Fault Detection Methods Based On Local And Global Structure Information

Modern industrial production process is so complex that the operators tend to be in the face of high-dimensional complex data.As a result,we need to use the effective dimension reduction techniques and corresponding process monitoring methods to help operators to monitor the running status of process.By adding the constraints of the global structure on the basis of locality preserving projection(LPP),the paper studies the fault detection methods based on local and global structure information.The main research content is as follows.In order to make full use of the feature information of data in the chemical process and improve the performance of fault detection,a fault detection method based on orthogonal nonlocal structure constrained locality preserving projections(ONSC-LPP)is proposed.In the method,the objective function of LPP is used to preserve the local structure and the nonlocal structure is preserved by maximizing the distance between nonlocal data points.Moreover,projection vectors are designed to be orthogonal through iterative computation.The feature information of local and nonlocal structure in the process data can be extracted sufficiently by using ONSC-LPP and the orthogonal projection vectors have better effect in preserving the overall shape of the distribution of data and constructing monitoring statistics.ONSC-LPP-based T~2 and Q Statistics are constructed in the feature space and residual space for fault detection after dimensionality reduction respectively.The simulation results of Tennessee Eastman(TE)process demonstrate that the proposed method detects faults more quickly and achieves lower fault missing alarm rate than traditional monitoring methods.In order to deal with the problem that LPP does not take into account the global structure and dynamic characteristic of process data,a fault detection method based on dynamic sparse locality preserving projections(DSLPP)is proposed.In the study,the original data matrix is firstly extended to a time-delay augmented matrix.Then,a sparse coefficient matrix which can represent global sparse reconstructive relationship of data is gotten by solving an optimal problem of sparse representation(SR).The sparse coefficient matrix combines with the objective function of LPP to form a new objective function for dimensionality reduction.The DSLPP algorithm can not only preserve the local neighbor structure of the original data space,but also have better effect in preserving the global sparse reconstructive relationship.At last,DSLPP-based T~2 and Q statistics are constructed respectively in the feature space and residual space for fault detection.The simulation results of TE process demonstrate that the proposed method detects faults more quickly and achieves lower fault missing alarm rate than the LPP method.In view of the problem that nonlinear relationship between variables exists in the actual industrial process,the sparse locality preserving projection(SLPP)is extended to nonlinear field by using kernel method and a nonlinear fault detection method called kernel sparse locality preserving projection(KSLPP)is proposed.In the study,the original data is firstly projected into a high-dimensional space by using the kernel function,which makes the original linear inseparable data become linear separable.Then the sparse representation is carried out in the high-dimensional space and the kernel sparse reconstructive weight matrix is constructed.And the kernel sparse reconstructive weight matrix combines with kernal LPP algorithm for feature extraction.Finally fault detection statistics are constructed in each subspace for fault detection and the simulation results of TE process validate the effectiveness of the proposed method.Finally,the GK06 process control experimental device is used to validate the effectiveness of the ONSC-LPP,DSLPP and KSLPP in the actual device.