Research On Data-driven Fault Prediction And Maintenance Decision-making For Multi-granularity Faults

As machinery is supposed to be even more sophisticated,automated and intelligent,the connection between component parts inside a particular mechanical equipment is becoming more and more complex and close-linked.The failure of a key component in the equipment may lead to damage to the performance of the overall mechanical equipment,thereby affecting the reliability of the entire production line,and even lead to the shutdown of the production line in severe cases,resulting in unpredictable economic losses and even casualties.Therefore,in order to ensure the safe and stable operation of mechanical equipment,it is of great significance to accurately predict the state of mechanical equipment and make scientific and effective maintenance strategies in time.Bearings,as one of the most widely used mechanical components in various rotating machinery equipment,play a key role in supporting the mechanical rotating body during the operation of the equipment.The function and working conditions of the bearing itself determine its fault-prone characteristics.Therefore,this paper studies the fault prediction of bearings,which are the key parts of the equipment.The traditional bearing fault prediction model usually focuses on the prediction and diagnosis of fault types and ignores the difference of fault degree.Once the bearing is detected to be faulty,the unified maintenance mode will usually be followed.However,the fault degree is different,and the exact same maintenance decision will lead to waste of resources caused by "over-maintenance".Therefore,the study of the fault degree is the basis of cost-effective maintenance strategies.The research on the remaining useful life provides theoretical support for the fault degree.If we can predict the fault type of the bearing during the working process of the rolling bearing,and at the same time realize the quantification of the fault degree based on the analysis of the remaining useful life,identify potential faults as soon as possible,and formulate targeted maintenance strategy,which can effectively prevent the occurrence of abnormal equipment fault and achieve efficient and economical active equipment maintenance.In this paper,the key component bearing in mechanical equipment is taken as the research object,and the vibration signal data is used as the basis to study the extraction and analysis method of bearing fault characteristics.And this paper further studies the fault type of bearing and the quantification method of fault degree based on remaining useful life analysis,and predicts fault type and fault degree of bearing at the same time;Based on the comprehensive prediction results of bearing faults,the differential maintenance decision-making is studied.The main contributions of this paper are as follows:(1)This paper analyzes the fault information of bearings,and conducts signal feature extraction research according to the characteristics of vibration signals in the time domain,frequency domain and time-frequency domain,to provide a basis for fault prediction based on signal characteristics.For the extracted features of time domain,frequency domain and wavelet packet energy,the features are further selected by combining L1-SVM((L1-Support Vector Machine))and variance filtering method,so as to effectively select the features containing more fault information and remove the redundant features further improve the accuracy of fault prediction.(2)Aiming at the problem that traditional bearing fault prediction models usually focus on the prediction of fault types or simply carry out research on remaining useful life prediction,this paper proposes a parallel prediction method for fault type and fault degree,and proposes BP-LSTM(Back Propagation-Long-Short Term Memory)model,which can not only predict fault types,but also predict fault degree based on the remaining useful life,providing more complete fault information for maintenance decisions.(3)Aiming at the problem that the existing methods are based on a single fault mode and the discrete division of degraded states,this paper uses the characteristics of reinforcement learning DQN to deal with continuous infinite states,establishes the corresponding maintenance decision model for different component faults based on Markov decision process.And this paper designs DQN algorithm to solve the Markov decision maintenance models of different components at the same time.Thus,economical and effective maintenance schemes can be given for different types and different degrees of faults.