The Utilization of Multi-Layer Perceptron (MLP) as a Diagnostic Tool for Dielectric Testing on Stator Winding Insulation to Improve the Generator's Reliability

The generator is a critical component in power plants, with winding insulation failure accounting for approximately 56% of all generator failures. Ensuring insulation health is therefore essential for maintaining operational reliability. Remaining Life Assessment (RLA) is a diagnostic approach used to evaluate the generator's condition and estimate its remaining lifespan. While conventional RLA methods often rely on heuristic or rule-based evaluations, this study introduces a data-driven approach using a Multi-Layer Perceptron (MLP) to predict insulation life expectancy. Unlike previous studies that depend on simulated or idealized data, this research uses 22 real datasets collected from practical field assessment, incorporating four key dielectric tests - insulation resistance (IR), polarization index (PI), dissipation factor (tan d), and partial discharge (PD), along with operational years as input features. The proposed MLP model consists of one input layer, three hidden layers, and one output layer, and achieves an average Mean Absolute Error (MAE) of 0.6802. The integration of actual diagnostic test data into a predictive neural model for generator insulation aging, enabling more accurate lifespan estimation and proactive maintenance planning. This approach offers significant improvements over traditional methods by enhancing reliability and reducing unplanned outages in power plant operations. Increasing the volume of training data further enhances the model's predictive capabilities, ensuring more reliable decision-making in generator maintenance.