Research On Data-driven Remaining Useful Life On-line Prediction Method With Information Entropy Measurement

Remaining useful life(RUL)prediction can assess the system operation condition and further support faults identification,timely reparations and condition-based maintenance.Thus,it can effectively reduce the fault probability and maintenance costs.It is the key technology of system prognostics and health management.With the rapid development of sensor technology,intelligent instruments,industrial internet of things,etc.,the types and quantity of system sensing data are becoming even more abundant.Sufficient data resource support prompts the development of the data-driven RUL prediction approaches,which are becoming the research focus of RUL.However,when RUL predictions are applied to the long-term autonomy systems(e.g.,satellite and unmanned aerial vehicle),besides the basic data modeling algorithms,more attention should be paid on the on-line continuous prediction approaches.The on-line sensing data selection and analysis are difficult to be carried out automatically due to the flaws of traditional modeling data selection methods(e.g.,artificial observation and prediction involved approaches).On this basis,the sensing data anomaly detection needs to be considered which are caused by sensor faults,failures or broken transmission lanes to ensure the on-line continuous RUL prediction.The relevant handling precautions are also needed to deal with the anomaly data.In this way,the on-line data-driven RUL prediction for solving the practical application challenge can be achieved.In view of this,the information entropy measurement with the feature of quantitative characterization is adopted in this research.The system degradation is quantitatively analyzed from the point view of sensing data to eliminate the couple between the prediction approach and sensing data.In this way,the basic method for machine automatic processing and on-line application can be realized.Meanwhile,the anomaly detection and recovery of RUL modeling data are also studied.They can help to decrease the performance loss of RUL prediction when the system is working normally but the sensing data are anomalous.The whole research work provides a novel solution for on-line continuous RUL prediction.The main contributions of this dissertation are as follows.(1)To satisfy sensing data automatic processing for data-driven RUL on-line prediction,this work proposes a RUL prediction approach based on permutation entropy.This approach utilizes the improved two order permutation entropy to mine the system degradation information contained in the sensing data.This information is further utilized to quantitatively select the RUL prediction modeling data.Therefore,the quantitative criterion for optimized selecting input data for RUL on-line prediction operation can be satisfied.Compared with the random selection method,the artificial observing method and the grey correlation analysis method,the experimental results indicate that the proposed approach has higher prediction precision and stability.Meanwhile,the quantitative analysis result of the system degradation information contained in the sensing data can be directly obtained.It provides the basic premise condition for RUL prediction on-line application.(2)To resolve the anomaly data detection of input sensing data for data-driven RUL on-line prediction,this work proposes a data anomaly detection approach based on mutual information.By quantitatively analyzing the information between the detected data and other data,the training data set which has the maximal correlation with the detected data is determined.With this process,the metric of anomaly detection approach can be improved.The anomalous sensing data can be detected when the system is working properly.Experimental results show that the proposed approach can achieve lower false positive ratio with the single dimension and multiple dimension training data.(3)To recover the anomalous data for data-driven RUL on-line prediction,this work proposes an adaptive data recovery approach based on mutual information for single dimension anomalous data.The independent variable data of anomalous data recovery approach are determined by modeling the correlation among sensing data.The recovered single anomalous parameter can support on-line continuous RUL prediction with acceptable performance loss.Experimental results demonstrate not only the lower relative error and root mean square error of recovered data can be achieved,but also the performance loss of RUL prediction can be effectively controlled by utilizing the recovered sensing data.