the value corresponding to the second specific column among the plurality of columns, the number of first records including the first value for the first specific column and the number of second records including the second value for the first specific column satisfy the pre-set configuration ratio criteria.

[0023] In the embodiment, in the step of configuring the training data set and the verification data set, a first record including the first value for the first specific column and a second record including the second value for the first specific column may be included in the training data set and the verification data set, respectively, based on the value corresponding to the second specific column.

[0024] In an embodiment, the configuration of the training data set and the validation data set may be performed by simultaneously considering the value corresponding to the first specific column and the value corresponding to the second specific column.

[0025] In the embodiment, the artificial intelligence model may include a model based on the GBDT (Gradient Boosting Decision Tree) algorithm.

[0026] In the above training step, among a plurality of records included in the above training data set, a weight is assigned to a second record containing the second value for the first column, and the artificial intelligence model can perform training on the above training data set by reflecting the weight assigned to the second record containing the second value for the first column.

[0027] In an embodiment, the learning data set includes a plurality of inter-process time interval data and investigation time data extracted using at least one of the process data and the manual data, and the artificial intelligence model can perform learning on the plurality of inter-process time interval data and the investigation time data.

[0028] A control method for a process defect prediction system according to the present invention may include: receiving input data corresponding to a product to be determined as abnormal among products being produced; inputting the input data into an artificial intelligence model trained with a training data set divided based on a preset configuration criterion among process data composed of normal product data and defective product data; obtaining a prediction value for the input data from the artificial intelligence model; and determining whether the product is abnormal using the obtained prediction value.

[0029] A process defect prediction system according to the present invention, comprising a memory configured to store executable instructions and one or more processors configured to perform operations by executing one or more instructions, specifies process data configured to include data corresponding to a produced product, constructs a training data set and a verification data set using the process data, trains an artificial intelligence model to be trained using the training data set, obtains an artificial intelligence model trained with the training data set based on the training, inputs the verification data set to the trained artificial intelligence model, obtains a prediction value for the verification data set from the trained artificial intelligence model, and determines whether at least one product included in the verification data set is defective using the prediction value for the verification data set.

[0030] A program according to the present invention is a program that is executed by one or more processes in an electronic device and can be stored on a computer-readable recording medium, and may include instructions for performing steps such as specifying process data configured to include data corresponding to a produced product; configuring a learning data set and a verification data set using the process data; training an artificial intelligence model to be trained using the learning data set; acquiring an artificial intelligence model trained with the learning data set based on the learning; inputting the verification data set into the trained artificial intelligence model; obtaining a predicted value for the verification data set from the trained artificial intelligence model; and determining whether at least one product included in the verification data set is defective using the predicted value for the verification data set.

[0031] As described above, according to the process defect prediction method and system of the present invention, an artificial intelligence model can be constructed to determine whether a product is defective by utilizing various process data collected during the product production process. In other words, the present invention enables cost reduction and process optimization by saving product inspection costs and time through process defect prediction using an artificial intelligence model.

[0032] In addition, according to the process defect prediction method and system of the present invention, by constructing an anomaly detection model considering class imbalance and determining in advance the possibility of defect occurrence during the production process using the constructed anomaly detection model, the burden of the inspection process can be reduced and overall manufacturing quality control costs can be reduced. That is, the present invention can create an environment where bubble formation, thermal deformation, and crack formation can be prevented by finding appropriate process time and inspection time for each production line in the display process and applying them appropriately through PID control.

[0033] Furthermore, according to the process defect prediction method and system of the present invention, by configuring a training data set and a validation data set based on pre-set configuration criteria, the ratio of each class is maintained equally in the training and validation sets, and the model can learn normal data and defective data evenly in work order units.

[0034] Furthermore, according to the process defect prediction method and system of the present invention, when configuring the training data set and the verification data set, the work order information of the product and the defect status are considered simultaneously, thereby ensuring not only the balance of normal and defective data in the training and verification data sets but also the diversity of process conditions. That is, the present invention enables the model to learn the diversity of data characteristics between processes without being biased toward a specific process, thereby improving generalization performance under various production conditions when applied to actual processes.

[0035] FIG. 1 is a conceptual diagram illustrating a process defect prediction system according to the present invention.

[0036] FIGS. 2A, FIGS. 2B, FIGS. 2C, and FIGS. 2D are conceptual diagrams for illustrating an embodiment of a display production process.

[0037] FIGS. 3a and FIGS. 3b are flowcharts illustrating a process defect prediction method according to the present invention.

[0038] FIGS. 4a, FIGS. 4b, FIGS. 4c, FIGS. 4d, FIGS. 4e, FIGS. 4f, FIGS. 4g, FIGS. 4h, FIGS. 4i, FIGS. 4j, FIGS. 5a, FIGS. 5b, FIGS. 5c, FIGS. 6a, FIGS. 6b, FIGS. 6c, FIGS. 6d, FIGS. 7a, FIGS. 7b, FIGS. 8a, FIGS. 8b, FIGS. 8c, FIGS. 9a, FIGS. 9b, FIGS. 10a, FIGS. 10b, FIGS. 11a, FIGS. 11b, and FIGS. 12 are conceptual diagrams for explaining a process defect prediction method according to the present invention.

[0039] Hereinafter, embodiments disclosed in this specification will be described in detail with reference to the attached drawings. Identical or similar components are assigned the same reference number regardless of the drawing symbols, and redundant descriptions thereof will be omitted. The suffixes "module" and "part" used for components in the following description are assigned or used interchangeably solely for the ease of drafting the specification and do not have distinct meanings or roles in themselves. Furthermore, in describing the embodiments disclosed in this specification, if it is determined that a detailed description of related prior art could obscure the essence of the embodiments disclosed in this specification, such detailed description will be omitted. Additionally, the attached drawings are intended only to facilitate understanding of the embodiments disclosed in this specification; the technical concept disclosed in this specification is not limited by the attached drawings, and it should be understood that they include all modifications, equivalents, and substitutions that fall within the spirit and technical scope of the present invention.

[0040] Terms including ordinal numbers, such as first, second, etc., may be used to describe various components, but said components are not limited by said terms. These terms are used solely for the purpose of distinguishing one component from another.

[0041] When it is stated that one component is "connected" or "connected" to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. On the other hand, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between.

[0042] A singular expression includes a plural expression unless the context clearly indicates otherwise.

[0043] In this application, terms such as “comprising” or “having” are intended to specify the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0044] The present invention relates to a method and system for predicting process defects. The process defect prediction system according to the present invention may be a method and system for predicting process defects capable of analyzing various production data collected during a product production (or manufacturing) process using an artificial intelligence model and determining whether there is an abnormality in the product.

[0045] The process defect prediction system according to the present invention includes an artificial intelligence model, and the present invention aims to solve the class imbalance problem and predict whether a product is normal or defective. In particular, the present invention aims to provide an artificial intelligence model capable of detecting and predicting defects occurring in a sub-assembly line during the product production process in advance.

[0046] Hereinafter, the present invention will be examined in more detail with reference to the attached drawings. FIG. 1 is a conceptual diagram illustrating a process defect prediction system according to the present invention. FIG. 2a, FIG. 2b, FIG. 2c, and FIG. 2d are conceptual diagrams illustrating an embodiment of a display production process, and FIG. 3a and FIG. 3b are flowcharts illustrating a process defect prediction method according to the present invention. Furthermore, FIGS. 4a, 4b, 4c, 4d, 4e, 4f, 4g, 4h, 4i, 4j, 5a, 5b, 5c, 6a, 6b, 6c, 6d, 7a, 7b, 8a, 8b, 8c, 9a, 9b, 10a, 10b, 11a, 11b, and 12 are conceptual diagrams for explaining a process defect prediction method according to the present invention.

[0047] Meanwhile, the process defect prediction system (100) according to the present invention can determine whether there is an abnormality in the product to be manufactured (20). More specifically, the process defect prediction system (100) can identify defects occurring during the production process of the product to be manufactured (20) and predict whether there is a defect in the product to be manufactured (20). For example, in the present invention, the type of product to be manufactured (20) may be a display product. However, in the present invention, the type of product to be manufactured is not necessarily limited to the examples mentioned, and it is obvious that various other products may be included. For convenience of explanation, the present specification will be described on the premise that the product to be manufactured (20) is a display product.

[0048] In this regard, as illustrated in FIG. 2a, the production of a product proceeds sequentially through a component assembly line, which is a process at the component level, a sub-assembly line that combines components, and a total assembly line that assembles the final product.

[0049] At this time, the process defect prediction system (100) according to the present invention can detect and predict defects occurring in the sub-assembly line during the production process of a product in advance. For example, the sub-assembly line process of a display product may consist of continuous process steps (or processes) including resin application and semi-curing - bonding - degassing - full curing - inspection.

[0050] Resin application and semi-curing is a step of applying resin to bond the display and the glass, and bonding may be a step of bonding (or free-falling) the display and the glass upon completion of resin application and semi-curing. Additionally, degassing is a step of removing residual bubbles present between the display and the glass by raising the temperature within a vacuum chamber, and full curing may be a step of maximizing the bonding strength between the display and the glass by irradiating UV light over the entire surface. Furthermore, inspection may be a step of evaluating quality by conducting an inspection to determine whether the bonding between the display and the glass has been properly achieved.

[0051] In this regard, as illustrated in FIGS. 2b to 2d, the types of defects occurring in the sub-assembly line process of a display product may include at least one of bubble formation, misalignment, and crack formation. Here, “Disp.1” and “Disp.2” may represent two display units (or screens). For example, they may represent the arrangement of displays in a display manufacturing process related to a vehicle HUD, a wearable device, a multi-screen device, etc.

[0052] For example, bubble formation may be a type of defect where bubbles occur when two or more displays are laminated (see Fig. 2b). Additionally, misalignment may occur when the display and glass are misaligned during lamination, and in this case, it may be judged as an appearance defect (see Fig. 2c). Furthermore, crack formation may be a type of defect where thermal deformation occurs due to poor curing during the curing or lamination process (see Fig. 2d).

[0053] That is, the process defect prediction system (100) according to the present invention can identify defects occurring in the sub-assembly line during the product production process and predict major defects of the product.

[0054] Meanwhile, the process defect prediction system (100) according to the present invention may include at least one of an input unit (110), an output unit (120), a communication unit (130), a storage unit (140), a data processing unit (150), an artificial intelligence model (160), and a control unit (170).

[0055] A process defect prediction system (100) according to the present invention may include at least one processor and at least one memory containing computer program code. In this case, the memory may perform the role of the storage unit (140) above. In the present invention, the memory and the program code may cooperate with the processor to perform a series of processes described below.

[0056] Although not illustrated, the process defect prediction system (100) according to the present invention may include one or more processors, and such processors may include one or more general-purpose processors and / or one or more special-purpose processors (e.g., digital signal processor, tensor processing unit (TPU), graphics processing unit (GPU), neural network processing unit (NPU), application integrated circuit, application semiconductor (ASIC), field programmable gate array (FPGA), quantum processing unit (or quantum processor, QPU), etc.). One or more processors may be configured to execute instructions, computer-readable instructions, and / or other instructions described herein that are stored (or included) in the storage unit (140). The process defect prediction method and system according to the present invention may perform data processing described below in cooperation with memory and at least one processor. The processor may perform a series of operations and data processing using data and information stored in memory. In this case, memory may be a component of the storage unit (140).

[0057] In addition, the process defect prediction system (100) according to the present invention can perform data processing and calculation processes using quantum gates, quantum entanglement, and quantum superposition states, taking into consideration implementation in a quantum computer environment. For example, the present invention can perform parallel operations based on qubits, and such quantum operations can operate complementarily with existing classical computers.

[0058] Such quantum computers may include parallel computation using qubits and high-speed data processing devices utilizing quantum entanglement, and hardware-based computational optimization using FPGAs and ASICs is possible. In addition, quantum computers may utilize quantum processors capable of qubit-based parallel computation, and data processing efficiency can be improved through a hybrid structure with existing classical computers.

[0059] Meanwhile, the input unit (110) can be configured in various ways as a means of data input. For example, the input unit (110) can be configured to receive user input. The input unit (110) can be configured to receive user input from a user terminal (not shown). Here, “receiving input” may mean receiving an input signal (or selection signal) corresponding to the user’s input based on input made by the user through the configuration of the input unit provided in the user terminal.

[0060] Here, the user terminal may include at least one of a mobile phone, a smartphone, a notebook computer, a laptop computer, a slate PC, a tablet PC, an ultrabook, a desktop computer, a digital broadcasting terminal, a PDA (personal digital assistants), a PMP (portable multimedia player), a navigation device, and a wearable device (e.g., a smartwatch, a smart glass, a head-mounted display).

[0061] In addition, the input unit (110) in the present invention does not necessarily mean a hardware means, but can be understood as a channel for receiving input from a user.

[0062] The input unit (110) may also be referred to as a user interface module. The input unit (110) may include a touch screen, a computer mouse, a keyboard, a keypad, a touchpad, a trackball, a joystick, a voice recognition module, or other similar devices. However, the present invention does not limit the type of input unit (110).

[0063] Here, user input may include documents, text, images (or videos), voice, etc. In this case, the process defect prediction system (100) may further include a module that converts voice into text.

[0064] Next, the output unit (120) can output information through an output unit configuration (e.g., a display unit, a touch screen, a speaker, etc.) provided in a user terminal linked to the process defect prediction system (100) according to the present invention. For example, the output unit (120) can output a page (or service page) linked to the process defect prediction system (100) according to the present invention to the display unit of the user terminal. In addition, the output unit (120) does not necessarily mean a hardware means, but can be understood as a channel for outputting results to a user.

[0065] Next, the communication unit (130) may be connected via a wireless or wired network to a user terminal, a server (e.g., a central server, an external server, etc.), a device, and at least one network, etc., to receive or transmit overall data and information necessary for the operation of the process defect prediction system (100) according to the present invention.

[0066] The communication unit (130) can support various communication methods depending on the communication standard of the communicating device.

[0067] For example, the communication unit (130) may be configured to communicate with a communication target using at least one of the following technologies: WLAN (Wireless LAN), Wi-Fi (Wireless-Fidelity), Wi-Fi (Wireless Fidelity) Direct, DLNA (Digital Living Network Alliance), WiBro (Wireless Broadband), WiMAX (World Interoperability for Microwave Access), HSDPA (High Speed ​​Downlink Packet Access), HSUPA (High Speed ​​Uplink Packet Access), LTE (Long Term Evolution), LTE-A (Long Term Evolution-Advanced), 5G (5th Generation Mobile Telecommunication), Bluetooth (Bluetooth™ RFID (Radio Frequency Identification), Infrared Communication (Infrared Data Association; IrDA), UWB (Ultra-Wideband), ZigBee, NFC (Near Field Communication), Wi-Fi Direct, and Wireless USB (Wireless Universal Serial Bus).

[0068] Next, the storage unit (140, or memory) serves to store various data related to the present invention and may include one or more non-transient computer-readable storage media that can be read and / or accessed by at least one of one or more processors.

[0069] One or more computer-readable storage media may include volatile and / or non-volatile storage components, such as optical, magnetic, organic, or other memory or disk storage devices. In some examples, the storage unit (140) may be implemented using a single physical device (e.g., one optical, magnetic, organic, or other memory or disk storage device), whereas in other examples, the storage unit (140) may be implemented using two or more physical devices.

[0070] The storage unit (140) may include computer-readable instructions and additional data. The storage unit (140) may include a storage necessary to perform at least some of the methods, scenarios, and techniques described herein and / or at least some of the functions of the device and network.

[0071] Furthermore, at least a portion of the storage unit (140) may be a cloud storage or a cloud server. The storage unit (140) may store at least some of the data corresponding to user input received from the input unit (110) (e.g., process data (200, or production data), handwritten data, etc.) and learning data.

[0072] That is, the storage unit (140) is sufficient as a space where information necessary for the operation of the process defect prediction system (100) according to the present invention is stored, and it can be understood that there are no restrictions on the physical space.

[0073] Furthermore, the storage unit (140) may store a computer program including computer program instructions. Furthermore, the storage unit (140) may store a computer program including computer program instructions that control the operation of the system (100) or control the operation of the control unit (170) when loaded into the processor of the system (100).

[0074] Next, the data processing unit (150) may be configured to perform preprocessing on various process data (200, or production data) collected during the production process of a product to be manufactured.

[0075] The data processing unit (150) can clean the process data (200) to process errors, missing values, and single values ​​in the process data (200), and detect (or identify) and remove outliers or duplicate records (or data). For example, the data processing unit (150) can remove columns that interfere with calculations among the data (or columns) included in the process data (200).

[0076] Additionally, the data processing unit (150) checks the process data (200) based on the confirmed row (or reference row), and if there is data (or value) that is shifted in the process data (200), it can align the shifted data to a normal position. For example, the confirmed row can be understood as a record (or row) that contains both the explicit reference value and the value of the data (or the value of a specific column).

[0077] Furthermore, the data processing unit (150) can extract (or generate) various derived variables (or derived columns, derived data, etc.) and / or features (or features, characteristics, etc.) by using at least one of the process data (200) and handwritten data stored in the storage unit (140, or memory). This process expands (or augments) the data that the artificial intelligence model (160) can learn, and can also be understood as a feature engineering process. By using at least one of the process data (200) and handwritten data, at least one of the time interval between multiple processes (e.g., a first process (ex: “Dam”), a second process (ex: “Fill1”), a third process (ex: “Fill2”), a fourth process (ex: “AutoClave”)), an investigation time, and a total process time can be extracted.

[0078] Meanwhile, the data processing unit (150) can construct a learning data set and a verification (or test) data set by using process data (200) configured to include data corresponding to products produced for each of the multiple different work orders. Here, a work order refers to a production order and / or work data of a process unit, and may be a basic unit for identifying each process data. For example, when manufacturing a display product, the processes proceed sequentially in the order of a first process (ex: “Dam”) - a second process (ex: “Fill1”) - a third process (ex: “Fill2”) - a fourth process (ex: “AutoClave”), and these processes can be grouped into a single work order.

[0079] The data processing unit (150) can divide (or distribute) data in a balanced manner to each of the training data set and the validation data set through stratified distribution so as to resolve the class imbalance (unbalancing) problem and maintain a balance of data distribution among multiple different work orders. More specific details regarding this will be described later.

[0080] Next, the artificial intelligence model (160) may be configured to perform the role of determining whether there is an abnormality in the product by analyzing production data collected during the product production process. The artificial intelligence model (160) may be configured to identify (or detect) defects occurring in the sub-assembly line during the product production process and to predict major defects.

[0081] The artificial intelligence model (160) can determine whether a product is normal or defective based on the characteristics of the input data and output the probability of the product being normal or defective. In this case, the artificial intelligence model (160) can output an anomaly score (or score) that represents the probability of the product being normal or defective as a value between a first value (e.g., “0”) and a second value (e.g., “1”). For example, as a result of the judgment (or determination), the artificial intelligence model (160) can output an anomaly score close to the first value if the product is normal, and output an anomaly score close to the second value if the product is defective.

[0082] Furthermore, in the present invention, the artificial intelligence model (160) may be a model based on the Gradient Boosting Decision Tree (GBDT) algorithm, and may partition data and perform learning based on the decision tree. For example, the artificial intelligence model (160) may be a model based on the CatBoost algorithm. The artificial intelligence model (160) may prevent target leakage problems that may occur in categorical data by using normalization techniques (or methods) such as “Ordered Target Statistics” and / or “Ordered Boosting.” Additionally, the artificial intelligence model (160) may distribute balanced data at each level of the tree using a symmetric tree structure. In the present invention, the artificial intelligence model (160) may also be named a “process defect prediction model,” an “anomaly detection model,” an “anomaly determination model,” or a “CatBoost model.”

[0083] Next, the control unit (170) can perform the role of controlling the overall operation of the process defect prediction system (100) related to the present invention. The control unit (170) can process signals, data, information, etc. that are input or output through the components of the process defect prediction system (100) described above, or perform a series of data processing to provide or process appropriate information and functions to the user. The control unit (170) can be physically implemented by the processor described above. Meanwhile, the process data described above is a set of product unit records, and each record includes multiple columns. The columns include one or more of the following: process name, equipment identifier, time information (timestamp), environmental variables (temperature / humidity / pressure / current, etc.), inspection result, defect status (first specific column), and work order identifier (second specific column). The process data can be stored in a DB table or CSV format, and each record can have a unique product ID.

[0084] Furthermore, the defect status (first specific column) can be expressed as a binary value. In this embodiment, '0' is defined as normal and '1' as defective (the opposite symbol system is also possible, but is applied consistently throughout the specification).

[0085] Furthermore, the work order (second specific column) may be a string or integer identifier identifying identical / similar production conditions or batches. A single work order may contain multiple product records.

[0086] Manual data is data collected from operator inputs or external inspection equipment (e.g., field inspection results, manual measurements) and can be mapped to process data based on product ID or time. Mapping may be performed using one or more of the following: (i) product ID matching, or (ii) a time synchronization method using a production time ±Δt tolerance (e.g., ±5 minutes, ±300 seconds).

[0087] The predicted value is an abnormal (defective) probability or score calculated by an artificial intelligence model, and can be binary determined as abnormal / normal depending on the threshold θ.

[0088] Furthermore, in the present invention, the pre-set configuration criteria may include one or more of the following rules: - Class ratio criterion: Sampling is performed so that the normal:defective ratio in both the training and validation sets satisfies the target ratio (R*) (e.g., R*=80:20, adjustable to 95:5, etc. according to the data distribution).

[0089] - Work Order Separation Criteria: Records belonging to the same work order are assigned to only a single set (training or validation) (preventing data leakage).

[0090] - Time-based: By assigning past periods to learning and future periods to verification based on the production timeline, it ensures generalized evaluation similar to the actual operating environment.

[0091] The above criteria are not mutually exclusive, and the present invention divides by considering at least two of the above criteria simultaneously. That is, priority can be given to a division that satisfies the target class ratio, has no cross-assignment of work orders, and preserves chronological order.

[0092] Furthermore, in the present invention, if necessary to achieve a class ratio, sample weighting or oversampling (SMOTE) may be applied to minority classes (bad) in the training set. Data augmentation is not applied to the validation set (prevention of bias).

[0093] Furthermore, in the present invention, the time interval data between processes is defined as the difference ΔT_i = T_{i+1} - T_i between the time T_i when the same product finishes process i and the time T_{i+1} when the next process i+1 starts. This can be composed of temporal features that reflect physical phenomena such as bottlenecks, waiting, and deviations.

[0094] Investigation time data refers to the time required from input to completion of a specific inspection or investigation process or the investigation cycle / frequency, and is correlated with physical requirements such as thermal / pressure stabilization, curing, and cooling of the process. ΔT_i and investigation time can be characterized as scalar values ​​or statistics (mean / standard deviation / percentile).

[0095] Environmental variables, equipment status, operator input, and event logs (alarms / stops / resets) are aggregated into product records (e.g., average / maximum / outlier counts within a window) after time synchronization and added as columns. Missing values ​​can be handled using process specification-based imputation (e.g., median of sensor valid range) or in-model missing value handling functions (such as missing branch handling in GBDT).

[0096] The artificial intelligence model can be implemented based on GBDT (Gradient Boosting Decision Tree) (e.g., XGBoost / LightGBM). The input can be the feature vector of Section 3, and the target can be the first specific column (defect status).

[0097] Meanwhile, the present invention aims to provide a process defect prediction method and system capable of determining whether a product is abnormal by analyzing various production data collected during the product production process. More specifically, the present invention aims to provide a process defect prediction model capable of resolving class imbalance problems and predicting whether a product is normal or defective. Below, we will examine in more detail the learning method of the process defect prediction model (or system (100)) and the process defect prediction method using the learned process defect prediction model.

[0098] In the present invention, process data configured to include data corresponding to a produced product is specified (S310), and a process of configuring a training data set and a verification data set using the process data can be carried out (S320, see FIG. 3).

[0099] The control unit (170) can specify process data (200) to be used for training the artificial intelligence model (160) to be trained (S401, see FIG. 3b). In this case, the criteria for specifying the process data may vary. The control unit (170) can specify process data (200) to be used for training the artificial intelligence model (160) based on various criteria.

[0100] In one embodiment, the control unit (170) stores various data collected during the production process of a product to be manufactured in the storage unit (140), and can specify the data stored in the storage unit (140) as process data (200).

[0101] In another embodiment, the control unit (170) collects (or receives) data related to a product to be manufactured from at least one of various sources (e.g., a database (DB), a server linked to the process defect prediction system (100), a central server, an external server, etc.) and can specify the collected data set as process data to be used for training an artificial intelligence model (160).

[0102] The process data (200) may include various types of data. For example, as illustrated in FIG. 12, the process data (200) may include at least one of the following: i) process name (process of sub-assembly work during display production (Dam, Fill1, Fill2, AutoClave, etc.), ii) time information (operation Tact Time, Collect Time, etc. of each process), iii) environmental variables (coordinate values ​​of each sensor, speed during resin dispensing, Temp of the Chamber, etc.), and iv) defect status (indicated as Normal in the target column if normal, and Abnormal if defective). Such process data may be in various forms, such as numerical data, categorical data, text data, etc. However, the form of data included in the process data (200) is not necessarily limited to the examples mentioned, and may include additional data in various forms.

[0103] Process data (200) may be configured to include multiple records having values ​​for multiple different columns.

[0104] Here, a record represents at least one unit of data, and a record may contain data values ​​from multiple columns (i.e., multiple fields or attributes). In a database, this can also be referred to as a "row." For example, in an Excel sheet, each row represents a single record, and each column within a row can represent the data values ​​of various columns within the record.

[0105] That is, each piece of data included in a single data set, or a single unit of data containing multiple categories of data values, can be named a “record” or “sample.”

[0106] Furthermore, in the present invention, the column may also be named as a “category,” “categories,” “feature,” “variable,” or “element.”

[0107] Meanwhile, the control unit (170) can perform preprocessing on the process data (200) (S403, see FIG. 3b). The control unit (170) can perform preprocessing on the process data (200) using the data processing unit (150).

[0108] The data processing unit (150) can remove columns that interfere with computation among the columns included in the process data (200). For example, the data processing unit (150) can remove columns (missing values, single values, etc.) that are deemed unnecessary during the learning process of the artificial intelligence model (160). At this time, the data processing unit (150) can perform column removal by increasing the dimensionality of the data to account for interference with learning and slowdown in speed.

[0109] In one embodiment, as illustrated in FIGS. 4a and 4b, the data processing unit (150) can perform column removal when the value of the column data is NaN or when the value of the column data is unique.

[0110] Referring to Fig. 4a, as variables related to the CURE process, “CURE END POSITION X Unit Time_Dam”, “CURE END POSITION Z Unit Time_Dam”, and “CURE END POSITION θ Unit Time_Dam” represent unit time values ​​indicating the position (X, Z, θ coordinates) at the time of the end of the CURE process, and “Unit Time” may represent a relative time or time interval from the time when the position information is recorded. “CURE END POSITION X Judge Value_Dam”, “CURE END POSITION Z Judge Value_Dam”, and “CURE END POSITION θ Judge Value_Dam” may represent the result (quantified judgment value) of determining whether the corresponding position coordinates meet the process reference value or specifications. “CURE SPEED Unit Time Dam” refers to the unit time reference value of speed-related values ​​within the CURE process, such as resin dispensing (e.g., movement speed, curing speed, etc.), and “CURE SPEED Judge Value_Dam” may refer to the result of judging whether the CURE speed meets the reference value. “CURE STANDBY POSITION X Unit Time_Dam” refers to the unit time value of the X-coordinate of the CURE process standby position, and “CURE STANDBY POSITION X Judge Value_Dam” may refer to the result of judging whether the X-coordinate of the standby position meets the reference.

[0111] Additionally, as variables related to the Fill2 process, “Machine Tact time Unit Time_Fill2” refers to the unit time value (Tact Time) required to complete a single task in the Fill2 process, and “Machine Tact time Judge Value_Fill2” may refer to the value determining whether the corresponding Tact Time falls within the standard. “Pallet ID Unit Time_Fill2” refers to the point in time or time interval information where the Pallet ID is recorded, and “Pallet ID Judge Value_Fill2” may refer to the result of determining whether the Pallet ID matches the standard information. “Production Qty Unit Time_Fill2” refers to the unit time value where production quantity information in the Fill2 process is recorded, and “Production Qty Judge Value_Fill2” may refer to the value determining whether the production quantity is appropriate compared to the standard production quantity. “Receip No Unit Time_Fill2” refers to an identification number (based on unit time) capable of identifying production quantity or input information per batch, and “Receip No Judge Value_Fill2” may refer to a value used to determine whether the corresponding Receipt No matches the standard. “WorkMode Unit Time_Fill2” refers to the unit time standard value of the work mode (e.g., Auto / Manual, etc.) of the Fill2 process, and “WorkMode Judge Value_Fill2” may refer to a judgment on whether the work mode meets the set standard.

[0112] In other words, variables containing “Unit Time” are sensor and log data collected over time, variables containing “Judge Value” are qualitative / quantitative indicators that determine whether there is an anomaly relative to a reference value, and “Dam,” “Fill2,” etc. represent process names; even the same variable can be recorded separately by process.

[0113] Also, referring to FIG. 4b, as variables related to the Dam process, “Wip Line_Dam” refers to the WIP line number (based on the Dam process) currently in progress, and “Process Desc_Dam” may refer to a process description or a description of the work type. “Insp. Seq No._Dam” refers to the inspection sequence number, and “Insp Judge Code_Dam” may refer to the inspection judgment code (normal or defective, etc.). “CURE STANDBY POSITION X Collect Result Dam”, “CURE STANDBY POSITION Z Collect Result Dam”, and “CURE STANDBY POSITION θ Collect Result Dam” refer to the actual measured values ​​of the standby position coordinates (X, Z, θ directions), and “CURE START POSITION Z Collect Result Dam” may refer to the measured values ​​of the CURE start position Z-axis. “HEAD NORMAL COORDINATE X AXIS(Stage1) Judge Value_Dam” may represent a judgment value regarding whether the X-coordinate of HEAD at Stage1 is normal.

[0114] In addition, as variables related to the AutoClave process, “Wip Line_AutoClave” refers to the WIP line number of the AutoClave process, and “Process Desc_AutoClave” and “Equipment AutoClave” may refer to the process description and the name of the equipment used. “Insp. Seq No. AutoClave, Insp Judge Code_AutoClave” refers to the inspection sequence and judgment code of the AutoClave process, and “1st Pressure Judge Value AutoClave”, “2nd Pressure Judge Value AutoClave” and “3rd Pressure Judge Value AutoClave” may refer to whether the 1st, 2nd, and 3rd pressure measurements meet the criteria. “GMES_ORIGIN_INSP_JUDGE_CODE Collect Result, Judge Value_AutoClave” may refer to the inspection code result and whether the criteria are met in the Manufacturing Execution System (GMES).

[0115] Additionally, as variables related to the Fill1 process, “Wip Line_Fill1, Process Desc_Fill1” may refer to the WIP line and process description of the Fill1 process, and “Insp. Seq No. Fill1, Insp Judge Code_Fill1” may refer to the inspection sequence and judgment code of the Fill1 process. “HEAD NORMAL COORDINATE X AXIS(Stage) Judge Value_Fill1” may refer to the judgment value (whether it meets normal / defective criteria) of the HEAD X-coordinate (Stage-based).

[0116] Furthermore, as variables related to the Fill2 process, “Wip Line_Fill2, Process Desc_Fill2” may refer to the WIP line and process description of the Fill2 process, and “Insp. Seq No. Fill2, Insp Judge Code_Fill2” may refer to the inspection sequence and judgment code of the Fill2 process. “CURE END POSITION θ”, “CURE STANDBY POSITION X” and “CURE STANDBY POSITION θ”, “START POSITION θ Collect Result_Fill2” may refer to the measurement results of the CURE end, standby, and start positions (based on each axis) in the Fill2 process. “DISCHARGED SPEED OF RESIN Collect Result_Fill2” refers to the measured value of the resin dispensing speed, and “DISCHARGED TIME OF RESIN(Stage1) Collect Result_Fill2”, “DISCHARGED TIME OF RESIN(Stage2) Collect Result_Fill2”, and “DISCHARGED TIME OF RESIN(Stage3) Collect Result_Fill2” may refer to the resin dispensing time (measured value per process stage). “Dispense Volume (Stage1) Collect Result_Fill2”, “Dispense Volume (Stage2) Collect Result_Fill2”, “Dispense Volume (Stage3) Collect Result_Fill2” represent the resin dispensing volume measurement results for each stage, and “HEAD NORMAL COORDINATE X AXIS(Stage1) Judge Value_Fill2” may represent the judgment of normality regarding the Stage1 X-coordinate of the HEAD.

[0117] In other words, Collect Result refers to sensor values ​​or coordinate values ​​measured in the actual process, Judge Value refers to the result of determining normal / abnormal status relative to a reference value, and Insp Judge Code may refer to the final defect judgment in visual and sensor inspections. Additionally, Stage 1, Stage 2, and Stage 3 refer to the distinction of each stage within the multi-stage processing, Wip Line refers to unit information managing the production line flow, and AutoClave may refer to processes such as curing performed in high-temperature and high-pressure environments.

[0118] Such column removal can be performed for each of the data related to multiple processes (e.g., first process (Dam), second process (Fill1), third process (Fill2), fourth process (AutoClave), etc.) included in the process data (200).

[0119] For example, as illustrated in FIGS. 4c and 4d, the data processing unit (150) can remove missing value columns and / or single value columns that are to be removed among the columns corresponding to the first process (e.g., “Dam”). Additionally, as illustrated in FIGS. 4e and 4f, the data processing unit (150) can remove missing value columns and / or single value columns that are to be removed among the columns corresponding to the second process (e.g., “Fill1”).

[0120] As another example, as illustrated in FIGS. 4g and 4h, the data processing unit (150) can remove missing value columns and / or single value columns that are to be removed among the columns corresponding to the third process (e.g., “Fill2”). Additionally, as illustrated in FIGS. 4i and 4j, the data processing unit (150) can remove missing value columns and / or single value columns that are to be removed among the columns corresponding to the fourth process (e.g., “AutoClave”).

[0121] And, the data processing unit (150) checks the process data (200) based on the confirmed row, and if there is data (or value) that is shifted in the process data (200), it can align the shifted data to a normal position. For example, as illustrated in FIGS. 5a to 5c, if there is a value that is shifted in the process data (200) (see FIG. 5b), the data processing unit (150) can align the shifted value to a position corresponding to each of the multiple columns (see FIG. 5c).

[0122] Referring to Fig. 5c, the aligned data represents the results of data collection for the X, Y, and Z axes by stage. These values ​​may be collected during the Dam process and may include sensor data for precisely tracking the position of the HEAD during the process. Here, the “HEAD NORMAL COORDINATE X AXIS (Stage 2) Collect Result_Dam” column contains the X-axis coordinate values ​​of the HEAD in Stage 2 of the Dam process, and the “HEAD NORMAL COORDINATE X AXIS (Stage 3) Collect Result_Dam” column may contain the X-axis coordinate values ​​of the HEAD in Stage 3 of the Dam process. The “HEAD NORMAL COORDINATE Y AXIS (Stage 1) Collect Result_Dam” column contains the Y-axis coordinate values ​​of the HEAD in Stage 1 of the Dam process, and the “HEAD NORMAL COORDINATE Y AXIS (Stage 2) Collect Result_Dam” column may contain the Y-axis coordinate values ​​of the HEAD in Stage 2 of the Dam process. The column “HEAD NORMAL COORDINATE Y AXIS(Stage3) Collect Result_Dam” contains the Y-axis coordinate value of HEAD at Stage 3 during the Dam process, the column “HEAD NORMAL COORDINATE Z AXIS(Stage1) Collect Result_Dam” contains the Z-axis coordinate value of HEAD at Stage 1 during the Dam process, and the column “HEAD NORMAL COORDINATE Z AXIS(Stage2) Collect Result_Dam” can contain the Z-axis coordinate value of HEAD at Stage 2 during the Dam process.These data can be used to i) analyze the positional accuracy of the HEAD (check whether the HEAD has reached a normal coordinate position at each process stage), ii) compare reference positions for defect prediction (determine whether it is within the error range by comparing with the Judge Value), and iii) evaluate process stability (detect equipment abnormalities based on deviations or outliers in coordinate values).

[0123] Furthermore, the data processing unit (150) may specify handwritten data configured to include data corresponding to a produced product, and map the handwritten data to process data (200) based on the production time (date and time) of the produced product. For example, as illustrated in FIGS. 7a and 7b, the data processing unit (150) may specify handwritten data including data corresponding to a produced product stored in a storage unit (140, or memory), and perform mapping (or merging) on ​​the handwritten data based on the date and time. Here, “Equipment_Dam” means the name of the equipment, “start_datetime” means the start time of the process, and “end_datetime” may mean the end time of the process.

[0124] The data processing unit (150) can extract (or generate) various derived variables and / or features using at least one of process data (200) and handwritten data (see FIG. 6d).

[0125] Referring to Fig. 6d, among the various derived variables and / or features extracted, “Workorder_Dam” refers to the work order information (work order number) of the Dam process, and “Workorder_Dam_slice_6” may refer to a specific string slice value of the work order. “PalletID Collect Result_Dam” refers to the pallet ID of the Dam process, and “process_time” may refer to the time taken for the entire process flow. “HEAD NORMAL COORDINATE X AXIS(Stage2)” refers to the HEAD X-coordinate (based on Stage2) of the Dam process, and “Production Qty Collect Result_Dam” may refer to the production quantity of the Dam process. “Collect Date_Dam” refers to the collection date and time of the Dam process, and “Machine Tact Time Collect Result_Fill2” may refer to the Tact Time of the Fill2 process. “Chamber Temp. Collect Result_AutoClave” refers to the measured temperature of the AutoClave chamber, and “1st Pressure Collect Result_AutoClave” may refer to the measured first pressure of the AutoClave. “DISCHARGED TIME OF RESIN(Stage3) Collect Result_Fill1” refers to the resin dispensing time in Fill1 Stage3, and “Machine Tact Time Collect Result_Dam” may refer to the Tact Time of the Dam process. “Collect Date_Fill1” refers to the date and time of collection for Fill1, and “Machine Tact Time Collect Result_Fill1” may refer to the Tact Time of the Fill1 process.“1st Pressure AutoClave” refers to the 1st pressure reference value or Judge value, and “HEAD NORMAL COORDINATE Y AXIS (Stage 2)” may refer to the Y-coordinate of the Dam process HEAD (Stage 2 reference).

[0126] For example, time-related variables (e.g., Time_diff_3_4, Collect Date, process_time, etc.) can indicate that the time interval between processes is strongly associated with defects. Additionally, physical variables such as coordinate information, pressure, and temperature are included as important variables, indicating that equipment position accuracy and the maintenance of process conditions influence the occurrence of defects. Furthermore, variables such as Workorder and PalletID can be used to detect whether problems with a specific lot or equipment are recurring.

[0127] In one embodiment, as illustrated in FIGS. 6a and 6c, the data processing unit (150) can extract a plurality of time interval data between processes (or time interval data for each process) using at least one of process data (200) and manual data. The plurality of time interval data between processes may include at least one of a first time interval data (ex: “Time_diff_1_2”) between a first process (ex: “Dam”) and a second process (ex: “Fill1”), a second time interval data (ex: “Time_diff_2_3”) between a second process (ex: “Fill1”) and a third process (ex: “Fill2”), and a third time interval data (ex: “Time_diff_3_4”) between a third process (ex: “Fill2”) and a fourth process (ex: “AutoClave”).

[0128] Here, “Collect Date Dam” may refer to the data collection time of the Dam process, and “Collect Date_Fill1” may refer to the data collection time of the Fill1 process. “Collect Date_Fill2” may refer to the data collection time of the Fill2 process, and “Collect Date_AutoClave” may refer to the data collection time of the AutoClave process. “Time_diff_1_2” may refer to the time taken (in hours) between the Dam → Fill1 process, “Time_diff_2_3” may refer to the time taken between the Fill1 → Fill2 process, and “Time_diff_3_4” may refer to the time taken between the Fill2 → AutoClave process. These derived variables and / or features can be utilized to i) detect process anomalies or bottlenecks by detecting cases where the time taken between each process is too long or too short, ii) detect time-based causes of defects in the process flow, or iii) improve the accuracy of normal or defect predictions by using them as important input variables for a model.

[0129] In another embodiment, as illustrated in FIG. 6b, the data processing unit (150) can extract an investigation time (e.g., “investigation_time”) using handwritten data. The data processing unit (150) can merge the handwritten data with process data (200) and extract features (e.g., investigation time) in light intensity and light quantity using a photometric relationship. In this case, the investigation time can be extracted through the calculation of “investigation energy (Fill cure energy (mJ))” and “investigation output (or power, Fill cure power (mW))”.

[0130] Here, “Fill cure energy (mJ) - Every break time” refers to the Fill curing energy consumed in each break (unit process short circuit), and “Fill cure power (mW) - Every break time” refers to the power (in milliwatts) used for Fill curing in that section. “investigation_time” refers to the curing time required for each break ((in seconds) = Energy / Power), which can represent the investigation time. The investigation time can be utilized i) to determine the accurate curing time through energy-power-based time calculation, ii) to analyze the correlation with defect occurrence conditions (if the curing time is too long or too short, the probability of defects increases), or iii) as a process optimization factor based on energy efficiency.

[0131] In another embodiment, as illustrated in FIGS. 8a to 8c, the data processing unit (150) can additionally extract various features in addition to the features mentioned above.

[0132] Referring to FIGS. 8b and 8c, “Workorder_Dam_slice_2” may refer to the first two digits of the Workorder extracted, and “Workorder_Dam_slice_6” may refer to the first six digits of the Workorder extracted. “1st Pressure AutoClave” may refer to the first pressure measurement value (kPa, etc.), and “2nd Pressure AutoClave” may refer to the second pressure measurement value. “3rd Pressure AutoClave” may refer to the third pressure measurement value, and “Chamber Temp Time” may refer to the temperature maintenance time inside the chamber (unit: minutes or hours). “Stage1 Circle value”, “Stage2 Circle value”, and “Stage3 Circle value” may refer to the Circle motion control values ​​of each Stage, and “Stage1 Line value”, “Stage2 Line value”, and “Stage3 Line value” may refer to the Linear motion control values ​​of each Stage. “THICKNESS value” refers to the thickness of the resin or lamination layer (unit: mm or μm), and “DIS_per_TIME1” may refer to the transport distance (or speed) per unit time.

[0133] Furthermore, the features used (or utilized) for training the artificial intelligence model (160) in the present invention may include various features such as “Equipment”, “Model_suffix”, “Workorder”, “Collect_date”, “CURE coordinate value”, “CURE speed”, “DISCHARGED speed and time value”, “Dispense Volume value”, “HEAD Purge coordinate value”, “Machine Tact Time”, “PalletID”, “Product QTY”, “Receip No”, “WorkMode”, “Pressure time and unit value”, “Chamber Temp and unit value”.

[0134] Through this, the process data (200) used for training the artificial intelligence model (160) may further include values ​​corresponding to the derived columns (multiple inter-process time interval data, investigation time data, etc.) generated through the process described above, as well as various features and / or data and / or variables related to pressure, temperature, total process time, etc.

[0135] Meanwhile, the control unit (170, or data processing unit (150)) can configure a learning data set and a verification data set by using process data (200) configured to include data corresponding to products produced for a plurality of different work orders.

[0136] In this case, the process of configuring the training data set and the verification data set in the present invention can also be understood as a process of dividing the data (or records) included in the process data (200) and including (or distributing) the divided data in each of the training data set and the verification data set (see FIG. 9a).

[0137] In addition, the process of configuring the training data set and the verification data set in the present invention may divide (or distribute) the normal data and defective data included in the process data (200) according to a characteristic (or feature) called a work order, and distribute the number of normal data and defective data included in the training data set and the verification data set evenly to satisfy a predetermined configuration ratio standard.

[0138] The control unit (170) can configure the training data set and the validation data based on the preset configuration criteria. Here, the preset configuration criteria may relate to configuring the training data set and the validation data set based on the value corresponding to at least one column among a plurality of columns.

[0139] The control unit (170) can specify at least one column among a plurality of columns that serves as a standard for constituting a training data set and a verification data set. For example, at least one column may include at least one of a first specific column (e.g., “Target”, 201) and a second specific column (e.g., “Workorder_Dam_slice_2”, 202) (see FIG. 9b). In this case, the first specific column (201) may be a column indicating whether the produced product is defective, and the second specific column (202) may be a column indicating work order information of the produced product (or work order information of the product processed in at least one of the plurality of processes).

[0140] The value corresponding to the first specific column (201) may have a first value or a second value depending on whether the produced product is defective. More specifically, normal data (i.e., normal product data) corresponding to a normal product among the produced products may have a first value (e.g., “0”) in the process data (200), and defective data (i.e., defective product data) corresponding to a defective product among the produced products may have a second value (e.g., “1”) in the process data (200). As another example, normal data may be data with a normal label (or label) assigned to it corresponding to (or corresponding to) a normal product, and defective data may be data with a defective label assigned to it corresponding to a defective product.

[0141] Additionally, the value corresponding to the second specific column (202) may be configured to have different identification values ​​(or work order identification values) depending on the work order information of the produced product. For example, the work order information may include information related to at least one of the following: the production time of the product (production year, production date and time, etc.), equipment information, process name, product name, sensor value (temperature, pressure, speed, etc.), resin dispensing speed, location, time, identification code (or classification code), work quantity, production line (or manufacturing line), process conditions, product characteristics, product type, and whether there is a defect.

[0142] In this regard, the process defect prediction system (100) according to the present invention can slice work order data and train an artificial intelligence model (160) using the sliced ​​work order data.

[0143] Slicing may include a data processing process that divides work order data into units (or intervals) according to time (date) and product characteristics to form training data sets and validation data sets, and extracts or applies specific patterns when training a model and / or inferring (predicting) based on this.

[0144] For example, in the present invention, work order data is sliced ​​according to criteria based on date and / or product characteristics to identify patterns of defect rates occurring in each slice unit (e.g., time, product combination, etc.), and based on this, it can be utilized to train machine learning models such as CatBoost models (i.e., a prediction model reflecting the pattern for each slice can be built). In addition, in the present invention, when new work order data is input into a trained model, the probability of defect occurrence can be predicted by applying feature combinations of the same work order type (e.g., date range, product combination, etc.) that were previously sliced ​​and trained.

[0145] The process defect prediction system (100) can slice multiple different work order data according to pre-set criteria (e.g., production time, product characteristics, etc.) and train an artificial intelligence model (160) for each sliced ​​work order unit. That is, the process defect prediction system (100) can slice process data (200) into work order units and train an artificial intelligence model (160) based on the production time and characteristics of the products included in each work order unit.

[0146] For example, data sliced ​​by multiple different work order units may include at least one sequence flow and may generate a feature that includes process data such as the process flow of a specific work order, time information (time sequence information, time interval between each process, etc.), sensor values, environmental variables, and whether there is a defect.

[0147] In this case, when new process data (or input data corresponding to a product to be determined as abnormal) is input into the process defect prediction system (100), the new process data can be sliced ​​into work order units and input into the artificial intelligence model (160). Even when new process data is input, the trained artificial intelligence model (160) can predict whether the product is defective based on the pre-trained work order slicing features. That is, the feature extraction method generated during training is applied as is during inference, thereby ensuring consistent performance.

[0148] As such, the present invention enables the model to learn by reflecting the time flow between each process through work order slicing, thereby allowing the model to process individual work orders as a single data point without being biased by specific products or process combinations, and to detect patterns of abnormal situations more accurately. That is, the present invention allows for the construction of a prediction model specialized for each segment by slicing work order data based on time and / or product information. Through this, the model can learn and predict quality abnormality patterns more precisely based on the time series of the data or product characteristics.

[0149] In one embodiment, the value corresponding to the second specific column described above is a value representing work order information of a product processed in a specific process among a plurality of processes, and may include an identification value configured to distinguish or identify products produced in the same process unit.

[0150] In another embodiment, the value corresponding to the second specific column may include a work order identification value of the process step in which the product is processed, so as to be able to identify the manufacturing process or production line flow containing the produced product.

[0151] Meanwhile, the control unit (170) can analyze process data (200) to form a learning data set and a verification data set.

[0152] Here, “analyzing process data” can be understood as understanding multiple records (or data) included in process data (200) and determining (or analyzing) what value each record has based on the results of understanding.

[0153] As seen above, the records may contain data values ​​corresponding to each category. More specifically, the control unit (170) analyzes a plurality of records included in the process data (200) based on the first column, and based on the analyzed results, classifies a first record (i.e., normal data) containing a first value for the first column and a second record (i.e., defective data) containing a second value for a specific column among the plurality of records.

[0154] For example, let us assume that there are 40,506 records included in the process data (200), as illustrated in FIG. 9b. Based on the analysis results of the process data (200), the control unit (170) classifies a record containing a first value for a first specific column (e.g., “Target”, 201) among the records included in the process data as a first record, and a record containing a second value for a first specific column (201) as a second record, and can determine the number of classified first records (e.g., “32,360”) and the number of second records (e.g., “8,200”), respectively.

[0155] As seen above, based on the fact that products are produced for multiple different work orders, the normal data and defective data included in the process data (200) may have different distributions for each different work order. That is, the multiple records (or data) included in the process data (200) may have different distributions depending on the work order information of the produced products.

[0156] For example, as illustrated in FIG. 9b and FIG. 10a, among the records included in the process data (200), the number of records having a first identification value (e.g., “3G”) based on the first production year (e.g., 2023) may be 4,631, and the number of records having a second identification value (e.g., “3M”) may be 4,574. In this case, among the plurality of records having a first identification value based on the first production year, there may be 4,197 first records containing a first value for the first specific column (201), and 434 second records containing a second value for the first specific column (201). Additionally, among the plurality of records having a second identification value based on the first production year, there may be 4,240 first records containing a first value for the first specific column (201), and 334 second records containing a second value for the first specific column (201).

[0157] As another example, among the records included in the process data (200), the number of records having a first identification value (e.g., “4B”) based on the second production year (e.g., 2024) may be 3,062, and the number of records having a second identification value (e.g., “4E”) may be 2,742. In this case, among the multiple records having a first identification value based on the second production year, there may be 2,832 first records containing a first value for the first specific column (201), and 230 second records containing a second value for the first specific column (201). Additionally, among the multiple records having a second identification value based on the second production year, there may be 2,529 first records containing a first value for the first specific column (201), and 213 second records containing a second value for the first specific column (201).

[0158] However, the identification value examined above may also be understood as representing a specific work order type, and in the present invention, the term “record having an identification value” may also be understood as representing data included in (or belonging to) a specific work order (or a specific work order type, a specific work order item, a specific work order group, etc.). For example, based on a first production year (e.g., 2023), a record having a first identification value (e.g., “3G”) may be a record included in the first work order, and a record having a second identification value (e.g., “3M”) may be a record included in the second work order.

[0159] That is, multiple records included in the process data (200) may have different distributions for different work orders (or work order types) (i.e., the total number of records, the number of first records and second records, etc., may have different distributions).

[0160] When the analysis of the process data (200) is completed, the control unit (170) may include at least some of the plurality of records in each of the training data set (210) and the verification data set (220) according to a preset configuration standard, such that the composition ratio of the first record containing a first value for the first specific column (201) and the second record containing a second value for the first specific column (201) among the plurality of records included in the process data (200) satisfies a preset configuration ratio standard.

[0161] Here, the pre-set configuration criteria may be related to configuring the training data set (210) and the validation data set (220) such that, based on the value corresponding to the first specific column (201) and the value corresponding to the second specific column (202) among a plurality of columns, the number of first records containing the first value for the first specific column (201) and the number of second records containing the second value for the first specific column satisfy the pre-set configuration ratio criteria.

[0162] Additionally, the pre-set configuration ratio criteria can be set in various ways. For example, the pre-set configuration ratio criteria can be understood as a ratio of 8:2, with the training data set “8” and the verification data set “2”. However, the pre-set configuration ratio criteria are not necessarily limited to this and can be changed in various ways by the manager (or user) of the process defect prediction system (100) or the process defect prediction system (100) itself.

[0163] The control unit (170) can enable a first record containing a first value for a first specific column (201) and a second record containing a second value for a first specific column to be included in the training data set (210) and the verification data set (220), respectively, based on the value corresponding to the second specific column (202).

[0164] Specifically, in the process of configuring the training data set (210) and the verification data set (220), the control unit (170) may simultaneously consider the value corresponding to the first specific column (201) and the value corresponding to the second specific column (202) so that the balance (or ratio) of the first record containing the first value and the second record containing the second value for the first specific column (201), as well as the distribution of records (first record and second record) included in each of the plurality of different work orders, satisfy the pre-set configuration criteria.

[0165] More specifically, the control unit (170) can configure the training data set (210) and the verification data set (220) based on the value corresponding to the first specific column (201) and the value corresponding to the second specific column (202) so that the ratio of the first record having a first value and the second record having a second value for the first specific column (201) satisfies a preset configuration standard (e.g., preset configuration ratio standard), while maintaining the distribution (or ratio) of the multiple records included in each of the multiple different work orders evenly (or uniformly, without imbalance).

[0166] That is, the control unit (170) can configure the training data set (210) and the verification data set (220) based on the combination unit of the first specific column (201) and the second specific column (202) so that the distribution of records having different values ​​according to the combination of the first specific column (201) and the second specific column (202) is maintained evenly (or uniformly) in each of the training data set (210) and the verification data set (220).

[0167] For example, as illustrated in FIG. 9b and FIG. 10a, the control unit (170) can specify, based on a preset configuration standard, the number of first records (e.g., “25845”) and the number of second records (e.g., “6560”) to be included in the learning data set (210) among a plurality of records (e.g., 40506) included in the process data (200). Then, based on the number of first records and second records to be included in the learning data set (210) being specified, the control unit (170) can configure the learning data set (210) by including the first records and second records in the learning data set (210) in the specified number.

[0168] As another example, as illustrated in FIG. 9b and FIG. 10b, the control unit (170) can specify, based on a preset configuration standard, the number of first records (e.g., “6461”) and second records (e.g., “1640”) to be included in the verification data set (220) among a plurality of records (e.g., 40506) included in the process data (200). Then, based on the number of first records and second records to be included in the verification data set (220) being specified, the control unit (170) can configure the verification data set (220) by including the first records and second records in the verification data set (220) in the specified number.

[0169] In this way, the present invention configures a training data set and a validation data set based on pre-set configuration criteria, thereby maintaining a uniform ratio of normal and defective classes in each of the training and validation sets, and enabling the model to uniformly learn normal and defective data on a work order basis. Furthermore, when configuring the training data set and the validation data set, the present invention simultaneously considers a first specific column and a second specific column, thereby ensuring not only the balance of normal and defective data in the training and validation data sets but also the diversity of process conditions. This allows the model to avoid bias toward a specific process and to robustly learn the diversity of data characteristics between processes, thereby improving generalization performance under various production conditions when applied to actual processes.

[0170] Meanwhile, in the present invention, a process may be performed to train an artificial intelligence model to be trained using a training data set (S330), and based on the training, to obtain an artificial intelligence model trained with the training data set (S340, see FIG. 3).

[0171] The control unit (170) can train the artificial intelligence model (160) to be trained using the training data set (210) (S405, see FIG. 3b). As seen above, the artificial intelligence model (160) may be a model based on the Gradient Boosting Decision Tree (GBDT) algorithm, and may split data and perform training based on the decision tree. For example, the artificial intelligence model (160) may be a model based on the CatBoost algorithm. However, it is obvious that the training method according to the present invention is not necessarily limited to models based on the GDBT algorithm and can be applied to various models.

[0172] Here, the control unit (170) can set parameters of the artificial intelligence model (160) so that the artificial intelligence model (160) can properly process (or learn) the training data set (210).

[0173] In one embodiment, the control unit (170) sets the maximum depth of the tree (depth, the maximum depth of the tree is set by assigning a value of “4” so that the artificial intelligence model (160) learns patterns from the training data set (210) with an appropriate amount of branches and does not overfit), ii) sets the number of boosting rounds to train (iterations, training is performed using many rounds with a high value of “1137”. To prevent overfitting, the “early_stopping_rounds” value is set to 300 so that training is stopped at an appropriate point if performance does not improve), iii) sets the number of boundaries (bins) used when processing continuous variables (border_count, data is processed finely with a relatively high value of “250”), iv) sets the minimum number of data that can be included in the leaf node (min_data_in_leaf, an intermediate value of “13” is set so that too little data is included in the leaf and overfitting is prevented), v) the complexity level when generating cross-entropy-based features for categorical variables Settings(max_ctr_complexity, increases complexity with a high value of “10” to learn more interactions. Using a high value of “10” allows for the generation of features combining coordinates, time, environment, etc., to identify detailed data patterns), vi) L2 Regularization Factor Settings(l2_leaf_reg, assigns a slightly higher value of “8.2227” to reduce model complexity and prevent overfitting), vii) Importance of the Positive Class when dealing with class imbalance issues(scale_pos_weight, “7.viii) Set to 4505” to apply to an imbalanced dataset with a very small number of positive classes, allowing the model to demonstrate performance without ignoring them; setting a high weight of “7” increases the loss for defective data, encouraging greater focus and helping to better detect defective data), viii) Set the randomness level for feature splitting (assigning a relatively high value of “0.9301” to introduce more randomness and prevent overfitting), ix) Set the proportion of features to use in each boosting step (using “rsm” of “0.6077” to randomly select and use approximately 60% of the total features, preventing overfitting with an appropriate level of randomness), ix) Set the degree to which randomness is added during data sampling, xi) Set the boosting method (set to “Plain’” to use gradient boosting for faster computation speeds), xii) Set the loss function to optimize (set to “Logloss” to suit binary classification problems and optimize prediction probabilities), xiii) Criteria for evaluating model performance xiv) Set the setting (set to “'F1’ to align with the competition rules”), xiv) set a fixed seed for randomness reproduction (“random_seed”, '42’ to ensure the same result with the same data and parameters), xv) set the learning rate during the boosting phase (learning_rate, “0.027481”, a low value to expect high generalization performance), xvi) specify categorical variables (cat_features, applied to reflect the characteristics of the competition data where there are many data points judged to be categorical variables with the corresponding parameter setting).For example, model parameters can be configured such as: xvii) setting the tree growth method (grow_policy, set to “SymmetricTree” to generate a balanced tree and increase computational efficiency), xviii) setting the maximum number of categories when using One-Hot Encoding (one_hot_max_size, set to a low value of “5” to use One-Hot Encoding only when the number of categories is 5 or less, and use CTR (Cross-Entropy) otherwise. This is set to reduce computational costs when processing high-dimensional categorical variables), and xix) specifying the best-performing model during training (use_best_model, provides the optimal model and allows it to be used immediately after training is complete without further adjustments). The parameter configuration of such artificial intelligence models (160, or the CatBoost algorithm) can be carried out with a focus on preventing overfitting due to the characteristics of the data.

[0174] In this case, the control unit (170) may assign a weight to a second record containing a second value for a first specific column (201) among a plurality of records included in the learning data set (210). More specifically, in the present invention, a specific parameter (e.g., “scale_pos_weight”) among the parameters of the artificial intelligence model (160) may be used to increase the contribution of defective data (i.e., the second record), thereby allowing the artificial intelligence model (160) to learn the defective data with greater importance.

[0175] Accordingly, when the artificial intelligence model (160) learns the training data set (210), it can perform training on the training data set (210) by reflecting the weight assigned to the second record containing the second value for the first column. For example, the artificial intelligence model (160) can perform training by reflecting a higher importance on the second record containing the second value for the first column (201) among a plurality of records included in the training data set (210).

[0176] Meanwhile, as seen above, the training data set (210) may further include values ​​corresponding to multiple inter-process time interval data (or columns) and investigation time data (or columns), and various features (or data, columns, etc.) related to pressure, temperature, total process time, etc.

[0177] The artificial intelligence model (160) can perform learning on multiple time interval data between processes and investigation time data included in the training data set (210). That is, in the present invention, by adding the time required for each process, the model can learn situations where defects occur with a higher probability when the process takes a long time, thereby improving performance.

[0178] Additionally, the artificial intelligence model (160) can perform learning on combinations of variables such as time, pressure, and temperature between processes. For example, the artificial intelligence model (160) can learn the probability of defects occurring according to pressure and duration. That is, in the present invention, bubble defects in the display can be detected better by the characteristics of the combination of temperature and pressure of the vacuum chamber.

[0179] Furthermore, the control unit (170) can obtain an artificial intelligence model (160) trained with a training data set (210).

[0180] In this regard, as illustrated in FIGS. 11a and 11b, the control unit (170) may perform (or proceed with) StratifiedKFold verification on the learned artificial intelligence model (160). StratifiedKFold is a cross-validation method used in classification problems, which may be a method of dividing the data such that the proportion of classes in each fold is similar to the proportion of classes in the original entire dataset. For example, the control unit (170) may divide the entire data (validation data) into five folds (e.g., Fold 1 to Fold K) based on the entire data, and for each split, use one fold as the verification dataset (220) and the rest as the training dataset (210) to train the artificial intelligence model (160). After repeating the training and verification a total of N times (e.g., 5 times), the control unit (170) may finally average the results of all splits or evaluate the performance of the artificial intelligence model (160) based on the best split result.

[0181] Meanwhile, the control unit (170) processes the verification data set (220) as input to the learned artificial intelligence model (160) and obtains a predicted value for the verification data set (220) from the learned artificial intelligence model (160) (S407). Then, the control unit (170) can determine whether there is an abnormality in at least one product included in the verification data set (220) using the predicted value for the verification data set (220). That is, the control unit (170) can verify the performance of the learned artificial intelligence model (160) in determining whether there is an abnormality in a product using the verification data set (220) (S409, see FIG. 3b). This learned artificial intelligence model (160) can be configured to predict a value (e.g., defective or normal) for a first specific column (201).

[0182] For example, as illustrated in FIG. 12, a learned artificial intelligence model (160) may infer a verification data set (220) and output an anomaly score representing the probability of a normal or defective product included in the verification data set (220) as a value between a first value (e.g., “0”) and a second value (e.g., “1”). In this case, if the product is normal, it may be represented as Normal corresponding to the first value, and if the product is defective, it may be represented as Abnormal corresponding to the second value.

[0183] Here, “Production Qty Collect Result_Fill2” refers to the actual collected production quantity, and “Production Qty Unit Time_Fill2” may refer to the time interval in which the quantity was recorded. “Production Qty Judge Value_Fill2” refers to the value determining normality relative to the reference quantity, and “Receip No Collect Result_Fill2” may refer to the product identification number or batch number. “Receip No Unit Time_Fill2” refers to the time interval in which the recipe number was recorded, and “Receip No Judge Value_Fill2” may refer to a judgment of normality or defect (whether the recipe criteria are met). “WorkMode Collect Result_Fill2” refers to the process operation mode (manual or automatic, etc.), and “WorkMode Unit Time_Fill2” may refer to the time information in which the operation mode was applied. “WorkMode Judge Value_Fill2” may refer to a judgment on whether the mode setting is suitable for the criteria.

[0184] Meanwhile, in the inference process according to the present invention, the normality or defect status of a product subject to defect determination can be determined through the following steps: receiving input data corresponding to a product among the products being produced that is subject to defect determination (not shown); inputting the received input data into an artificial intelligence model trained with a training data set divided based on a pre-set configuration criterion among process data composed of normal product data and defective product data (S350); obtaining a predicted value for the input data from the trained artificial intelligence model (S360); and determining whether the product is abnormal using the predicted value for the input data (S370).

[0185] In the inference process, the artificial intelligence model (160) can determine whether the product is normal or defective based on the characteristics of the input data and output the probability of the product being normal or defective. In this case, the artificial intelligence model (160) can output an abnormality score that represents the probability of the product being normal or defective as a value between a first value (e.g., “0”) and a second value (e.g., “1”). For example, as a result of the judgment (or determination), the artificial intelligence model (160) can output an abnormality score close to the first value if the product is normal, and output an abnormality score close to the second value if the product is defective.

[0186] The process defect prediction method according to the present invention is described in detail to a level that can be implemented by a typical manufacturing data analysis technician.

[0187] For example, process data may include data collected from a Manufacturing Execution System (MES), a Programmable Logic Controller (PLC), equipment logs, or a sensor network. This data may consist of process name, time information, temperature, humidity, pressure, current, work speed, equipment ID, operator information, etc.

[0188] This data is structured into a standard database table or CSV format, configured so that each row represents an individual product unit record and each column represents a variable (feature).

[0189] The construction of training and validation datasets can be implemented using Python-based scikit-learn or LightGBM / XGBoost frameworks, and if the number of defective data is relatively small, class weight or oversampling (SMOTE) techniques can be applied.

[0190] Based on this configuration, a person skilled in the art can reproduce the entire process of the invention based on the description in the specification.

[0191] Meanwhile, according to the embodiment, the artificial intelligence model can be implemented as a Random Forest, XGBoost, LightGBM, or a neural network-based multilayer perceptron (MLP) in addition to a Gradient Boosting Decision Tree (GBDT).

[0192] Methods for handling class imbalance can also be replaced not only with weight-based learning but also with learning methods using SMOTE (Synthetic Minority Over-sampling Technique), undersampling, or focal loss.

[0193] The composition of process data is not limited to display processes and can be applied to all manufacturing domains where process sensor data can be collected from production facilities, such as semiconductors, secondary batteries, and precision metal processing.

[0194] In addition, the mapping criteria for manual data can be implemented using production sequence numbers, timestamps, or equipment serial numbers in addition to product IDs.

[0195] Furthermore, each step of the present invention can be implemented in hardware (industrial server, controller) or software (cloud AI platform, process monitoring system), and each step can be executed independently or integrally.

[0196] Therefore, the “data configuration,” “training,” and “prediction” stages can be separated into independent functional modules, and each module can be executed on the same system or on different network nodes.

[0197] Furthermore, the present invention relates to a method and system for predicting whether a product is defective using data collected in a manufacturing process, and in particular, to a technology for predicting whether a product is defective by configuring process-specific data into a learning and verification data set considering class imbalance problems and training an artificial intelligence model based on the GBDT (Gradient Boosting Decision Tree) algorithm.

[0198] The process data used in the present invention includes production process information at the individual product level and consists of a plurality of columns and records including process name, production time, equipment information, environmental variables (temperature, humidity, pressure, etc.), inspection results, worker identification information, defect status, and work order information. Each record corresponds to one product, and each column corresponds to a specific feature. Process data can typically be extracted from a Manufacturing Execution System (MES), a Programmable Logic Controller (PLC), or a sensor log database.

[0199] The predefined configuration criterion mentioned in the present invention refers to a data partitioning rule for configuring a training data set and a validation data set, and includes one or more of (i) a criterion for maintaining a constant class ratio (class ratio rule), (ii) a criterion for ensuring that data belonging to the same work order are not simultaneously included in the training set and the validation set (workorder separation rule), and (iii) a criterion for maintaining chronological order and assigning past data to training and subsequent data to validation (time-based rule). Such configuration criteria are intended to reduce bias in the training data and improve the generalization performance of the model.

[0200] The first specific column according to the present invention is a target variable indicating whether a product is defective, and has a first value (e.g., 0) for a normal product and a second value (e.g., 1) for a defective product.

[0201] The second specific column represents product work order identifier information and enables products within the same work order to share the same production conditions. In this case, work orders can be classified based on one or more criteria among batch, production line, and equipment setup conditions.

[0202] In addition, the present invention may utilize manual data in addition to process data. Manual data consists of operator input data, measurements from external inspection equipment, etc., and is mapped to process data based on the product's production time or product ID. This allows actual field operation variables or unstructured input data to be utilized together in the analysis.

[0203] Using process data and manual data, the present invention extracts process interval data and inspection time data.

[0204] Inter-process time interval data refers to the difference in time required for the same product to move between consecutive processes and is used as an indicator to detect process bottlenecks or abnormal delays.

[0205] Inspection time data refers to the difference between input and completion times in an inspection or investigation process and can be used as an indirect indicator of process defects such as thermal deformation, bubbles, and cracks.

[0206] The artificial intelligence model of the present invention is implemented as a learning model based on the GBDT algorithm. In the training phase, to correct class imbalance when the proportion of defective data is low, higher weights may be assigned to the defective class, or oversampling (SMOTE), undersampling, or cost-sensitive learning methods may be applied. Model training may be performed using k-fold cross validation or holdout validation, and F1-score, Recall, Precision, or AUC are used as evaluation metrics.

[0207] The artificial intelligence model of the present invention is trained to predict whether there are defects from a plurality of records included in process data. When a verification data set is input to the trained model, the model outputs a predicted value for each product record. The predicted value is expressed as a probability of defect occurrence or a score, and the abnormality of the product is determined based on a threshold.

[0208] This threshold can be set to minimize the cost function or to satisfy the lower limit of recall. For example, in a process that prioritizes defect detection, the threshold can be adjusted to maintain recall at 0.9 or higher.

[0209] In addition, the method of the present invention can feed back to process control using results from verification data. For example, defect prevention control is possible by activating a PID control loop that automatically adjusts the temperature or pressure of the production line for work orders with a high probability of defect occurrence.

[0210] In addition to the GBDT model, the present invention may be replaced with a Random Forest, XGBoost, LightGBM, or Multilayer Perceptron (MLP)-based deep learning model, which is included in the technical concept of the invention. Furthermore, the present invention is not limited to display manufacturing processes but can also be applied to sensor data-based quality control processes such as semiconductors, batteries, automotive parts, and precision machining.

[0211] The process defect prediction system according to the present invention includes one or more processors and memory, and is composed of a data acquisition module, a set configuration module, a learning module, a verification module, and a prediction and discrimination module.

[0212] The data acquisition module collects data from the MES, PLC, or sensor network, and the set configuration module configures training / validation sets according to previously defined configuration criteria. The training module trains a GBDT-based model, and the validation module evaluates the model's performance using a validation data set. The prediction and discrimination module determines whether the input product data is defective and provides the results to the process control system.

[0213] By the above configuration, the present invention can (1) improve the recall rate of minority classes (defective data) through data partitioning and learning that considers class imbalance, (2) prevent data leakage through data partitioning at the work order level, and (3) detect physical process abnormalities early by utilizing inter-process time intervals and investigation time data as features.

[0214] Therefore, the present invention goes beyond simple data classification to provide a technical effect that improves manufacturing equipment control efficiency, thereby solving the technical problem of reducing process quality control costs and defect rates.

[0215] As described above, according to the process defect prediction method and system of the present invention, an artificial intelligence model can be constructed to determine whether a product is defective by utilizing various process data collected during the product production process. In other words, the present invention enables cost reduction and process optimization by saving product inspection costs and time through process defect prediction using an artificial intelligence model.

[0216] In addition, according to the process defect prediction method and system of the present invention, by constructing an anomaly detection model considering class imbalance and determining in advance the possibility of defect occurrence during the production process using the constructed anomaly detection model, the burden of the inspection process can be reduced and overall manufacturing quality control costs can be reduced. That is, the present invention can create an environment where bubble formation, thermal deformation, and crack formation can be prevented by finding appropriate process time and inspection time for each production line in the display process and applying them appropriately through PID control.

[0217] Furthermore, according to the process defect prediction method and system of the present invention, by configuring a training data set and a validation data set based on pre-set configuration criteria, the ratio of each class is maintained equally in the training and validation sets, and the model can learn normal data and defective data evenly in work order units.

[0218] Furthermore, according to the process defect prediction method and system of the present invention, when configuring the training data set and the verification data set, the work order information of the product and the defect status are considered simultaneously, thereby ensuring not only the balance of normal and defective data in the training and verification data sets but also the diversity of process conditions. That is, the present invention enables the model to learn the diversity of data characteristics between processes without being biased toward a specific process, thereby improving generalization performance under various production conditions when applied to actual processes.

[0219] Meanwhile, the present invention described above can be implemented based on a quantum computer. The present invention implemented based on a quantum computer may include a qubit-based quantum processor and quantum memory, and may include software and hardware interfaces optimized for quantum computation.

[0220] Quantum processors in quantum computers utilize qubits to efficiently process complex operations through parallel computation, quantum entanglement, and quantum superposition, which cannot be performed by the binary bits of classical computers. Quantum processors process data using quantum gates and can provide exponential speed improvements for specific problems.

[0221] Meanwhile, the present invention described above can be implemented as a program that is executed by one or more processes on a computer and can be stored on a computer-readable medium (or recording medium).

[0222] Furthermore, the present invention described above can be implemented as computer-readable code or instructions on a medium on which a program is recorded. That is, the present invention can be provided in the form of a program.

[0223] Meanwhile, computer-readable media include all types of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable media include HDD (Hard Disk Drive), SSD (Solid State Disk), SSD (Silicon Disk Drive), ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.

[0224] Furthermore, the computer-readable medium may be a server or cloud storage that includes a storage and is accessible to an electronic device via communication. In this case, the computer may download the program according to the present invention from the server or cloud storage via wired or wireless communication.

[0225] A computer program may reach the system (100) through various suitable transmission mechanisms. The transmission mechanism may be, for example, a computer-readable storage medium, a computer program product, a memory device, a recording medium such as a CD-ROM or DVD, or a product that tangibly embodies the computer program. The transmission mechanism may be a signal configured to reliably transmit the computer program through air or an electrical connection. The system (100) may propagate or transmit the computer program as a computer data signal.

[0226] Furthermore, references to 'computer-readable storage media,' 'computer program products,' 'computer programs embodied in a tangible form,' etc., or to 'controller,' 'computer,' 'processor,' etc., should be understood to include not only computers with various architectures such as single / multi-processor architectures and sequential (Von Neumann) / parallel architectures, but also specialized circuits such as Field-Programmable Gate Arrays (FPGAs), Application Specific Circuits (ASICs), signal processing units, and other devices. References to computer programs, instructions, code, etc., should be understood to include software for programmable processors or firmware, such as programmable content for hardware devices, whether it is instructions for a processor or configuration settings for a fixed-function device, gate array, or programmable logic device.

[0227] Furthermore, in the present invention, the computer described above is an electronic device equipped with a processor, namely a CPU (Central Processing Unit), and no special limitations are placed on its type.

[0228] Meanwhile, the above detailed description should not be interpreted restrictively in all respects but should be considered exemplary. The scope of the invention should be determined by a reasonable interpretation of the appended claims, and all modifications within the equivalent scope of the invention are included within the scope of the invention.

Claims

1. In a computerized method comprising the following, A step of specifying process data configured to include data corresponding to a produced product; A step of constructing a training data set and a validation data set using the above process data; A step of training a target artificial intelligence model using the above training data set; Based on the above learning, a step of obtaining an artificial intelligence model trained with the above learning data set; A step of inputting the above verification data set into the above-mentioned trained artificial intelligence model; A step of obtaining a predicted value for the verification data set from the above-mentioned learned artificial intelligence model; and A process defect prediction method characterized by including a step of determining whether at least one product included in the verification data set is abnormal using a predicted value for the verification data set.

2. In Paragraph 1, The above process data is, It is configured to include multiple records having values ​​for multiple different columns, and In the step of constructing the above training data set and the above validation data set, A method for predicting process defects characterized by configuring the training data set and the verification data set based on pre-set configuration criteria.

3. In Paragraph 2, The above process data is, It is configured to include at least one of a process name, time information, environmental variables, and defect status, and The above-mentioned pre-established configuration criteria are, A method for predicting process defects characterized by being related to configuring the training data set and the verification data set based on a value corresponding to at least one column among the plurality of columns.

4. In Paragraph 3, The above at least one column includes at least one of a first specific column and a second specific column, and The above-mentioned first specific column is a column indicating whether the produced product is defective, and A method for predicting process defects, characterized in that the second specific column above is a column representing work order information of a produced product.

5. In Paragraph 4, The value corresponding to the above-mentioned first specific column is configured to have a first value or a second value depending on whether the produced product is defective, and The above artificial intelligence model is, A process defect prediction method characterized by being trained to predict the value for the first specific column.

6. In Paragraph 4, A process defect prediction method characterized in that the value corresponding to the second specific column above is configured to have a different identification value according to the work order information of the produced product.

7. In Paragraph 5, Normal data corresponding to a normal product among the produced products is configured to have the above-mentioned first value, and A process defect prediction method characterized by defect data corresponding to defective products among produced products having the second value.

8. In Paragraph 1, A step of specifying handwritten data configured to include data corresponding to a produced product; and A method for predicting process defects characterized by further including the step of mapping the manual data to the process data based on the production time of the produced product.

9. In Paragraph 8, A process defect prediction method characterized by further including the step of extracting a plurality of inter-process time interval data and investigation time data using at least one of the above process data and the above manual data.

10. In Paragraph 9, The above process data includes, A process defect prediction method characterized by further including the above-mentioned multiple process time interval data and the above-mentioned investigation time data.

11. In Paragraph 5, In the step of constructing the above training data set and the above validation data set, A method for predicting process defects characterized by including at least a portion of the plurality of records in each of the training data set and the verification data set, according to the above-mentioned configuration criteria, such that the configuration ratio of a first record containing a first value for a first specific column and a second record containing a second value for a first specific column among the plurality of records satisfies the above-mentioned configuration ratio criteria.

12. In Paragraph 11, The above-mentioned pre-established configuration criteria are, Based on the value corresponding to the first specific column and the value corresponding to the second specific column among the plurality of columns above, A method for predicting process defects characterized by configuring the training data set and the verification data set such that the number of first records containing the first value for the first specific column and the number of second records containing the second value for the first specific column satisfy the preset configuration ratio criteria.

13. In Paragraph 11, In the step of constructing the above training data set and the above validation data set, A method for predicting process defects characterized by including a first record containing a first value for a first specific column and a second record containing a second value for a first specific column in each of the training data set and the verification data set based on a value corresponding to the second specific column.

14. In Paragraph 12, The composition of the above training data set and the above validation data set is, A process defect prediction method characterized by being performed by simultaneously considering the value corresponding to the first specific column and the value corresponding to the second specific column.

15. In Paragraph 1, The above artificial intelligence model includes a model based on the GBDT (Gradient Boosting Decision Tree) algorithm. A process defect prediction method characterized by 16. In Paragraph 5, In the above-mentioned training step, Among a plurality of records included in the above training data set, a weight is assigned to a second record containing the second value for the first column, and The above artificial intelligence model is, A method for predicting process defects characterized by performing learning on a learning data set by reflecting the weight assigned to a second record containing the second value for the first column.

17. In Paragraph 8, The above training data set includes, A plurality of inter-process time interval data and investigation time data extracted using at least one of the above process data and the above manual data are included, The above artificial intelligence model is, A process defect prediction method characterized by performing learning on the above-mentioned multiple process time interval data and the above-mentioned investigation time data.

18. A step of receiving input data corresponding to a product among the products being produced that is subject to determination of whether it is abnormal; A step of inputting the above input data into an artificial intelligence model trained with a training data set divided based on pre-set configuration criteria among process data consisting of normal product data and defective product data; A step of obtaining a predicted value for the input data from the artificial intelligence model; and A process defect prediction method characterized by including a step of determining whether the product is abnormal using the above-mentioned predicted value.

19. A system comprising memory configured to store executable instructions and one or more processors configured to perform operations by executing one or more instructions, The above system is, Specifying process data configured to include data corresponding to the produced product, and Using the above process data, a training data set and a validation data set are constructed, and Train the target artificial intelligence model using the above training data set, and Based on the above learning, an artificial intelligence model trained with the above learning data set is obtained, and Input the above verification data set into the above-mentioned trained artificial intelligence model, and From the above-mentioned trained artificial intelligence model, a predicted value for the above-mentioned verification data set is obtained, and A process defect prediction system characterized by determining whether at least one product included in the verification data set is abnormal using a predicted value for the verification data set.

20. A program that is executed by one or more processes in an electronic device and stored on a computer-readable recording medium, The above program is, A step of specifying process data configured to include data corresponding to a produced product; A step of constructing a training data set and a validation data set using the above process data; A step of training a target artificial intelligence model using the above training data set; Based on the above learning, a step of obtaining an artificial intelligence model trained with the above learning data set; A step of inputting the above verification data set into the above-mentioned trained artificial intelligence model; A step of obtaining a predicted value for the verification data set from the above-mentioned learned artificial intelligence model; and A program stored on a computer-readable recording medium characterized by including instructions for performing a step of determining whether at least one product included in the verification data set is abnormal using a predicted value for the verification data set.

Images