Condition monitoring for the condition of the cutting table of a laser cutting machine

Abstract

The present invention relates to a training method (T), a detection method (D), and a system (20) for detecting the condition of a support table (2) of a sheet metal laser processing machine (18), in particular defects in the support table (2), by using a trained condition detection network (sdnw). An optical sensor (1) is installed above the support table (2), so that the optical sensor (1) and the support table (2) are movable relative to each other. An output interface (200, 200', 200") is configured to provide a prediction result.

Description

[Technical Field]

[0001] The present invention relates to the field of laser cutting and automatic or digital monitoring of the state of a support table for a laser processing machine, and in particular to a training method, a detection method, a system for detecting the state of a support table for or in a laser processing machine, and a computer program. [Background technology]

[0002] In laser cutting, monitoring the condition of the support table is important to ensure high-performance cutting results and / or high reliability of the laser machine or system as a whole. The cutting table supports the workpiece during cutting and is therefore subject to wear and tear.

[0003] For example, if a workpiece, such as a sheet metal workpiece to be cut, is placed on a faulty support, the flat workpiece may no longer be supported uniformly by the support, ultimately resulting in poor cutting results. For example, the support may include multiple slats, each with multiple tips. All tips of all slats form a two-dimensional plane as a support structure. If one of these tips is burned out, for example, due to wear and tear, the entire set of tips may no longer form a precise plane, leading to inaccurate support and poor cutting results.

[0004] When cutting, especially when cutting small pieces and / or holes on a support table with grid-like slats, small cut pieces, also called puzzle pieces, may become welded to the slats. After the cutting is completed, the cut pieces are removed and sorted, after which the welded puzzle pieces cannot be removed and remain on the support table. Unfortunately, these puzzle pieces can prevent the subsequent metal sheet from being supported flatly, resulting in poor cutting. If even one puzzle piece is on the slats, it can cause the workpiece sheet to tilt or wobble.

[0005] Another situation that needs to be avoided is when molten metal and metal slag adhere to the slats below the cut. After prolonged cutting (especially flame cutting), such molten metal and slag can accumulate to a significant and disturbing slag pile, which can prevent flat support, which again can lead to poor cutting. The slag pile can also become large enough to significantly reduce the gap between the slats, which can prevent the forks of the loading / unloading system from loading / unloading the sheet.

[0006] Defects such as "welded parts", "slag buildup" and "burnt tips" are primarily related to the cutting results, while defects such as "chips between slats" and "bent and missing slats" are primarily related to the automation system, especially the automatic sorting and / or removal system.

[0007] Therefore, it is necessary to monitor the condition of the cutting table to provide high quality cutting results.

[0008] Furthermore, it should be noted that machine development generally strives for greater machine autonomy, enabling production to be carried out with fewer personnel. Therefore, autonomous monitoring of machines, systems, and processes must constantly be improved. Appropriate sensor technology and intelligence are becoming increasingly important. Flatbed cutting machines equipped with loading and unloading systems must also follow this trend and become increasingly autonomous. In recent years, cutting processes in particular have become more autonomous and unmanned. Furthermore, the loading and unloading processes of sheet metal cutting machines should also be improved in terms of autonomous monitoring. In particular, it is necessary to be able to take into account possible defects in the support table during loading / unloading or other associated processes, such as cleaning the support table. The detected state of the support table should be taken into account in order to appropriately control these processes. In contrast, for example, currently available cleaning systems are not automated and / or are "blind" regarding the state of the support table (especially with regard to wear and defects), which may lead to inappropriate cleaning cycles and times (too long or too short). Furthermore, the refurbishment or replacement of the support base or its parts (such as slats) should be carried out appropriately according to the actually detected state of the support base (not too early or too late).

[0009] In the state of the art it is known to monitor the support table by means of sensors.

[0010] Patent document 1 describes a method and a device for determining the actual condition of the support bars of a workpiece support. For this purpose, it is proposed to determine the shape of the support protrusions on the support bars using a light section method. In one possible embodiment, it is proposed to process an image of the entire workpiece support by a convolutional artificial neural network in order to detect improper shapes of the support bars.

[0011] [1] compares conventional image processing methods for locating support slats in a single image, explicitly ignoring neural networks, as it is considered that too few images are available for adequate training.

[0012] Non-Patent Document 2 describes some general approaches for surface defect detection using deep learning techniques.

[0013] US Pat. No. 5,699,499 discloses a method for detecting the position of the support bars or slats of a cutting table by means of structured light and a camera.

[0014] Furthermore, US Pat. No. 5,629,999 proposes to determine the actual shape of the support protrusions (tips) of the support bars using a light section method while a light beam is scanned longitudinally across the platform.

[0015] Finally, in Patent Document 4, the table is imaged before processing in the first step, and the table is imaged after processing in the second step, and the state of the table before and after processing is compared to detect more subtle structural changes in the cutting table, thereby addressing wear of the cutting table.

[0016] In addition to the above-mentioned cutting table monitoring system, there are several known documents dealing with cleaning and removing slag and defects on the cutting table. However, these proposed systems basically operate blindly and without visual sensors. For example, Patent Document 5 proposes a slag removal system.

[0017] A drawback of these systems is that they are unable to learn based on empirical data.

[0018] US Pat. No. 6,299,499 discloses a camera-based detection of a support member for a particular defect, namely slag buildup, where the camera is fixed to the support base.

[0019] As mentioned above, machines and their loading / unloading systems are becoming increasingly autonomous, requiring more sensors for condition monitoring and the intelligence that comes with it.

[0020] Currently available cleaning systems are not unattended and / or blind, resulting in inadequate (too long or too short) cleaning cycles and times.

[0021] A monitoring system capable of detecting wear and defects in the cutting table allows for deliberate cleaning of the table, replacement of the table or replacement of elements / slats of the table. [Prior art documents] [Patent documents]

[0022] [Patent Document 1] US Patent No. 2023 / 0001522A1 [Patent Document 2] German Patent Invention No. 102019104649A1 [Patent Document 3] International Publication No. 2021 / 185899A1 Brochure [Patent Document 4] International Publication No. 2021 / 215429A1 Brochure [Patent Document 5] US Patent No. 2007 / 0215250A1 [Patent Document 6] JP 2021-171786 (A2) [Non-patent literature]

[0023] [Non-Patent Document 1] F. Struckmeier et al. “Methods for the localization of supporting stats of laser cutting machines in single images”, Forum Bildverarbeitung2020,1,11,2020, ISBN978-3-7315-1053-8 [Non-patent document 2] P. Bhatt et al. “Image-Based Surface Defect Detection Using Deep Learning: A Review”, Journal of Computing and Information Science in Engineering, vol.21, no.4, 1.08.2021, ISSN:1550-9827, DOI10.115 / 1.449535 Summary of the Invention [Problem to be solved by the invention]

[0024] SUMMARY OF THE INVENTION It is therefore an object of the present invention to provide a solution for improving the automatic detection of the state of the support table and avoiding the above-mentioned drawbacks of the state of the art. [Means for solving the problem]

[0025] In a first aspect, the invention relates to a training method for training a neural network as a state detection network according to claim 1.

[0026] The training method may include at least one convolutional neural network to detect the state of a support table of a laser processing machine that is a sheet metal laser processing machine and defects in the support table. The method may include the following method steps: Receiving an image from an optical sensor located above the support platform, the optical sensor and the support platform being movable relative to one another. Labeling the received image or a portion thereof with a label and analyzing the state of the support platform represented in the received image. Using a condition detection network, the labeled image or a portion thereof is input and a class label detection and / or prediction vector is provided as a result, which represents the condition of the support base, i.e., a defective condition or a normal condition. During use of the state detection network, gradually adjusting the weights of the state detection network to minimize the loss function and provide a trained state detection network.

[0027] The sheet metal laser processing machine is preferably a flatbed laser processing machine, and in this application is also referred to as a laser processing machine for short. The laser processing machine may be a laser cutting machine or a drilling machine. The laser processing machine may be used for punching and / or engraving. It may also be a flatbed laser cutting machine for processing flat or two-dimensional workpieces. The laser processing machine may be operated at a laser power exceeding 1 kW, preferably in the range of 1 to 100 kW, particularly 6 to 30 kW. The laser processing machine may preferably be part of a system. In other words, the system may include the laser processing machine. The system may include an automation system. The automation system may include an automatic sorting and / or loading / unloading system. The optical sensor may be attached or installed in the automation system or part thereof, and may not be attached to the laser processing machine or its support base. The system, particularly the automation system, may include a rack that functions as a gantry structure for a robot arm. In one embodiment, the optical sensor may be attached to the robot arm and provided on the rack.

[0028] The workpiece may consist of a metal, an alloy, a synthetic material, or a combination thereof. The workpiece is a flat workpiece, in particular a sheet metal workpiece.

[0029] The optical sensor is configured to provide at least one image. The optical sensor may be or may include a camera, particularly a CMOS and / or CCD camera. The optical sensor may be a 2D camera. Alternatively or additionally, the optical sensor may be a 3D camera, such as a stereo vision camera or a light field camera. The optical sensor preferably has a field of view covering the entire width of the support platform. Alternatively or additionally, the optical sensor may have a field of view covering the distance between each two slats of the support platform or at least one-third of the length of the support platform. For particularly long support platforms (e.g., in the range of 10 to 15 meters, particularly 12 meters), the field of view needs to be selected appropriately. Typically, the field of view (FoV) is in the range of 1000 x 1000 mm to 8000 x 8000 mm, with 2000 x 2000 mm being preferred. The FoV can be selected depending on the length and width of the support platform. At least the width of the platform (1500 mm / 2000 mm) needs to be covered. The FoV does not necessarily have to be square (e.g. a 2000x1000mm FoV would work well).

[0030] The optical sensor is disposed above the support table. The optical sensor is disposed facing the support table. The optical sensor is in data exchange with at least a (central) processing unit, where the (central) processing unit may be implemented in, among others, the laser processing machine, an automation system, and / or a server-based processing unit.

[0031] The optical sensor is provided so as to be movable relative to the support base.

[0032] In a first embodiment, the optical sensor itself is movable. In particular, the optical sensor itself may be movable, or may be attached to or coupled to a system or part thereof that is movable relative to the support base, for example, an automated system. Alternatively, the optical sensor may be mounted on a rack, in particular a frame rack, which is positioned above the support base. The rack may include a horizontal bar extending along the longitudinal axis of the support base. The horizontal bar may cover at least most of the length of the support base. The optical sensor may be mounted on a slidable carrier mounted on the bar. The optical sensor may be moved along the longitudinal direction of the bar via the slidable carrier. The longitudinal direction of the bar corresponds to the longitudinal direction of the support base.

[0033] In a second embodiment, the light sensor may be provided in a fixed position (eg in an automated system) and the support platform is movable relative to the light sensor, eg a camera, for example by translational movement in a horizontal plane.

[0034] Alternatively or additionally, the optical sensor is configured such that the received image acquired by the optical sensor covers at least the full width of the support table. Alternatively or additionally, the optical sensor is configured such that the full width of the support table is not covered by the field of view. In particular, a central processing area of ​​the workpiece being processed may be covered, in which case the width of the processing area is narrower than the width of the support table.

[0035] Alternatively or additionally, the optical sensor is configured such that a single image capture is sufficient to represent the entire support platform.

[0036] Alternatively or additionally, only one single photosensor may be used for image acquisition.

[0037] Alternatively or additionally, an optical sensor already installed in the laser processing machine or its environment may be used. The environment of the laser processing system may in particular be an automated system such as a sorting system and / or an input (output) system. In other words, preferably, a camera or optical sensor that is already part of or installed in the automated system may be used.

[0038] The optical sensor can be installed inside the cutting cell of the laser processing machine, particularly on the ceiling of the cutting cell, and can be controlled to detect an image when the support table is not supporting the workpiece to be cut, and the support table or the structure of the support table can be detected using the optical sensor.

[0039] Alternatively, an optical sensor can be installed outside the cutting cell to monitor the support table before a new workpiece, e.g., a metal plate, is introduced to be processed. The resolution and camera settings (including the distance from the camera to the support table) should be selected so that a resolution of 1-2 mm is achievable. In one embodiment, a camera with an effective resolution of 24 MP (6000 x 4000) can be used, mounted 76 cm from the support table, using a 16 mm focal length lens.

[0040] The support table may be a cutting table for supporting the workpiece, in particular during the cutting phase and when the workpiece is loaded onto the support table (before cutting) and removed from the support table (after cutting). The support table may also be a shuttle table, which consists of two table elements that are used alternately, one element being inside the cutting cell when the other element is outside, with the elements of the cutting table alternating inside and outside. One table element outside the cutting cell is intended to load and / or unload a workpiece, for example a metal plate, and is intended to be imaged by an optical sensor, in particular when the support table is unloaded (no workpiece is placed thereon and the slats of the support table are visible).

[0041] The support base may include a plurality of slats (also called lamellae) arranged parallel or in series to form a two-dimensional support surface for the workpiece being processed by the laser processing machine. The slats function as support structures and may exhibit longitudinal length extensions. The slats of the support base may include tooth-like support structures extending upward. In one embodiment, the slats may be arranged essentially parallel. Alternatively or additionally, the slats may exhibit curved length extensions and may be arranged in series. The slats may be curved. In side view, the slats may have a quasi-rectangular shape, and their upper side may not be linearly formed and / or may have tips and / or may be formed according to a regular or irregular sawtooth shape. The tips are subject to wear and damage.

[0042] A neural network is a digital computing system for performing specific tasks, such as object detection and / or condition detection related to the support platform. In particular, neural networks are useful for detecting the condition of the support platform. Neural networks can be configured to generate predictions of the support platform condition, particularly defects such as unintentional welded parts, slag buildup, burned slat tips, parts between slats, bent or missing slats in the support platform, and / or others.

[0043] The neural network in the form of a condition detection network may include an object detection portion and a classification portion. In particular, the detection of welded parts, burnt tips, and chips between the slats of the support base is performed by applying or using a trained object detection network. The detection or identification of slag buildup, bent slats, or missing slats is performed by applying or using a trained classification network.

[0044] The neural network can be trained by a training method, in particular a supervised training method (using labels), and provided as a trained network, which can be used or applied in this trained form in the inference phase. It should be noted that in a preferred embodiment, the training of the state detection network and the application of the trained state detection network (in the inference phase) are preferably performed on different computing entities. The training can be performed on a central server, which may be data-connected to a set of laser processing machines. The inference can be performed on another computing unit, in particular a coupled or integrated computer such as a controller, which may be located locally in the environment of the laser processing machines.

[0045] Neural networks can be implemented as convolutional neural networks and attention-based neural networks by integrating convolution and self-attention. For more information, see "CoAtNet: Marrying Convolution and Attention for All Data Sizes," September 15, 2021, Zihang Dai, Hanxiao Liu, Quoc V. Le, Mingxing Tan, Google Research, Brain Team.

[0046] The main technical advantage of the CNN architecture is its isovariance of transformations, which allows it to take advantage of data augmentation techniques by using less input data for training, making it more widely applicable to unknown cases (i.e., other or unknown conditions of the support).

[0047] In one embodiment of the present invention, the CoAtNet architecture may be used for state detection networks, particularly in classification tasks. The state detection networks described herein are provided in untrained and trained states.

[0048] The CoAtNet architecture combines the advantages of CNN and attention mechanisms to process input data. The main features of the CoAtNet architecture include: Hybrid Model: The CoAtNet model combines convolutional layers, which effectively capture local features of the input data, with attention layers, which capture long-term dependencies and relationships between features. Multi-stage architecture. CoAtNet models typically have a multi-stage architecture consisting of multiple blocks. Each block contains a combination of convolutional and attention layers, which allows the model to capture more complex and abstract features at each stage. Feature fusion. The CoAtNet model contains a feature fusion block that combines features learned from previous stages. This block helps the model learn a unified representation of the input data, which may improve performance on downstream tasks. Scalability. The CoAtNet model can be scaled up or down to accommodate different computational resources and input data sizes. This scalability allows the model to be used for a wide range of applications, from small-scale tasks such as image classification to large-scale tasks. Outstanding performance. The CoAtNet model achieves excellent performance across a wide range of tasks, including image classification and object detection.

[0049] To employ attention-based mechanisms, transformers are used (which are known to require large amounts of data to prevent overfitting). In particular, the CoAtNet architecture uses transformers to calculate attention weights between different regions of the input data. These attention weights are then used to combine features from the different regions in a weighted sum, allowing the model (network) to focus on the information most relevant to a given task.

[0050] Incorporating transformers into the CoAtNet architecture allows the model to capture both local and global relationships between features in the input data, which can lead to improved performance on a variety of tasks. Overall, the use of transformers in the CoAtNet architecture is one of its key features.

[0051] The CoAtNet architecture uses downsampling of the input image. The CoAtNet architecture can include four different stages or blocks. 1. A first stage with a single-layer or multi-layer network that downsamples or scales down the input to reduce the resolution. 2. A second stage with one or more convolution blocks. 3. A third stage with one or more transformer blocks with attention computation. 4. A fourth stage with a feature fusion block that combines features learned in previous stages. In particular, the fourth stage takes the output from the third stage and applies a series of convolutional and attention layers to learn a unified representation of the input data. The output of the fourth stage is typically passed to a final classification or regression layer to generate the model's prediction, in this case the state of the support (faulty or non-faulty).

[0052] The architecture exhibits improved generalization capabilities and, in parallel, improved model capacity. In particular, the model can be used with two or more convolution blocks and two or more transformer blocks. Preferably, CoAtNet-0 is used. The CoAtNet architecture can include 2+2+3 convolution blocks followed by 5+2 transformer blocks.

[0053] As a learning algorithm, a gradient descent algorithm can be used to find a local minimum of a function, resulting in the neural network algorithm converging to the local minimum. In particular, the Adam optimization algorithm can be used, which is an algorithm for first-order gradient-based optimization of a stochastic objective function based on adaptive estimates of low-order moments.

[0054] The loss function may be implemented as a cross-entropy loss. The cross-entropy between two probability distributions based on the same event set measures the average number of bits required to distinguish an event from the set when the encoding scheme used for the set is optimized for the estimated probability distribution rather than the true probability distribution.

[0055] In general, a training method (also called a machine learning algorithm) may process a dataset used to train a neural network using a machine learning algorithm. The dataset consists of example output data (in this case, labels or labeled images) and a corresponding input dataset (in this case, image data) that influences the output. Using the training method, the input data is passed through the algorithm and the processed output is correlated with the example output. The training method may involve adapting the weights of the neural network so that a loss function is minimized. This may be implemented by a gradient descent algorithm. Alternatively or additionally, other iterative optimization algorithms may be applied.

[0056] According to the present invention, detecting the state of the support table is selected from the group consisting of: A) Detecting welded parts; B) Detecting burnt tips; C) detecting small pieces (also called "parts") between the slats of the support; D) Classifying slag deposits; E) Classifying the bent slats, and F) Classifying missing slats.

[0057] Training is performed so that a set of different conditions, and in particular a set of different defects, can be detected in a single run or simultaneously with the condition detection network.

[0058] The trained condition detection network is trained to provide predictive results for multiple defects or defect conditions in the support platform. The trained condition detection network can detect all or some of conditions A) through F), and can specifically detect welded parts and detect and / or classify slag buildup. The trained condition detection network is comprehensive and encompasses defect identification for a variety of defects in the support platform.

[0059] Thus, in the inference phase, the trained neural network may be used in a detection method to provide predictions that include predictions for all of the above conditions in a single run. In this manner, the trained condition detection network may be configured to provide comprehensive condition detection that includes the above conditions or a subset of different conditions, particularly defects, including two or more defects, such as chip defects or slag buildup, as are known in the art. The condition detection network may be trained to provide predictions for two, three, four, or even a subset of the above set of A)-F), e.g., A) and D). Alternatively or additionally, the entire set of A)-F) or various subsets thereof may be calculated and provided.

[0060] The detection of various conditions, in particular defects, of the support platform by the trained condition detection network can be performed based on a single image acquisition. The image acquisition may include one image. This single image can show one condition of the support platform. This single image can have two or more sections, each of which can represent a part of the support platform, so that all sections together can cover the entire platform, typically with dimensions of 1 m x 3 m. Therefore, it is possible to identify defects from only one image. Alternatively or additionally, an image sequence (e.g., images from a video stream) can be used.

[0061] Generally, detecting the condition of a support platform refers to detecting a series of different defect or normal (non-defective) conditions.

[0062] Detecting the condition includes detecting various conditions in parallel, such as detecting welded parts on the support base (use case 1), burnt tips of the support base slats (use case 3), chips or parts between the support base slats (use case 4), slag buildup (use case 2), bent slats (use case 5), and / or missing slats (use case 6), among others.

[0063] Detecting the condition may include object detection, for example, as described above, the object may be or include a welded part, a burnt tip and / or slat, etc.

[0064] Classifying refers to classifying the detected state into different, preferably two, classes (e.g., slug or no slug, bent slat or not bent slat, missing slat or not missing slat).

[0065] The results (also called "prediction results") are a digital data set. The results may include class label detection and / or prediction vectors. The results represent the condition of the support platform, defective or normal. In the case of the object detection portion of the condition detection network, the results may also include a location indication representing information about where the defect is located.

[0066] When object detection is used or an object detection network is used, a class label is output. In general, different objects can be detected by the same network. Therefore, for each detected or discovered object, both a bounding box and a class label are output by the object detection network. The class label detection can be detection of a welded part (yes / no) in use case 1, detection of a chip between slats (yes / no) in use case 4, detection of a bent slat (lamella) (yes / no) in use case 5, or detection of missing slat(s) (yes / no) in use case 6. The prediction vector can include object detection for, for example, a welded part (use case 1), a burnt tip (use case 3), or a chip between slats (use case 4). The object detection can preferably include a position dataset.

[0067] The predicted vector is preferably output only to the object detection network. The predicted vector contains the bounding box information (x, y, width, height) of each identified object. The object detection network is used in use case 1 (detection of welded parts), use case 3 (detection of burnt edges), and use case 4 (detection of small pieces between slats).

[0068] The results representing the detected condition, whether classified or detected, may include a location dataset representing the location where the defect condition was detected. For example, the location dataset may indicate the location where a welded part is welded on the support platform, particularly the slats of the support platform, or the location where a slug is deposited on the support platform or its structure (e.g., slats).

[0069] The results may be presented as a graphical display, showing the support platform and any defect conditions (including location information) detected thereon.

[0070] Alternatively or additionally, the results may be provided as a textual display, for example in the form of a message, which may be provided on a user interface and / or forwarded by data communication means (e.g., a network connection) to an external device (e.g., a handheld or mobile device).

[0071] The main advantage of the solution presented here is that different types of defect conditions are shown in a single way and / or in particular in a common display, in particular a graphical display.

[0072] Alternatively or additionally, if the results indicate a fault condition, a post-processing algorithm may be applied, the post-processing algorithm being adapted to calculate control instructions based on or depending on the results.

[0073] Control instructions can be generated such that the cleaning module is controlled by these control instructions to automatically clean the support platform by removing elements if a buildup of welded parts or slag is detected.

[0074] If a burnt tip is detected, the control command may function to replace the defective slat or lamella with a new one, which has a defective tip.

[0075] If a part or debris is detected between the slats of the support platform, the control instructions function to remove the part. Part removal may be performed automatically using an ejection system and / or by a robot. Alternatively or additionally, part removal may be performed manually.

[0076] Alternatively or additionally, the control instructions may function to control the discharge system, in particular to issue commands to ensure that neither the discharge system or parts thereof, nor the workpiece or cut part, is damaged due to the detected part between the slats of the support table. Alternatively or additionally, the removal of the part may be performed by cutting out the part using a laser processing machine. In this case, the control instructions may function to control the laser processing machine to cut off the detected piece between the slats and remove it therefrom.

[0077] The post-processing algorithm may be adapted to issue a warning message, which, like the result message, may be sent to an external device.

[0078] The post-processing algorithm may be adapted to provide control instructions to control the laser processing machine to prevent cutting in areas detected as defective. In other words, the post-processing algorithm may be adapted to modify the cutting plan, i.e., to use other areas of the workpiece that are free of defects. Thus, the post-processing algorithm may provide control instructions to the laser processing machine and / or the cleaning module (or machine).

[0079] Alternatively or additionally, the provided trained neural network is tested to avoid overfitting, which may improve the network's ability to respond to new examples.

[0080] Alternatively or additionally, the labeled images are applied to the neural network after being augmented, the augmentation being performed by an augmentation algorithm.

[0081] A data augmentation algorithm is an algorithm that provides additional input data calculated from received images for data enrichment of the training data. The data augmentation algorithm may use varying image parameters selected from the group consisting of brightness, saturation, rotation, translation, zoom, color jitter, inversion, blur, and contrast. Alternatively or additionally, a mosaic data augmentation algorithm may be applied to generate a synthetic composite image from a set of received images.

[0082] Alternatively or additionally, labeling may involve graphically drawing bounding boxes within the image that represent the defects or defect regions around the defects, which may be used in an object detection network or object detection task.

[0083] Preferably, the bounding box completely surrounds the defect, especially the welded part (also called the welded puzzle piece). The bounding box is preferably rendered such that its sides are always parallel to the x and y axes of the image.

[0084] Alternatively or additionally, the support base comprises a set of slats arranged consecutively (sequentially, not necessarily parallel to each other) or parallel.

[0085] In the case of welded part detection, the prediction vector includes the coordinates of the predicted bounding box of each detected welded part, as well as its width and height. The bounding boxes can be used in an object detection network or for object detection (task).

[0086] For welded part detection (use case 1), the condition detection network may include or be an object detection network.

[0087] Images received from the optical sensor are labeled by a labeling algorithm. During labeling, the images are first loaded into the software and saved to a local repository. A human then locates the fused puzzle pieces on the image and draws a bounding box (or "box" for short) around each puzzle piece. Alternatively, this can be performed algorithmically. This box must completely surround the puzzle piece, and its edges are always parallel to the x and y directions of the image. After all puzzle pieces have been found and labeled, the software creates an .xml, .txt, or .json file containing the bounding box information. This file can be downloaded and saved. Images may be cropped from their original resolution (6000 x 4000 pixels) to a resolution (1000 x 1000 pixels), which corresponds to the input resolution supported by the network used in this analysis. The resulting batch of small images can then be augmented using data augmentation. In this particular use case, data augmentations that change the position of the puzzle pieces in the image also need to be translated into image labels. These augmentation methods include flipping, rotation, and translation. The (labeled) image can be used as input for an object detection network. The output of an object detection network consists of a numerical value defining the class label of the detection (in this case "puzzle piece" or welded part) and a prediction vector containing associated spatial information of the predicted bounding box. This vector may consist of four elements corresponding to the center coordinates of the bounding box and its width and height. For a given image, multiple predictions may be made by the network, each corresponding to a unique detection.

[0088] In all of the use cases 1-6 described herein, images are labeled.

[0089] Alternatively or additionally, the support platform comprises a set of slats arranged in parallel or succession. For slug buildup detection (use case 2), the method comprises at least one of the following steps: In a step of pre-processing the received image, the pre-processing step includes at least: Each slat of the support base is cut out from the received image. Cutting the processed image in the direction of extension of the slat, in particular at the center between two peaks, or according to a predefinable division scheme or pixel pattern, for example every predefinable number of pixels (for example every 128 pixels), which pixel pattern may be independent of the position of the tip in the image. The step of labeling the extracted images into two classes: dirty and clean. Using a data augmentation algorithm for data augmentation of the labeled images. Using a convolutional neural network and an attention-based neural network (CoAtNet), we generate as output a one-dimensional vector with two elements representing the clean and dirty classes. Applying a softmax function to the output to produce a normalized result that corresponds to a probability distribution. Thresholding the normalized result by applying a configurable threshold.

[0090] For slag buildup detection (use case 2), the condition detection network may include a classification network that classifies conditions into two classes: a normal or clean class and a dirty class (with slag buildup detected).

[0091] A binary image can be a black and white image, or a dark and light gray image.

[0092] In use case 2, i.e., slag pile detection, parts of the received image are labeled, specifically the cut parts (e.g., "10cm pieces").

[0093] Alternatively or additionally, the extraction of each slat of the support base from the received image (in the image representation) is performed by: A masking operation is further applied by converting the received image to a grayscale image and applying a Fourier transform to the grayscale image to provide an intermediate image representation, where the masking operation is configured to distinguish between low and high frequency portions in the received image, and then the intermediate image representation is transformed back to the spatial domain of the received image to provide a processed image. The processed image is thresholded to provide a binary image. Each pixel in a row of a binary image is analyzed and the discrete values ​​are interpolated to provide a continuous function, where the analysis may be performed by summing over all pixels in the row. By applying a peak detection algorithm to the provided continuous function, each peak in the continuous function corresponds to a slat of the support member in the received image.

[0094] Alternatively or additionally, the support base includes a set of slats arranged in parallel or succession, each slat having multiple tips arranged along the slat's length. For burnt tip detection (use case 3), the prediction vector includes the coordinates of the predicted bounding box of each detected burnt tip, as well as its width and height. For use case 3, the same procedure as for use case 1 may be performed.

[0095] In burnt tip detection (use case 3), an object detection network is used. Thus, the state detection network may include an object detection network.

[0096] Burned-out tips refer to the state where the tip of a slat is burned and lost. Burned-out tips of slats are different from slag buildup. Burned-out tips of slats can also refer to tips where the top of the tip is missing (burned away).

[0097] Alternatively or additionally, the support base comprises a set of slats arranged parallel or successively at a predetermined distance apart, and in order to detect the particles between the slats of the support base, the prediction vector comprises the coordinates of the predicted bounding box of each detected particle between the slats, as well as its width and height. This may include inputting the received image into a state detection network trained to detect particles between the slats.

[0098] An object detection network is used to detect parts or pieces between the slats (use case 4), so the state detection network may include an object detection network.

[0099] Alternatively or additionally, the support base includes a set of slats arranged parallel or successively at predetermined intervals. To detect bent and / or missing slats of the support base, the method may include: Applying a Fourier transform to the received image. By segmenting each slat in the image and inputting the segmented slats into a state detection network, the state detection network is trained to detect bent and / or missing slats on the support base.

[0100] A classification task is applied to detect bent slats (use case 5) and / or missing slats (use case 6). A classification network may be used to detect bent and / or missing slats.

[0101] To detect bent and / or missing slats, the same procedure as in Use Case 2 (detection of slag buildup) may be applied. In particular, the CoAtNet architecture may be used.

[0102] In another aspect, the present invention relates to a method for detecting the condition of a support table of a sheet metal laser processing machine, in particular a defect in the support table, by using a trained neural network, in particular at least one trained convolutional neural network, as a condition detection network, the neural network having been trained using the training method as described above, taking into account the aforementioned alternative embodiment. The method may comprise the following method steps: Receiving an image from an optical sensor located above the support platform whereby the optical sensor and the support platform are movable relative to one another. Applying the trained state detection network based on the received image. Providing a prediction result by applying a state detection network using the class label detection and / or prediction vector.

[0103] The detection method, like the training method, is preferably executed on a separate computing unit.

[0104] The optical sensors used in the detection method may be installed in an automation system, in particular in an input / output system and / or a sorting system.

[0105] Alternatively or additionally, the received image may be subjected to a crop operation to adjust the received image in terms of size and / or resolution to serve as an acceptable input in the trained state detection network.

[0106] Up to now, the present invention has been described with reference to the claimed methods (training method and inference method). Dependent claims, features, advantages, or alternative embodiments of this specification may be assigned to other claimed objects (e.g., computer programs, systems, or computer program products), and vice versa. In other words, a system / apparatus may be improved using features described or claimed in the context of a method, and vice versa. In this case, functional features of the method are embodied by structural units of the apparatus, device, or system, and vice versa. Generally, in computer science, a software implementation and a corresponding hardware implementation (e.g., an embedded system) are equivalent. Thus, for example, a method step for "saving" data may be performed using a storage device and respective instructions for writing the data to the storage device. To avoid redundancy, a device may also be used in alternative embodiments described with reference to a method, but these embodiments will not explicitly describe the device again.

[0107] In another aspect, the present invention relates to a system for detecting the state of a support table of a sheet metal laser processing machine, in particular, defects in the support table, using a trained state detection network, and includes the following components: Sheet metal laser processing machine with support table, an optical sensor, in particular at least one camera, mounted above a support base, such that the optical sensor and the support base are movable relative to each other; a central processing unit comprising the following components: an interface to a trained state detection network trained in the manner described above (considering the alternative embodiments discussed above); An output interface for providing prediction results.

[0108] In a preferred embodiment, the system interacts with or includes an automated system, and the optical sensor may be provided in or attached to the automated system.

[0109] The optical sensor may be a camera as described above and / or may be installed in an automation system. The automation system has the function of processing the cut parts. The automation system may be or may include a sorting system and / or an input / output system. In particular, the optical sensor of the automation system may be used as an optical sensor for the condition detection network.

[0110] Preferably, the system, particularly, but not necessarily, the automated system, may include a rack, which may serve to support a robotic arm, and the optical sensor may be mounted on the robotic arm.

[0111] It may be preferred that the optical sensor is movable relative to the support base. Alternatively or additionally, the support base is movable relative to the optical sensor. The optical sensor may then be fixed or located in a fixed position.

[0112] The optical sensor may be installed inside the laser processing machine. In this case, the table must be empty (without workpieces on it) to pass underneath. Alternatively or additionally, the optical sensor may be provided on or next to the cutting head. In particular, a camera already installed on or next to the cutting head may be used. The optical sensor is configured to scan the entire support table. Alternatively or additionally, the optical sensor may be provided on the frame of an infeed / outfeed system that moves across the support table. Alternatively or additionally, the optical sensor may be provided on a sorting system. The sorting system may comprise a robot arm that can move across the support table.

[0113] The trained condition detection network may be part of the system, in which case it is provided in a local storage of the system, which may be accessible via a communications link (wired or wireless). Alternatively or additionally, the condition detection network may not be provided locally, but may be accessible via an external communications link (e.g., a WLAN or internet communications link).

[0114] Alternatively or additionally, the system may include an automation system, which is preferably configured to be controlled in response to control commands based on or calculated from the provided prediction results, in particular by issuing a continue signal to the automation system if no defect is predicted in the support base and an interrupt signal to the automation system if a defect is predicted.

[0115] A continue signal triggers the system to continue processing (e.g., may trigger an additional workpiece to be loaded onto the support table for cutting). An abort signal triggers the system to abort processing. Alternatively or additionally, post-processing algorithms may be executed.

[0116] The post-processing algorithm is configured to calculate control commands based on or in response to the predicted results, which may be useful, for example, to control an automatic cleaning system, in particular to clean the areas where defects have been detected.

[0117] Alternatively or additionally, the optical sensor is provided in or attached to the automation system.

[0118] Alternatively or additionally, the system is configured to automatically initiate a cleaning operation of the support base portion if a defect is detected, in particular if slag deposits and / or welded components are detected.

[0119] In another aspect, the present invention relates to a computer program loadable into a memory unit of a computing unit, the computer program comprising program code sections for causing the computing unit to perform the above-mentioned method for training a state detection network when the computer program is executed on the computing unit.

[0120] In yet another aspect, the present invention relates to a computer program loadable into a memory unit of a computing unit, the computer program comprising program code sections for causing the computing unit to perform the detection method (inference phase) as described above when the computer program is executed on the computing unit.

[0121] In another aspect, the invention relates to a computer program product comprising a computer program as above.

[0122] The computer program product may also be provided for downloading, for example, via a wireless or cellular network, the Internet, and / or a host computer. Alternatively or additionally, the method may be encoded in a field programmable gate array (FPGA) and / or an application specific integrated circuit (ASIC), or the functionality may be provided for downloading in a hardware description language.

[0123] In another aspect, the invention relates to a computer-readable medium on which program code sections of a computer program are stored or saved, the program code sections being loadable into and / or executable by a computing unit, such that when the program code sections are executed by the computing unit, the computing unit performs the training method and / or the detection method described above.

[0124] In the context of the present invention, a "computing unit" or a "processor" may be understood to mean, for example, a mechanical or electronic circuit. In particular, a processor may be a central processing unit (CPU), a microprocessor, or a microcontroller, for example, an application-specific integrated circuit or a digital signal processor, possibly combined with a memory unit for storing program instructions or the like. A processor may also be, for example, an IC (integrated circuit), in particular an FPGA (field-programmable gate array) or an ASIC (application-specific integrated circuit), or a multi-chip module, for example, a 2.5D or 3D multi-chip module, in particular in which several so-called dies are connected to each other directly or via an interposer, or a DSP (digital signal processor) or a GPU (graphics processing unit). A processor may also be a virtualized processor, a virtual machine, or a soft CPU. It may also be, for example, a programmable processor with configuration steps for executing the method according to the invention, or configured with configuration steps such that the programmable processor implements features of the method, component, module, or other aspects and / or sub-aspects according to the invention.

[0125] The above-mentioned characteristics, features, and advantages of the present invention, as well as the manner in which they are realized, will become clearer and easier to understand with reference to the following description and embodiments, which are described in more detail in the context of the drawings. The following description does not limit the present invention to the embodiments contained therein. The same components or parts may be labeled with the same reference numerals in different figures. Generally, the figures are not drawn to scale.

[0126] It is understood that a preferred embodiment of the invention can also be any combination of the dependent claims or the above embodiments with the respective independent claim.

[0127] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiment(s) described hereinafter. [Brief explanation of the drawings]

[0128] [Figure 1] 1 illustrates a system setup according to one embodiment of the present invention. [Figure 2] 1 shows an exemplary schematic of welded parts on a slat. [Figure 3] 1 shows a schematic diagram of a portion of a support platform with an illustration of a slag buildup and bent slats. [Figure 4] 1 shows a schematic diagram of an example of a burnt tip of a slat. [Figure 5] 1 shows a schematic diagram of an example of an inter-slat part. [Figure 6] 1 shows a schematic diagram of an example of a support table that interacts with loading / unloading of sheet metal. [Figure 7] 1 illustrates an overview of the architectural workflow between hardware and software components according to one embodiment of the present invention. [Figure 8] This is the image labeling workflow. [Figure 9] We show the workflow of the training method for training the state detection network and the workflow of the detection method in the inference phase. [Figure 10]1 is a workflow for detecting the location of slats in an image according to an embodiment of the present invention. [Figure 11] 1 illustrates a portion of a system for detecting the state of a support table in an embodiment including a system for loading and unloading a laser processing machine. [Figure 12a] 1 is a schematic diagram of a loading and unloading system and sorting of a laser processing machine in a system according to an embodiment of the present invention, in which an optical sensor is used as a tool and applied. [Figure 12b] FIG. 1 is a schematic diagram of a loading and unloading system and sorting for a laser processing machine in a system according to an embodiment of the present invention, in which an optical sensor is fixed to a robot arm. [Figure 13] FIG. 1 is a schematic diagram of a system according to some embodiments. [Figure 14] 1 is a workflow of a training method for detecting the state of a support table. [Figure 15] 1 is a workflow of a detection method for detecting the state of a support base. DETAILED DESCRIPTION OF THE INVENTION

[0129] Any reference signs in the claims should not be construed as limiting the scope.

[0130] 1 is a structural diagram showing part of a system for detecting the state of a support table 2 of a laser processing machine 18 (hereinafter referred to as a laser processing machine) for sheet-like workpieces, in particular a laser cutting machine. The support table 2, also referred to herein as a cutting table, is a shuttle table including two parts, one part of which can be moved within the cutting machine or cutting cell, and the other part of which is outside the cutting cell, through which sheet-like workpieces, in particular sheet metal, are loaded and unloaded.

[0131] The support base 2 includes a plurality of slats 8, which may be arranged in parallel (as in the example shown in FIG. 1) or may be curved or patterned and arranged in series. The material of the slats 8 may be selected from the group consisting of carbon steel, copper, brass, stainless steel, mild steel, aluminum, and metal alloys.

[0132] The support base is subject to wear and damage from the cutting process and may therefore exhibit defects which may affect future cut quality.

[0133] The main defects on the support base 2 are selected from the following group: 1. Detecting welded puzzle parts (use case 1), 2. Detecting slag buildup (use case 2), 3. Detection of burnt slat tips (use case 3), 4. Detection of small pieces or parts between slats (Use Case 4), 5. Detecting bent slats (use case 5), and, 6. Missing slat detection (Use Case 6).

[0134] These defects can be detected by a method or system including an optical sensor 1, in particular a camera, mounted or placed on the support base 2. The optical sensor 1 can be movable relative to the support base 2.

[0135] The system further includes an interface to a trained condition detection network that is trained to automatically detect the condition of the support table, and an output interface for providing a prediction result, which may include a class label detection (e.g., for use case 1, welded part detected / welded part not detected, for use case 2, slag buildup detected / slag buildup not detected, etc.) and / or a prediction vector.

[0136] In general, the state detection networks described herein are provided in an untrained state (a state not yet trained by training method T) and a trained state that has been trained by training method T and can be applied to inference by detection method D. In the figures and descriptions, these two states (trained and untrained) are not distinguished by reference numerals, and the reference numeral sdnw is used for both types.

[0137] Figure 1 shows a preferred setup. A typical cutting machine typically has two cutting tables or two sections, with one or more sections inside the cutting cell and the other or other section outside, alternating between the inside and outside of the cutting table or sections of the cutting table. One cutting table, located outside the cutting cell, is intended to load and / or unload metal sheets as workpieces to be cut. Ideally, a camera is placed outside the cutting cell to monitor the cutting table before new metal sheets are loaded. The resolution and camera settings (including the distance from the camera to the table) should be selected so that a resolution of 1-2 mm is achievable. In the setup for this method, a Sony Alpha 6300 camera with an effective resolution of 24 MP (6000 x 4000) was installed 76 cm from the table, and a 16 mm focal length lens was used. Therefore, three photographs are typically required to monitor the entire 1.5 x 3 m cutting table. Other resolutions for cutting tables of other sizes are, of course, within the scope of this invention.

[0138] In principle, a normal 2D camera is suitable for all the envisaged tasks mentioned above. However, in other embodiments, more complex 3D cameras (stereo vision cameras, light field cameras, etc.) can also be used. It is also possible to use a video stream comprising an image sequence.

[0139] The laboratory camera support shown in Figure 1 is a first example of a possible embodiment. In the embodiment of Figure 1, the camera is movable above a fixed support base 2. In the embodiment of Figure 1, a camera, such as an example of an optical sensor 1, is movable relative to the support base 2. In an exemplary embodiment, the camera, or other optical sensor 1, is mounted on a rack above the support base 2. The optical sensor 1 may be movable using a slidable carrier, and the support base 2 may be fixed. Alternatively, the optical sensor 1 may be fixed in position above the support base 2, and the support base 2 may move relative to the optical sensor.

[0140] The slats 8 of the support base 2 may extend on an axis parallel to the longitudinal extension of the support base 2. Alternatively, the slats 8 may have a longitudinal axis rotated by 90° relative to the longitudinal extension of the support base 2.

[0141] The longitudinal extension of the slats 8 may be parallel or perpendicular to a bar on which the light sensor 1 may be mounted. The bar may be part of a frame-like rack, for example. The light sensor may be mounted on a slidable carrier that moves along the bar of the rack.

[0142] Particularly suitable are loading / unloading robots, where the camera is mounted and can be moved around on a platform, and then simply picked up as a tool when needed. Preferably, the camera is already integrated into the system, e.g., the loading / unloading system.

[0143] The above defects 1 to 3 mainly hinder cutting performance. For optimal cutting performance, the sheet support needs to be perfectly flat, but the presence of welded puzzle parts, large amounts of slag buildup, or burnt tips will hinder cutting performance.

[0144] Defects 4 to 6 above can interfere with automated loading / unloading systems to some extent. These systems are usually CNC-controlled and use fork-like elements between the slats to load and unload sheets. If a small piece gets caught between the grid slats or if the grid slats are bent, the fork will not be able to fit in. If a slat is missing, the broken piece will be prone to tilting, which can affect the unloading fork.

[0145] Below, various defects 1-6 are described in detail.

[0146] 1. Defects When cutting, especially when cutting small pieces and / or holes on slats, especially grid slats, small cut pieces, also called puzzle parts, may be welded to the slats.

[0147] FIG. 2 is a side view of a slat 8 with such puzzle pieces 3, showing a schematic cross section of the cutting grid. After cutting is complete, the cut pieces are removed and sorted. The welded puzzle pieces then remain on the table. Unfortunately, these puzzle pieces can interfere with the flat support of the subsequent metal plate, resulting in poor cutting.

[0148] 2. Defects During cutting, especially flame cutting, molten metal and metal slag can adhere to the slats below the cut surface, mainly on the (grid) slats. After cutting for a long time, such molten metal and slag can accumulate and form a significant obstructive slag pile, preventing flat support, which can also lead to poor cutting. The slag pile can become so large that it significantly reduces the gap between the slats, preventing the forks of the loading / unloading system from loading / unloading the sheet.

[0149] FIG. 3 shows diagrammatically a part of a cutting table with (grid) slats 8 and a defined distance 4 between the slats 8 and the interfering adhesive metal slugs 5 .

[0150] 3. Defects While cutting on the tip of the slat, it is possible for the tip to burn. A burnt or burned tip 7 is illustrated in Figure 4. The burnt tip 7 can prevent flat support and lead to poor cutting.

[0151] 4. Defects When a sheet metal part is cut by the cutting machine 18, a piece or part of the cut workpiece may become lodged between the slats 8. This often occurs because the piece tilts during cutting. It is also possible that a cutout may be cut from the workpiece that falls between the slats 8. However, in many cases the piece / object does not fall completely through the cutting table 2, but becomes lodged or trapped between the slats 8. Such a lodged or trapped piece or object 9 is shown diagrammatically in FIG. 5.

[0152] Any objects 9 left between the slats 8 can be a problem when the cut pieces and / or residual sheet metal are subsequently removed, since during the removal process, fork-like devices 10 of the automatic loading and unloading system are typically inserted between the grid slats to remove the cut and / or residual sheet from the cutting table 2.

[0153] The forked device 10 is shown in Fig. 6 during an unloading operation (using ByTrans as an example). If an object gets stuck between the slats 8, a collision with the forked device 10 will occur during unloading. This collision will usually lead to a system shutdown or even failure.

[0154] 5. Defects Due to exposure to high heat generated through the cutting process and / or for further mechanical reasons, the slats 8 may bend.

[0155] 3 shows, by way of example, a bent slat 6. Such a bent slat 6 may again prevent the fork-like device 10 of an automated loading and unloading system (not shown in FIG. 6) from properly loading or unloading the metal sheet. Automated loading and unloading systems are well known in the art and include, among other things, controllers and actors for controlling the movement of the fork-like device 10.

[0156] 6. Defects When cutting parts with the laser cutting machine 18, the cut parts may become welded to the cutting table 2. Automatic loading / unloading (removal) systems are usually able to remove parts even when they are welded to the support table or the cutting table 2. The large force required to remove the part from the cutting table 2 can lead to irreversible damage. This damage can accumulate over time and lead to partial or complete breakage of the slats 8. On the other hand, it is possible that the welded parts are not removed by the automatic removal unit, and instead the entire (grid) slat (or part of it) is removed along with the welded parts. Missing grid slats can significantly hinder the cutting performance of the machine and result in unsatisfactory cutting results.

[0157] These six major defects listed above are automatically detected and identified by the system and method according to the present invention, and can be identified in a single run.

[0158] Due to the probabilistic nature of these defects and the fact that they are always slightly different, their identification is ideally achieved using artificial intelligence, and more specifically, deep learning.

[0159] Deep neural networks act as multidimensional global approximators that can learn any mapping from one high-dimensional distribution to another.

[0160] This problem setup requires the use of multiple networks responsible for different computer vision tasks: object detection, object classification, and image classification. The input to each network must be either a full-resolution image or a cropped region of the full image.

[0161] The process of working with deep learning consists of two main parts: training 11 and inference 12, and is illustrated in Figure 9. Training of neural networks deserves special attention and will be described in detail in the next section.

[0162] For training 11 and inference 12, an image acquisition process is required and is shown in Figure 7. The imaging hardware may include at least one optical sensor, such as a camera. Preferably, a camera from an automated system 16, including a sorting system and / or an input / output system, is used. After the images are acquired, pre-processing software is applied, which may include labeling, cutting, and / or image pre-processing. The pre-processed images are then used for the training method.

[0163] A high-level description of the steps in these use cases includes image capture, image labeling, image pre-processing, and network training. First, an image is captured by a vision-based sensor 1 (in this case a 2D camera) after the metal plate is removed from the cutting table 2. This camera is placed above and facing the cutting table 2. The camera needs to be placed so that its focal plane is parallel to the cutting table.

[0164] In the next step, the captured images are labeled with the corresponding labels. Figure 8 provides a more detailed visual representation of this process. The labeling process differs depending on whether the images are processed for an object detection network or a classification network. In the case of object detection (e.g., welded parts / use case 1), a bounding box is inserted around the detected objects to enclose each object. In the case of image classification (e.g., slag deposits / use case 2), each image is labeled with a binary class label (e.g., 1 for detection, 0 for non-detection).

[0165] Furthermore, the image is cropped to a resolution supported by the current network, and individual cuts of the image are augmented using data augmentation. Data augmentation may be performed using a data augmentation algorithm and / or may consist of (small) modifications such as brightness, saturation, rotation, translation, zoom, color jitter, inversion, and / or contrast applied to the image. Data augmentation is selected by sampling different combinations that follow a uniform distribution. It is worth noting here that the data used in any application will involve the use of different cutting tables 2 in different locations in different environments. Integrating this fact with data augmentation results in a network that is robust to environmental changes and can therefore be used with any cutting table in any location. In the final step, these images are used as input for training the network. A visual representation of this process can be seen in Figure 7.

[0166] During network training, the network passes images through its layers and learns to extract semantic information from the images. This is repeated at each layer of the network, eventually condensing the information so that it can be used as a prediction. A visual representation of this process can be seen in Figure 9.

[0167] This process can be further explained by the fact that a network can be simplified and described as a composition of numbers, or network weights. These numbers are randomly sampled according to a standard distribution. During training, the network incrementally changes the values ​​of certain numbers, changing the network's behavior. Given enough data and training time, the model can find the optimal numbers that lead to a perfect analysis of the input. In this case, the input is an image, which is analyzed with either a convolutional or attention-based network. During training, the network uses only labeled data.

[0168] Labeling may include automatic or manual labeling. Labeling may be represented by assigning a label to an image. The label depends on the use case. For example, for use case 1, the label may be "Welded parts detected (yes / no) - with bounding box", or for use case 2, the label may be "Slag buildup detected (yes / no)".

[0169] After training is complete, the network's performance is measured by its accuracy on unseen data. Once a satisfactory level of accuracy is reached, the network weights are fixed and the network is implemented in this fixed state.

[0170] In any application, the weights of the network are no longer changed, and this network is used as a pre-trained tool for the analysis of any cutting table.

[0171] In a preferred embodiment, the customer does not have access to the weights of the state detection network (also called the model), but interacts with the model through an interface. The necessary computing power can be provided locally by a computing unit or via a cloud computing service such as Microsoft Azure or Google Cloud.

[0172] During inference 12, an image of the cutting table 2 is obtained and cropped to an appropriate resolution. When inferring an arbitrary input (i.e., an image of the support table 2), no labeling of the data is required. The output of the state detection network is a prediction result in the form of a class label detection and / or prediction vector. The image crop is then input into the trained network, which makes a prediction based on the input. Since the network weights are fixed, the network does not use this data to update its weights. After the network makes a prediction for a given input, the next input can be fed into the network. From this point on, the network can be used as a mapping tool, mapping inputs to predictions.

[0173] For a more detailed description of the training procedure, the first two use cases are described in more detail in the next sections.

[0174] In use case 1 (detecting welded puzzle pieces), the entire cutting table is imaged by an imaging device placed above and facing the cutting table. These images are then labeled using open-source labeling software. During labeling, the images are first loaded into the software and saved to a local repository. A human then locates the welded puzzle pieces on the image and draws a box around each puzzle piece. This box must completely surround the puzzle piece, and its edges are always parallel to the x and y directions of the image. After all puzzle pieces are found and labeled, the software creates an .xml, .txt, or .json file containing the bounding box information. This file can be downloaded and saved. In the next step, the image is cropped from its original resolution (6000 x 4000 pixels) to a resolution (1000 x 1000 pixels), which corresponds to the input resolution supported by the network used in this analysis. The resulting batch of small images is then augmented using data augmentation. In this particular use case, data augmentations that change the position of the puzzle pieces within the image also need to be converted into image labels. These augmentation methods include flipping, rotation, and translation. In the final step, the image can be used as input for an object detection network. The output of the object detection network consists of a prediction vector containing a numerical value defining the class label of the detection (in this case, "Puzzle Piece") and associated spatial information of the predicted bounding box. This vector consists of four elements corresponding to the center coordinates of the bounding box, as well as the width and height of the bounding box. For a given image, multiple predictions can be made by the network, each corresponding to a unique detection.

[0175] In the analysis of Use Case 2 (slag accumulation), the entire cutting table is imaged by the imaging device, which is installed above the cutting table and faces the cutting table.

[0176] The analysis is performed by cutting out each slat (also called grid slat) from the image. After cutting out each slat, each of these images is again cut along its main direction to generate smaller images showing different parts of the grid slat. To identify the grid slats and cut them out of the image, the slats need to be located within the given grid base image. This is done via a Fourier transform (see Figure 10). First, the image is converted from an RGB image to a grayscale image. In a next step, a Fourier transform is applied to the grayscale image to generate a Fourier representation of the image. Masking is applied to the Fourier representation to suppress the low and high frequency parts of the image. This is indicated by reference number 13 in Figure 10.

[0177] The image is then transformed back into the spatial domain, indicated by the box "Apply Inverse Fourier Transform" in Figure 10. Here, the image contains only the mid-frequency components relevant to identifying the grid slats. The image may then be thresholded, where a threshold is defined and pixels with intensity values ​​below the threshold are mapped to 0 and pixels with intensity values ​​above the threshold are mapped to 1. A detailed representation of the thresholding process is indicated by reference numeral 14 in Figure 10.

[0178] After the image has been thresholded, it is presented in binary format (black and white image). Pixel function generation can be applied, which is shown in FIG. 10 by reference numeral 15. Here, for each horizontal pixel row, the number of white pixels in that row is determined and plotted against the row index, which corresponds to the y-coordinate (vertical) of the row in pixel space. These discrete values ​​are then interpolated to form a continuous function.

[0179] A peak detection algorithm is then applied to this continuous function, as shown in Figure 10. Each peak in the function corresponds to a grid slat in the image. Using this analysis, the grid slats can be found in the image of any support platform. Following this analysis, the locations of these grid slats are used to extract each grid slat from the image.

[0180] Additionally, the image is cropped along the major direction to create a small portion of each grid slat.

[0181] In the next step, the images are labeled, for example, using the open-source labeling software LabelStudio. In this particular use case, the images were labeled in a binary format with only the options "Clean" and "Dirty."

[0182] In the next step, the images were augmented using a data augmentation algorithm and used as input for the classification network. The network accepts (128x128) pixel images as input, so all images were resized to fit this resolution. The network used in this analysis is a convolutional and attention-based network called CoAtNet. The output of this network is a two-dimensional vector. The numerical values ​​of this prediction can range from minus infinity to infinity and need to be mapped to values ​​between 0 and 1 to correspond to a probability distribution. This mapping is performed by applying SoftMax to the values ​​of the vector. After this mapping, each numerical value corresponds to a probability prediction that a given image belongs to a particular class.

[0183] In the final step, a threshold between 0.5 and 1 is defined, and predictions above the threshold are adopted as the class predicted by the model. The threshold does not interact with the underlying network and can be freely changed by customers to suit their desired prediction quality needs.

[0184] For the remaining use cases 3-6, a brief description of the detection process is given in the next section.

[0185] In one embodiment, when detecting burnt or heavily used tips (use case 3), a Fourier transform may be used to identify the location of each slug tip. In a next step, the slug tips need to be segmented and used as input for a classification network, which classifies the slug tips into condition classes (e.g., defective and non-defective classes). Alternatively, if an object detection network is used, the labeling procedure may be performed in a manner similar to or similar to use case 1.

[0186] Small pieces between the slats (use case 4) cause problems during production as they can prevent the automatic plate removal machine from removing the plate or sheet metal (see Figure 6). Therefore, it is recommended to install a camera (or other optical or visual sensor) above cutting table 2, facing it. The images taken by this camera system are used as input for an object detection network, which is trained to detect objects between the slats.

[0187] Use cases 5 and 6 are analyzed together using the same procedure. Here, images are again taken from a camera placed on and facing the support table. After the plate is removed, the camera takes an image of cutting table 2. Fourier analysis is applied to these images to extract each slat from the image, as illustrated in Figure 10. These slats are then input into a classification network, which is trained to find bent or missing areas of the slats.

[0188] All six different defect types mentioned above (Use Cases 1-6) can be identified by one detection method from the same image using one and the same detection device.

[0189] Preferably, an existing system module, such as an automated system like a sorting system, is selected for the installation of the optical sensor 1. This is cost-effective. By moving the optical sensor 1 on the platform 2, high resolution of the imaged platform can be achieved even with a moderate camera performance.

[0190] The above described methods and systems may be used to detect imperfections in the cutting table that would prevent a good cut and / or to prevent collisions with the loading / unloading system.

[0191] Alternatively or additionally, if a fault condition is detected, a post-processing algorithm may be executed, which may be adapted to provide control instructions.

[0192] The control instructions may include control instructions for subsequent laser cutting. In particular, in use cases 1 to 3, the subsequent cutting process may be controlled to avoid cutting above the detected defect. Alternatively, a cleaning process and / or a removal process may be initiated based on the control instructions. The control instructions may include position information (defect location). The control instructions may in particular cause an automated system, such as a sorting system or similar system, to clean dirt (a welded puzzle / use case 1 or slug buildup / use case 2) or autonomously remove detected debris (a stuck object) between the slats.

[0193] Alternatively or additionally, the control instructions may include control instructions for a loading / unloading system. For example, a loading / unloading system with fork-like elements may be used, which engage the support table and remove the cut parts and / or workpiece remnants from the table. The control instructions cause the fork-like devices 10 (FIG. 6) of the loading / unloading system to disengage at least at the detection location where a small piece between the slats (use case 4) and / or a bent slat (use case 5) is detected. Applying these control instructions ensures that no collisions occur.

[0194] Alternatively or additionally, a warning message may be generated to indicate that the detected defect needs to be troubleshooted and / or that the detected part (use case 4) or slat (use case 5) needs to be removed.

[0195] 11 is a schematic diagram of an exemplary embodiment of an input / output system 16 for use on or with a laser processing machine 18. The input / output system 16 is adapted to load and / or unload workpieces onto and / or from the support table 2. The direction of movement in FIG. 11 is indicated by reference numeral 17. An optical sensor 1 is attached to or located on or as part of the input / output system 16. It should be noted that the exact orientation and position of the optical sensor 1 is preferably selected so that its field of view corresponds to the entire support table 2, or at least the width of the support table 2.

[0196] FIG. 12a is a schematic diagram of an input / output system 16 with a sorting function for a laser processing machine 18, in which an optical sensor 1 is used and applied as a tool. The input / output system 16 may include a tool storage unit 19. The tool storage unit 19 provides various tools for use in processing the workpiece to be processed or on already processed workpieces (e.g., cut parts). The tools may include grippers, such as suction grippers or electromagnetic grippers, with different shapes, sizes, and / or distances between the suction cups, or other tools. One of these tools may be an optical sensor 1. The optical sensor 1 may be provided in the tool storage unit 19, and thus the optical sensor may be provided as a tool used in the system. The optical sensor 1 as a tool is indicated by the reference numeral 1 in FIG. 12a. The tool storage unit 19 is located somewhere within the range of the robot arm, preferably on the floor near the support table (as shown in the figure).

[0197] The system 20 according to claim 12 may include a rack r, which may have a frame-like structure. The rack r may be configured to surround the support table 1 and / or the loading / unloading system 16. The rack r may include a bar b. The bar b is configured to support a robot arm or a carrier c on which the robot arm is mounted. The carrier c may be slidable on the bar b in the longitudinal direction of the bar b. Various tools, including the optical sensor 1, may be selected to be mounted on the robot arm. The slidable carrier c allows the optical sensor 1, when mounted, to be moved along the longitudinal direction of the bar b.

[0198] Alternatively or additionally, the optical sensor 1 may be provided in a fixed position on the robot arm, as shown in Figure 12b. The optical sensor 1 may be fixed directly on the robot arm, for example at a position low enough on its side, so that the field of view of the optical sensor 1 is not obstructed by the carrier c and / or the robot arm (to avoid occlusion).

[0199] The robot arms may be mounted on a movable carrier c or bridge and are movable in the X and / or Y directions. The rack r serves as a gantry structure for the robot arms. Therefore, preferably, a gantry robot is used. What is important is that the robot arm(s) can reach every part of the support table 2. For productivity reasons, there may be more than one robot arm. Preferably, the robot can reach a certain height from the support table 2 to have a field of view across the entire width of the support table 2.

[0200] FIG. 13 is a schematic diagram of a system 20 for detecting the state of the support table 2. The system 20 may include a laser processing machine 18 with the support table 2 and an automation system 16, such as an automatic loading / unloading system and / or an automatic sorting system 161, which may be implemented as a sorting robot. In a preferred embodiment, the loading / unloading system and the sorting system may be integrated into one common system (e.g., a robotic system). The system 20 may include a rack r, which provides a gantry for a set of various tools. One of these tools may be an optical sensor 1. The rack r may include a horizontal bar b, which engages with a sliding or movable carrier c and supports a robot arm with a set of various tools.

[0201] The optical sensor 1, e.g. a camera, may be fixed or attached to the automation system 16. The optical sensor 1 exchanges data with a central processing unit CPU-16 of the automation system 16, a central processing unit CPU-18 of the laser processing machine 18, and / or a processing unit PU, which may be implemented on a central server.

[0202] The laser processing machine 18 includes a central processing unit CPU-18 for machine control. Optionally, the automatic automation system 16 may include a dedicated central processing unit CPU-16 for machine control. Images may be forwarded to an input interface 201 of the central processing unit CPU-18 of the laser processing machine 18. The central processing unit CPU-18 may include an interface 202 to a status detection network sdnw, which may be provided on the processing unit PU.

[0203] In the training phase, images from the optical sensor 1 can be sent directly to the processing unit PU for training of the state detection network sdnw. The training of the state detection network sdnw can be performed on a processing unit PU, which is preferably provided external to the laser processing system. The detection method (inference phase) can be performed locally on the central processing unit CPU-18 of the laser processing machine 18 and / or on the central processing unit CPU-16 of the automation system 16. In the latter case, the central processing unit CPU-16 of the automation system 16 includes an output interface (not shown in FIG. 13) for providing the prediction results.

[0204] In yet another embodiment, the sorting system 161 may include additional processing units and may be adapted to perform an inference phase and provide predicted results as well.

[0205] In the inference phase, the images from the optical sensor 1 may be forwarded to the processing unit PU to provide prediction results, which may be output via the output interface 200' of the processing unit PU, as in the training phase. Alternatively, the images may be forwarded to the central processing unit CPU-18 of the laser processing machine 18. The central processing unit CPU-18 of the laser processing machine 18 may use the received images or a preprocessed representation thereof to access the state detection network sdnw and provide the results.

[0206] In the first embodiment, the trained state detection network sdnw may be locally provided on the central processing unit CPU-18 of the laser processing machine 18. The prediction results may be provided via the output interface 200.

[0207] In the second embodiment, the trained state detection network sdnw may be provided to the processing device PU rather than being provided locally, and the processing device PU may be accessible via a network connection between the processing device PU and the central processing unit CPU-18 of the laser processing machine 18. In this case, the prediction results may be provided via the output interface 200′ of the processing device PU.

[0208] In the third embodiment, the trained state detection network sdnw may be provided on a processing unit PU accessible via a network connection, rather than being provided locally. However, compared to the second embodiment, in the third embodiment, the prediction results may be sent back to the central processing unit CPU-18 of the laser processing machine 18 and provided "locally" via the output interface 200. Also, the control instructions may be generated locally on the central processing unit CPU-18 of the laser processing machine 18, on the processing unit PU, or on another computing unit.

[0209] Control instructions may be calculated based on the predicted results. The control instructions may also include sections for controlling the loading / unloading system or other automated systems or parts thereof. The predicted results may be provided on a mobile device, which is networked to the central processing unit CPU-18.

[0210] The control instructions can be calculated by applying post-processing algorithms, which can be implemented in the processing unit PU, the central processing unit CPU-18 of the laser processing machine and / or the central processing unit CPU-16 of the automation system.

[0211] Alternatively or additionally, the control instructions may function to control the laser processing machine itself (e.g., not cutting at defective locations for later cutting), an automation system (e.g., not engaging at defective support table locations), and / or a cleaning system (e.g., cleaning locations where contamination has been detected, particularly in the form of slag residue).

[0212] In a preferred embodiment shown in Fig. 13, the processing unit PU may be provided on a cloud-based entity accessible via network access (based on a local area network LAN or a wireless access network, or an Internet Protocol connection). The processing unit PU may be located outside the central processing unit CPU-18 of the laser processing machine 18 and / or outside the central processing unit CPU-16 of the automation system 16. The processing unit PU is configured to store and / or execute the trained state detection network sdnw. The processing unit PU may exchange data with the central processing unit CPU-18 of the laser processing machine 18 and / or the central processing unit CPU-16 of the automation system 16 (indicated by a double arrow with diagonal lines in Fig. 13, if necessary).

[0213] Alternatively or additionally, the processing unit PU equipped with the state detection network may be provided locally on the central processing unit CPU-18 of the laser processing machine 18. In the latter case, the state detection network sdnw is provided locally.

[0214] The central processing unit CPU-18 of the laser processing machine 18 may include an output interface 200 for providing predicted results and / or control instructions. The control instructions are calculated based on the predicted results and serve to control the automation system 16. For this purpose, the control instructions may be forwarded to the central processing unit of the automation system CPU-16.

[0215] The system 20 preferably includes a laser processing machine 18 and an automation system 16. In another embodiment, the system 20 may include only the laser processing system 18 with its support base 2 and optical sensor 1, where the optical sensor 1 and support base 2 may be provided and arranged to be movable relative to each other (in a first option, the optical sensor 1 is movable, in a second option, the support base 2 is movable and the optical sensor may be stationary).

[0216] The interfaces described herein may be internal or external interfaces.

[0217] 14 shows an example workflow of a training method T for training a state detection network sdnw. After starting the method, in step T1, images are received from an optical sensor 1, in particular a camera. The camera may be installed in an automated automation system 16, which may be an in-feed / out system or a sorting system. In step T2, the images are labeled by applying a labeling algorithm. Alternatively or additionally, the labeling may be performed manually.

[0218] In step T3, the state detection network sdnw is used or applied to provide a result. The result (also referred to herein as "prediction result") may include a class label detection (normal state - faulty state) and / or a prediction vector. The provided result is compared with the respective label. This results in a loss (high if the difference between the prediction result and the label is large, and low if the difference between the prediction result and the label is small), which is gradually reduced by updating the weights of the state detection network sdnw. The weights are updated using backpropagation and gradient descent algorithms.

[0219] In step T4, the state detection network sdnw is provided in a trained (fixed weighted) state. The trained state detection network sdnw can then be tested and then used in the inference phase. After this, the method can end.

[0220] FIG. 15 shows an example workflow of the detection method D. After starting the method, in step D1, an image is received from an optical sensor 1, in particular a camera. The camera may be installed in an automated automation system 16, which may be an input / output system or a sorting system. Step D11 is optional and therefore indicated by diagonal lines in FIG. 15, but the received image may be cropped or otherwise preprocessed. In step D2, the state detection network sdnw is applied in a trained form based on the received image. In step D3, a prediction result is provided. The detection method may then be terminated.

[0221] The trained state detection network sdnw may be provided on the processing unit PU and / or the central processing unit CPU-18, CPU-16. As mentioned above, the processing unit PU may be located inside or outside the laser processing machine 18 and / or the automated machine 16.

[0222] Unless explicitly stated herein, individual embodiments, or individual aspects and features thereof, may be combined with and interchanged with one another without limiting or expanding the scope of the invention, as is typical for computer-implemented inventions, where such combinations or interchanges are meaningful and within the spirit of the invention.

[0223] Advantages described with respect to a particular embodiment of the invention or a particular figure are also advantages of other embodiments of the invention, where applicable. [Explanation of symbols]

[0224] 1 Optical sensors, especially cameras, As a separate tool, attached to a robotic arm 2 Support table, especially cutting table 3 Welded parts (puzzle parts) 4 Distance between grid slats 5. Slag accumulation 6 curved slats 7 Burnt tip 8 slats 9 Puzzle between grid slats (in-place objects) 10 Fork-shaped device for loading / unloading system 11 Training Workflow 12 Inference Workflow 13 Masking 14 Thresholding 15 Pixel Function Generation 16 Automatic automation systems, especially loading and unloading systems 17 Direction of movement of loading and unloading systems 18 Laser Cutting Machine 19 Tool storage for sorting system 20. Systems for detecting conditions 161 Automatic automation systems, especially sorting robots 200 Output Interface 201 Input Interface 202 Interface to SDNW sdnw State Detection Network CPU-18 Central processing unit of laser processing machine CPU-16 Central Processing Unit for Automation Systems PU: A processing unit with a memory for the trained state detection network. T Training Method (Training) T1 Receives images from the optical sensor Labeling T2 images Use the T3 status detection network T4 Provide a trained state detection network D. Detection Method (Inference) D1 Receives images from the light sensor D11 Cropping received images D2 Provide a trained state detection network based on the received image D3 Providing prediction results

Claims

1. A training method (T) for training a neural network as a condition detection network (sdnw), comprising at least one convolutional neural network for detecting a condition of a support table (2) of a laser processing machine (18) that is a sheet metal laser processing machine and a defect of the support table (2), wherein detecting the condition of the support table (2) comprises: Detecting welded parts; Detecting burnt tips; Detecting small particles between the slats (8) of the support base (2); Classifying slag deposits; Sorting the curved slats, and and classifying the missing slats. In the method steps: receiving (T1) an image from an optical sensor (1) installed above the support base (2), the optical sensor (1) being movable relative to the support base (2); labelling (T2) said received image or a part thereof with a label and analysing the state of said support base (2) represented in said received image; using the state detection network (sdnw) (T3) by inputting the labeled image or a part thereof and providing a class label detection and / or prediction vector as a result, the result representing the state of the support table (2), i.e. a defective state or a normal state; (T4) a step in which the weights of the state detection network (sdnw) are gradually adjusted during use of the state detection network (sdnw) to minimize the loss function and provide a trained state detection network (sdnw); Training Method (T).

2. The training method (T) of claim 1, wherein labeling comprises graphically drawing a bounding box in the image representing the defect or a defect region around the defect.

3. 3. A training method (T) according to claim 1 or 2, wherein the support table (2) comprises a set of slats (8) arranged in series or in parallel, and / or for welded part detection, the predicted vector comprises the coordinates of the predicted bounding box of each detected welded part, as well as its width and height.

4. The support base (2) comprises a set of slats arranged in parallel or succession, and / or for slag accumulation detection, the method comprises the following steps: Pre-processing the received image, the pre-processing comprising: Extracting each slat (8) of the support base (2) from the received image; - cutting out said processed image in the direction of the extension of the slats, in particular in the centre between two peaks, or according to a predefinable pixel pattern, labeling the preprocessed images into two classes: a dirty class and a clean class; using a data augmentation algorithm for data augmentation of the labeled images; Using a convolutional neural network and an attention-based neural network, CoAtNet, to generate as output a one-dimensional vector with two elements representing the clean and dirty classes; applying a softmax function to the output to produce a normalized result corresponding to a probability distribution; thresholding the normalized result by applying a configurable threshold. A training method (T) according to any one of claims 1 to 3.

5. Extracting each slat (8) of the support base (2) from the received image includes: converting the received image to a grayscale image and applying a Fourier transform to the grayscale image, further applying a masking operation to provide an intermediate image representation, wherein the masking operation is configured to distinguish between low and high frequency portions within the received image, and then converting the intermediate image representation to the spatial domain of the received image to provide a processed image; thresholding the processed image to provide a binary image; analyzing each pixel in a row of said binary image and interpolating discrete values ​​to provide a continuous function; applying a peak detection algorithm to the provided continuous function, whereby each peak in the continuous function corresponds to a slat (8) of a support member in the received image. A training method (T) according to claim 4.

6. the support base (2) comprises a set of slats (8) arranged in parallel or succession, each slat (8) having a plurality of tips, and / or for burnt tip detection, the prediction vector comprises the coordinates of a predicted bounding box of each detected burnt tip (7), as well as its width and height, and / or the method comprises: inputting the received image into the condition detection network (sdnw) trained to detect burnt tips (7), A training method (T) according to any one of claims 1 to 5.

7. The support base (2) comprises a set of slats (8) arranged parallel or successively at a predetermined distance (4), and / or for detecting small pieces between the slats (8) of the support base (2), the prediction vector comprises the coordinates of a predicted bounding box of each detected small piece between the slats (8), as well as its width and height, and / or the method comprises: inputting the received image into the state detection network (sdnw) trained to detect particles between slats (8), A training method (T) according to any one of claims 1 to 6.

8. The support base (2) comprises a set of slats (8) arranged parallel or successively at a predetermined distance (4) apart, and / or in order to detect bent slats (6) of the support base (2) and / or missing slats of the support base (2), the method comprises: applying a Fourier transform to the received image; Segmenting each slat (8) in the image and inputting the segmented slats into the state detection network (sdnw), the state detection network being trained to detect bent slats (6) and / or missing slats of the support base (2). A training method (T) according to any one of claims 1 to 7.

9. A detection method (D) for detecting a state of a support table (2), in particular a defect of said support table (2) of a sheet metal laser processing machine (18), by using a trained neural network, in particular at least one trained convolutional neural network, as a state detection network (sdnw), said trained neural network being trained by a training method according to any one of claims 1 to 8 of the training method, comprising the following method steps: receiving (D1) an image from an optical sensor (1) installed above the support base (2), the optical sensor (1) being movable relative to the support base (2); (D2) applying the trained state detection network (sdnw) based on the received image; (D3) providing a prediction result by applying the state detection network (sdnw) using the class label detection and / or prediction vector; Detection method (D).

10. 10. The detection method (D) of claim 9, wherein the received image is subjected to a cropping operation (D11) to adjust the received image in terms of size and / or resolution and to serve as an acceptable input in the trained state detection network.

11. A system (20) for detecting a state of a support table (2) of a sheet metal laser processing machine (18), in particular a defect of the support table (2), using a trained state detection network (sdnw), comprising: The sheet metal laser processing machine (18) equipped with the support table (2), an optical sensor (1), in particular a camera, located above the support base (2), which is movable relative to the support base (2); a central processing unit (CPU-18, CPU-16) comprising: an interface to the trained state detection network (sdnw), trained with the method according to any one of claims 1 to 8 of the training method; an output interface (200, 200', 200") for providing the prediction result; System (20).

12. The system (20) may include an automation system (16), which is configured to be controlled in response to control commands calculated from the provided prediction results, in particular by issuing a continue signal to the automation system (16) if no defect is predicted in the support platform and by issuing an interrupt signal to the automation system (16) if a defect is predicted.

13. 13. The system (20) according to claim 11 or 12, wherein the system (20) comprises an automation system (16), and the optical sensor (1) is provided in or attached to the automation system (16).

14. The system (20) according to any one of claims 11 to 13, wherein the system (20) is configured to automatically initiate a cleaning operation of the support base portion if a defect is detected, in particular if slag deposits and / or welded components are detected.

15. 9. A computer program loadable into a memory unit of a computing unit, said computer program comprising program code sections for causing said computing unit to carry out the method (T) for training a state detection network (sdnw) according to any one of claims 1 to 8, when said computer program is executed on said computing unit.

16. 11. A computer program loadable into a memory unit of a computing unit, said computer program comprising program code sections for causing said computing unit to carry out the detection method (D) of any one of claims 9 or 10, when said computer program is executed on said computing unit.

Images