Bildbasierte Charakterisierung und Regelung von Laserschweißprozessen

Abstract

Das Laserstrahltiefschweißen ist ein weit verbreitetes Verfahren in der industriellen Fertigung und obwohl bereits seit Jahrzehnten erfolgreich im Einsatz, mangelt es bis heute an Möglichkeiten der Prozessregelung. Verfügbare Regelsysteme beschränken sich meist auf die Positions- oder Abstandsregelung, lassen den eigentlichen Schweißprozess jedoch in aller Regel außen vor. Bisherige Ansätze zur Regelung des Schweißprozesses scheiterten regelmäßig an zu geringer Messgeschwindigkeit oder nicht robust messenden Integraldetektoren. Obgleich die Prozessüberwachung von Laserschweißprozessen bereits in vielen Bereichen Anwendung findet, handelt es sich auch hierbei meist um Verfahren mit integral messenden Detektoren, deren Messkurven lediglich über Korrelationsverfahren mit der erreichten Nahtqualität in Verbindung stehen. Kamerabasierte Verfahren zur Prozessüberwachung wurden zwar in den vergangenen Jahren massiv weiterentwickelt, abgesehen von Systemen zur Positionsüberwachung und -Regelung, kommen jedoch auch bei diesen meist Algorithmen zum Einsatz, die den Prozess auf Helligkeitsschwankungen hin untersuchen. Die Verwendung von Bildverarbeitungsalgorithmen, welche auf der Auswertung von geometrischen Formparametern beruhen, ermöglichen eine weit robustere und aussagekräftigere Beurteilung des Prozesszustandes, als es die eingangs genannten helligkeitsbasierten Algorithmen vermögen. Der notwendige hohe Rechenaufwand verhindert jedoch bis dato die Nutzung solcher Algorithmen für ein echtzeitfähiges System zur Prozessreglung. In dieser Arbeit wird basierend auf spektroskopischen Untersuchungen der elektromagnetischen Prozessemissionen und der Störeinflüsse durch Metalldampf und Schweißrauchpartikel, ein spektrales Fenster identifiziert, welches optimale Bedingungen für die Beobachtung der thermischen Prozessemission mit siliziumbasierten Kameras ermöglicht. Grundlagenuntersuchungen mittels kombiniertem Einsatz von Röntgenvideotechnik und Hochgeschwindigkeitskameras im nahen und mittleren Infrarot, erlauben einen dreidimensionalen Einblick in den Schweißprozess, auch unterhalb der Schmelzebadoberfläche. Die gewonnenen Erkenntnisse bilden die Basis für die Entwicklung einer kamerabasierten Prozessüberwachung, welche über eine koaxial zum Bearbeitungslaserstrahl angeordnete Kamera, die thermische Strahlungsemission des Prozesses erfasst und die entstehenden Bilder anhand geometrischer Bildmerkmale auswertet. Die identifizierten Bildmerkmale beschreiben die jeweiligen transienten Fehler eindeutig und liefern eine Charakterisierung des Prozesszustandes. Aus den evaluierten geometrischen Bildmerkmalen wird das Merkmal Durchschweißloch ausgewählt, um mittels eines geschlossenen Regelkreises den Durchschweißgrad von Lasertiefschweißprozessen zu regeln. Die Regelung wird dabei mittels einer neuartigen Rechnerarchitektur realisiert, der sogenannten Cellularen Neuronalen Netze (CNN). Die CNN-Architektur integriert dabei ein Netzwerk analoger Prozessoren direkt auf dem Kamerachip. Jeder einzelne Pixel verfügt bei diesem System über einen eigen simplen Prozessor. Diese Architektur ermöglicht es durch die Vernetzung der einzelnen Pixel eine Bildverarbeitung direkt auf dem Kamerachip durchzuführen, deren Berechnungen innerhalb eines Belichtungszyklus abgeschlossen sind. Auf diese Weise wurde ein Regelsystem implementiert, welches mit Regelfrequenzen bis zu 14 kHz bei minimaler Latenz, eine robuste Regelung der Durchschweißung und Einschweißung an I-Naht-Überlappverbindungen ermöglicht.

Deep-penetration laser welding is still an emerging application in the world of metal joining. It is increasingly replacing traditional resistance spot welding, particularly in automotive body construction, because of its higher productivity, lower costs and better quality. However, while process monitoring and control have found widespread usein classical joining processes, monitoring or closed-loop control of deep-penetration laser welding has only been established in a few applications so far.

Previously developed methods for direct in-process monitoring of laser welding processes were usually based on photo diodes. While allowing very high sampling rates, the spatial resolution of such systems is very limited. Camera-based systems on the other hand, offer a high spatial resolution but usually a very limited sampling rate, due to the limited bandwidth of the underlying data processing system. Both approaches are usually based on application-specific correlations between certain measuring signals and typical weld defects. Hence, the systems involved have to be calibrated to the specific application.

The goal of this work was the development of a fast, robust, camera-based, closedloop control system for the penetration depth during laser welding. The emphasis of the work is on the overlap joint geometry commonly used in car body construction. This work includes process diagnostics and camera-based process monitoring to build up an accurate picture of the three-dimensional geometry of the welding process. This knowledge is crucial in order to achieve the goal of a non-correlation based measurement of the process characteristics, necessary for the development of a robust closedloop control system.

Examination of spectral process emissions in the near infrared range below 1 μm enabled the identification of suitable wavelength ranges for process observation with silicon-based cameras detecting the thermal emission of the process itself. It is demonstrated within this work that conclusions drawn in many older publications have been based on measurements from uncalibrated spectrometers, leading to misinterpretation of various optical effects. By calibrated spectral measurement of the process emissions, two spectral windows free of characteristic line-emissions were identified for steel as well as for aluminium. Furthermore, it is shown that the spectral window be low 500 nm is not usable because of strong scattering effects in the welding plume. These scattering effects increase strongly with shorter wavelengths and cannot be avoided by using external illumination. Consequently the only practically usable spectral window for silicon-based cameras is between 650 nm and 1000 nm. It should also be noted here that the dynamic range of silicon-based cameras is usually not sufficient for a simultaneous detection of the thermal emissions originating from both the keyhole and the weld bead. This is only possible in the infrared spectral range on the longwavelength side of the thermal emission maximum of the welding process (above 3 μm).

Investigation of capillary and melt pool dynamics in deep penetration laser welding processes was the next step. High-speed cameras in the visual and infrared spectrum range offer excellent image quality and high frame rates but access to the process details is limited due to the small surface area of the weld zone. With these conventional diagnostic techniques it is thus not possible to observe the key mechanisms inside the volume of the material which essentially determine the behaviour of the welding process. To gain insight into process phenomena such as the shape and movement of the capillary or the melt flow behaviour in the weld bead, X-ray videography is the ideal instrument. In this work the development and implementation of a high-speed X-ray video system is described, which enables the observation of internal process phenomena with high frame rates combined with outstanding spatial resolution.

A combined time-synchronous measurement system involving X-ray video and near infrared high-speed video was used to gather three-dimensional information about the geometry of the weld zone and its dynamic behaviour. It was possible to precisely measure the capillary depth and its dynamic movement as well as the direction and velocity of fluid flows inside the molten pool and their influence on pore formation. These findings are crucial for the understanding of possible limitations on the monitoring and controllability of deep-penetration laser welding in an industrial setting.

Investigations with high-speed cameras revealed important geometric information on the formation of different welding failures and possibilities for their distinct detection with a passive camera mounted on the welding head coaxially to the laser beam. The geometric image properties observed are directly linked to specific failure mechanisms and do not relay on simple correlations between standard measures and observed welding failures. The automatic recognition of characteristic geometric image-properties was performed by software algorithms developed in Matlab as a proof-of-concept. The robustness of the algorithms developed was tested in an extensive experimental study to identify possible candidates for the development of the closed-loop control system. The so called full penetration hole (FPH) was identified as the most important image feature characterising the state of the process in terms of welding depth for a full penetration weld.

Using the aforementioned results it was possible to build up a camera-based closedloop control system to ensure stable welding results even under changing welding conditions. Optical integration into the laser welding head was realised by means of a dichroitic beam-splitter to ensure a coaxial configuration of the camera’s line of sight with the laser beam. This enables the system to be combined with two- or threedimensional scanning heads in remote welding processes.

The key-technology to overcome the above mentioned performance issue of camera based control systems is a novel architecture called a “Cellular Neural Network” (CNN). With “Cellular Neural Networks” it is possible to integrate basic processor elements in the electronic circuitry of a CMOS camera resulting in a Single-Instruction-Multiple-Data (SIMD)-architecture on the camera chip itself. Such pixelparallel systems provide extremely fast real-time image processing, since there is no need to transfer image data from the camera to a processor. The closed-loop control system developed in this work uses a CNN based camera surveying the contour of the full-penetration hole with a control frequency of up to 14 kHz for linear welding processes and up to 9 kHz for processes with variable welding trajectory, whereas the latency of the system is in the range of only one single frame.

An extensive experimental program was performed to validate the capabilities of the closed-loop control system. It was shown that the system is able to control the degree of full penetration in overlap joints with different steel grades. The closed-loop control successfully compensated for various external disturbances such as variations in material thickness and welding speed, as well as defocusing and the contamination of protective windows. These phenomena could be controlled with either linear or varying direction welding trajectories. Controlled full penetration welds are also possible in aluminum, but not with all alloys. In particular, the alloy AA5182 produces thermal camera images that are strongly disturbed by extensive fluctuations in brightness, which prevents reliable control. On the other hand, a closed-loop control of full penetration with the alloy group AA6000 is possible without difficulty. In closed-loop controlled full penetration welding conditions there is no need for the otherwise obligatory 10 % excess of laser power. This directly leads to a better quality of the root surface due to reduced spattering and smoke residue, as well as an energy saving or a corresponding gain in welding speed. As well as optimizing full penetration, the control system is also able to maintain a stable penetration depth in a partial penetration condition during welding of overlap joints in different steel grades. This is due to the fact that the image feature FPH is also visible when the laser beam reaches the gap between the two welding parts, enabling the system to use this working point as a datum for closed-loop control. Stable closedloop control of partial penetration welds in aluminum alloys could not be achieved in this work. Although the image feature itself was visible, the rate of false detection by the software algorithm was too high to ensure a stable operation. In the partial penetration mode the weld seam does not penetrate through the lower sheet bottom surface. This reduces the need for subsequent machining and also provides higher resistance against corrosion in car body welding. The further decreased laser power in comparison to full penetration mode offers the possibility of higher welding speeds, leading to higher productivity and cost efficiency. The limits of the camera-based closed-loop control system described in this work are basically due to the fact that the image feature of the full penetration hole (FPH) must be visible in the desired parameter field. This means also that a closed-loop control of an arbitrary welding depth, independent of the presence of any boundary surface, is generally not possible. The first welding results using this closed-loop control system were presented at the ICALEO-conference 2008 in Temecula, CA, USA, and the fully developed control system was honored with the third place in the “Berthold Leibinger Innovation Award 2012” and the third place in the “Steel Innovation Award 2012”.