Research On Compound Fault Diagnosis Based On Multi-label Learning And Extreme Learning Machine

In the real industrial process and life,there are inevitably different degrees of failure,and some even directly led to serious disasters.Therefore,it is of great practical importance to actively respond to the national call to conduct research on fault diagnosis and try to reduce all kinds of faults.Based on this,many scholars have also proposed many effective fault diagnosis methods for various conditions.Nevertheless,in real life,the working conditions are complex and one or more faults may occur at the same time.Sometimes,simple fault diagnosis cannot solve the actual problems.Therefore,it is imperative to put forward an effective compound fault diagnosis method.At present,some methods have been proposed in compound fault diagnosis,yet there is still a large amount of room for improvement to adapt to different practical conditions and requirements.A desired compound fault diagnosis method should not only ensure the safe and reliable operation of the equipment,but also obtain more economic and social benefits as far as possible.Of course,the speed and complexity of solving problems should also be as fast and simple as possible.Multi-label learning is a new supervised learning method in machine learning,which is different from traditional supervised learning(single label learning),and has been highly praised by many scholars in recent years.The multi-label learning method is mostly used in the research of image processing,recommendation system and other directions,but few people use it in fault diagnosis.In order to solve the problem of compound fault diagnosis,divide different state stages and provide early warning in advance,this thesis uses the idea of multi-label learning to simultaneously judge the state stages and fault types of bearings.In this thesis,the method of multi-label learning is applied to compound fault diagnosis,and the health state,degradation state,critical state,inner ring fault,outer ring fault,and cage fault are classified and diagnosed as labels,so that the label of each group of data is no longer a value,but a vector representing different information.Due to the rapid-speed and high-efficiency of the extreme learning machine method and the simplicity and convenience of the binary relevance algorithm,this study adopts the extreme learning machine as the basic classifier to carry out the compound fault diagnosis under the basic framework of the binary relevance algorithm.The comparative experiment between ELM-Label Powerset algorithm and ELM-BR algorithm proves the superiority of the proposed method.What’s more,the generalization ability of extreme learning machine is not stable,and the correlation between labels is not considered in the binary relevance algorithm.Therefore,this thesis improves the original method,a new classifier chains algorithm using kernel extreme learning machine as classifier is proposed.This method allows for the hidden relationship among different labels in compound fault diagnosis,and the generalization performance is more stable,which makes the compound fault diagnosis quickly,accurately and stably.