This study employs predictive maintenance to enhance the Explainable XGBoost Model for predicting failures in industrial components. The research utilizes a model that employs the adaptive sliding window approach to extract features. The timeframe for this approach is set at 24 hours. This method leverages multi-device sensor data to extract the features. The BO-XGBoost model is assessed using accuracy, MCC, F1-Score, and G-mean. The measures achieved 99.87%, 0.988, 0.990, and 0.0989, respectively. The SHAP analysis method also identifies the characteristics of target variable prediction. The variable “rotatemean_24” is the most significant predictor. The model is easily understandable, aiding in comprehending relationships between features and outcomes. Therefore, it can assist in making decisions regarding predictive maintenance. Research has shown that implementing the optimized Explainable XGBoost Model can enhance factory maintenance efficiency and cost-effectiveness. The model's predictive capabilities enable proactive maintenance by identifying machine faults in advance.
Top1. Introduction
The advent of Industry 4.0 and the Industrial Internet of Things (IIoT) (Pramanik, S. et al., 2023) has transformed the operational landscape of businesses, highlighting the importance of data-driven decision-making and automation (Kumar Pandey, B. et al., 2021). The incorporation of machine learning (ML) algorithms has resulted in significant advancements in the domain of predictive maintenance (PdM) (Nunes et al., 2023). The primary goal of Predictive Maintenance (PdM) is to predict equipment malfunctions in advance, allowing for timely maintenance interventions and optimal resource allocation. Abidi et al. (2022) reported that this phenomenon reduces downtime, improves operational efficiency, and decreases maintenance costs.
PdM approaches are increasingly replacing traditional maintenance strategies like reactive maintenance and preventive maintenance (Hashemian & Bean, 2011). Predictive maintenance (PdM), also known as condition-based maintenance, involves generating maintenance recommendations by utilizing data acquired through condition monitoring, often in the form of time series data (Cheng et al., 2020). It is commonly defined by one of the following definitions: This study has two primary objectives: first, to identify and detect a problematic state in the monitored equipment that indicates an impending failure, and second, to estimate the remaining useful life (RUL) of the machine. The scientific literature includes two distinct approaches: diagnostics and prognostics. Equipment failure in manufacturing can result in significant financial losses due to downtime, negatively impacting productivity and profitability. Advancements in sensor technology (Iyyanar, P. et al., 2023), increased data accessibility (David, S. et al., 2023), and improved computing capabilities of machine learning algorithms have contributed to the development of more accurate and efficient predictive maintenance (PdM) solutions (Coandǎ et al., 2020).
Machine learning methods are frequently employed in predictive maintenance (PdM) to analyze large volumes of data (Babu, S. Z. D. et al., 2022) collected from various sensors (Anand, R. et al., 2022), such as vibration, temperature, and acoustic emission sensors. The aim is to detect data patterns and anomalies indicating potential equipment failure. Ayvaz and Alpay (2021) utilized different machine learning (ML) techniques, including supervised, unsupervised, and reinforcement learning, to create prediction models in the domain of prediction Maintenance (PdM).
Recent advancements in machine learning, specifically deep learning (Pandey, D. et al., 2021), have shown promising results in predictive maintenance (PdM). These methodologies have effectively handled complex and large datasets (Lee & Mitici, 2023; Liu et al., 2022). Machine learning (ML) has enhanced the capabilities of predictive maintenance (PdM) systems by integrating with advanced technologies like digital twin technology and edge computing (Panagou et al., 2022).