Abstract: Because batch process has the feature of multi-modality and data Gaussian and non-Gaussian mixing, so that the accuracy of fault detection often is not high and the monitoring performance is affected. Therefore, an improved NPE algorithm is proposed in this paper. Firstly, in this algorithm, the local k-nearest neighbor set of each sample is found, the local neighbors are normalised by seeking their mean and standard deviation, and, therefore, a single mode is generated and the normalized data is made to obey approximately the multivariate Gaussian distribution. Then, neighborhood preserving embedding (NPE) algorithm is corporated to perform dimensional reduction for new data, the local information between the neighbors of data sample and sample information are fully exploited, and the local structural information of data is extracted. Finally, support vector data description (SVDD) is used to construct monitoring statistics Ω and control limits for fault detection, so that a faster and better result than standard statistic SPE detection is realized. By means of test on penicillin fers montation simulation plat form and compared with the two types of faults of the SPE and Ω statistics of NPE, the effectiveness of the proposed LNSNPE-SVDD algorithm is verified.

Key words: batch process, local neighbor normalization, process monitoring, multimode