Multi-mode Anomaly Detection Methods Based On Data Feature Extraction

Large automation system structure is increasingly complex, researching the effective anomaly detectionmethods can enhance the security of the system is running, and reduce the effects of downtime. Due to theinfluence of such as equipment wear and tear, process load change, production process there will be multipleoperation mode, transition between modes has obvious dynamic characteristics. Considering stable modal andtransitional modal characteristics, the effective feature extraction and anomaly detection method can ensure thesafety, economic and efficient operation of the multimodal system.On the basis of analyzing the existing research methods, used statistical analysis theory such as theprincipal component analysis, partial least squares, Gaussian mixture model as the tool, carry out multi-modeanomaly detection method research based on the data feature extraction technique. The paper mainly does thefollowing several aspects:(1) Reviewing several offline mode and online mode identification methods, anglicizing Gaussianmixture model and typical EM algorithm in-depth. The incremental EM algorithm based on BYY offline modeis realized, online modal identification is realized by using Bayesian inference strategy. To carry out featureextraction and anomaly detection research under the different modes.(2) The transition mode feature extraction and anomaly detection method based on differential geometryfeature extraction is proposed in this paper. The method by extracting differential geometry features such asthe location slope, curvature of the transition mode, to depict dynamic characteristic of the transition mode,establish the anomaly detection model of the transition mode based on the rolling ball, in order to realizethe anomaly detection of transition mode.(3) To enhance anomaly detection results visual, in the MATLAB environment, design the GUI ofmultimodal anomaly detection system, providing a reference for engineering and technical personnel.