DEVELOPING A PREDICTIVE MODEL FOR HYDRAULIC MACHINE MAINTENANCE

Abstract

The hydraulic system is an essential infrastructure element of all kinds of heavy machinery used in several applications such as manufacturing, mining, construction, aviation, and processing industries. There are several types of degradations that occur in these hydraulic systems (for example, pump cavitation, seal leakage, valve erosion, accumulator failure, fluid contamination, among others). This leads to unexpected malfunctions accompanied by lengthy downtime periods, posing safety threats and causing costs. Traditionally, there have been no successful proactive maintenance strategies applied to address these challenges. The current research suggests a novel approach to predictive maintenance for hydraulic systems. The suggested approach relies on the Hydraulic Condition Monitoring Dataset (Hewli et al., 2015), containing 2,205 operating cycles collected via 17 sensors. Such sensors collect different types of physical variables, namely pressure, flow, temperature, vibration, and efficiency. The considered case study focuses on seven distinct machine learning algorithms, including Isolation Forest, One Class SVM, K-Means Clustering, DBSCAN, Autoencoder neural networks, 1D-CNN, and XGBoost. The analysis of predictive maintenance involves a set of steps to be taken. The first step refers to data preprocessing, including such aspects as handling missing values, normalization, and filtering out outliers. The second step relates to feature engineering. This stage is characterized by calculating rolling statistics and creating health indexes. The next step refers to data analysis, such as creating a correlation heat map and performing principal component analysis. Afterward, each of the algorithms is applied to the dataset, and a comparative analysis is conducted based on a 5 fold cross-validation. Moreover, bibliometric analysis is presented within the current research .Sources for bibliometric analysis are taken from three databases – Scopus, Web of Science, and IEEE Xplore. The obtained data provide an overview of global research trends in the area of interest, pointing at prominent researchers, popular methodologies, and certain limitations. Bibliometric analysis indicates China to be the leading contributor, generating 30.1% of all publications under consideration. Besides, it becomes evident that deep learning approaches have dominated this field starting from 2021. The experimental evaluation proves that XG Boost provides the best predictive performance: Accuracy 94.2%, Recall 92.8%, F1 Score 91.5%, and AUC-ROC0.971, with inference time of 158 milliseconds. 1D CNN iv achieves second-best accuracy (91.7%), yet its inference time is threefold higher compared to XGBoost, thus limiting its suitability for deployment at the edge. A hierarchical framework with three levels of alertness for maintenance (Normal/Alert/Critical) is proposed for translating XG Boost's predictions into a maintenance strategy. Most important sensors are Hydraulic Pressure (PS1), Volume Flow (FS1), and Motor Power, identified via a feature importance analysis. These results contribute significantly to a reproducible algorithmic benchmark for India's manufacturing sector aiming to shift from reactive to predictive maintenance practices. Keywords: predictive maintenance, hydraulic systems, machine learning, XG Boost, anomaly detection, condition monitoring, Industry 4.0, sensor fusion, fault prognosis, UCI Hydraulic Dataset, bibliometric analysis .