Research On Residual Life Prediction Of Degraded Data Driven

Prognostic and Health Management(PHM) based technologies focus on predicting equipments’ remaining life and on determining maintenance strategy to make expenses of maintenance and support as well as risk of failure minimized, by using equipment condition monitoring data. Apparently, maintenance decision making for equipment is based on remaining lifetime(RUL) prediction. As for special equipment with characteristic of high-reliability and long-life but small sample- and poor information-type data, an effective method of RUL prediction is how to make use of real-time status monitoring data. Therefore, researching RUL prediction with degradation data-driven has both theoretical significance and application value.Taking classification of equipments RUL prediction in mind, this paper studied typical data-driven RUL prediction methods, based on computational intelligence and probability statistics respectively.As for a fault being characterized as a variety of characteristic signal, namely RUL of equipment associated with a number of performance parameters, this paper studied Multi-variable Grey Model(MGM(1,)). Firstly, MGM(1,) with optimized background value was constructed, which used the functions with non-homogeneous exponential law to fit the accumulated sequences for original data and reconstructed the calculating formula of background value. Then initial predictive values of original data sequences were obtained. Secondly, the mapping relation between the residual sequences and original data sequences was established, training RBF neural networks. Finally, the multi-variable grey compensating RBF neural network prediction model was constructed, combining improved MGM(1,) and RBF neural network. This paper takes relay as example with three performance degradation parameters to conduct study. Results indicate that prediction accuracy is improved by using the presented method.As for equipment with dynamic changes of running environment and load, differences in degradation among same type of equipments and randomness of single equipment can be found. Thus a prediction method of RUL which combined prior and current degradation data was proposed. Firstly, the equipment degradation model was constructed, conforming to a nonlinear Wiener process. Then the RUL distribution of the equipment was determined. Secondly, the unknown parameters were estimated by using the maximum likelihood estimate(MLE) method. Parameters were updated by using the Bayesian method when new degradation data was available. After that,the real-time RUL was further evaluated. Numerical simulation and case study of laser generator were conducted. Results indicate that the presented method reflect differences between individual equipment well and significantly reduce uncertainty of the RUL distribution.