What is it about?
Manufacturing processes are usually monitored by a large number of sensors, but it can be almost impossible to understand high-level behavior from the enormous amounts of data without taking time and effort to manually label it, which often is necessary to learn about different aspects of the production. We propose an unsupervised learning pipeline to enable manufacturers to discover high-level abstractions from sensor data such as the start and stop of machining, milling, tool wear anomalies, and so on. We do so without requiring the tedious and time-consuming process of labeling them with domain experts. The tool labels the sensor data without human intervention, which can be used for detection of deviations from normal production and automatic segmentation of various phases of a production cycle.
Featured Image
Photo by Possessed Photography on Unsplash
Why is it important?
The quality of each product in an assembly line relies heavily on continual observation and error compensation during the production. While sensor technology has enabled manufacturers to collect massive amounts of data, it is still a challenge to use it in an efficient manner and optimizing manufacturing. The use of unsupervised learning for deviation detection greatly expands the amount of sensor data that can be used for detecting errors and monitoring the processes. This can lead to higher quality and fewer defects, eliminating unnecessary scrap and waste of resources.
Perspectives
It has been inspiring to look at ways to realize zero-defect manufacturing, and there is obviously a lot of potential for using sensor data to improve these types of processes. While society ultimately needs less production and consumption in order to be sustainable, one step on the way is to reduce the amounts of waste we generate in the current manufacturing industry.
Erik Johannes Husom
SINTEF
Manufacturing is a complex orchestration of several machines and operators giving rise to potentially hundreds if not thousands of sources of time series data from sensors. Therefore, it is necessary to discover higher-level abstractions of activities that occur in manufacturing such as when machining starts, when it stops, when the milling operation happens, when it fails, when are there anomalies, and so on. Unsupervised learning can cluster manufacturing behavior of the same kind and help discover higher-level abstractions that a human can understand at a glance. Furthermore, it is easy to understand deviations in behavior based on deviations in these abstractions rather than raw time series data.
Sagar Sen
SINTEF
Read the Original
This page is a summary of: UDAVA, May 2022, ACM (Association for Computing Machinery),
DOI: 10.1145/3522664.3528603.
You can read the full text:
Contributors
The following have contributed to this page
