Effective, accurate and adequately early detection of any potential defects in power transformers is still a challenging issue. As the acoustic method is known as one of the noninvasive and nondestructive testing methods, this paper proposes a new approach of the classification method for defect identification in power transformers based on the acoustic measurements. Typical application of acoustic emission (AE) method is the detection of partial discharges (PD); however, during PD measurements other defects may also be identified in the transformer. In this research, a database of various signal sources recorded during acoustic PD measurements in real-life power transformers over several years was gathered. Furthermore, all of the signals are divided into two groups (PD/other) and in the second step into eight classes of various defects. Based on these, selected classification models including machine learning algorithms were applied to training and validation. Energy patterns based on the discrete wavelet transform (DWT) were used as model inputs. As a result, the presented method allows one to identify with high accuracy, not only the selected kind of PD (1st step), but other kinds of faults or anomalies within the transformer being tested (2nd step). The proposed two-step classification method may be applied as a supplementary part of a technical condition assessment system or decision support system for management of power transformers.

Results of a comparative analysis of two non-dimensional coefficients, calculated based on signals emitted by partial discharges (PD) are presented in the paper. The PD were generated using setups for modeling of three basic types of PD sources, which can occur in paper-oil insulation in power transformers. The following measuring methods were considered: electrical, ultra-high frequency, acoustic emission, and optical spectrophotometry. All measurements, for each of the PD-source configuration using the various methods, were conducted simultaneously in a laboratory under same measurement conditions. Based on the gathered results it was stated that it is possible to recognize the partial discharges form based on the non-dimensional coefficients, which differ in value for all of the considered methods except for the acoustic emission. The main contribution of the research performed lies in the simplicity of the applied signal processing procedure, which can be adopted to a device or system for diagnosis of partial discharges occurring in paper-oil insulation.