Implementation of Predictive Maintenance Models on Serial Machines Using Temporal Fusion Transformers

Abstract

The operational stability of the PCB Depaneling Machine at PT XYZ is currently compromised by an erratic failure pattern and the reliance on a corrective maintenance strategy, leading to significant unplanned downtime and production losses. This study aims to address these challenges by developing a Deep Learning-based Predictive Maintenance system using the Temporal Fusion Transformer (TFT) architecture to estimate the Remaining Useful Life (RUL) of the machine. Furthermore, this research evaluates the potential economic impact of shifting from a reactive to a proactive maintenance approach. The methodology involves processing historical time-series data from sensor readings and maintenance logs to train the TFT model. The model's predictive performance is evaluated using statistical metrics, including Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the Coefficient of Determination (R-squared). To validate the practical feasibility, a cost sensitivity analysis is conducted to determine the optimal decision threshold for maintenance intervention. The research findings demonstrate that the TFT model effectively captures the degradation trend of the machine, achieving a high prediction accuracy with an MAE of 1.3888, an RMSE of 1.4013, and an R-squared of 0.7620. In terms of economic efficiency, the simulation results indicate that implementing the predictive strategy with an optimal threshold of 2.0 shifts can reduce operational costs significantly. The proposed model offers a potential cost avoidance of Rp 196,464, or equal to 54,82% efficiency per failure cycle if compared to the existing corrective maintenance approach. These results confirm that the TFT- based model is not only technically robust but also financially viable for industrial implementation.