| dc.description.abstract | Based on time series operational data gathered from PT Toba Pulp Lestari Tbk, this study proposes a hybrid deep learning model for detecting cavitation levels in a centrifugal pump (Wernert GMBH NEPO 50-32-200). Cavitation causes mechanical damage and a reduction in efficiency. It is typified by the formation and collapse of vapour bubbles inside the pump. This study uses a variety of operational parameters, such as inlet/outlet pressure, fluid temperature, flow rate, rotational speed, and vibration, which are recorded every five minutes throughout 2024, in contrast to conventional vibration-based techniques. The suggested model incorporates Bidirectional Long Short-Term Memory (Bi-LSTM) networks for classification, Recursive Feature Elimination (RFE) for choosing the most pertinent features, and Uniform Manifold Approximation and Projection (UMAP) for nonlinear dimensionality reduction. Two baselines were used to assess performance: a UMAP-Bi-LSTM model without feature elimination and a standard Bi-LSTM model. With an accuracy of 95.09% as opposed to 93.83% and 93.57%, respectively, the hybrid UMAP-RFE-Bi-LSTM model performed better than both. Inlet pressure, flow rate, and shaft rotational speed were found to be the most predictive factors for cavitation severity based on feature importance analysis. This hybrid approach offers a scalable and economical solution for predictive maintenance in industrial pump systems by increasing detection accuracy and model robustness. According to the findings, deep learning in conjunction with supervised feature selection and nonlinear dimensionality reduction efficiently extracts intricate patterns from operational data, enabling early cavitation detection and minimising possible downtime. | en_US |
