The GIS Partial Discharge Signal Feature Learning And Pattern Recognition Method Based On Dual Views

Gas-insulated switchgear is widely used in contemporary power systems and closely related to the safety and stability of power grid.Partial discharge is an important indicator of insulation defects.Therefore,it is of great significance to study the feature learning and pattern recognition method for partial discharge signal.Partial discharge pattern recognition task has distinct characteristics.For example,the probability of real partial discharge accidents is low,and the real data is difficult to obtain.The current research mainly relies on physical simulation data,which still has problems such as insufficient sample number and less diversity.Besides,there is complex interference in substation under real conditions,which places higher requirements for pattern recognition algorithm on generalization performance.Most of the current methods are based on feature engineering or deep learning.While making great progress,there are still many problems to be solved.For example,feature engineering is time-consuming and labor-intensive and it relies on expert experience with weak expression ability.Besides,the deep learning-based methods have a large risk of overfitting under small sample conditions.When test on complex partial discharge data,both methods expose the disadvantage of insufficient robustness.In response to above problems,we propose the GIS partial discharge signal feature learning and pattern recognition method based on dual views.Its main work and innovations are as follows:First:At the data processing,we propose a data augmentation method based on the partial discharge data,which enriches the data diversity while preserving the main information.At the same time,we further convert the partial discharge data into the form of partial discharge grayscale which uses the pixel information and spatial structure information to unify the phase,amplitude and discharge times.Second:At the model design,we propose the partial discharge feature learning and pattern recognition network based on dual views.The input is different types of partial discharge grayscale under the same label,we set them as two views.Besides,the unsupervised representation learning method based on maximum mutual information is used to extract high-quality feature.In addition,inspired by human’s multi-perception,we further maximize the mutual information between the dual-view features to achieve feature regularization,which is conducive to enhancing the expressive capabilities of features.Finally,based on the dual-channel structure,we introduce an adaptive integration strategy and set adaptive weights to measure the contribution of each view channel.This strategy can achieve more performance improvement.Experiments show that our method is effective,especially when it faces the partial discharge data with complex interference.Our method can obtain higher quality features and better generalization performance,which has strong practical value.