Based On The Offline Time Series Data Of Sudden Failure Prediction

The paper is based on research project from2010National Natural Science Foundation-Self-organized Critical State Identification and Risk Research of Machine Sudden Large Failure (Grant No.:51075060).Because of the complexity of modern electromechanical machine, the foreshadow of some unexpected accident sometimes becomes very subtle, in this condition it is very difficult to predict the machine sudden large failure, which does great harm to the production. In semiconductor assembly and test factory which is capital-intensive, the sudden failure of the overloaded bottleneck machine is especially serious, and because of the reasons all above, the prediction of machine sudden failure has a very important significance. The main contents include:(1) The SOC study of machine failure time series. Because machine sudden large failure has high hidden characteristic and does great harm to the machine, it is hard to predict the sudden large failure with the traditional method. The Self-organized criticality theory is one of complexity science theories, which focuses on the system complexity and complex of systems and because of the analysis and processing capabilities of various types of emergencies, the SOC theory is used to study the SOC of machine failure which can be used in prediction stage. Because SOC theory in the analysis of data does not require data stationary, it ensures the sudden failure data from being removed as abnormal data.(2) The selection of definition methods for machine sudden large failure points. Before the prediction of machine sudden large failure, the definition of machine sudden large failure points becomes critical. The kurtosis method is adopted to define the sudden large failure points which can be used in the prediction stage based on extremely value theory.(3) The study of machine sudden large failure prediction model. Depending on whether having power law distribution characteristics or not, extreme value statistical method or calamities grey prediction method is used to the establishment of a sudden large machine failure prediction model separately. At the same time the off-line time series from semiconductor assembly and test factory is used for case study.Compared to traditional machine failure prediction methods, the machine sudden failure prediction method described in this paper does not remove unexpected machine failure data as a "noise points" or "outliers" in the data pre-processing. And through distribution characteristics analysis of the machine failures off-line time series with SOC theory, the required distribution characteristics can be gotten which is needed when the prediction of unexpected events with extreme value theory. The method proposed in this paper does not require the data to have linearity character which is needed in regression analysis and time series model, therefore the proposed methos has advantage to predict sudden machine failure and it also provides new ideas and methods for machine failure prediction.