Purpose/Aim
Power transformer is one of the most essential equipment in power system. Acoustic signals, produced by the operation of power transformer, are closely related to its normal conditions and some typical failures, including overheating, partial discharge, and variations of mechanical structures. With the advantage of time-efficient and convenient to be measured, acoustic signals have becoming a hot spot in the field of condition monitoring of transformer. In views of the difficulties to obtain the abundant data samples of acoustic signals for the recognition of different failures and variations of operation environment of transformer, this paper proposes an on-line condition monitoring method of power transformer by the deep Q-network with the features parameters of acoustic signals. The acoustic signals of a 500kV power transformer are measured and calculated to verify the correctness of the propose method.
Experimental/Modeling methods
The cochlear filter cepstral coefficients (GFCC) of acoustic signals are calculated to obtained the feature parameters of power transformer with the proper preprocessing of framing and windowing and the design the gamma filter banks. Then the data samples of training set and testing set are built according to the GFCC parameters of acoustic signals of a 500kV power transformer. With the input of GFCC parameters of deep network and the interaction of Q-learning strategy with the environment, the data with reward value is obtained to update the parameter of on-line network until the completion of iterative process.
Results/discussion
It is seen that the acoustic signals are changed with the variations of load currents. According to the frequency spectrum of acoustic signals of power transformer, a 40-channel gammatone filter bank is designed to calculate the GFCC parameters. With the GFCC parameters as the input, the operation flow and the corresponding network of DAQ are designed to recognize the operation condition of power transformer. The results of the regular reinforcement learning algorithm are also given to further illustrate the effectiveness of the proposed method.
Conclusions
The GFCC parameters of acoustic signal of power transformer are more clear and effectively compressed compared to its time-frequency spectrum. The designed DAQ are converged fast with high learning performance and exploring ability, which can better recognize the operation condition of transformer with acoustic signals.
 

acoustic signal, condition monitoring, deep Q-network, power transformer