Fault Prognostics For Key Equipment In Thermal System Of High-end Generating Units

Due to the complicated structure and the harsh operating conditions of key equipment in thermal system of high-end generating units,it inevitably suffers from fault degradation.Once a fault seriously degrades,it may lead to unplanned shutdown,causing serious economic losses and even casualties.In order to ensure the safe and reliable operation of high-end generating units,it is necessary to research on fault prognostics methods for key equipment in thermal system of high-end generating units.Reasonable maintenance time can be determined based on prognostics result,thus eliminating safety hazards and reducing the risk of unplanned shutdown of the unit.However,the key equipments in thermal system of high-end generating units have numerous measuring points,and complex process characteristics,which bring severe challenges to the research of fault prognostic methods.This artical focuses on the fault prognostic related issues such as faulty variable selection,health status assessment,fault degradation trend prediction and so on,and proposes effective algorithms with practical application value.The specific research content of this article is as follows:(1)Considering the existence of huge measurement variables in key equipment in thermal system of high-end generating units,most of which are redundant variables that are not related to faults,this article proposes an algorithm for selecting faulty variables by block refinement.This method first uses the missing rate filtering method and the variance filtering method to remove the noise variables and uses the k-Nearest Neighbor algorithm to fill in the missing values.Then,the variables are divided into linear and non-linear sub-blocks based on the Pearson correlation coefficient and mutual information.The refined characteristics of these two sub-blocks are analyzed Lasso-Logistic regression algorithm is used to select fault variables from linear sub-blocks,and XGBoost is used to select fault variables from non-linear sub-blocks(2)Considering that there are a large number of closed-loop control loops in the thermal system of high-end generating units,the dynamic characteristics of the process may change under the feedback regulation.Therefore,a health state assessment method for key equipment in thermal system of high-end generating units based on closed-loop information analysis is proposed.First,the method uses Canonical Variate Analysis to extract time-related features,and then uses slow feature analysis to further extract static and dynamic fault features.Through the two-step progressive fault feature extraction strategy,the extracted features can not only distinguishing faults from normal operating conditions,but also fully characterize the potential fault degradation process under the action of closed-loop control systems.Finally,the Continuous Hidden Markov Model is used to monitor the changes of static and dynamic characteristics,so as to realize the health status evaluation under the closed-loop control system.(3)Taking into account the large-scale non-stationary characteristics of the thermal system operation process of high-end generating units,and the complex and diverse fault characteristics,a fault prognostic method based on collaborative analysis of multiple fault characteristics is proposed.The method first uses the ADF test to distinguish between stationary and non-stationary variables.For stationary variables,Slow Feature Analysis and Kernel Principal Component Analysis are applied to extract features that reflect the slowly varying and non-linear characteristics of the fault degradation process.For non-stationary variables,Cointegration Analysis is used to distinguish the trend of fault degradation from the normal non-stationary trend of variables,and extracts the features of non-stationary characteristics that reflect the process of fault degradation.Thus the fault degradation characteristics of key equipment in thermal system of high-end generating units can be fully characterized.However,the extracted features contain a large number of redundant features,so the key fault features are comprehensively filtered based on the three indicators of monotonicity,robustness and correlation to improve the accuracy of fault prediction.Then Dissimilarity Analysis is applied to fuse the selected key fault characteristics into DISSIM health indicators.Finally,a fault prognostic model is built based on the health indicators to achieve accurate prediction of fault degradation trends.