Development of AI-driven Predictive Maintenance for Steel Manufacturing Processes

Thesis Abstract

In the highly competitive and energy-intensive steel manufacturing industry, unplanned equipment failures and maintenance downtimes significantly contribute to production losses, increased operational costs, and equipment longevity challenges. Traditional preventive maintenance approaches often fail to accurately predict machinery failures due to their reliance on fixed schedules or reactive fault detection, underscoring the urgent need for more intelligent, real-time maintenance solutions. This study aims to develop and validate an AI-driven predictive maintenance model tailored for steel manufacturing processes, with the overarching goal of enhancing equipment reliability and operational efficiency. The specific objectives are (1) to identify key failure modes and associated sensor data pertinent to steel production machinery; (2) to design and implement a machine learning-based predictive model utilizing historical operational data; (3) to evaluate the model’s predictive accuracy through rigorous validation techniques; and (4) to assess the potential operational benefits and challenges associated with deploying AI-driven predictive maintenance systems in steel manufacturing environments. The study adopts a mixed-method research design, integrating quantitative data analysis with qualitative insights from industry experts. The population of the study comprises data collected from 15 steel manufacturing plants with similar operational profiles, encompassing a combined total of approximately 50,000 machine fault logs and sensor readings gathered over five years. A stratified random sampling technique was employed to select a representative sample of 200 machinery units, ensuring coverage across different types of critical equipment such as rolling mills, furnaces, and conveyor systems. Data collection instruments included embedded sensor systems, maintenance logs, and process control records, complemented by semi-structured interviews with maintenance engineers to contextualize sensor data and failure patterns. To ensure the robustness of the data, pre-processing steps involved data cleaning, normalization, and feature extraction, with an emphasis on identifying relevant indicators such as temperature fluctuations, vibration patterns, and power consumption anomalies. The primary analytical technique employed was supervised machine learning, specifically utilizing Random Forest and Support Vector Machine classifiers to develop predictive models. Model performance was validated through cross-validation methods, with metrics such as accuracy, precision, recall, and F1-score guiding the evaluation process. Additionally, root cause analysis was conducted to interpret model outputs and understand failure precursors. Key findings are expected to demonstrate that AI-based predictive models can achieve classification accuracies exceeding 85%, considerably outperforming traditional threshold-based monitoring systems. The models are anticipated to enable early fault detection with lead times sufficient for scheduled maintenance interventions, potentially reducing unplanned downtimes by up to 30% and maintenance costs by approximately 20%. The study also highlights critical factors influencing model effectiveness, such as sensor data quality and feature selection strategies, which are vital for successful real-world implementation. This research contributes to the body of knowledge by integrating advanced machine learning techniques within the context of steel manufacturing, providing evidence-based insights into how AI can transform maintenance paradigms in heavy industry. It extends previous empirical studies by demonstrating the practical deployment potential of predictive analytics in complex manufacturing environments, guided by the theoretical framework of predictive failure modeling and maintenance optimization theories. The main conclusion indicates that AI-driven predictive maintenance systems can significantly improve operational reliability and cost efficiency in steel manufacturing plants, provided that data quality and system integration challenges are adequately addressed. Recommendations include adopting a phased implementation approach, investing in sensor technology standardization, and fostering cross-disciplinary collaboration between data scientists and industrial engineers. Future research should explore the integration of IoT platforms and cloud computing to facilitate real-time monitoring and decision-making, along with longitudinal studies assessing long-term impacts on equipment lifespan and overall plant productivity.