MVControl - Generating process feedback from heterogeneous data sources in quality control

Machines learn from Production Data. Quality control is a crucial factor in multi-stage production processes. Adaptive data analysis recognizes problems and provides a competitive advantage.

Short Description

During quality control large quantities of data are generated, that today mainly aim at detecting defects and identifying parts to reject. However, in the long run the goal of quality control is to prevent the manufacturing of reject parts.

In today’s production environments the necessary data analysis is only rarely done, because of the time and effort that are required. Short-term reaction to quality problems is thus not possible. Through a combination of product, process and quality data, knowledge about the production process can be automatically acquired and process improvements can be automatically deduced.

The project aims at the development of data analysis methods that accumulate knowledge about the process and enable its use in several situations: During re-starting of the process after a break the rate of reject parts is temporarily increased. Data analysis will allow a targeted modification of the process parameters to reduce the reject-rate during that time period.

Currently, modifications to the process are only possible after reject parts show up during end-of-line quality control. Inline data analysis will allow a quick reaction and prevent the manufacturing of reject parts. Process improvements that are targeted at avoiding occasional defects are currently very difficult to identify, because of a lack of analysis tools. The project aims at providing exactly such tools.

To provide a solution for these situations, the project aims at the development of data-driven machine learning models that merge data from many different and heterogeneous sources. The models will cover the whole multi-stage process and merge process-, design- and quality data over the different stages to discover causal relationships.

Technologically this requires research in fields such as

  • time-series prediction,
  • incremental learning,
  • drift analysis and
  • process optimization.

A key challenge is the optimization of the prediction time horizon, making robust distinctions between intended and un-intended changes and automatically obtaining proposals for adjustments of the process.

The results of the project will allow manufacturing companies to perform an automatic and very detailed analysis of production data. It will be specifically beneficial for manufacturing of small lot sizes or high number of product variants.