Research On Multiple Models Based State Estimation And Its Combined Method For Fault Prognosis

With the trend of large-scale,interconnected and complicated modernization systems,the cost of regular maintenance and breakdown maintenance is getting higher and higher.It is hoped that fault prognosis technology can be used to realize early state prediction and remaining useful life estimation,so we can identify problems in time and take appropriate maintenance decisions to avoid security incidents.Fault prognosis usually implements state estimation and state trend prediction based on system degradation model or historical monitoring data,including model-based fault prognosis methods and data-based fault prognosis methods.These two methods have their own merits and cannot be replaced by each other.Therefore,making full use of known data information and limited mechanism knowledge to develop combined fault prognosis methods has become a new way to improve system reliability and safety.In order to avoid the inaccuracy of a single model to characterize the actual motion pattern,this paper firstly studies the interacting multiple model estimation algorithm based on extended Kalman filter from the perspective of state estimation,and then proposes a guidance law identification algorithm based on interacting multiple model extended Kalman filter.The algorithm uses the output of model probability within interacting multiple model to characterize the levels of usage of the candidate models,so the guidance law is identified by looking for the model corresponding to the maximum probability,and the estimation of the guidance law parameters is realized at the same time.Considering that support vector regression has better generalization ability for small sample sizes,this paper also studies a class of early fault prognosis algorithm based on support vector regression to achieve health condition prediction.As the parameters of support vector regression seriously affect its prediction accuracy,this paper proposes to use the whale optimization algorithm and the sine cosine algorithm to select the parameters of the support vector regression based on the optimization theory.These algorithms effectively balance the global search and local exploration of the parameter space,avoiding the problem of falling into local optimum,reducing the influence of the selection of the built-in parameters of the optimization algorithm on the prediction accuracy,and improving the accuracy of time series prediction of support vector regression.A reliable method of fault prognosis needs to be aware of the trend of future degradation state as much as possible,so that sufficient time can be taken to take measures to avoid the occurrence of failures or reduce the losses caused by failures.As the prediction step size increases,the accuracy of the multi-step prediction model will decrease rapidly,and the reliability in the uncertainty environment is more unacceptable.Aiming at the problem that the weight of the long short-term memory model cannot be updated in the multi-step prediction process,this paper uses support vector regression to model and predict the training error,and then compensates the residual signal of the long short-term memory model during the multi-step prediction process.Therefore,the accuracy of health state prediction based on multi-input multi-output multi-step strategy is improved.In view of the non-stationary signal in practical situations,the empirical mode decomposition is used to decompose the non-stationary signal,and then the least square support vector regression machine is used to model and forecast in each decomposition component.Finally,the prediction results are integrated to improve the multi-step prediction accuracy and efficiency of the non-stationary signals.At present,the fault prognosis research based on support vector regression mainly focuses on system state prediction,and realizes the point estimation of remaining useful life by calculating the time when the failure threshold is reached for the first time.The actual degradation process becomes more complicated due to the influence of uncertain factors,and the degradation states of different stages may correspond to different degradation models.Based on interacting multiple model,this paper uses the particle filter to realize the uncertainty expression of fault prediction.At the same time,the support vector regression is used to model and predict the measurement values,and the residual signal of the filter is compensated during the multi-step prediction process,thereby improving the accuracy of the multi-step remaining useful life estimation and obtaining the distribution information of the probability density function.In summary,the fault prognosis based on multiple model estimation and its combination method has good scalability,which can fully exploit the known data information and limited degradation knowledge to improve the accuracy of multi-step fault prognosis.At the end of this paper,the research work is summarized,and the future research direction is prospected.