Hybrid Classification Network And Its Application In Transformer Diagnosis

Intelligent classification networks are currently widely used in supervised classification problems,but in practice,more sample data is unlabeled samples,and the number of labeled samples is small.The transformer diagnosis problem is tha t typical unlabeled sample data has many in the case of a small number of label sample data,based on the study of intelligent classification algorithms,this paper combines the DGA data commonly used in transformer diagnosis to carry out a diagnostic analysis to achieve the optimization of the classification algorithm.A hybrid classification network based on a combination of deep contractive autoencoder(DCAE)network and kernel semi-supervised extreme learning machine(KSSELM).Firstly,the DCAE network is trained with unlabeled samples to obtain the network structure parameters,and then the DCAE-KSSELM hybrid classification network parameters are fine-tuned with the labeled samples.Finally,the labeled samples and unlabeled samples are input into the mixed classification network for classification.This method not only makes full use of a large amount of unlabeled sample data,improves the feature extraction capability of the autoencoder,but also introduces the principles of semi-supervised learning and kernel functions into the ELM,and also enhances the classification stability of the ELM.DCAE-KSSELM hybrid classification network optimize based on improved firefly algorithm(FA).Firstly,the traditional FA algorithm is improved in three aspects of particle initial position,position update and step factor,which improves the optimization ability of the algorithm.Then,the improved FA algorithm is used to optimize the structural parameters of DCAE network in DCAE-KSSELM hybrid classification network,which further improves the diagnosis performance of hybrid classification network.The actual data of power plant transformers are used to test the performance of the above two hybrid classification algorithms,and the diagnostic results are compared with the diagnostic results of commonly used classification algorithms.The experimental results show that the two hybrid classification networks constructed in this paper have high accuracy in fault diagnosis.The stability is good,and it is also suitable for dealing with two-class and multi-class problems in public data sets.