# U-AttentionFlow: A Multi-Scale Invertible Attention Network for OLTC Anomaly Detection Using Acoustic Signals

## TL;DR

This paper introduces a new deep learning model called U-AttentionFlow for detecting faults in power transformer components using sound data.

## Contribution

The novel U-AttentionFlow model combines invertible multiscale coupling with attention mechanisms for anomaly detection in OLTC acoustic signals.

## Key findings

- U-AttentionFlow achieved 99.15% accuracy in detecting OLTC anomalies using real-world acoustic data.
- The model uses normal data to learn patterns and detect deviations, making it suitable for one-class anomaly detection.
- Integration of SE blocks, CBAM, and MHSA improves feature focus and temporal learning in acoustic signals.

## Abstract

The On-Load Tap Changer (OLTC) in power transformers is a critical component responsible for regulating the output voltage, and the early detection of OLTC faults is essential for maintaining power grid stability. In this paper, we propose a one-class deep learning anomaly detection model named “U-AttentionFlow” based on acoustic signals from the OLTC operation. The proposed model is trained exclusively on normal operating data to accurately model normal patterns and identify anomalies when new signals deviate from the learned patterns. To enhance the ability of the model to focus on significant features, we integrate the squeeze-and-excitation (SE) block and Convolutional Block Attention Module (CBAM) into the network architecture. Furthermore, static positional encoding and multihead self-attention (MHSA) are employed to effectively learn the temporal characteristics of time-series acoustic signals. We also adopted a U-Flow-style invertible multiscale coupling structure, which integrates features across multiple scales while ensuring the invertibility of the model. Experimental validation was conducted using acoustic data collected under realistic voltage and load conditions from actual ECOTAP VPD OLTC equipment, resulting in an anomaly detection accuracy of 99.15%. These results demonstrate the outstanding performance and practical applicability of the U-AttentionFlow model for OLTC anomaly detection.