System and method for error control of laser welding processes

Abstract

The invention relates to a system for error control of laser welding processes on a component machined by a laser machining installation, having an image recording device for producing two-dimensional image data of the component and an evaluation unit for the image data, wherein the evaluation unit is designed in such a manner that it calculates, by means of a convolutional neural network, an associated height value for the two-dimensional image data produced by the image recording device and produces a height profile of the component.

Description

[0001] System and method for error control of laser welding processes

[0002] In laser welding processes, monitoring the quality of a weld seam is essential to ensuring high-quality components. Therefore, after production, the weld seams are subjected to a post-process inspection, in which their quality is determined by evaluating image data of the weld seams. Furthermore, the component configuration before the welding process can be used to assess whether a stable process can be achieved. This is also determined by evaluating image data. In both applications, the height information of the component offers added value and can cover additional requirements for component configuration and quality assurance.

[0003] One example where weld monitoring is important is the manufacture of electric motors and generators whose stators are wound using hairpin technology. Adjacent hairpin ends are welded together using a laser according to a circuit diagram to establish electrical contact between the pins. Welds where the hairpin ends have been incorrectly welded together are reworked in a second step—where possible—to prevent the entire stator from becoming unusable. Weld quality is influenced by several parameters, such as material quality, beam deviations, and environmental factors. Defects can be identified, for example, by excessive material ejection or insufficient penetration of the laser beam into the component material.When welding hairpins, defects can be identified by factors such as excessively large or too small weld beads, spatter, or a mutual height offset between the pins before the welding process. Therefore, in addition to a two-dimensional image of a component's surface, the component's height profile is also important for detecting and characterizing defects.

[0004] Optical quality control of weld seams is typically performed either by a 3D scan or optical coherence tomography using a dedicated OCT (Optical Coherence Tomography) scanner on the laser processing system, or by grayscale photography using a camera that creates a two-dimensional grayscale image of the component. 3D scans produce a very precise three-dimensional image of the component, but require long acquisition times and are therefore difficult to integrate into an ongoing manufacturing process. With two-dimensional grayscale images, however, an analysis of the height structure of a weld seam is only possible with great inaccuracy and / or requires the use of multiple cameras.

[0005] DE 10 2018 129 425 A1 discloses converting image data and height data of a workpiece into information about the machining quality using a transfer function. The transfer function is formed by a deep convolutional network that can be adapted to changing situations, such as other workpieces, using transfer learning. The image and height data are acquired using an OCT scan, a stereo camera system, and / or a triangulation system. This known error control system is therefore relatively complex and requires a long evaluation time.

[0006] DE 10 2010 017 316 A1 describes a welding system equipped with multiple cameras that capture the component from several different points and create a stereoscopic image of a weld bead. From this image, the height and width of the weld bead can be calculated.

[0007] The present invention is based on the object of enabling a reliable and rapid error control of laser welding processes.

[0008] This object is achieved by a system for error control of laser welding processes on a component machined with a laser processing system, with an image recording device for creating two-dimensional image data of the component to be machined or machined and an evaluation unit for the image data, which is characterized in that the evaluation unit is set up in such a way that it calculates associated height values ​​for the two-dimensional image data created by the image recording device by means of a folded neural network and creates a height profile of the component.

[0009] From a single two-dimensional image, a height profile of the component can be generated using a convolutional neural network, which is preferably a deep convolutional neural network. The calculated height values ​​approximate the actual height values ​​of the component, and the deviation between the calculated and actual height values ​​can be kept very small by a well-trained neural network. The calculated three-dimensional image is a type of height map of the component, from which not only the location of a defect in a weld seam but also the type of defect can be determined. Creating the height map requires little computing time and can therefore be integrated into the component manufacturing process. A time-consuming three-dimensional scan of the component is not required for this. Such scans are only necessary during the network training phase.

[0010] Preferably, the image acquisition device can create two-dimensional image data for pixels of a pixel matrix of the image of the component it has captured, and the evaluation unit can calculate an associated height value for each of the pixels. The resulting height profile thus has an identical resolution to the two-dimensional image acquisition. Processing the two-dimensional image data by the neural network thus does not result in any loss of accuracy in the defect detection.

[0011] The evaluation unit can create a three-dimensional representation of at least a section of the component surface that is relevant for defect detection. This can result in the number of pixels in the two-dimensional representation being higher than in the three-dimensional representation of the component. Nevertheless, the resolution in the two-dimensional and three-dimensional representations can be identical.

[0012] To acquire the two-dimensional image data, the image acquisition device can have at least one camera aligned coaxially with a laser processing head of the laser processing system. This camera is preferably a grayscale camera. Color images of the component are, of course, also possible. The use of multiple cameras capturing images of the component from different positions can also be provided. The use of multiple cameras is particularly advantageous for training the convolved neural network and can replace a scanner.

[0013] However, the system can also incorporate a 3D scanner, particularly an OCT scanner, which creates a height profile of the component and transmits it to the evaluation unit. The camera images and height profiles serve as training data for the neural network. The scans are only required for training the network and are not constantly required for error control. Their high accuracy can be used for very effective training of the neural network.

[0014] Furthermore, if manufacturing conditions change, it is possible to create a scan of the component, which the evaluation unit uses to retrain the convolved neural network to calculate the corresponding height values ​​for the component's two-dimensional image data. This allows the neural network to adapt to new situations through transfer learning. A new manufacturing condition could, for example, be a new position of the component or a switch to a new component type.

[0015] The component's height profile generated by the evaluation unit can be compared with stored three-dimensional defect representations. This allows for immediate categorization of a detected defect and thus a corresponding quality assessment of the component.

[0016] The convolutional neural network can preferably be a modified UNet. These networks have proven their worth in image processing in general, and in semantic segmentation in particular.

[0017] The invention also includes a laser processing system with a laser processing head for producing weld seams on a component, which is characterized in that it comprises a system according to one of claims 1 to 8.

[0018] Furthermore, the invention relates to a method for error control of laser welding processes on a component machined with a laser processing system, comprising the steps:

[0019] - Acquisition of two-dimensional image data of the component surface to be processed or being processed,

[0020] - Calculating corresponding height values ​​for the two-dimensional image data using a convolutional neural network,

[0021] - Creating a height profile of the component, whereby the neural network was previously trained with image data and associated height profiles to calculate the height values.

[0022] The neural network can be retrained at any time if further refinement of the results is desired. The convolved neural network can also be adapted to changed manufacturing conditions, such as different components, component positions, or changed manufacturing parameters, through transfer learning of the network.

[0023] A use case of the system and its results are described in more detail below.

[0024] They show:

[0025] Fig. 1 ag shows various representations of a hairpin after a laser welding process; Fig. 2 ac shows a grayscale image, a 3D scan, and a 3D reconstruction with a system according to the invention of a first welded hairpin;

[0026] Fig. 3 ac a grayscale image, a 3D scan and a 3D reconstruction with a system according to the invention of a second welded hairpin.

[0027] Hairpin 1 shown in Fig. 1a has two copper rod ends that must be welded together to establish electrical contact. Hairpin 1 is part of a winding of an electric motor stator (not shown here). Hairpins are regularly laser welded, and checking the quality of the welds is particularly important here, as all hairpins of a stator must be welded together in such a way that the electrical contacts are reliably established.

[0028] Fig. 1b shows a good weld of the copper rods 10, 11 of the hairpin 1. The weld seam 12 has a smooth, slightly curved surface.

[0029] Figs. 1c to 1g, on the other hand, show various possible errors when welding hairpin 1. In Fig. 1c, hairpin 1 was not in the focus of the laser during welding. Fig. 1d shows a weld seam that was produced with insufficient laser power, while in Fig. 1e, hairpin 1 was welded with too much power. In Fig. 1f, copper rods 10, 11 are offset from one another, and in Fig. 1g, copper rods 10, 11 are not stripped. Figs. 2 and 3 illustrate how a weld seam 12 of a hairpin 1 can be represented and evaluated three-dimensionally using a system and method according to the invention.

[0030] Fig. 2a shows a two-dimensional grayscale image of a good weld seam 12 of a hairpin 1 (see Fig. 1b). Fig. 2c shows the three-dimensional representation of the weld seam 12 reconstructed from the grayscale image of Fig. 2a using a convolutional neural network. The convolutional neural network is also referred to as a convolutional neural network. For this purpose, the convolutional neural network calculates approximate values ​​for the actual height values ​​of the component for each pixel of the grayscale image from Fig. 2a. The lighter the color in the representation in Fig. 2c, the larger the height value, as the scale next to Fig. 2 illustrates. For comparison, Fig. 2b shows a three-dimensional OCT scan of the same hairpin. The computational reconstruction in Fig.2c deviates only insignificantly from the OCT scan and illustrates that with the system and method according to the invention, an assessment of the quality of a weld seam 12 is just as reliable as with the aid of a much more complex OCT scan.

[0031] Fig. 3a shows a grayscale image of a defective weld seam. The three-dimensional representation reconstructed from this in Fig. 3c, with its lighter color compared to Fig. 2c, clearly shows that the weld seam is too high. From this and the still clearly visible contours of the two copper rods 10, 11 of hairpin 1, it can be concluded that the hairpin in Fig. 3 was welded with insufficient laser power, similar to the hairpin shown in Fig. 1c. Here, too, Fig. 3b shows a three-dimensional OCT scan of the same hairpin for comparison, which again proves that the height values ​​in Fig. 3c were calculated very accurately by the convolved neural network. The convolved neural network used to calculate the three-dimensional reconstructions of the hairpins shown in Figs. 2c and 3c was trained using three-dimensional OCT scans of hairpins.The network is part of an evaluation unit (not shown), which can also store comparison images of faulty welded hairpins. This allows the evaluation unit to directly evaluate and categorize the welds based on the three-dimensional reconstructions of the hairpins in Figs. 2c and 3c.

[0032] It is understood that the system and method according to the invention can also be applied to weld seams of components other than hairpins. Furthermore, the convolved neural network can be modified so that other laser processing steps of a component, such as laser cutting, can also be evaluated by three-dimensionally reconstructing the component from grayscale images.

Claims

Claims: System for error control of laser welding processes (12) on a component (1) to be processed or processed with a laser processing system, comprising an image acquisition device for generating two-dimensional image data of the component (1) and an evaluation unit for the image data, characterized in that the evaluation unit is configured such that it calculates corresponding height values ​​for the two-dimensional image data generated by the image acquisition device using a convolutional neural network and generates a height profile of the component (1). System according to claim 1, characterized in that the image acquisition device generates two-dimensional image data for pixels of a pixel matrix of the image of the component (1) it captures, and the evaluation unit calculates a corresponding height value for each of the pixels.System according to claim 1 or 2, characterized in that the image acquisition device comprises at least one camera aligned coaxially to a laser processing head of the laser processing system. System according to claim 3, characterized in that the at least one camera is a grayscale camera. System according to any one of the preceding claims, characterized in that the system comprises a scanner, in particular an OCT scanner, which creates a scan of the component (1) and transmits it to the evaluation unit, which derives the following from the scan: A folded neural network is trained to calculate the associated height values ​​for the two-dimensional image data. A system according to claim 5, characterized in that, upon a change in manufacturing conditions, it creates a scan of the component (1), with the aid of which the evaluation unit retrains the folded neural network to calculate the associated height values ​​for the two-dimensional image data of the component (1). A system according to any of the preceding claims, characterized in that the evaluation unit creates a three-dimensional representation for at least a section of the component surface. A system according to any of the preceding claims, characterized in that the evaluation unit compares the generated three-dimensional representation of the component (1) with three-dimensional defect representations stored in the evaluation unit.System according to one of the preceding claims, characterized in that the folded neural network is a modified UNet. Laser processing system with a laser processing head for generating weld seams on a component (1), characterized in that the laser processing system comprises a system according to one of the preceding claims. Method for error control of laser welding processes on a component (1) to be processed or processed with a laser processing system, comprising the steps of: - Capturing two-dimensional image data of the machined component surface, - Calculating corresponding height values ​​for the two-dimensional image data using a convolutional neural network, - Creating a height profile of the component (1), whereby the neural network was previously trained with image data and associated height profiles of the component (1).

Images