COINTEGRATION ANALYSIS METHOD FOR FAULT DETECTION BASED ON SENSOR DATA

Abstract

Sensors are a popular source of information about the operation of complex dynamic technical systems. Considering data from sensors as a multidimensional time series is also used to describe cyber-physical systems. The article proposes a method for detecting system malfunctions based on the method of analyzing cointegration dependencies. It is determined that in the data for analysis it is possible to reveal cointegration dependences as facts of interdependence of data from different sensors. Calculations are given on the example of a system with 52 parameters. Out of 1,326 data pairs, 75 are cointegrated. The conducted analysis shows that the proposed method enables one to clearly illustrate situations with changes in behavior. Having identified cointegrated pairs, we can follow them, and if cointegration has ‘disappeared’, that is, at some new time interval we can no longer talk about the presence of a cointegration ratio, then something has changed in the process itself. In practice, this means either a change in technology (which the operator knows about), or a breakdown/accident/failure, due to equipment errors, changes in some parameters of the resources used. In the latter case, such information (that the process has changed) can be used to attract attention in general, which may ultimately lead to the need for equipment repair or maintenance or readjustment, etc. The analysis shows that the proposed method enables one to clearly illustrate situations with changes in behavior. As an example of using the method, we used the ready-made Tennessee Eastman Process (TEP) data set. Different pairs of data may have the ability to identify different errors. All errors cause a change in the behavior of one or several pairs of data, thus tracking the behavior of the value of random component enables identifying cases of deviation of the process from long-term equilibrium (in terms of cointegration), that is, cases of failure from the normal system operation. The results obtained are clear and objective and can be used by process operators or by a source for automatic process control.