This study proposes a novel transformer oil micro-water detection method based on the ultrasonic pulse-echo technique

optimised by a sparrow search algorithm (SSA) to enhance the prediction performance of a random forest (RF) model. Initially

finite element simulations were conducted to select optimal ultrasonic frequencies of 2 and 2.5 MHz. An accelerated thermal ageing experiment was performed using #25 Karamay oil samples

and ultrasonic pulse-echo signals were collected via a custom-built detection platform. Variational mode decomposition was employed to extract effective echoes from the raw pulse-echo signals. Temporal and frequency domain analyses yielded 162 dimensional features

which were subsequently filtered to 88 key parameters using the maximum information coefficient method. A transformer oil micro-water detection model was then developed by integrating the SSA with RF and trained using K-fold cross-validation. The model achieved an impressive average prediction accuracy of 97.34% over 10 cross-validation runs. The testing set demonstrated a prediction accuracy of 96.40%

a remarkable improvement of 16.53% compared to the unoptimised RF model. The findings provide a solid foundation for the rapid detection of micro-water content in transformer oil using the ultrasonic pulse-echo method.