The complex internal structure of a transformer introduces difficulties in identifying the acoustic propagation paths from partial discharge (PD) sources to sensors. As a result

time-delay information alone is insufficient for intuitively reconstructing the spatial location of PD sources

ultimately increasing localization errors. To improve the accuracy of acoustic PD localization in complex multipath propagation conditions

this paper proposes a method for distinguishing direct and non-direct acoustic waves in transformer acoustic signals

and analyzes the impacts of these two propagation modes on localization errors. First

the physical mechanisms of direct and non-direct waves formation are analyzed

and their differences in waveform morphology are described. Direct waves exhibit a steep rising edge at the moment of arrival

with dominant high-frequency content and a rapid amplitude decay after the peak. To capture these differences

four representative features are constructed from both time and frequency domains: the peak-to-rise time ratio

signal kurtosis

energy decay rate

and the ratio of high-frequency to low-frequency energy. These features are then used to train a support vector machine (SVM) classifier to accurately identify direct and non-direct wave signals. Subsequently

acoustic PD signal acquisition and localization experiments are conducted on a laboratory-scale 110 kV transformer platform. Experimental results demonstrate that the proposed SVM-based classification method achieves recognition rates of 88.3% and 84.6% for direct and non-direct waves

respectively. When using non-direct waves for PD localization

the average positioning error reaches 75.23 cm with a standard deviation of 32.34 cm. In contrast

localization based on direct waves yields a significantly lower average error of just 6.66 cm with a standard deviation of 2.97 cm

reducing the positioning error by 68.57 cm and the standard deviation by 29.37 cm.