Pattern matching in multivariate time-series data

A new pattern matching strategy is proposed for multivariate time-series data based on statistical techniques, especially principal component analysis (PCA). The new approach is both data driven and unsupervised because neither training data nor a process model is required. Given an arbitrary set of multivariate time-series data, the new approach can be used to locate similar patterns in a large historical database without the knowledge of the start and end times of disturbances. The proposed pattern matching strategy is based on two similarity factors: the standard PCA similarity factor and a new similarity factor that characterizes the pattern of alarm violations. Extensive simulation studies for a chemical reactor, batch fermentation and the Tennessee Eastman challenge process demonstrate that the proposed pattern matching strategy is more effective than existing PCA methods and can successfully distinguish between different operating conditions.; The effect of data compression on pattern matching is also investigated. Several popular data compression techniques including a wavelet based method and the commercial PI(TM) software are compared for their effect on pattern matching.; A new methodology for clustering multivariate time-series data is proposed. The methodology is based on calculation of the degree of similarity between multivariate time-series datasets using similarity factors. The standard K-means algorithm is modified to cluster multivariate time-series datasets using similarity factors. Data from a highly nonlinear batch acetone-butanol fermentation example are clustered to demonstrate the effectiveness of the proposed methodology. Comparison with existing clustering methods for multivariate time-series data show several advantages of the proposed methodology.; A new dynamic data rectification methodology based on the Kalman filter is developed, that rectifies noise as well as outliers in measurements. The filter equations are formulated in the form of probability distributions. Then the expectation-maximization algorithm is used to find the maximum-likelihood estimates of the true values of the measurements based on the process model, past data and current observations. The methodology can be used with any dynamic process model and can be implemented online. The proposed technique can also be used to provide diagnostic information about process changes.