Research On Fault Prediction Algorithm Based On State Estimation And Multi-methods Integration

Dictated by the need for higher safety and reliability of system operations, it is expected that the way system faults may develop and mutate is know, given minor indications of abnormalities identified in a system. This can thus inhibit the aggravation of system faults and subsequently avoid latent accidents. Compared with traditional planned maintenance strategies, fault prediction based predictive maintenance techniques can not only improve equipment utilization, but also reduce maintenance and production costs. Therefore, research and development of fault prediction technologies have huge theoretical significance and application values.By monitoring system status and gathering historical operational data, fault prediction algorithms usually make state estimation and then predict the trends of the state. Based on this, the contents of this thesis can be divided into two parts. The first part analyzes the nonlinearities of a system, model uncertainties, multiple missing measurements, sampling methods of the sensor and other factors on system performance. Then it studies the state estimation problems for several classes of random nonlinear dynamic systems; the second part investigates the problem where with a known system analytical model, model-based approaches can effectively estimate the system state, but struggle to establish a precise mechanism based mathematical model for complex industrial systems. In contrast, data-driven fault prediction methods require both historical data from a system under normal and abnormal operations. To obtain the system's operational data under fault conditions ofter incur a very high cost, even result in catastrophic. Overall, this thesis explores an appropriate fusion framework, which makes full use of both existing physical knowledge and historical data. It also studies the dynamic data-driven state estimation method in order to improve prediction accuracy. The mian contents of this thesis can be outlined as follows:Targeting at the estimation of time-varying random nonlinear systems, multiplicative noises, parameter uncertainties and multiple missing measurements have been taken into consideration concerning the performance of the filter. This filter takes a recursive form in the sense of minimum variance, such that it is robust to the uncertainties of system parameters as well as being non-fragile to the changes of filter parameters.Regarding a type of nonlinear system which is with event-triggered sampling control, multiple missing measurements and uncertainties, the art of filter design has been investigated in the sense of minimum variance. By designing an appropriate gain matrix for the filter, at each sampling instant, an upper bound for the filter error covariance can be minimized. The resulting filter is iterative and recursive and it is capable of performing online computation. Additionally, the system estimation error has been analyzed, while it has been proved by the stochastic analysis theory that the estimation error is mean square bounded under certain conditions.An integrated unscented Kalman filter (UKF) and the relevance vector regression (RVR) approach has been proposed for prediction. They aim to solve the problems that model-based filtering methods are overly reliant on the physical model, while data-driven methods have poor stability and yield low prediction accuracy. Taking the influence of both the long-term and short-term data on the future data trends into account, the predicted value by the RVR method and the latest real residual value constitute the future evolution of the residuals with a time-varying weighting scheme. Next, the future residuals are utilized by UKF to recursively estimate the system state for predicting, which realize the dynamic adjustment and prediction updating for the filter as well as improving the accuracy of predictions.In order to solve the problem of low prediction accuracy with regard to medium/long time span predictions, and to improve the adverse effect of prediction errors caused by data-driven method, which increasing with the prediction steps on the fusion method. An adaptive weighting term has been introduced into the fusion algorithm, which can be used to constrain the influence of predicted future measurements in the filter updating step. In addition, considering that the measurement noise in a system is changing along with the predicted measurements during the prediction period, a dynamically updating algorithm has been adopted to update the system model to improve prediction accuracy.Finally, the research outcomes have been summarized, and further works have been prospected.