Research On Fault Detection Of Multimodal Industrial Process Based On Local Algorithm

In recent years,the rapid progress of science and technology and modern people's pursuit of high quality products,production safety and stable production process leads to more and more frequent failures.Monitoring of industrial production process is very important for safe production.In order to improve the safety of the system,the performance of the monitoring system problems have attracted more and more attention.The fault detection technology based on data-driven can improve system safety and reduce losses.In order to improve the performance of fault detection in multi-modal process with obvious variance difference,this thesis studies the fault detection algorithm of multimodal process based on local algorithm.The main work and contributions are as follows:(1)In order to improve the monitoring effect of k-nearest neighbor(kNN)method in multimodal data with large sparse differences,a multi-modal process fault detection algorithm of kNN based on markov distance(MD-kNN)is proposed.First,k-nearest neighbor samples are found for the training data.The square sum of the markov distance is calculated,and the control limit is determined by the kernel density estimation.For the new sample,the square sum of markov distance between the new sample and the k-nearest neighbor samples in the training set is calculated,and compared with the control limit for process monitoring.Simulation results of a numerical example and the semiconductor process data verify the effectiveness of the algorithm.(2)To improve the fault detection performance of the locality preserving projections(LPP)algorithm in multimodal process where the dispersion degree of different modes varies greatly,a fault detection method of multimodal process based on second order difference quotient LPP(SODQ-LPP)is proposed.First,the second order difference quotient was used to preprocess the training data of multimodal process to eliminate the variance difference between modes.Secondly,the LPP algorithm is used for dimensionality reduction and feature extraction.The statistics of the samples is calculated,and kernel density estimation(KDE)is used to determine the control limits.The new validation sample is projected onto the LPP model after the second order difference quotient preprocessing.The statistics of the new data are calculated and compared with the control limits for fault detection.Finally,simulation results of a multimodal numerical example and the semiconductor process data verify the effectiveness of the algorithm.(3)In order to improve the monitoring effect of NMF method in multimodal process with obvious variance difference,a process monitoring method of multimodal process based on LPD-NMF.First,the training data of the multimodal process are preprocessed by local probability density(LPD),and the multi-modal characteristics of the variance difference are eliminated.Then,the non-negative matrix factorization(NMF)algorithm is used for dimensionality reduction and local feature extraction.The statistics D~2of the samples is calculated,and kernel density estimation(KDE)is used to determine the control limits.After the local probability density is used to preprocess the new validation data,it is projected onto the NMF model.The statistics of the new data are calculated and compared with the control limits for process monitoring.Simulation results of a numerical example and the semiconductor process data verify the effectiveness of the algorithm.