Pattern Recognition Based On Deep Neural Networkfor Partial Discharge

Gas insulated substation has developed very quickly and has been in use all over the world.Its compact size,high reliability,low maintenance have made it an attractive option in many circumstance.However,operation experiences show that intrinsic defects in GIS ofen cause accidents though its high reliability.It is well known that insulation breakdown is often preceded by partial discharge(PD)activities.Several types of defects ofen exist in GIS,such as metal protrusions,free metal particles,surface metal contamination on the insulator,gap existence in the insulator and so on.Therefore,the analysis and pattern recognition for PD pulse shape are very important to estimate GIS insulation condition.PD caused by GIS internal insulation defects has complicated and highly dispersed characteristics,susceptible to operation environment.Traditional methods of PD pattern recognition based on statistical characteristics are strongly subjective in feature extraction,easy to lose some characteristic information and have low recognition rate for free metal particles.Therefore,based on the analysis of detection and pattern recognition method for partial discharge in GIS,the depth study of the characteristics of PD signal of different insulation defects was done.And a new method for the pattern recognition for PD signals was put forward.The main work of this paper are:1.Four typical insulation defects were simulated in the GIS equipment in our laboratory by utilizing existing online monitoring and fault diagnosis technologies.They were in different dimensions,put in different positions and under different voltage levels to get large amounts of experimental PD data,then to establish a PD database.A foundation for the following research work had been laid.2.The deep belief nets(DBN)were applied to pattern recognition for the PD faults in the GIS device.DBN can learn the higher features from the data,avoiding the subjective influence of the feature selection.The results show that the DBN algorithm could identify the defects of free metal particles very well,within the time for recognition much lower than that of S VM or BP neural network.3.Simulated annealing algorithm was used to optimize the training process of deep neural networks,making the training error can be changed in two directions,which solved the problem that the BP algorithm cannot get the global optimal solution in the training process.The learning rate was adjusted according to the error size,which improved the training speed compared with the traditional choosing method for learning rate.In this paper,the two methods were used in the training process of deep neural network,which not only improved the recognition accuracy,but also improved the training efficiency.4.The recognition accuracy of SDAE for partial discharge within a large number of interference was not ideal.The training method of SDAE was improved and extended to the DBN,constructed the stacked noising auto encoders(SNAE)and noising deep belief nets(NDBN).Results show that the recognition accuracy of the two algorithms for the PD data with low quality or some distortion improved significantly.