Non-similarity index-based fault classification diagnosis method

Abstract

The invention discloses a non-similarity index-based fault classification diagnosis method, which aims to solve two key problems, that the number of available training samples of reference fault types is limited and sampling data of different faults has an overlapping phenomenon in spatial distribution, confronted in carrying out a fault classification diagnosis method in an actual industrial process. The method comprises the steps of firstly, selecting out a characteristic variable, most distinguished from normal data, of a fault by performing characteristic variable selection on each reference fault type; and secondly, performing paired comparison on distribution non-similarity of an online fault data window and each reference fault data window by utilizing the characteristic variable, wherein an online detected fault type corresponds to a reference fault type with a minimum non-similarity index. Compared with a conventional classification diagnosis method, the non-similarity index-based fault classification diagnosis method has the advantages that the variable dimension is reduced through the variable selection, the conditionality of training data insufficiency is greatly reduced, and the ''interference'' influence of non-characteristic variables can be eliminated. In addition, the fault diagnosis is implemented through similar matching of window data in spatial distribution, so that the misclassification situation of overlapping data can be avoided to the maximum extent.

Description

technical field [0001] The invention relates to an industrial process fault diagnosis method, in particular to a fault classification diagnosis method based on a non-similarity index. Background technique [0002] With the complexity and large-scale trend of the modern industrial process, the requirement for continuous normal operation of the production process is increasing, and more and more attention is paid to timely and accurately diagnosing the faults in the production process. In modern industrial processes, due to the widespread adoption of DCS control systems and advanced measuring instruments, a large amount of sampled data can be stored and measured online in real time. These sampling data contain important information such as whether the production process is normal and whether the product quality is qualified, which provides a solid foundation for the data-driven process monitoring method. Generally speaking, process monitoring mainly includes two aspects: faul...

AI Technical Summary

Problems solved by technology

Compared with the hundreds of measurement points in the modern industrial process, it is conceivable that the data samples of a certain fault type that are allowed to be collected in the actual process are difficult to meet the minimum number of samples required for the sufficient condition of the number of samples

At the same time, due to the coupling between the production process and the control system, there is a large degree of correlation between the measured variables, and different faults may cause the same abnormal changes in some variables

This will lead to overlapping in the spatial distribution of sampled data of different fault types, and the classification model method for classification and diagnosis with a single data point will cause a large number of misclassification and diagnosis phenomena

Images