Fault Detection Algorithm For Batch Process Based On Neareast Neighbor Rule

Batch process as a kind of important mode of production in modern industry, is widely used in the production and processing of high-value, high-precision products, therefore the safe and reliable operation of batch process has increasingly become the focus of attention。 With the constant improvement of the detection accuracy of batch process, the real time fault detection is also put forward higher request. The research subject of this paper is a fault detection method using KNN method, for the purpose of optimization of real time fault detection for batch process, the following research work is carried out and the research content is divided into three part.Account for the feature that the lenghth of data for batch process is usually unequal, a research on a rajectory synchronization method using DTW is carried out. For the reason that it will lead to large computation and storage when DTW method face high-dimensional data, a parallelogram window is used to limit the search path for DTW. It greatly improves the synchronous speed for DTW, and the simulation experiments using TE process data are used to verify the effectiveness of the algorithm.Based on the method of FD-KKNN, with the consideration of the nonlinear feature of batch process data, a fault detection method based on Kernel-KNN is proposed. By substituting gauss kernel distance for euclidean distance to measure similarity between samples, the normal samples close to each other and fault samples close to each other, then the classification surface is extended, so the detection accuracy is improved and the simulation experiments using TE process data are used to verify the effectiveness of the algorithm.Account for the disadvantages of mass compulation and storage where KNN method in the face of a large sample data or high-dimensional data, proposed using fast nearest neighbor method for fault detection. Firstly, the clip neighbor method is used to remove the fuzzy samples on the classification surface of the training samples, then the condensed neighbor method is used to remove the redundant samples at the ends of the classification surface, in can improve the operational efficiency of the algorithm, and also ensure the fault detection accuracy. Account for the problem that FD-KKNN method leads a high error rate due to the diversity of the batch process, a further research on fault detection method using K-Means and Kernel-KNN is in progress, and an improved K-Means method which selects the clusters center by density and distance is proposed. The improved K-Means method can classify the samples correctly, and the simulation experiments in TE process prove the fault detection method based on K-Means and Kernel-KNN improves the accuracy and reduces the time expended in fault detection.