Research On Method For Equipment Health Prognosis Based On Auto-Correlated Hidden Markov Models

With the development of modern industry,advances in science and technology have driven equipment to become more complex,efficient and automated.As an important guarantee for productivity,the health status of equipment directly affects the operation and economic benefits of enterprises.Therefore,research on equipment health prognosis is of great significance.For the requirements of a gas turbine parts processing co.,Ltd.on operation and maintenance,this paper proposes a method of equipment health prognosis based on auto-correlated hidden Markov model.Algorithm improvement and Model optimization are carried out after analyzing related research.Through signal analysis and feature extraction,the method uses performance degradation assessment to monitor the current state of the equipment and distinguishes the three phases in its lifecycle via adaptive alarms.When implementing health predictions,a trigger mechanism is introduced.According to the results of performance evaluation,monitor the health status of the equipment if it in the normaluse-phase,while perform RUL(Remaining Useful Life)estimate on equipment which entering the initial deterioration and rapid failure phases.Performance degradation assessment can provide a qualitative reference for maintenance decisions.The idea of staged forecasting can improve the accuracy of health prediction,thus effectively solve the problem of equipment health management.The main research results in this dissertation are as follows:(1)To perform health predictions,the deterioration trend of equipment during its whole life needs paying attention to.The vibration signal directly collected contains noise and is plain.Therefore,firstly extract multiple sets of eigenvalues,such as peak-to-peak value,RMS,skewness,center frequency,and wavelet packet energy,from time domain,frequency domain and through time-frequency joint analysis.Those features reflect different characteristics of the vibration signal.Then select sensitive ones strongly correlated with the running time of equipment via correlation analysis.Considering the contrast strength and mutual relationship of sensitive features,apply weighting method to construct combined feature,which can better describe the monotonous change of equipment performance,and also guarantee the integrity and effectiveness of the vibration signal.(2)The health of equipment will gradually deteriorate during its life and deviate from the normal state.Acquire the signal of equipment in normal operation,and establish an evaluation model by training HMM(Hidden Markov Model).Then plug characteristic parameters of the studied device into the model,and use forward algorithm to output likelihood probability,after which a health index HI can be constructed by eliminating influences of the length of observation sequences.Furthermore,set an adaptive alarm to distinguish the different stages of degradation.For data loss problems that may occur during monitoring,improve Viterbi algorithm,searching for the maximum observation probability in each hidden state,within the range of the observations’ value,to fill the missing positions.In such way,the output of HI is guaranteed.The method of performance evaluation is verified by data measured in the factory of S company and data in bearing accelerated degradation test performed by IMS(Intelligent Maintenance System,Cincinnati University).It shows that the accuracy of the assessment results is high.(3)As the health of the equipment monitored,the characteristic parameters of the signal obtained through feature extraction tend to be auto-correlated in time domain.Aiming at this specialty,optimize HMM and propose the Auto-correlated Hidden Markov Model(AHMM).The first-order autocorrelation function is introduced into the mean of the probability distribution of the observed sequences.Furthermore,the microstate of equipment health is defined,under which the core variables of AHMM have practical meanings,so as to accurately express the physical process of equipment degradation.Then considering two end-of-life criteria—the device deteriorating to the worst microstate of its health,and the characteristic parameters of the signal reaching the failure threshold,the estimation methods of RUL are established respectively.At this stage,the data measured in the factory is so little to perform life prediction of the bearing,thus the RUL evaluation method proposed in this paper is analyzed and verified based on bearing lifecycle data of the accelerated degradation test,and the results show that the prediction reliability is high.