Data-based quality analysis in machining production: A case study on sequencing time series for classification

Abstract: The introduction of AI applications is accelerated by an on-going development phase at present, creating a technology push. The application of machine learning methods not only in manufacturing research but also in manufacturing practice is still a growing field. Against a backdrop of increasingly volatile markets, highly customised products and increasingly complex production processes with consistently high-quality requirements, manufacturing is facing growing challenges. The rush of new providers of Industrial Internet of Things (IIoT) solutions and machine learning applications to the market is opening up new possibilities for data collection and analysis that go beyond the classic approach of model- and empirical-based process analysis and challenge traditional approaches to production management, such as lean management.However, at the value stream level, certain problems in lean production resulting from a lack of data can be addressed precisely with IIoT applications. As an extension to the challenges described, classic production tasks should be critically reviewed for their relevance. Reducing waste is one of the key principles in lean manufacturing for which data analytics can be used to increase labour productivity by either supporting or eliminating existing processes.We present the underlying goal of replacing one step in the value chain, the measurement of product quality, with the prediction of product quality using machine tool signals from the machining process. This idea has the potential for great cost reduction in manufacturing, especially if the use of existing data sources such as machine tool signals is sufficient as a data source. Although more data is being generated in companies than ever before, its potential can only be utilised if it is selected and analysed. Collecting high-quality data in machining practice often requires installing sensors and performing time-consuming data preparation steps. Given the sparse availability of useful data and the additional cost of new data sources, we use machine tools as data providers for machining processes.The advanced analytics goal derived from this business objective is the classification of time series, in particular strong sequence classification. Although there are several approaches to this classification problem, we opted for a feature vector-based approach, which has shown both potential and limitations in previous studies.In this case study, we demonstrate the potential of time series sequencing of machine tool control data for quality prediction. A comparison of optimised feature vector-based random forest classification models trained on different sequences of a real drilling time series database is conducted. The results suggest that while sequence length has a subordinate effect, sequence overlap offers great potential for effective classification without exhaustive feature engineering, which in practice is limited by computational constraints.

Keywords: Quality Assessment, Machine Learning, Data Preparation, Machining

DOI: 10.54941/ahfe1001021

Cite this paper:

​