Anomaly detection and early warning via a novel multiblock-based method with applications to thermal power plants

The accurate and timely detection of anomalous conditions are essential for the safe and economical operation of complex thermal power plants (TPPs). However, the development of an excellent anomaly detection model without sufficient fault data is difficult in practice. In addition, global-based detection methods can submerge local anomalous behavior, causing serious delays in providing early warning of anomalous conditions. To solve this issue, a multiblock detection method based on the framework of evidence theory is proposed in this study. Measured variables collected from different units are automatically divided into several subblocks by using mutual information (MI)-based spectral clustering. Then, an evidential k-nearest neighbors algorithm (EKNN) is developed in each block, and local detection results are calculated. To provide an intuitionistic indication, the Dempster-Shafer rule is adopted to fuse the detection results of all the subblock EKNN models. The proposed approach can be applied to linear and nonlinear processes on the basis of MI and the nonparametric k-nearest neighbors procedure. To confirm its effectiveness, the proposed method is validated on samples collected from an ultra-supercritical TPP in China.