Acoustic Emission-Based Anomaly Detection and Statistical Analysis for Predicting Failures in Cast Resin Transformers

This paper proposes an acoustic emission-based anomaly detection and statistical analysis method to detect partial discharges caused by insulation degradation inside cast resin transformers. While high-frequency current transformers and ultra-high frequency sensors offer high accuracy, they are expensive and require complex signal interpretation. In contrast, acoustic emission sensors can measure elastic waves generated within the insulation at a relatively low cost. By applying artificial intelligence techniques, data-driven anomaly detection can be achieved without complicated analysis procedures. In this study, an acoustic emission sensor was attached to the outer enclosure of a 22.9 kV cast resin transformer to acquire signals under both normal and faulty conditions. The Hilbert transform was applied to enhance impulsive characteristics, and a one-dimensional convolutional long short-term memory autoencoder model was developed for anomaly detection. Using a threshold based on the interquartile range, normal and fault signals were classified with 100% accuracy in the laboratory environment. The proposed method provides an efficient data-driven analytical framework for diagnosing insulation degradation and offers potential applications in condition monitoring and predictive maintenance of cast resin transformers.