Nguyen Xuan Huy - CH2002007
Welding defects are flaws resulting from deviations in the external shape, severity, and metal structure compared to the design during the welding process. Welding defects can impact the quality and aesthetics of the weld joint. In the production and construction of prefabricated steel for various projects, failure to detect welding defects can compromise the quality and safety of the structure, posing a threat to human lives. In the manufacturing sector, there are two methods for inspecting welding defects: destructive testing and non-destructive testing. Destructive testing is typically carried out in laboratories, while non-destructive testing is prioritized in the production process.
There are various non-destructive testing methods to inspect welding defects, such as ultrasonic testing (UT), X-ray testing (RT), gamma ray testing (GT), eddy current testing (ET), and visual testing (VT). Computer vision, a field applied in sectors like healthcare, security, and business, can also be employed for classifying welding defects based on images.
Image data related to welding defects can vary significantly due to the production process. However, utilizing deep learning methods for classifying welding defects from increasingly large datasets poses several challenges. One of the major challenges is image quality when captured from cameras or smartphones, influenced by factors like the angle of the image, brightness, or obscured weld joints. This can lead to missing information and reduced classification model accuracy. Additionally, the training cost for deep learning models is another challenge. Some deep learning methods require a large volume of data and time for training, and collecting suitable data can be a demanding task. Moreover, building a reliable and accurate welding defect classification model requires careful consideration and diligence in data collection and processing.
To address these challenges, this thesis proposes a method for classifying welding defect images based on big data processing technology. This method utilizes pre-trained models for transfer learning (TL) and applies distributed parallel training to the training data. The data processing framework of Apache Spark combined with the BigDL library helps expedite the training for welding defect detection, making it faster and more accurate.
From the results of Experiment 1 on a thermal welding dataset consisting of 9,058 images, the EfficientNetB0 model showed the lowest results (0.4394, 0.2035, and 0.2683 for Accuracy, Macro F1-score, and Weighted F1-score), while ResNet101 and VGG16 performed better. VGG16 yielded the best results among the five model configurations (with 0.8230 Accuracy, 0.8205 Macro F1-score, and 0.8222 Weighted F1-score). In the task with 7 classes, EfficientNB0 and VGG16 both exhibited poorer results, with the performance of models in the 7-class task being approximately 20% lower than in the 3-class task.
Based on the results of Experiment 2, the average training time for the model was evaluated at xxx seconds for training on 9,058 images over a LAN network compared to xxx seconds over VPN, achieving an accuracy of xx%. With the Data Parallelism method, the results indicated that training on a single GPU was only slightly faster than on multiple GPUs (xxs/epoch versus xxs/epoch). However, when training the model with a large dataset, the processing capabilities of a single GPU may be overwhelmed. The optimal solution in this case is to use parallel training and data distribution. This is one of the most effective solutions for dealing with large datasets.
From these results, it can be concluded that big data processing technology is a suitable approach for classifying images of steel weld defects. Applying this technology enhances accuracy, significantly reduces training data processing time, and helps reduce training costs. Moreover, this method has proven to be effective and practical in real-world applications.
Further information: https://fit.uit.edu.vn/index.php/tin-tuc/goc-hoc-tap/6476-phan-lo-i-hinh-nh-m-i-han-thep-d-a-tren-cong-ngh-x-ly-d-li-u-l-n
Written by: Ha Bang
Translated by: Ngoc Diem
