Application of Recurrent Neural Networks for Time-Series Analysis of Low-Frequency Signals Generated by Power Transformers

Traditional diagnostics of power transformers heavily rely on signal transformations, such as Welch’s method, to analyze low-frequency noise signals. This study proposes a novel approach using Recurrent Neural Networks (RNNs), specifically Long Short-Term Memory (LSTM) networks, for direct time-series analysis of raw low-frequency signals without frequency-domain transformation. By training and testing multiple LSTM architectures on transformer vibroacoustic data, the proposed approach achieved approximately 86% accuracy in the fine-grained multi-class benchmark and up to 95.54% in the broader grouped categorization scenario. The model further demonstrated near-perfect classification accuracy in distinguishing transformer types (normal vs. overload) using a simplified RNN architecture. These findings illustrate that RNN-based models can streamline transformer diagnostics and improve accuracy in identifying operational states and types, potentially advancing non-invasive monitoring techniques in power system infrastructure.