Page 116 - Kỷ yếu hội thảo quốc tế: Ứng dụng công nghệ mới trong công trình xanh - lần thứ 9 (ATiGB 2024)

th
HỘI THẢO QUỐC TẾ ATiGB LẦN THỨ CHÍN - The 9 ATiGB 2024 107

operate facilities efficiently, decision-making and analyze them. Select appropriate data, learn it,
problems are modeled based on deep reinforcement and evaluate accuracy and loss.
learning to enable appropriate decisions. In addition,
we seek optimization methods that can continuously
guarantee product quality even when the process
environment changes and develop them so that they
can be applied anywhere in the process. This requires
top-level AI technology. Each focus point for
equipment, process, and product is determined through
AI, and process data consisting of various types and
factors are utilized and analyzed. Build an AI-based
platform capable of explanation/prediction/
optimization.
3. APPEARANCE DEFECT DETECTION
TECHNOLOGY
Fig 2. Deep learning process
For deep learning-based object detection, an object When detecting a defect in the appearance of a
detection model using YOLO (Single Short Object product, information such as the location and shape of
Detection)-based network and tracking technology is the part presumed to be defective is obtained from the
used. The YOLO object detector has a high frame acquired image, and the cause and analysis are
processing speed per second, making it suitable for performed. In this process, the preprocessing process
real-time object detection [1]. Based on the field is an essential process for image recognition rate [2].
customized dataset, objects are effectively guessed and Duplicate inspection is performed to improve
applied to tracking and detection algorithms. Effective recognition rate, and ID tracking is performed to
defect detection is possible by applying a customized check whether the same object or defect is recognized.
learning routine through building and learning The figure below shows an example of image quality
negative data through continuous data collection and improvement through image preprocessing.
verification. The figure below shows the application of
a deep learning-based object detection algorithm.

Fig 1. Example of application Fig 3. Image quality improvement
of deep learning-based object detection algorithm through image preprocessing
In order to develop an object detection algorithm,
validation of defect impact prediction modeling is
necessary. To this end, the data model of the primary
collected data is identified and the modeling
effectiveness and consistency are verified through
iterative learning for each data type. To select an
easily understandable visualization model, related
visualization functions and pattern analysis are
repeatedly used, and after selecting a deep learning
suitable model, multi-learning is performed using a
Python program. The photo below is an example of
the deep learning process. Set the analysis period and Fig 4. Object detection and tracking
acquire the data set, visualize key quality factor items, Pre-processing may be performed to improve the
recognition rate as above, but there is also a method of

ISBN: 978-604-80-9779-0