Fault Diagnosis Based On Manifold Feature Extraction For Chemical Processes

Fault diagnosis technology has the important theory and application value in improving the process safety and ensuring the stationary continuity of production. Large amounts of data are collected and stored in the modern chemical processes, which provide research opportunity for data-based fault diagnosis. In this paper, for multivariable, nonlinear and multimode characteristics of continuous chemical processes, some novel fault diagnosis methods are proposed combined with manifold learning theory. The main work includes:(1) In order to consider the nonlinear cross correlations and serial correlations in the multidimensional observed data, two fault diagnosis methods are proposed based on nonlinear manifold features extraction. One is Differential Evolution Dynamic Autoencoder(DDAU)and the other is fault diagnosis based on Improved Dynamic Isometric Mapping(IDISOMAP).DDAU firstly establishes an autoregression model through correlation analysis considering serial correlation of the variable itself. Meanwhile, the weights of network can be gotten by the adaptive differential evolution algorithm. The low-dimensional nonlinear manifold features can be obtained and two monitoring statistics are constructed for process monitoring.IDISOMAP establishes the augmented matrix considering the serial correlation of the variable firstly. Instead of euclidean distance, manifold distance is used to calculate geodesic distance.The neighbor parameter can be acquired adaptively according to data distribution. The monitoring statistics can be set up in the low-dimensional manifold space and residual space.Simulation results of Continuous Stirred Tank Reactor(CSTR) process and two-tank liquid level control system show that the proposed method can obtain the nonlinear manifold features, and it can improve the fault detection rate effectively.(2) Aiming at the problem that both the local and global structure information coexist in the process data, a fault diagnosis method is proposed based on structure preservation manifold feature extraction. Locally Linear Embedding(LLE) algorithm extracts the local structure features and neglects the global structure information. The objective function of LLE is improved and the local and global structure information can be acquired simultaneously byregulating the proportion parameter. Simulation results of Swiss Roll data sets and two-tank liquid level control system show that the proposed method can obtain the global and local manifold features information more fully. Besides, fault detection can be realized more accurately and timely.(3) The inherent slow features exist in the data of chemical processes, and the fault diagnosis methods based on the improved slow feature manifold extraction are proposed in the paper. The incremental covariance matrix is introduced to describe the slow feature information of high-dimensional data. A new objective function is built combined the maximizing deviations manifold features extraction. The low-dimensional manifold features can be calculated by eigenvalue decomposition and the fault variable identification is carried out based on contribution plots. Slow feature analysis is extended to nonlinear processes through kernel transformation method, and slowly varying information can be extracted from the nonlinear observation data. The simulation of two-tank liquid level control system illustrates that the proposed method can obtain the slow features information and find the fault more accurately.(4) Considering the multimode operation in the chemical processes, a fault diagnosis method of the multimode manifold features extraction is proposed based on Relative Isometric Mapping(RISOMAP). It utilizes relative geodesic distance instead of geodesic distance to establish distance matrix for extracting the low-dimensional manifold features.Meanwhile, kernel ridge regression is used to estimate the mapping of input observed data and the multimode manifold information. The monitoring statistic can be calculated based on the low-dimensional outputs. The simulation of numerical example and CSTR process illustrate that, the proposed method does not need judge the modes and it can monitor the multimode process effectively.Lastly, the Tennessee Eastman(TE) chemical process is used to verify the effectiveness of the proposed methods in the paper. The results show that the proposed methods can obtain the low-dimensional manifold features from the high-dimensional data and they can provide the higher process monitoring performance than the traditional fault diagnosis methods.