Anomaly Detection Method of Rotating Machinery Based on Dirichlet Mixture Model

Abstract

The invention discloses a rotating machine operation state anomaly detection method based on a Dirichlet mixture model. the method comprises the following steps of: 1) independently collecting the normal working condition operation data and the real-time working condition operation data of a rotating machine; 2) extracting a data feature set, and constructing a feature phase space; 3) setting the initial parameter value of the Dirichlet mixture model; 4) carrying out training by normal data to obtain a statistical distribution model based on the Dirichlet mixture model, wherein the self-learning result of a model number is T; 5) calculating a distance between the feature phase space models of the normal working condition data, and carrying out self-learning on an alarm threshold; 6) setting the model number as T, and training a real-time data statistical distribution model; 7) calculating the distance between the feature phase space models of the normal working condition operation data and the real-time working condition operation data; and 8) judging whether the distance exceeds a set alarm threshold or not, giving an alarm if the distance exceeds the set alarm threshold, and otherwise, continuously collecting data. The method has the advantages of being high in practicality and high in accuracy, an alarm time point can be drastically brought forward, and the method is suitable for the rotating machine anomaly detection.

Description

technical field [0001] The invention belongs to the field of mechanical abnormality detection, and relates to an abnormality detection method for the running state of a rotating machine, in particular to a method for detecting an abnormality in the running state of a rotating machine based on a Dirichlet hybrid model. Background technique [0002] Chemical machinery pervades all aspects of chemical production and plays a vital role in chemical production. Once a failure occurs, the production will be affected at the slightest, and the machine will be destroyed at the worst. Therefore, how to effectively realize the real-time intelligent state monitoring of rotating machinery and discover the abnormality of the unit in time has become a hot spot of current research. [0003] At present, a lot of research has been done on the abnormal detection of rotating machinery in China. The abnormal detection methods are mainly divided into two types: 1. Early warning of single eigenval...

AI Technical Summary

Problems solved by technology

Traditional anomaly detection methods based on statistical models have many limitations. For example, the condition monitoring data in actual production is very complex, and it is difficult to determine its distribution form. In order to simplify calculations, most of the data assume a Gaussian distribution. Subjectivity, for complex data distribution problems, the analysis results will deviate greatly from the actual; in the training process, the model parameters are directly calculated from the sample data, ignoring the role of prior knowledge on anomaly detection; the number of models is based on experience setting etc.

These all greatly reduce the accuracy of anomaly detection

Images