Method and system for identifying and / or quantifying pore formation on at least one workpiece during a laser welding process

Abstract

The invention relates to a method and a system (10) for identifying and / or quantifying pore formation on at least one workpiece (13) during a laser welding process, wherein spatter (20) during the laser welding process is detected by means of a spatially and temporally resolving optical sensor (2) and a conclusion is drawn as to a pore increase or decrease from the correlation of the spatter number with the pore number, wherein a convolutional neural network (3) is used to evaluate the spatter data (4) provided by the optical sensor (2), and the convolutional neural network (3) carries out a spatially resolving qualification with a resolution of at least 2 pixels.

Description

[0001] Method and system for identifying and / or quantifying pore formation during a laser welding process on at least one workpiece

[0002] The present invention relates to a method and a system for identifying and / or quantifying pore formation during a laser welding process on at least one workpiece.

[0003] In the field of laser welding of metallic materials, particularly using continuous wave (CW) or pulsed laser technology, methods and systems for identifying and quantifying porosity are of central importance for quality assurance. The use of single-spot or 2-in-1 multimode or single-mode lasers has become established in industrial practice for producing precise and efficient welds. Despite these technological advances, however, significant challenges remain in the detection and quantification of porosity during the welding process.

[0004] Traditionally, pore detection is primarily performed using computed tomography (CT) or X-ray analysis after welding. While these methods offer high accuracy and detail in pore detection, they have significant drawbacks. CT and X-ray analysis are time-consuming and expensive, as they require extensive post-processing of the workpieces. Furthermore, the application of these techniques to large-scale production is often not economically viable due to long lead times and the need for specialized equipment.

[0005] Another significant disadvantage of CT and X-ray analysis is that they only allow for a retrospective inspection of the welds. This means that any quality defects can only be detected after the welding process is complete. This leads to increased scrap and

[0006] Post-processing costs are incurred because defective workpieces are only identified late in the production process. In addition to CT and X-ray analysis, destructive testing using metallographic images for porosity detection is also employed. This method involves cutting and polishing the weld seam to visualize the internal structures. Although this technique provides detailed information about porosity, it also has significant drawbacks. Destructive testing is time-consuming and results in the destruction of the workpiece, making it unsuitable for continuous quality control. Furthermore, it is labor-intensive and requires the use of qualified personnel, which increases production costs.

[0007] A significant disadvantage of all the aforementioned methods is that they do not allow for real-time monitoring of porosity during the welding process. This means that potential causes of porosity, such as incorrect welding parameter settings (BLW), faulty focus position (FL), focus shift due to contamination, material and batch variations, as well as positional or fixture deficiencies, cannot be immediately detected and corrected. The lack of real-time monitoring thus increases the likelihood of quality defects and reduces the efficiency of the welding process.

[0008] The object of the invention is therefore to provide a method and system for identifying and / or quantifying pore formation during a laser welding process on at least one workpiece, which avoids or at least reduces the aforementioned disadvantages.

[0009] This problem is solved by a method for identifying and / or quantifying pore formation during a laser welding process on at least one workpiece, wherein spatter is detected during the laser welding process using a spatially and temporally resolved optical sensor, and an increase or decrease in porosity is determined by correlating the number of spatter with the number of pores, wherein a convolutional neural network is used to evaluate the spatter data provided by the optical sensor, and the convolutional neural network performs a spatially resolved qualification with a resolution of at least pixels.

[0010] This method offers the advantage that the detection and analysis of spatter during the laser welding process, using a spatially and temporally resolved optical sensor in conjunction with a convolutional neural network (CNN), enables precise and efficient identification and quantification of porosity. By correlating the number of spatter with the number of pores, conclusions can be drawn about the increase or decrease in porosity. This allows for continuous monitoring of porosity growth and decrease during welding and / or between multiple welds. As a result, quality defects can be detected and corrected early, leading to a reduction in scrap and rework. Furthermore, the system provides early warnings of increasing porosity, enabling preventative measures to avoid quality problems.An additional advantage is that the CNN analysis results allow for controlled intervention to achieve low-porosity welding, significantly improving the overall weld quality. The spatially resolved qualification of the CNN, with a resolution of at least 2 pixels, ensures high accuracy in the acquisition and analysis of spatter data.

[0011] First, the individual elements of the claimed invention are explained in the order in which they are mentioned in the claim set, and subsequently, particularly preferred embodiments of the invention are described.

[0012] Optical sensor

[0013] For the purposes of this patent application, an optical sensor is a device that detects light and converts this signal into electrical data that can be used for further processing. An optical sensor consists of one or more light-sensitive elements, such as photodiodes or CMOS sensors, capable of detecting light in the visible or infrared spectral range. The optical sensor in this invention is used to detect spatter generated during the laser welding process. Its main function is to detect the light emissions of the spatter and convert this information into electrical signals that can then be evaluated by a convolutional neural network (CNN). The detection is spatially and temporally resolved, meaning that the sensor can determine both the precise position and the time of each detected spatter.

[0014] The optical sensor assembly comprises light-sensitive elements, preferably photodiodes or CMOS sensors, which are known for their high sensitivity and fast response times. These elements are capable of detecting even weak light emissions from splashes. The sensor's optical system includes lenses or other optical components that focus the light onto the light-sensitive element. These may also contain filters to select specific wavelength ranges and improve contrast. A signal processing unit, which amplifies and digitizes the signals generated by the light-sensitive elements, is also integrated. This unit may also include noise reduction and signal amplification functions.To improve the contrast between the detected splashes and the process zone, the optical sensor is advantageously equipped with a bandpass filter in the range of 800 nm to 1000 nm or alternatively with a broadband filter with a spectral width of <200 nm.

[0015] Preferably, the optical sensor is aligned coaxially to the laser beam. This means that the sensor is aligned directly along the axis of the laser beam, which provides an optimal view of the weld zone and ensures consistent recording conditions. Alternatively, the sensor can also be arranged non-coaxially, which can offer advantages in certain applications, such as better detection of spatter flying off to the side.

[0016] There are several possible designs for the optical sensor. Single-camera systems use a single camera with a suitable optical filter and a signal processing unit. This design is compact and cost-effective, making it particularly suitable for applications where space is limited. Multi-camera systems, which use several cameras to view the weld zone from different angles, allow for more comprehensive spatter detection and can further improve pore detection accuracy. CMOS and CCD sensors both offer high sensitivity and resolution. CMOS sensors are preferred for their fast readout times and low power consumption, while CCD sensors offer high image quality and sensitivity.Specialized sensors for infrared ranges, specifically designed for the detection of infrared light, can be useful in applications where the splashes mainly emit in the infrared spectral range.

[0017] Convolutional Neural Network

[0018] For the purposes of this patent application, a Convolutional Neural Network (CNN) is a special type of artificial neural network particularly suited for processing image data. A CNN is designed to analyze and classify input data, such as images, by applying multiple layers of convolution and pooling. The convolutional layers apply various filters to the input data to extract features such as edges, textures, or complex patterns. These features are further processed in subsequent layers to generate deeper and more abstract representations of the original inputs.

[0019] The function of a CNN within the scope of this invention is to analyze and qualify the spatter data acquired during the laser welding process. By using convolution operations, the CNN can identify specific features in the spatter data that indicate pore formation. Analyzing these features enables precise identification and quantification of pore formation, which is crucial for quality assurance and process optimization.

[0020] A CNN is structured in several layers, each performing specific tasks. First, the input layer, which receives the raw data, is followed by one or more convolutional layers. These convolutional layers use various filters to extract different features. Following the convolutional layers are pooling layers, which reduce the dimensionality of the data and highlight the most important features, thereby increasing the computational efficiency and robustness of the network. After several convolutional and pooling layers, at least one fully connected layer, also called a fully connected layer, is added, which combines the extracted features into a final classification or prediction.

[0021] Preferably, the CNN features an ll-Net architecture specifically designed for image segmentation. This ll-Net architecture consists of a contracting path that extracts features and progressively reduces the image data, and an expanding path that enlarges the data again, enabling precise segmentation. This architecture is particularly advantageous for pore detection because it considers both global and local information within the image, thus allowing for accurate and detailed analysis.

[0022] Possible implementations of the CNN include not only the ll-Net architecture but also other architectures such as ResNet (Residual Network), which, through the use of residual blocks, enables a very deep network structure and is therefore particularly efficient. Another possible implementation is DenseNet (Densely Connected Convolutional Network), in which each layer has direct connections to all subsequent layers, improving information flow and gradient propagation. Each of these architectures offers specific advantages and can be used depending on the requirements and complexity of the image data to be analyzed.

[0023] Spray data

[0024] For the purposes of this patent application, spatter data refers to information that captures and describes the characteristics of spatter generated during a laser welding process. This data is acquired by a spatially and temporally resolved optical sensor and includes parameters such as the number, size, position, speed, and trajectory of the spatter. Splash data enables precise analysis of the welding process by revealing the extent and distribution of spatter, thus allowing conclusions to be drawn about the weld quality and the formation of pores.

[0025] The function of spatter data is to serve as a basis for analyzing and evaluating weld quality. By analyzing spatter data, a correlation can be established between the number of spatter particles and porosity formation. This enables the precise identification and quantification of pores during the welding process. The spatter data comprises both spatial and temporal components, which are captured by the optical sensor. Spatial data describes the position and distribution of spatter on the weld seam, while temporal data captures the dynamics and development of spatter over time.

[0026] Preferably, the spatter data is analyzed by a Convolutional Neural Network (CNN) that performs spatially resolved qualification with a resolution of at least 2 pixels. This ensures high accuracy in the acquisition and evaluation of the spatter data. One possible implementation of the spatter data is the acquisition of sequential images, which enables detailed temporal analysis. Alternatively, maximum value images of an entire process recording can be created to identify and evaluate critical phases of the welding process.

[0027] Advantageously, the spatter data is further optimized by using bandpass or broadband filters to improve the contrast between the spatter and the process zone. This facilitates the accurate acquisition and evaluation of the spatter data. The application of semantic segmentation for pixel-accurate resolution of the spatter data enables precise identification and classification of individual pixels in the image, which further increases the accuracy of pore detection. This comprehensive and precise spatter data forms the basis for effective monitoring and control of the laser welding process to ensure high weld quality. Advantageous embodiments of the invention

[0028] According to an advantageous embodiment of the invention, the optical sensor can be aligned coaxially to the laser beam. This alignment allows a direct and unobstructed view of the process zone, leading to improved detection of spatter and thus a more accurate determination of porosity. The constant recording conditions contribute to the stability and reliability of the measurements, further improving the quality and reproducibility of the welding results.

[0029] An additional advantage arises from equipping the time-resolved optical sensor with a bandpass filter in the range of 800 nm to 1000 nm or, alternatively, with a broadband filter with a spectral width of <200 nm. This significantly improves the contrast between the detected splashes and the process zone, thereby increasing the accuracy and reliability of splash data detection and analysis. The improved differentiation of splashes from the background minimizes misinterpretations, leading to a more precise determination of pore formation.

[0030] Furthermore, according to an advantageous embodiment of the invention, the convolutional neural network can perform spatter evaluation based on sequential single images or maximum value images of an entire process recording. This enables a detailed analysis of the welding process over time, allowing for the precise detection of temporal changes in spatter behavior and thus in porosity formation. The use of maximum value images also allows for the identification of particularly critical phases of the welding process, enabling targeted quality control and optimization of the welding parameters.

[0031] According to a further particularly preferred embodiment of the invention, semantic segmentation of the recorded image material can be performed to achieve pixel-precise resolution. This technique allows for extremely precise identification and classification of the individual pixels in the image, enabling the determination of the exact position and size of the spatter. Pixel-precise analysis improves the accuracy of pore detection and allows for a detailed evaluation of weld quality. This enables even the smallest changes in pore formation to be detected and addressed accordingly.

[0032] Semantic Segmentation

[0033] For the purposes of this patent application, semantic segmentation is the technique by which each pixel of an image is assigned to a specific class or category. This method enables pixel-accurate classification and analysis of the image material, thereby allowing precise information about the objects and structures in the image to be obtained. Semantic segmentation is used in the present invention to identify and classify spatter and other relevant features in the image material of the laser welding process.

[0034] The function of semantic segmentation is to divide the welding process image data into individual segments and assign each segment to a specific class. This is achieved using a convolutional neural network (CNN) with a specific architecture, preferably an ll-net architecture. The CNN processes the image data and generates an output mask in which each pixel is assigned to a class, such as spatter or background. This pixel-precise resolution enables detailed analysis and evaluation of the spatter, which is crucial for identifying and quantifying porosity.

[0035] Semantic Segmentation involves several stages of image processing and analysis. First, the raw image data is processed by the CNN, using various filters and layers to extract features at different levels of abstraction. The U-Net architecture consists of an encoder and a decoder. The encoder progressively reduces the spatial resolution of the image, extracting increasingly abstract features. The decoder restores the original resolution and combines the abstract features to generate precise segmentation. A key advantage of the U-Net architecture is the use of skip connections, which allow detailed information from the early layers of the encoder to be transferred directly to the decoder, thus increasing the accuracy of the segmentation.

[0036] Semantic segmentation often employs additional techniques such as data augmentation and transfer learning to improve the robustness and accuracy of the segmentation. Data augmentation includes methods like rotating, scaling, and mirroring the image data to increase the diversity of the training data and make the model robust against variations in the image material. Transfer learning allows pre-trained models to be used as a starting point on large datasets, which reduces training time and improves model performance.

[0037] Semantic segmentation can advantageously be performed in real time, enabling continuous monitoring and analysis of the welding process. This real-time capability is particularly important for the early detection of quality problems and for implementing control interventions to optimize the welding process. Alternative implementations of semantic segmentation could involve the use of other network architectures, such as fully convolutional networks (FCNs) or DeepLab, which are also suitable for pixel-accurate image segmentation. Such alternatives could be advantageous in specific application scenarios, depending on the specific requirements for image processing and analysis in the respective welding process.

[0038] Advantageously, laser welding can be performed on at least two metallic workpieces with weld penetration depths ranging from <20 mm to >100 pm. This flexibility in weld penetration depth allows the process to be applied to a wide variety of welding tasks, from fine precision work to deep welds. This increases the versatility and applicability of the process in various industrial sectors, leading to wider acceptance and use of the technology.

[0039] Another advantage is the use of a spatially and temporally resolved camera to capture spatter, with an exposure time ranging from 1 ps to 20,000 ps and a frame rate greater than 20 Hz. These technical parameters ensure high-resolution and rapid acquisition of spatter data, thus increasing the accuracy and reliability of pore detection. Adjusting the exposure time and frame rate allows the camera to be optimally configured for different welding conditions, further enhancing the system's flexibility and performance.

[0040] It can also be advantageous to further develop the invention such that the Convolutional Neural Network has an ll-Net architecture. This specific network architecture is particularly well-suited for image segmentation and analysis, as it efficiently captures both global and local image features. The ll-Net architecture enables precise and detailed analysis of splash data, which further improves the accuracy of pore detection. Furthermore, the network's architecture allows for rapid and efficient training, reducing development and implementation times.

[0041] The object of the invention can further be achieved by a system for identifying and / or quantifying pore formation during a laser welding process, comprising:

[0042] - a spatially and temporally resolving optical sensor for detecting spatter during the welding process and providing spatter data representing the spatter during the welding process, wherein the optical sensor has a bandpass filter in the range of nm to nm or alternatively a broadband filter with a spectral width of < nm;

[0043] - A convolutional neural network for processing and qualifying the spatter data, wherein the convolutional neural network has a data connection to the optical sensor so that the spatter data can be transmitted from the optical sensor to the convolutional neural network, and the convolutional neural network performs spatially resolved qualification with a resolution of at least pixels; - A laser processing system comprising a laser for processing at least one metallic workpiece and a control unit for controlling the laser processing system, wherein the convolutional neural network is connected to the control unit of the laser processing system in such a way that processing parameters of the laser processing system can be changed depending on the analysis results of the convolutional neural network,

[0044] - Means for coaxial or non-coaxial alignment of the optical sensor to the laser beam;

[0045] The system for identifying and / or quantifying porosity during laser welding offers the advantages of an integrated and automated solution for monitoring and optimizing the welding process. By combining a spatially and temporally resolved optical sensor with a convolutional neural network (CNN) and a laser processing system, welding parameters can be adjusted in real time based on the CNN's analysis results. This leads to a continuous improvement in weld quality and a reduction in scrap and rework. The ability to align the optical sensor coaxially or non-coaxially with the laser beam provides additional flexibility in system integration and application.

[0046] The laser's optical fiber can be configured as either an active fiber laser or a passive disk laser, with the core fiber diameter ranging from 10 pm to 50 pm or from 50 pm to 400 pm. These configurations offer the advantage of covering a wide range of laser applications, from high-precision micro-welding to robust deep penetration welding. The various core fiber diameter options allow for optimal adjustment of the laser parameters to the specific requirements of the welding process, further improving weld efficiency and quality. Preferably, spatter detection is performed without additional external illumination. This configuration has the advantage of making the spatter stand out brightly against the dark background. This enables clear and high-contrast spatter detection, increasing the accuracy and reliability of pore detection.Another advantage of this design is that it eliminates the need for an illumination laser, reducing the system's complexity and cost. Alternatively, spatter detection can be achieved using an illumination laser in the 800 nm to 1000 nm wavelength range. This design offers the advantage of making dark spatter stand out against a light background. This high-contrast image also facilitates spatter detection and analysis, particularly in environments with varying lighting conditions. Furthermore, using an illumination laser can improve the consistency and quality of the images by ensuring uniform illumination of the weld zone. Both designs offer specific advantages and can be selected based on the requirements of the respective welding process and environmental conditions.The preferred option without additional lighting reduces energy consumption and system complexity, while the alternative option with illumination laser ensures optimal image quality in different lighting conditions.

[0047] According to a further advantageous embodiment of the invention, the optical fiber (OC) can also be provided with an additional protective coating that protects the fibers from environmental influences and mechanical damage, thereby increasing the service life and reliability of the system. The optical fiber (OC) can be either active (fiber laser) or passive (disk laser) with a core fiber surrounded by a ring fiber.

[0048] Advantageously, the LLK can also be equipped with variable core fiber diameters, which can be adapted according to specific welding requirements to improve the flexibility and adaptability of the welding process. The core fiber diameter d_core is in the range of 10 pm–50 pm (SM-like) or in the range of 50 pm–400 pm, particularly 50 pm–200 pm (MM). According to a further advantageous embodiment of the invention, the ring fiber of the LLK can be provided with a special coating that improves the reflective properties and thus increases the efficiency of light transmission, resulting in better weld quality. The outer diameter of the ring fiber d_ring is in the range of 40 pm–2000 pm, particularly 80 pm–800 pm (MM).

[0049] Advantageously, single-spot fibers with identical diameters as described above can also be used, as these allow for a consistent and uniform energy distribution on the workpiece, which increases weld quality and process stability.

[0050] A further advantageous embodiment of the invention provides that the core-to-ring diameter ratio can be dynamically adjusted to optimally respond to different material properties and welding requirements, thereby increasing the versatility and efficiency of the welding process. The core-to-ring diameter ratio is 1:2 to 1:10, particularly 1:4.

[0051] According to a further advantageous embodiment of the invention, the method can also be applied to alloys or coated metals in order to cover a wider range of applications and to improve the weldability of different materials. Laser welding can be performed on at least two metallic workpieces, e.g., based on iron, copper, and / or aluminum.

[0052] Advantageously, the workpieces can also be made of composite materials, which expands the application possibilities in modern manufacturing technology and increases the scope of the welding process. The workpieces can be made of the same or different materials.

[0053] According to a further advantageous embodiment of the invention, the weld penetration depth can be dynamically adjusted during the process to meet different welding depth requirements and to increase the flexibility of the welding process. The weld penetration depth is preferably in the range of <20 mm and >100 pm.

[0054] Advantageously, the aspect ratio (depth:width) of the weld can be adjusted depending on the application to achieve optimal mechanical properties of the weld and to ensure the strength and integrity of the joint. The aspect ratio (depth:width) of the weld is preferably in the range >=0.5:1 (deep penetration welding).

[0055] According to a further advantageous embodiment of the invention, other laser technologies with similar properties to TruDisk with BrightLine Weid or variMODE Star can also be used to further improve the adaptability and efficiency of the process. All TruDisk with BrightLine Weid or variMODE Star (the latter characterized by a ring fiber fed by several laser modules).

[0056] Advantageously, adaptive optics could be used to dynamically adjust the beam parameter product during the welding process in order to optimize weld quality under varying conditions. The beam parameter product of the core beam is preferably in the range of 0.38–16 mm*mrad, particularly <= 0.6 mm*mrad (single-mode similar) or <= 8 mm*mrad (multi-mode).

[0057] According to a further advantageous embodiment of the invention, the beam diameter could be adjusted by using variable focusing optics in order to respond flexibly to different workpiece geometries and welding requirements. The beam diameter of the core beam dwCore on the workpiece is preferably in the range of 10 pm to 300 pm, in particular 30 pm to 70 pm (single-mode) and 50 pm to 1200 pm (multi-mode).

[0058] Advantageously, lasers with wavelengths such as ultraviolet or X-ray lasers could also be used to address specific materials or applications requiring greater penetration depth or specific absorption properties. Infrared lasers with wavelengths in the 800–1200 nm range, particularly 1030 nm or 1070 nm, are preferred.

[0059] According to a further advantageous embodiment of the invention, a flying optic BEO with the same imaging ratios as the scanner optic PFO33-2 could also be used to increase the mobility and range of the welding system. Scanner optic PFO33-2 with an imaging ratio of 1:1 to 5:1, in particular 1.5:1 to 2:1.

[0060] Advantageously, other sensor technologies, such as hyperspectral cameras or thermographic sensors, could also be used to obtain additional information about the welding process and material properties. A spatially and temporally resolved area sensor, especially a camera with a CMOS or CCD sensor, enables highly precise recording and analysis of the welding process.

[0061] Advantageously, wider or more specialized filters could also be used to better capture specific features or material properties. The observed wavelength range is preferably in the 300–2000 nm range, particularly in the 800–1100 nm range, whereby the wavelength of the processing beam (e.g., 1030 nm or 1070 nm) must be blocked.

[0062] According to a further advantageous embodiment of the invention, the CNN could also be continuously improved through new data and learning algorithms to further increase its performance and accuracy. The Convolutional Neural Network (CNN) offers significant advantages in monitoring and optimizing the welding process due to its ability to precisely analyze and classify images.

[0063] Semantic segmentation can also be advantageously used in combination with other image processing techniques to achieve an even more comprehensive analysis and optimization of the welding process. Semantic segmentation enables pixel-accurate classification of the image data, resulting in extremely precise capture and analysis of the weld seam and spatter.

[0064] According to a further advantageous embodiment of the invention, other network architectures, such as DeepLab or Fully Convolutional Networks (FCNs), could also be used to achieve optimal results depending on specific requirements and conditions. The ll-Net architecture of the neural network, due to its special structure, offers a particularly efficient and

[0065] The invention will now be explained in more detail with reference to figures, without limiting the general concept of the invention.

[0066] It shows:

[0067] Figure 1 shows a schematic representation of a system for identifying and / or quantifying pore formation during a laser welding process.

[0068] Figure 1 shows a system 10 for identifying and / or quantifying pore formation during a laser welding process, comprising a spatially and temporally resolved optical sensor 2 for detecting spatter 20 during the welding process and providing spatter data 4 representing the spatter 20 during the welding process, wherein the optical sensor 2 has a bandpass filter 16 in the range of 800 nm to 1000 nm or alternatively a broadband filter 16 with a spectral width of <200 nm. In the embodiment shown, the optical sensor 2 is a camera 5 which has an upstream collimation lens 15.

[0069] The light emitted from the processing zone and the splashes 20 is directed to the camera 5 via an optical system, which includes, among other things, the semi-transparent mirrors 18, 19. Furthermore, the system 10 has a convolutional neural network 3 for processing and qualifying the splash data 4, wherein the convolutional neural network 3 has a data connection to the optical sensor 2, so that the splash data 4 can be transmitted from the optical sensor 2 to the convolutional neural network 3, and the convolutional neural network 3 performs spatially resolved qualification with a resolution of at least 2 pixels.

[0070] Furthermore, the system 10 includes a laser processing system 11 comprising a laser 12 for processing the workpieces 13a, 13b and a control unit 14 for controlling the laser processing system 11. The Convolutional Neural Network 3 is connected to the control unit 14 of the laser processing system 11 in such a way that processing parameters of the laser processing system 11 can be changed depending on the analysis results of the Convolutional Neural Network 3. This connection allows various causes for the formation of pores, such as incorrect BLW parameterization, faulty focus position (FL), focus shift due to contamination, changes in material and batch variations, as well as positional or fixture deficiencies, to be efficiently identified and corrected.

[0071] Incorrect BLW parameterization can lead to insufficient or excessive energy input into the workpiece, promoting porosity. By analyzing the welding data, the CNN can detect these deviations and instruct the laser processing system's control unit to adjust the parameters accordingly to ensure optimal energy distribution. Incorrect focus position (FL) can prevent the laser beam from being correctly focused on the weld zone, impairing weld quality. The CNN can detect focus deviations and provide the laser processing system's control unit with appropriate correction instructions to precisely align the focus with the weld area. Focus shift due to optics contamination can also affect weld quality by scattering or attenuating the laser beam.The CNN can identify such contamination by continuously monitoring the welding process and instruct the control unit to initiate cleaning procedures or adjust the optical parameters to restore optimal beam quality. Changes in material and batch variations can also lead to porosity, as different materials or batches may exhibit different welding properties. The CNN can detect these variations by analyzing the welding data and instruct the laser processing system's control unit to dynamically adjust the welding parameters to achieve consistent welding results. Positioning or fixture deficiencies, where the workpieces are not correctly aligned or fixed, can also cause porosity.The CNN can identify such deviations by detecting irregularities in the welding process and give instructions to the control unit to correct the positioning or adjust the fixtures to ensure optimal alignment and fixation of the workpieces.

[0072] Integrating the CNN with the laser processing system's control unit and enabling real-time modification of processing parameters allows for continuous monitoring and optimization of the welding process. This leads to a significant improvement in weld quality, a reduction in scrap and rework, and increased efficiency and reliability of the entire welding process.

[0073] A method for identifying and / or quantifying pore formation during a laser welding process on at least one metallic workpiece 13 is now designed such that spatter 20 is detected during the laser welding process by means of the spatially and temporally resolved optical sensor 2 and a conclusion is drawn about an increase or decrease in porosity by correlating the number of spatter with the number of pores, wherein a Convolutional Neural Network 3 is used to evaluate the spatter data 4 provided by the optical sensor 2, and the Convolutional Neural Network 3 performs a spatially resolved qualification with a resolution of at least 2 pixels.

[0074] The invention is not limited to the embodiments illustrated in the figures. The foregoing description is therefore not to be considered limiting, but rather explanatory. The following claims are to be understood as meaning that a named feature is present in at least one embodiment of the invention. This does not preclude the presence of further features. Where the claims and the foregoing description define 'first' and 'second' features, this designation serves to distinguish between two similar features without establishing any hierarchy.

[0075] List of reference signs

[0076] 1 workpiece

[0077] 2 optical sensors

[0078] 3 Convolutional Neural Network

[0079] 4 spray data

[0080] 5 cameras

[0081] 10 System

[0082] 11 Laser processing system

[0083] 12 lasers

[0084] 13 workpiece

[0085] 14 Control unit

[0086] 15 Collimation lens

[0087] 16 Optical filters

[0088] 18 mirrors

[0089] 19 mirrors

[0090] 20 splashes

Claims

- 22 - Claims 1. Method for identifying and / or quantifying pore formation during a laser welding process on at least one workpiece (13), characterized in that Splashes (20) during the laser welding process are detected by means of a spatially and temporally resolved optical sensor (2) and a conclusion is drawn about an increase or decrease in porosity by correlating the number of splashes with the number of pores, wherein a Convolutional Neural Network (3) is used to evaluate the splash data (4) provided by the optical sensor (2), and the Convolutional Neural Network (3) performs a spatially resolved qualification with a resolution of at least 2 pixels.

2. Method according to claim 1, characterized in that the optical sensor (2) is aligned coaxially to the laser beam.

3. Method according to claim 1 or 2 characterized in that the time-resolving optical sensor (2) is equipped with a bandpass filter in the range of 800 nm to 1000 nm or alternatively with a broadband filter with a spectral width of <200 nm.

4. Method according to one of claims 1 to 3 characterized in that the Convolutional Neural Network (3) evaluates the splashes (20) based on sequential single images or maximum value images of an entire conducts process recording.

5. Method according to one of claims 1 to 4, characterized in that a semantic segmentation of the recorded image material is carried out for pixel-accurate resolution.

6. Method according to one of claims 1 to 5, characterized in that the laser welding of at least two metallic workpieces (13) is carried out, the welding depth of which is in the range of <20 mm to >100 pm.

7. Method according to one of claims 1 to 6, characterized in that a spatially and temporally resolving camera (5) is used to detect the splashes (20), wherein the exposure time of the camera (5) is in the range of 1 ps to 20000 ps and the recording rate of the camera (5) is >20 Hz.

8. Method according to one of claims 1 to 7, characterized in that the Convolutional Neural Network (3) has an ll-Net architecture.

9. System (10) for identifying and / or quantifying pore formation during a laser welding process, comprising: - a spatially and temporally resolved optical sensor (2) for detecting spatter (20) during the welding process and providing spatter data (4) representing the spatter (20) during the welding process, wherein the optical sensor (2) incorporates a bandpass filter in range from 800 nm to 1000 nm or alternatively a broadband filter with a spectral width of <200 nm; - a Convolutional Neural Network (3) for processing and qualifying the splash data (4), wherein the Convolutional Neural Network (3) has a data connection to the optical sensor (2) so that the splash data (4) can be transmitted from the optical sensor (2) to the Convolutional Neural Network (3), and the Convolutional Neural Network (3) performs spatially resolved qualification with a resolution of at least 2 pixels; - A laser processing system (11) comprising a laser (12) for processing at least one metallic workpiece (13) and a control unit (14) for controlling the laser processing system (11), wherein the Convolutional Neural Network (3) is connected to the control unit (14) of the laser processing system (11) in such a way that processing parameters of the laser processing system (11) can be changed depending on the analysis results of the Convolutional Neural Network (3), - Means for coaxial or non-coaxial alignment of the optical sensor (2) to the laser beam; 10. System (10) according to claim 9, characterized in that the optical fiber of the laser (12) is either actively configured as a fiber laser or passively as a disk laser, wherein the diameter of the core fiber is in the range of 10 pm to 50 pm or in the range of 50 pm to 400 pm.

Images