AI-Based Defect Detection Model Development to Enable Zero Defect Manufacturing in Ethiopian Steel Industries
No Thumbnail Available
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Product quality stands as the primary competing factor in manufacturing industries, especially in the rapidly evolving digital landscape of the current era. Zero Defect Manufacturing (ZDM) has emerged as a state-of-the-art quality enhancement framework in response to Industry 4.0 technologies. This data-centric approach encompasses four crucial strategies: detection, prediction, repair, and prevention. Detection, the cornerstone of ZDM, operates through two distinct methods. The physical method relies on direct product measurements to identify defects, while the virtual approach, known as virtual metrology, conducts quality inspections without physical contact. This novel technology is revolutionizing quality inspection practices in the steel industry on a global scale. Ethiopian steel industries are facing a persistent challenge of producing a significant volume of defective products due to their reliance on manual inspection methods, a practice that is both costly and time-intensive.
The successful integration of virtual detection relies entirely on data accessibility (via robust data infrastructure) and advanced analytics. The pivotal components for achieving virtual defect detection in industries are organizational shifts toward digitalization and the adoption of standardized data analytics (DA) practices. Current literature on ZDM overlooks crucial factors related to industrial digital readiness and fails to analyze the interconnected impact of these factors on the digitization of manufacturing industries. While various digital maturity (DM) models are in use in industrial contexts, there is a lack of analysis regarding the intricate relationships among MM dimensions. Furthermore, the absence of a standardized approach to data analytics poses challenges. Although numerous data analytics frameworks exist beyond ZDM, customizing these methods for ZDM applications can complicate the development of effective AI models. Hence, the primary objective of this dissertation is to develop an AI-driven defect detection model. This will be achieved by initially addressing the challenges outlined in industrial digital readiness and standardized data analytics framework enabling the implementation of virtual defect detection.
To achieve the primary goal, a multi-stage research methodology is employed, encompassing diverse analytical dimensions to tackle different aspects of the research problem. The work begins with an exploration of the theoretical underpinnings of zero-defect manufacturing, leading to the identification of challenges in effective industrial data analytics in ZDM and the proposition of a standardized DA framework. Subsequently, it proceeds to identify the crucial factors for digital readiness in ZDM implementation, utilizing structural equation modeling (SEM) to explore the connections among these factors using Statistical Package for the Social Sciences (SPSS) and Analysis of Moment Structures (AMOS) as a tool. A digital maturity assessment of industries is then conducted to devise a robust ZDM strategy. Causal relationship analysis of the digital readiness factors has been performed by collecting a data from a sample of 149 steel industries in Ethiopia, while a detailed assessment of digital maturity is carried out on a specific industry case, employing a well-established assessment framework and the six maturity levels identified in recent publications. Finally, an AI-based defect detection model is developed and validated using one year of time-series quality inspection data collected from the case industry, using R programming language with RSTUDIO as a tool and three ensemble learning algorithms.
The finding from the relationship analysis of DM factors reveals that people and expertise play a crucial mediating role between the adoption of digital technology and digital maturity, underscoring the significance of human capital in digital transformation in manufacturing. The evaluation of digital readiness at the industry level indicates that the industry is presently positioned at level 2, representing the initial phase of connectivity (having attained computerization). The assessment exposes a notable digital maturity gap concerning people and expertise. Another significant empirical aspect of this dissertation is the development of a defect detection model, which exhibits enhanced accuracy in defect identification through the incorporation of comprehensive quality parameters. The modified Gradient Boosting model outperforms the other two ensemble learning models, which are Bagging and Random Forest in predicting the quality of the considered steel product (Rebar output quality parameters) based on the following R² values: 0.9538 for Yield Load, 0.9726 for Yield Stress, 0.9751 for Ultimate Load, 0.9718 for Tensile Strength, 0.9614 for Elongation after Fracture, 0.983 for Tensile strength to Yield strength ratio, and 0.964 for Bend Test. Comparative analysis with existing literature also demonstrates that this study achieves superior levels of prediction accuracy across most output variables.
The DD model developed in this dissertation tackles the shortcomings of manual quality inspection techniques leveraging AI and machine learning on recorded QI data. Traditional methods (physical inspection), resulting in high costs, time inefficiencies, and inconsistencies, especially when detecting concealed defects. Unlike these approaches, the AI-based DD model autonomously learns and detects defects. Prior to delving into model development, the dissertation also explores key prerequisites for effective AI implementation and propose solutions to realize virtual defect detection: industrial digital maturity assessment and the adoption of standardized data analytics. These enable steel industries to shift systematically from manual quality inspections to a data-driven method and progress towards the implementation of each ZDM strategies.
The research contributes to filling gap in the current literature of DM models by critically analyzing the relationships among the existing DM models dimensions. This enables industries to focus on the critical factors which highly contribute to DM. The proposed assessment guide provides practitioners with the tool to accurately determine their DM index and implement the appropriate ZDM strategy. The novel DA framework proposed also enables the development of AI model standardized (considering key elements of product quality control) which were fragmented in the ZDM literature. The developed AI-driven defect detection model addresses the current challenge for the steel industries quality inspection activities assisting operators in accurately estimating the occurrence of defects. Lastly, the proposed ZDM Continuous Improvement Cycle (CIC) provides a clear framework for continuous improvement, aligning the ZDM strategy with the industry's current digital maturity level. This approach enables industries to systematically progress through the four ZDM strategies, starting with detection as the baseline to prediction and prevention.
Description
Keywords
Zero defect manufacturing, Defect detection, Quality control, Quality inspection, Virtual defect detection, Data analytics, Industrial digital readiness, Digital maturity model, Artificial intelligence