Project: Cognitive Cloud for Laser Welding
Machine learning methods are nowadays implemented in a multitude of industrial applications and their field of application continuously grows , particularly for autonomous industrial production systems with increased flexibility, especially in countries with high labor costs. Because laser welding processes are individually different in optical setup, materials, or joint geometry, the current laser welding systems have to be configured with many manual trials by human experts. Once configured, industrial laser welding systems require costly manual reconfiguration for every process change. To expedite the setup and reconfiguration times, human experts often use tables and settings from previous work to take a good guess of initial process parameters. Even when fully configured, small process variations may have a large impact on the seam quality._x000D__x000D_We want to apply modern machine learning methods in order to improve laser welding quality, increase automation and flexibility as well as reduce costs of configuration and down times. Our recent research results indicate that cognitive laser welding systems equipped with machine learning can learn laser welding parameters from human expert feedback. The systems improve with every feedback iteration but need enough training data to improve processing. Therefore, this project’s aim is to jointly collect and analyze training data on a large scale how to weld individually different processes. Once the collaborative systems gain enough machine knowledge, they avoid repetitive configuration steps and may significantly reduce down times as well as increase product flexibility._x000D__x000D_All distributed systems are connected to each other via internet in order to facilitate the integration of several cognitive control systems The knowledge gained by these systems is shared, thus allowing a global database of process configurations, sensor setups and quality benchmarks. In order to share information between machines, all of them have to use a similar method of feature acquisition. Different laser welding scenarios are constantly being investigated within the labs of TUM, Precitec KG, and IREPA LASER. Within this consortia, we can acquire the necessary training data and processing knowledge for a locally distributed network of cognitive laser welding systems of the future. Further participants within this network to come, can contribute and benefit from the automatically growing machine knowledge. _x000D__x000D_As a first scenario to achieve these goals using cognitive data processing, approaches for combining the input data from multiple sensors are evaluated, in order to receive a good estimation of the process state. The systems will be composed of a coaxially mounted camera, photodiodes, and an optical topography sensor. The camera will provide information about the melt pool and keyhole geometries, while the photodiodes are giving a very high spectral resolution of optical emissions. The topography sensor can provide pre- and post-process data._x000D__x000D_Using cognitive dimensionality reduction techniques, unnecessary and redundant data from these sensors can be removed. The reduced sensor data is used to classify the state of the process. Clustering allows for identification of specific process states, even between different set-ups. If a significant difference from the references, and therefore an unknown process condition, is detected, the supervisor will be alerted. The expert can then teach the new state and countermeasures (if possible) to the system in order to improve its performance._x000D__x000D_The cognitive system to be developed should be able to learn to separate acceptable and unacceptable results and furthermore be able to avoid unacceptable results where possible. The usage of technical cognition eliminates the need for a complete physical model of the welding process. The system is able to stabilize the process by improving at least one steering variable. Distributed cognition allows for a central database between different manufacturing locations. The information gathered from one process can be transferred to a similar process at a different location. _x000D__x000D_The learning abilities of the system together with the ability to share and cluster the knowledge between manufacturing locations significantly reduces the expert time needed for calibration, leading to an improved throughput, higher agility and lower production costs._x000D__x000D_We are going to improve the efficiency in environments where laser material processing is already successfully used, while increasing the potential market of laser applications to areas where it has not been used due to quality and reliability concerns. The cognitive laser welding network will offer two significant advantages to industrial laser welding: it can autonomously process a broad set of different laser welding scenarios and the joint knowledge will exponentially improve over time for all future participants in this network.
|
Acronym
|
CCLW
(Reference Number: 6325)
|
|
Duration
|
01/10/2011 - 30/06/2014
|
|
Project Topic
|
This project aims to develop a laser welding system, able to acquire process relevant information from different sensors. By means of machine learning methods, the acquired process data is shared between multiple welding stations in order to improve their process monitoring and control capabilities.
|
|
Network
|
Eurostars
|
|
Call
|
Eurostars Cut-Off 6
|
Project partner
