Researches And Applications Of Data Mining Techniques In Diagnosis Of Fault Positions For Rotating Machineries

Rotating machineries are the most common machinery systems,the fault diagnosis of rotating machineries is always a hot research direction.Vibration is the main factor causing the structural fatigue and faults,vibration of rotating machineries can result to the abrasion,decreasing of performance and invalid of rotor components,finally contribute to major economic loss and casualties.For rotor systems with complicated structures,its vibration information usually becomes uneasy to identify after the transfer procession.Also,it's difficult to find the correspondence between features of signal and faults,which makes it harder to apply conventional fault diagnosis techniques in machineries with complicated structures.However,as the scale,specification and depth of database application in condition monitoring and fault diagnosis keeps enlarging,more and more monitoring data are accumulated.Proliferating data contains a great amount of important information which provides favorable conditions for data mining techniques based on data analysis.In data mining algorithms,the classification and regression tree(CART)algorithm is widely applied in artificial intelligence for its high speed,high accuracy,good robustness and intuitive classification rules.In the past thirty years after the CART algorithm is proposed,there have been optimized algorithms on its classification accuracy,relative algorithms are applied in fault diagnosis and brings breakthrough progress.In 2014,“China manufacturing 2025” is proposed to gradually connect to industry 4.0,which also puts forward a higher request for intelligent manufacturing.As an important progress in intelligent manufacturing,the intelligence and online detection of condition monitoring and fault diagnosis are the key problems,so a datamining algorithm with high classification speed is needed.Current fault diagnosis methods focus on the enhancement of classification accuracy and the optimization of application ways,while few of them commit to another important indicator,classification speed.There are also few effective CART optimized algorithms which can realize online real-time fault diagnosis.In this paper,a complete experimental system for fault diagnosis is set up,and four common fault of rotor systems such as unbalanced mass,bearing outer ring,bearing inner ring and bearing ball failure are simulated.Vibration information like displacement,speed,acceleration and rotating speed are collected to build datasets and be processed for generating training sets and testing sets.This paper also proposes a D-CART algorithm based on the CART algorithm,which modifies the underlying algorithm steps by changing the way of selecting the best splitting points using the dichotomy algorithm.This method can greatly improve the classification efficiency of the CART algorithm under the premise of ensuring classification accuracy.In the other hand,this paper modifies the way approaching the midpoints of the mainstream unsupervised learning algorithm,the K-means clustering algorithm.Traditional K-means algorithm uses regular profiles to approach midpoints,which is replaced by the centroid method of material mechanics in this paper.Between the two optimized algorithms,this paper programs and realizes the D-CART algorithm.Additionally,the training sets are used to train decision tree models,and then compared with other four novel CART optimized algorithms to show the superiority of the proposed D-CART algorithm.Finally,the D-CART algorithm is applied to successfully diagnose the fault types of rotor systems,as well as quantitative analysis of unbalanced mass.The results prove that the D-CART algorithm proposed in this paper can effectively complete all the online diagnosing missions of rotor systems,and has the potential for real-time fault diagnosis.