Supervised subgraph augmented non-negative matrix factorization for interpretable manufacturing time series data analytics

Data analytics has been extensively used for manufacturing time series to reduce process variation and mitigate product defects. However, the majority of data analytics approaches are hard to understand for humans who do not have a data analysis background. Many manufacturing conditions, such as trouble shooting, need situation-dependent responses and are mainly performed by humans. Therefore, it is critical to discover insights from the time series and present those to a human operator in an interpretable format. We propose a novel Supervised Subgraph Augmented Non-negative Matrix Factorization (Super-SANMF) approach to represent and model manufacturing time series. We use a graph representation to approximate a human’s description of time series changing patterns and identify frequent subgraphs as common patterns. The appearances of the subgraphs in the time series are organized in a count matrix, in which each row corresponds to a time series and each column corresponds to a frequent subgraph. Super-SANMF then identifies groups of subgraphs as features that minimize the Kullback–Leibler divergence between measured and approximated matrices. The learned features can yield comparable prediction accuracy (normal or defective) in case studies, compared with the widely used basis expansion approaches (such as spline and wavelet), and are easy for humans to memorize and understand.