A Vibration Detection and Fault Diagnosis Method for Dry-Type Transformers Based on Remote Laser Vibrometry

Abstract

Traditional contact-based vibration measurements can be affected by sensor installation and mechanical coupling, which complicates feature extraction for dry-type transformer core diagnosis. This study evaluates a fault-diagnosis workflow combining laser Doppler vibrometry (LDV), Gramian angular field (GAF) encoding and a deep residual network (ResNet-18). First, a finite-element multiphysics model was developed to analyse the vibration response of the core under normal operating conditions and to guide the selection of measurement points for the subsequent experiments. A laboratory fault-simulation platform was then established, and an LDV system was used to remotely acquire vibration signals at two measurement points under five core configurations and three low-voltage-side excitation levels (240, 320 and 380 V). The one-dimensional time-domain signals were converted into two-dimensional GAF images and classified using ResNet-18. At 380 V under normal operation, the simulated and measured responses both had a dominant frequency of 100 Hz, while the peak-to-peak displacement differences for the two paired regions were 9.7% and 24.2%, respectively. Under the current segment-level internal training-validation protocol, the model achieved diagnostic accuracies of 97.7% and 96.5% for measurement points 1 and 2, respectively. These accuracies describe discrimination among signal segments from the available acquisition records and should not be interpreted as estimates of performance on independently acquired records. The results provide a proof of concept for integrating remote LDV acquisition, GAF encoding and residual learning to classify the tested laboratory configurations.