Time-Series Similarity Application to Qualitative Process Trend Analysis

Abstract

Beside the widely applied quantitative statistical tools, qualitative methods get more and more popular in the field of data mining techniques. Qualitative results are often easier to understand to a user, but to achieve such results, these methods always claim for a priori knowledge of the object they analyze. This paper proposes a technique that is able to compare and qualify time series in an unsupervised way, whereto even a priori knowledge can be incorporated. The two main steps of our method are: it applies triangular episode segmentation proposed by Cheung and Stephanopoulos to get a symbolic trend representation, and secondly it compares episode sequences by pairwise sequence alignment, a known technique in bioinformatics for aligning amino acid sequences based on a dynamic programming matrix filled with transformation weights. An alignment is considered as optimal if sum of weights is minimal. Instead of weights, our technique applies a predefined similarity measure. The algorithm was made up with data preprocessing methods to handle multidimensional, noisy data as well: Principal Component Analysis and Gaussian-filter, respectively. It is shown that the presented technique is able to compare, classify or qualify time series to discover their similarities. The algorithm was tested on industrial process data as well to show how it works on process trends and how it supports the analysis of product transitions in a multi-product polymerization plant.