KR-20260065321-A - Deep learning-based quality inspection method for reducing overkill and quality inspection system using thereof

KR20260065321AKR 20260065321 AKR20260065321 AKR 20260065321AKR-20260065321-A

Abstract

The present invention relates to a quality inspection system and method for detecting product defects. According to the present invention, in order to solve the problems of conventional machine vision systems and rule-based quality inspection methods, which had limitations such as increased unnecessary costs and reduced overall process efficiency and yield due to overkill where normal materials or non-defective products are judged as defective, a deep learning-based quality inspection system and method for reducing overkill is provided. This is configured so that a process is performed to determine whether a product is good or defective based on a threshold input by a user through the ensemble of existing rule-based inspection and deep learning-based inspection. By compensating for the disadvantages of rule-based inspection and deep learning-based inspection, the system effectively solves the cold start problem of deep learning and reduces the occurrence of overkill. Consequently, the system is configured to reduce unnecessary costs by reducing human resources required for rework and to increase the overall production process efficiency and productivity.

Inventors

이경민
김무진

Assignees

디에스 주식회사

Dates

Publication Date: 20260508
Application Date: 20241101

Claims (14)

In a deep learning-based quality inspection system for reducing overkill, An input unit configured to perform processing for receiving images or videos of products subject to quality inspection; A quality inspection unit configured to perform deep learning-based quality inspection processing using a product image input through the above input unit; and A deep learning-based quality inspection system for reducing overkill, characterized by comprising an output unit configured to perform processing that outputs various information, including data input through the input unit, inspection results of the quality inspection unit, and the operation and current status of the quality inspection system, according to a predetermined setting.
In Article 1, The above quality inspection system is, A database unit configured to perform a process of constructing a database related to quality inspection by storing various information, including the overall processing process and processing results of the above-mentioned quality inspection system, according to predetermined settings; A communication unit configured to perform processing for transmitting and receiving various data by communicating with an external device, including a user terminal or a server, in at least one of wireless or wired communication; and A deep learning-based quality inspection system for reducing overkill, characterized by further comprising a control unit configured to perform processing that controls the overall operation of the quality inspection system.
In Article 1, The above input unit is, A deep learning-based quality inspection system for reducing overkill, characterized by being configured to directly receive video or images of a product captured through a camera or CCTV, or to receive pre-captured video or images from an external source.
In Article 1, The above quality inspection department, A deep learning-based quality inspection system for reducing overkill, characterized by being configured to perform training of a pre-established deep learning model for quality inspection using a product image input through the above input unit, and to perform processing to determine whether there is a defect from a product image input in real time based on the training result.
In Paragraph 4, The above quality inspection department, Rule-based quality inspection is performed on the product image input through the above input unit using a rule-based method, and If the inspection result of the above rule-based quality inspection is a Good Device, the inspection details are used as training data for the above deep learning model, and the next material is inspected. If the inspection result of the above rule-based quality inspection is a defective product (Bad Device), the inspection details are input into the above deep learning model to perform a re-inspection, and The inspection results of the above rule-based quality inspection and the judgment results of the above deep learning model are input into a pre-established ensemble classifier to determine whether the product is good or defective according to predetermined settings or criteria, and A deep learning-based quality inspection system for reducing overkill, characterized by being configured such that when the judgment result of the above ensemble classification model is defective, a process is performed to classify it as rework.
In Paragraph 5, The above quality inspection department, A deep learning-based quality inspection system for reducing overkill, characterized by being configured to input the inspection results of the rule-based quality inspection and the judgment results of the deep learning model into the ensemble classification model, and to perform a process for determining whether a product is good or defective based on a threshold input by a user.
In Paragraph 5, The above quality inspection department, A deep learning-based quality inspection system for reducing overkill, characterized by being configured to input the inspection results of the rule-based quality inspection and the judgment results of the deep learning model into the ensemble classification model and to perform a process for determining whether a product is good or defective based on a preset weighting algorithm.
In Paragraph 4, The above deep learning model is, A deep learning-based quality inspection system for reducing overkill, characterized by being configured using a VAE-based machine learning classifier (VAE-based ML Classifier) that performs processing to determine whether there is a defect by calculating the similarity between original image data and image data reconstructed through a Variational AutoEncoder (VAE) according to predetermined criteria or settings.
In Article 1, The above output unit is, A deep learning-based quality inspection system for reducing overkill, characterized by being configured to perform a process of generating and outputting output data by processing various information regarding inspection content and inspection results, including an image input through the input unit and the quality inspection results of the quality inspection unit, into a corresponding form or format according to a user's request or a predetermined setting.
In deep learning-based quality inspection methods, An input stage in which processing is performed to receive a video or image of a quality inspection target; A quality inspection step in which training of a deep learning model is performed using product images input through the above input step, and processing is performed to determine whether there is a defect from product images input in real time based on the training results; and It is configured to include an output step in which processing is performed to output various information, including data input through the above input step and inspection results of the above quality inspection step, according to a predetermined setting. The above quality inspection step is, A deep learning-based quality inspection method characterized by being configured to perform deep learning-based quality inspection processing using a deep learning-based quality inspection system for reducing overkill as described in any one of claims 1 to 9.
In the quality management system, A quality inspection processing unit comprising a plurality of quality inspection systems in which deep learning-based quality inspection is performed; User terminals for each user to request and receive desired services; and The system is configured to include a server that is linked to each quality management system and the user terminal of the quality inspection processing unit, receives various data including quality inspection results from each quality inspection system, and performs processing to provide a desired service to the user in accordance with a user request received through each user terminal. The above quality inspection system is, A quality management system characterized by being configured to perform deep learning-based quality inspection and transmit and receive various data with a server using a deep learning-based quality inspection system for reducing overkill as described in any one of claims 1 to 9.
In Paragraph 11, The above user terminal is, A quality management system characterized by being configured by installing a dedicated application linked to each of the quality inspection systems and the server on a personal portable information and communication terminal including a smartphone or tablet PC, or an information processing device including a PC or laptop.
In Paragraph 11, The above server is, A quality management system characterized by being configured to collect various information including quality inspection content and results from each of the above-mentioned quality inspection systems, build a database related to quality inspection for each product, and, when a user request is received, perform a process to provide customized information desired by the user through the user's terminal based on the information stored in the database.
In Paragraph 11, The above server is, A process for receiving information on the current status from each of the above quality inspection systems, monitoring the operation of each of the above quality inspection systems, and, in the event of an anomaly, delivering a notification regarding the relevant content to the user through a user terminal according to a predetermined setting, and Processing and providing monitoring information so that each user can check the operation and status of a specific quality inspection system through their user terminal A quality management system characterized by being configured to perform processing that enables remote control by transmitting control commands received from each user to the corresponding quality inspection system.

Description

Deep learning-based quality inspection system and method for reducing overkill The present invention relates to a quality inspection system and method for detecting defects in products produced through a manufacturing process. More specifically, it relates to a deep learning-based quality inspection system and method for reducing overkill, configured to apply deep learning-based quality inspection to existing production equipment with a relatively simple configuration and low cost, thereby reducing overkill and unnecessary costs and improving overall process efficiency and yield. This is achieved by utilizing artificial intelligence technology including deep learning to solve the problems of conventional machine vision systems and rule-based quality inspection methods, which had limitations such as frequent occurrence of so-called overkill—where normal materials or non-defective products are judged as defective during the quality inspection process for detecting defects or defects in products produced through manufacturing processes, for example, semiconductors, etc.—which led to increased unnecessary costs and reduced overall process efficiency and yield. Furthermore, to address the limitations of conventional rule-based machine vision systems and quality inspection methods, which, as mentioned above, suffered from increased unnecessary costs and reduced overall process efficiency and yield due to overkill, the present invention is configured to perform a process in which inspection is conducted using existing rule-based methods; if a device is good, it is used to train a deep learning model, and if it is bad, it is input into the deep learning model for re-inspection as it may be overkill; the two inspection results are input into an ensemble classifier to determine whether the device is good or bad based on a threshold entered by a user; and if the final result is bad, rework is performed. By configuring this process to complement the shortcomings of both rule-based and deep learning-based inspections through ensemble testing, the invention effectively resolves the cold start problem of deep learning and reduces the occurrence of overkill. Consequently, it reduces the human resources required for rework, thereby decreasing unnecessary costs and improving the overall production process efficiency and This relates to a deep learning-based quality inspection system and method for reducing overkill configured to increase productivity. Recently, machine vision systems that identify defects through product images are generally being applied in manufacturing processes for various products, such as semiconductors and various components, for quality inspection of the manufactured products. In addition, as presented in, for example, the "Inspection Method of Semiconductor Device" in Korean Published Patent Application No. 10-2024-0115610 and the "Vision Inspection System Using a CCTV Camera" in Korean Published Patent Application No. 10-2024-0109645, various devices and methods have been proposed in the past to improve the accuracy or performance of machine vision used for quality inspection of various products such as semiconductors. Therefore, as mentioned above, since using a machine vision system offers advantages such as automating existing quality inspection tasks that were previously performed by visual inspection or manual work and increasing the reliability of defect identification, machine vision systems have recently been widely used in quality inspection tasks across various fields; however, the conventional machine vision systems described above had the following limitations. More specifically, generally, conventional machine vision systems and methods as described above have a problem in which overkill frequently occurs, where normal materials are misidentified as defective, for example, when there are incorrect parameter settings during the teaching operation of the machine vision system or technical defects in rule-based inspection. Accordingly, in the conventional semiconductor manufacturing process, when a product is determined to be defective through machine vision, it is initially classified as rework. The classified defective products are then manually re-inspected by humans to classify good products, and rework is performed only on the remaining defective products, thereby mitigating the occurrence of unnecessary overkill. Therefore, the conventional machine vision systems and methods described above had limitations in that the process to compensate for the overkill problem was performed manually, requiring unnecessary human resources and time, which consequently led to a decrease in the overall yield of the semiconductor manufacturing process and an increase in manufacturing costs. Furthermore, conventional machine vision systems and methods require a prior teaching process in which the user manually sets dozens to hundreds of inspection parameters in advance. Additionally, since existing ru