Recognizing Phases in Batch Production via Interactive Feature Extraction

Abstract

Batch production is a manufacturing process, in which different components of a product are processed in a step-by-step procedure. Each step can be considered as a batch phase and each batch phase can be distinguished based on the process time series data. Identifying the batch phases from production signals can reveal useful information to analyze the production quality, and optimize the control and monitoring of the process. In many cases the timestamps of start and end of batch phases are not recorded in historical data, neither are the labels for batch operations. The conventional way of using machine learning techniques to determine batch phases in such a situation involves three major steps: 1) segmenting time series data into samples that correspond to batch phases, 2) labeling the obtained samples, and 3) building a machine learning classifier with labeled samples. Each step can be very tedious for realworld applications. Algorithms for segmenting process data into expected batch phases often need parameter tuning. Labeling such industrial data requires domain knowledge. Model selection and hyperparameter tuning are both necessary processes for building a classifier, which is also time-consuming. In this study, we introduce a workflow for extracting phase segments directly from time series data without following the three conventional steps. The proposed workflow starts with extracting distinguished shape features from time series in a semi-automated manner. Subsequently, user-desired shapes can be selected through an interactive interface. In the end, the corresponding segments can be identified and exported. The advantage of this method is that it requires limited human effort in data preparation and machine learning model building, and the workflow can be used for batch phase extraction, data exploration, etc.