Study On Fault Diagnosis Based On The Statistical Feature Extraction

With the development of science and technology, in modern industrial process,performance of large complex system is improved constantly, the scale of whichbecomes bigger and bigger, and the complicated degree higher and higher. Once thereis an accident in the complex system, great property damage and casualties can’t beavoided. It is difficult to build an accurate model since the system is becoming morecomplicated. In the practical industrial process, as the application of all kinds ofintelligent instruments, large amounts of data are collected and stored, but these dataare not effectively applied to the abnormal monitoring, and appeared "data rich,information poor" phenomenon. Therefore, in order to improve the monitoringcapability, based on the large quantities of available data, and this paper applies thecorresponding mathematical theory to set up a new fault detection and diagnosismethod of the system, which will further complete the abnormal control based ondata-driven theoretical system . The research work and innovative results of this paperhave important scientific significance. It mainly studied multivariate statistical featureextraction. The main research contents are as follows:(1)The traditional principal component analysis, relative principal componentanalysis, the designated component analysis and K-L expansion have something incommon. They firstly select corresponding orthogonal projection base, and thenreduce the data dimension through the coordinate transform. At the same time, theyalso have something in common when used in fault diagnosis. Therefore, this paperanalyses the common characteristics in a deep level.(2)There exists the model composite effect in the traditional principal componentanalysis, and it requires the data to satisfy normal distribution in fault diagnosis.When it uses the fixed model to detect time-varying data, it is neither reliable noradaptive. Relative principal component analysis and designated component analysisboth need to know the prior knowledge. Therefore, this paper proposes a method offault diagnosis based on the information incremental matrix. Through a series ofprocess of defining the information incremental matrix, information incremental meanand dynamic threshold, the method can detect abnormity and diagnose fault. Themethod can significantly reduce the rate of false and missed alarm, and it can beeffectively used in fault diagnosis based on the rate of new information contribution, and at the same time, it has good real-time characteristic.(3)The current method of alarm design requires the data to meet the certaintydistribution, whereas this paper presents an alarm designing method that based on thestatistics of actual output data. This method needn’t require the sampling data to obeythe certainty distribution anymore, and then using the methods of filtering and timedelay, the alarm is designed.