The Application Of Deep Learning On The Pattern Recognition Of Partial Discharges

Partial discharges, or PD, are common phenomena found in working power facilities. PD reflects the insulation strength, working condition and other important signals of facilities. Frequent discharges could aggravate defects and cause insulation breakdown. Thus, the correct and effectiveness of recognizing PD is meaningful in the monitoring and diagnosis of power facilities.The currently widely used BP neural networks have achieved quite effective recognition among PD patterns. While the training mechanism limited its performance and further enhancement. The recently emerging deep learning(DL) method is considered to give solution to such limitations and has been implementing to more successful applications. As to improve the implementation of general neural networks, this paper mainly introduces the DL method to PD pattern recognition.Concretely, as to study PD patterns, this paper established a PD experiment system and gathered PD to a sample library. For those three PD types of interior gaps, surface flashovers, and coronas, corresponding models are designed. And six subdivisions have covered situations like gaps in oil paper insulation and synthetic insulation, surface flashovers in the air and SF6, and coronas in the air and SF6. Data from on-developing PD are collected into the library, a set for the following training and testing. To improve the network performance, especially to avoid over fitting, this paper gives Monte Carlo simulation based on parameters’distribution, which serves as a supplement to PD experiments. By analyzing the statistical distributions of discharge capacity and phase, their union possibility distribution are derived. The simulation enlarged the sample amount scale, with parameters controllable and statistically dispersed.Based on the samples, this paper conducted three recognition tasks in terms of different targets, which are three types, six groups. BP neural networks and autoencoders of DL networks are involved as to compare their performance. The results shows that the autoencoders of DL have better performance because of higher accuracy and rubustness. And, it consumes longer training time too.As a result, this paper deems that the autoencoders of DL, due to the reform of its training mechanism, are better than traditional BP neural networks. Thus gives a reference for further implementation of DL networks in solving PD pattern recognitions.