Nonlinear industrial process fault detection method based on kernel principal component analysis

Abstract

The invention discloses a nonlinear industrial process fault detection method based on kernel principal component analysis. After conducting normalization processing on training data, a KPCA model isestablished, nonlinear features are extracted from the training data to serve as kernel components, and a normal threshold value is determined for each kernel component; a kernel principal component space (PCS) and a kernel residual error space (RCS) are divided out according to the number of the kernel principal components; a local outlier analysis algorithm is used for calculating correspondinglocal outlier value statistics LOF<PCS> and LOF<RCS>, and control limits are determined; test data is acquired, and a corresponding kernel principal component vector and a kernel residual error vectorare extracted by utilizing the KPCA model; weighting is carried out on the vectors by utilizing the normal threshold values of the kernel components, weighted local outlier value statistics WLOF<PCS>(h) and WLOF<RCS>(h) are calculated, and the control limits are used for monitoring. According to the method, a kernel component weighting technology and a local outlier factor technology are introduced into a KPCA method, nonlinear characteristic information in industrial process data can be measured accurately, and the fault detection rate is increased.

Description

technical field [0001] The invention belongs to the technical field of industrial process fault detection, and relates to a non-linear industrial process fault detection method, in particular to a non-linear industrial process fault detection method based on nuclear principal component analysis. Background technique [0002] Modern industrial processes tend to be larger and more complex, and fault detection and diagnosis technology has become an important technology to ensure the safety of industrial processes. With the wide application of modern computer measurement and control systems, a wealth of process operation data is collected and stored in industrial processes. Therefore, data-driven fault detection and diagnosis technology has gradually become a research hotspot. Typical data-driven fault detection and diagnosis methods include Principal Component Analysis (PCA), Independent Component Analysis (ICA), Partial Least Squares (PLS), Fisher Discriminant Analysis (FDA),...

AI Technical Summary

Problems solved by technology

However, in practical applications, traditional KPCA still has deficiencies: on the one hand, the above T 2 The two statistics of SPE and SPE treat each core component equally in the calculation process, but in fact only some core components can significantly highlight the fault information, so treating each core component equally cannot reflect the fault information effectively; on the other hand, Traditional KPCA is based on the assumption that the nuclear components obey the Gaussian distribution. However, the actual industrial process data distribution is complex, and it is impossible to ensure that the nuclear components strictly obey the Gaussian distribution. Therefore, the traditional T 2 The control limits of SPE and SPE cannot meet the needs of actual process monitoring, which will affect the effect of fault detection.

Images