Laser processing system for carrying out a processing operation on a workpiece by means of a laser beam and method for monitoring a processing operation on a workpiece by means of a laser beam

Abstract

A laser processing system for performing a processing operation on a workpiece by means of a laser beam, the laser processing system comprising: a laser processing head for causing the laser beam to be incident on a processing region on the workpiece; and a sensor unit for monitoring the processing operation by means of at least one hyperspectral sensor, wherein the sensor unit is configured to detect a hyperspectral image having N×M pixels from the region of the workpiece, wherein the hyperspectral image has two spatial dimensions x and y and one spectral dimension λ, wherein N specifies the number of pixels in the first spatial dimension x, M specifies the number of pixels in the second spatial dimension y, and L specifies the number of spectral bands in the spectral dimension λ of the hyperspectral image, wherein M, N and L are all natural numbers. In addition, a method for monitoring a processing operation on a workpiece by means of a laser beam is described.

Description

Technical Field

[0001] The present disclosure relates to a laser processing system for performing a processing operation on a workpiece, in particular a metal workpiece, by means of a laser beam and a method for monitoring a processing operation on a workpiece, in particular a metal workpiece, by means of a laser beam. The present disclosure relates in particular to a laser processing system having a sensor unit, which is configured to detect a hyperspectral image of an area of ​​the workpiece, and a method, in which a hyperspectral image of the area of ​​the workpiece is detected. Background Art

[0002] In a laser processing system for processing a workpiece by means of a laser beam, a laser beam emitted by a laser source or the end of a laser conductor fiber is focused or converged onto the workpiece to be processed by means of a beam guide and focusing optical device. The processing can, for example, include laser cutting, laser brazing or laser melting welding. The laser processing system can include a laser processing device, such as a laser processing head (such as a laser cutting head or a laser melting welding head). In particular, it is important to monitor the welding or brazing process in order to be able to judge and ensure the quality of the processing. This includes the identification and classification of processing errors. Current solutions for process monitoring and quality judgment include so-called "pre-process", "in-process" and "post-process" inspections or monitoring.

[0003] The task of pre-process monitoring is to detect the joint gap between two joining partners or workpieces in order to guide the laser beam to the appropriate position and determine the deflection of the joining partners. In most cases, a triangulation system is used for this purpose.

[0004] Post-process inspection is particularly used for process monitoring and quality judgment, for example, for identifying processing errors. The purpose of post-process inspection is to reliably locate and identify all defects or processing errors. Post-process inspection of weld seams or brazing seams is typically achieved by means of image processing, in which a 2D image showing the weld seams or brazing seams on the surface of the processed workpiece is analyzed. For example, in the case of laser welding, the produced weld seams can be inspected and measured or analyzed according to applicable standards (such as SEL100). Then, based on the detected image, features are extracted and classified, which represent or describe the processing quality, especially the quality of weld seams and brazing seams. Processing errors (such as holes or pores in the workpiece surface) are identified and classified based on the extracted and classified features, and the processed workpiece is labeled or classified as, for example, "good" (i.e. suitable for further processing or sale) or "bad" (i.e. as scrap). Which significant features can be used to assess the quality of the laser processing and what influence these features then have on the assessment of the quality can only be determined and implemented by experts in this field, because the complexity of these systems is very high due to the large number of parameters to be set. For this reason, up to 300 parameters are set in current systems (so-called "parameterization"). The use of post-process inspection for quality assessment and classification increases the costs of the laser processing system, the integration effort and the maintenance effort.

[0005] In-process inspection or monitoring is typically used for continuous monitoring of laser machining processes and is performed during the execution of the laser machining process. During monitoring, measurement signals of different measurement variables of the laser machining process are detected and evaluated, in particular beams emitted or reflected by the workpiece during the execution of the machining process, such as plasma beams, backscattered laser power and temperature beams. Typically, during laser machining, the molten pool emits or reflects beams in the visible range between 400nm and 850nm, the generated plasma emits or reflects beams in the range between 400nm and 1100nm, backscattered light is emitted or reflected by the laser in the range between 900nm and 1100nm, and temperature beams are emitted or reflected in the region greater than 1000nm. Thus, beams are emitted or reflected in a wide range between 400nm and 1800nm. Usually, in the case of in-process monitoring, the signals are detected or processed without spatial resolution and / or without wavelength resolution. Existing solutions for in-process monitoring use diodes, which detect the emitted or reflected beams in a narrowband wavelength range and output them as measurement signals. For example, Si diodes are used for detection in the range between 400nm and 800nm, InGaAs diodes are used for detection in the range between 800nm ​​and 1200nm, and other InGaAs diodes or Ger diodes are used for detection in the range between 1200nm and 2000nm. In addition, depending on the process, areas with corresponding optical filters can be sorted out from these wavelength ranges. For example, depending on the type of laser used, the wavelength range between 1020nm and 1090nm is filtered or attenuated. The wavelength range outside the detection range of the diode and the optical filter will not be detected. Intensity and spatial dependence of the presentation cannot be achieved using diodes either.

[0006] The intensity profiles recorded in this way by means of the diodes are filtered and checked for exceeding threshold values. The filter parameters and threshold values ​​are set individually for each wavelength range. Thus, the detection and evaluation of individual wavelength ranges is carried out independently of the other wavelength ranges. Inferences about specific error types (e.g. gaps between the components to be connected, lack of penetration, defective weld penetration, the formation of pores or a deviation of the weld relative to the joint edge) cannot be reliably drawn using a single assessment or classification of the intensity profiles in a specific wavelength range.

[0007] The regulation of laser processing steps based on individually detected and evaluated wavelength ranges is also usually not possible, because the connection between the individual measurement signals and the execution variables or processing parameters of the laser processing step is not yet sufficiently clear. In particular, the dependence of the execution variables on the detected signals is not yet clear. In other words, the connection between the execution variables of the laser processing step (such as the focus position, the distance and position of the laser processing head relative to the joint, the feed speed, the laser power, the gas supply, and, if necessary, the wire feed speed) and the measurement signal is not clearly known or recognizable. Complete information about the status of the laser processing step may not be contained in the measurement signal. Therefore, the laser welding process cannot be regulated.

[0008] Other solutions use spatially resolved sensors and image processing to measure the geometry of the weld pool and the keyhole in order to make a quality assessment from this. DE 10 2011 078 276 B3 describes a method in which the brightness profile of the workpiece surface is recorded using a CMOS sensor in the range between 450 nm and 800 nm. These profiles are compared with a model in order to draw conclusions about the machining quality from this.

[0009] Other systems use image sensors to provide spatially resolved images of the joint gap, the processing area itself and the weld bead. However, the CMOS sensors used integrate the detected radiation intensity over a large wavelength range, so that the interpretability of the processing sequence is limited and the processing sequence cannot therefore be adjusted. For this reason, these systems also require complex parameterization. Summary of the invention

[0010] It is therefore an object of the present invention to simplify the monitoring and control of laser processing sequences as well as the detection and classification of processing errors.

[0011] Furthermore, the object of the present invention is to enable the recognition and classification of processing errors and the monitoring and regulation of the laser processing process based on the detected emitted process beam and / or reflected laser beam.

[0012] Furthermore, it is an object of the present invention to enable spatially resolved and wavelength-resolved detection of the emitted process beam and / or the reflected laser beam of a laser processing process.

[0013] Furthermore, the object of the present invention is to detect and classify processing errors reliably, quickly and without complicated parameterization processes. Furthermore, the object of the present invention is to detect processing errors during an ongoing laser processing process, preferably in real time and automatically.

[0014] These objects are achieved by the subject matter of the independent claims. Advantageous embodiments and developments are the subject matter of the respective dependent claims.

[0015] The present invention is based on the concept that a laser processing process (especially a laser melting process or a laser cutting process) is monitored using a hyperspectral sensor (especially a hyperspectral camera). In other words, during the laser processing process, at least one hyperspectral sensor preferably continuously detects a hyperspectral image of a region of a workpiece. The detected hyperspectral image includes in each pixel a spectrum over a wide wavelength range of a beam emitted or reflected by the workpiece during the execution (which can be referred to as "emitted beam" for short). The emitted beam can especially include an emitted process beam (such as a temperature beam and a plasma beam) and the reflected light of the laser beam. The detected hyperspectral image completely describes or characterizes the laser processing process. The hyperspectral image can be used as input data for image processing by a deep neural network, which continuously, preferably in real time, calculates an output vector based on this, which contains information about the processing process, such as information about the state of the laser processing process or about the processing result or about processing errors on the workpiece. For example, the output vector can be used for quality judgment of the processing process, especially for judging weld seams and brazing seams. The output vector can also be used to adjust the laser processing process.

[0016] Therefore, the use of a hyperspectral sensor (such as a hyperspectral camera) simplifies the detection of images of the region of the workpiece in different wavelength bands or spectral bands. Thus, by spatially and spectrally or wavelength-resolved detection of the beam of the region of the workpiece, accurate data detection for monitoring and regulating the processing process is achieved. Thus, the use of multiple diodes, beam splitters and / or filters or laborious filter cascades can be avoided. Here, the hyperspectral image can be used as a so-called snapshot photo, wherein all data of the hyperspectral image are recorded simultaneously. Alternatively, the hyperspectral image can be detected in the so-called push-broom method, wherein, for example, all pixel rows of the hyperspectral sensor are recorded simultaneously in each other wavelength band and then the hyperspectral sensor and the workpiece region are moved relative to each other in the y direction or row direction of the sensor pixels, so that the pixel rows are detected in sequence in all wavelength bands. Here, the pixels of a pixel row are arranged along a first direction (also called x direction or column direction).

[0017] According to a first aspect of the present disclosure, a laser processing system for performing a processing operation on a workpiece by means of a laser beam is proposed, wherein the laser processing system comprises: a laser processing head for causing the laser beam to be incident on a processing area on the workpiece; and a sensor unit for monitoring the processing operation using at least one hyperspectral sensor, wherein the sensor unit is configured to detect a hyperspectral image having N×M pixels of an area of ​​the workpiece, wherein the pixels respectively include L different values ​​in a spectral dimension, wherein the hyperspectral image has two spatial dimensions x and y and a spectral dimension λ, wherein N indicates the number of pixels in a first spatial dimension x, M indicates the number of pixels in a second spatial dimension y, and L indicates the number of spectral bands in the spectral dimension λ of the hyperspectral image, wherein M, N and L are natural numbers.

[0018] According to a second aspect of the present disclosure, a method for monitoring a processing process on a workpiece by means of a laser beam is proposed, wherein the method comprises the following steps: causing a laser beam to be incident on a processing area on the workpiece; detecting a hyperspectral image having N×M pixels of the area of ​​the workpiece, wherein each pixel includes L different values ​​in a spectral dimension λ, wherein the hyperspectral image has two spatial dimensions x and y and a spectral dimension λ, wherein N indicates the number of pixels in a first spatial dimension x, M indicates the number of pixels in a second spatial dimension y, and L indicates the number of spectral bands in the spectral dimension λ of the hyperspectral image, wherein M, N and L are natural numbers.

[0019] The area of ​​the workpiece detected in the hyperspectral image can include at least one of the following areas on the workpiece: the current processing area of ​​the laser processing process, the area in the forward part (Vorlauf) of the laser beam, the area in the follow-up part (Nachlauf) of the laser beam, the area still to be processed and the processed area. The current processing area represents the area of ​​the workpiece into which the laser beam is irradiated. In particular, in the case of laser welding, the area of ​​the workpiece detected can include at least one of the following areas: a steam capillary, a molten pool, a solidified melt, a joint edge between two workpieces to be welded to each other, and a weld bead. The hyperspectral sensor can be aligned accordingly to the area of ​​the workpiece to be detected in the hyperspectral image. Depending on which areas are included in the detected workpiece area, the hyperspectral image can be used for pre-process, in-process and / or post-process monitoring.

[0020] Thus, the laser processing system comprises a sensor unit which is configured to detect and output a hyperspectral image of a region of a workpiece. The sensor unit can be configured to detect a beam emitted by the region of the workpiece, for example an emitted process beam and / or a reflected or backscattered beam (in particular a reflected or backscattered laser beam) and output it as a hyperspectral image. The sensor unit can be configured in particular to detect at least one of the following beam types: a temperature beam, a beam in the infrared range of light, a beam in the near infrared range of light, a beam in the visible range of light, a plasma beam, reflected or backscattered light of a (processing) laser beam, and light incident and reflected by an irradiation source (such as incident and reflected measurement light of a coherence tomography scanner).

[0021] The hyperspectral image comprises a plurality of pixels. The hyperspectral image comprises in particular N×M pixels, wherein the hyperspectral image has two spatial dimensions x and y and a spectral dimension λ, wherein N specifies the number of pixels in the first spatial dimension x and M specifies the number of pixels in the second spatial dimension y. Each of the pixels comprises L values ​​in the spectral dimension λ, wherein L specifies the number of spectral bands in the spectral dimension λ of the hyperspectral image. The first spatial dimension x and the second spatial dimension y can correspond to two Cartesian directions of the region of the workpiece. Therefore, each of the N×M pixels can correspond to a point in the detected region of the workpiece. Each of the L values ​​of the pixel corresponds to a detected beam intensity (in the corresponding spectral band of the beam emitted by the corresponding point of the detected region of the workpiece), referred to as “intensity” for short. In this way, the hyperspectral image can describe the spectral distribution of the beam emitted by the detected region of the workpiece in the wavelength range covered by the spectral band. Therefore, the detected hyperspectral image preferably corresponds to a spatially resolved and wavelength-resolved image of the workpiece region. In this way, the sensor unit can replace a plurality of individual sensors or an arrangement of individual sensors, in particular a plurality of diodes, according to the prior art.

[0022] A hyperspectral image can also be referred to as a hyperspectral cube or hyperspectral block. A hyperspectral image can also be considered as L individual frames, each of which has a resolution of N×M pixels and corresponds to the detected intensity of a beam emitted by an area of ​​the workpiece in one of the L spectral bands. This can be the case in particular for so-called mosaic sensors. Conversely, when using a push-broom method with a hyperspectral sensor, in which L different filters for L different wavelength bands are arranged on the pixel rows of the hyperspectral sensor, a hyperspectral image can be considered as M individual frames, each of which has a resolution of N×L, and each of the L rows of a single frame corresponds to the detected intensity of a beam emitted by a corresponding row of the area of ​​the workpiece in a corresponding spectral band of the L spectral bands. A hyperspectral image can also be referred to as a hyperspectral cube, wherein the hyperspectral cube has N×M×L or N×L×M elements or values. The data output of the sensor unit can be realized in the form of a 3D data cube with two spatial dimensions and one spectral dimension. A hyperspectral cube can also be called a hyperspectral data cube or an L-channel hyperspectral image similar to a 3-channel RGB image.

[0023] The number L of spectral bands can be equal to or greater than 16, preferably equal to or greater than 20, preferably equal to or greater than 25, particularly preferably equal to or greater than 100.

[0024] The spectral bands in the spectral dimension λ can be of equal size and / or be distributed uniformly over the wavelength range of the detected radiation and / or be adjacent to one another and / or follow one another. Preferably, the spectral bands do not overlap.

[0025] In order to enable continuous monitoring of the laser machining process, the sensor unit can be configured to continuously detect a hyperspectral image or to detect a hyperspectral image at predetermined time intervals.

[0026] The sensor unit can be designed to detect all L values ​​for all N×M pixels of the hyperspectral image substantially simultaneously. In this case, the hyperspectral image can be recorded as a snapshot or a snap shot, wherein the intensities in all spectral bands are detected simultaneously for all pixels. Therefore, the detected hyperspectral image preferably corresponds to a spatially resolved and wavelength-resolved image of the workpiece area at a specific moment. In other words, the intensity of the beam in all spectral bands and for the entire detected area of ​​the workpiece can be detected simultaneously. As a result, the laser processing process can be precisely monitored.

[0027] The hyperspectral sensor can include a mosaic filter having L different optical bandpass filters. Each optical bandpass filter can have a passband region corresponding to one of the spectral bands in the spectral dimension λ. A 4×4 or 5×5 optical mosaic filter can be arranged on the pixels of the hyperspectral sensor, wherein each individual bandpass filter of the mosaic filter is arranged on a corresponding pixel. Therefore, each pixel of the hyperspectral sensor detects the intensity of the emitted beam filtered by the corresponding bandpass filter. Thus, in the case of a hyperspectral sensor with a 4×4 mosaic filter, 16 bandpass filtered frames of a hyperspectral image are generated, and in the case of a hyperspectral sensor with a 5×5 mosaic sensor, 25 bandpass filtered frames of a hyperspectral image are generated.

[0028] Alternatively, the hyperspectral sensor can have a row-based filter with L different optical bandpass filters. Each optical bandpass filter can have a passband region corresponding to one of the spectral bands in the spectral dimension λ. L different optical bandpass filters can be arranged on the pixel rows of the hyperspectral sensor. The pixels of a pixel row are arranged in a first direction (also called the x direction or the column direction). For example, each bandpass filter can be arranged above at least one row, and L different optical bandpass filters are arranged successively in a second direction (also called the y direction or the row direction). The first direction is perpendicular to the second direction. Each of the L optical bandpass filters can correspond to a row of pixels or n rows of pixels (n>1). Therefore, the L different optical bandpass filters are preferably arranged so that the transmission or passband region changes after each row or after n rows. Therefore, each row or n rows of the hyperspectral sensor detects the intensity of the emitted beam filtered by the corresponding bandpass filter. Thus, the discrete spectrum of the emitted beam can be recorded. This method is also called the "push-broom" method and can be used in laser processing systems with high-precision material delivery. The synchronization of the line reading can be matched to the material conveying speed so that each area of ​​the surface of the workpiece is recorded multiple times (ie at different times with each of the different bandpass filters). This approach allows increasing the number of bandpass filters.

[0029] The at least one hyperspectral sensor can include a hyperspectral camera. In particular, the at least one hyperspectral sensor can include a CMOS camera, an infrared-enhanced CMOS sensor, a near-infrared (NIR)-enhanced CMOS sensor, an InGaAs-based sensor, a sensor array and / or a diode array.

[0030] The at least one hyperspectral sensor can have a spectral sensitivity range from 400 nm to 1800 nm and / or from 400 nm to 950 nm and / or from 400 nm to 1000 nm and / or from 1000 nm to 1700 nm and / or from 950 to 1800 nm and / or from 1200 nm to 2000 nm.

[0031] The sensor unit can in particular have at least two hyperspectral sensors with different sensitivity ranges. The sensor unit can be configured to convert the data detected by the respective hyperspectral sensors into a hyperspectral image with N×M pixels and output it, wherein each pixel includes L values ​​in the spectral dimension.

[0032] The sensor unit can include at least one beam splitter which is provided to split the radiation emitted by the region of the workpiece onto at least two hyperspectral sensors.

[0033] The sensor unit can be coupled to the laser processing head, in particular detachably coupled. The sensor unit can be arranged on the outside, in particular on the side, of the laser processing head. The optical axis of the sensor unit can extend at least partially parallel and / or coaxially with the propagation direction of the laser beam.

[0034] The laser processing system can further include a computing unit configured to determine an input tensor based on the hyperspectral image and to determine an output tensor based on the input tensor by means of a transfer function, wherein the output tensor contains information about the processing step, wherein the transfer function between the input tensor and the output tensor is formed by a deep neural network, in particular a deep convolutional neural network. The computing unit can be configured to form the output tensor in real time and / or to output control data based on the output tensor to a control unit of the laser processing system.

[0035] The method can also include the following steps: determining an input tensor based on the hyperspectral image, and determining an output tensor based on the input tensor and with the aid of a transfer function, the output tensor containing information about the processing step, wherein the transfer function between the input tensor and the output tensor is formed by a learned neural network, such as a deep neural network or a deep convolutional neural network. The formation of the output tensor can be performed in real time.

[0036] The information about the machining process can include information about machining errors and / or machining areas of the workpiece. The output tensor can contain in particular one of the following information: information about the joint gap, information about the offset between the joint partners, information about the weld pool, information about the steam capillary, information about the weld bead, information about the state of the machining process, information about machining errors, the presence of at least one machining error, the type of machining error, the location of the machining error on the workpiece, the probability of a specific type of machining error, and the spatial and / or planar extension of the machining error.

[0037] The detected hyperspectral images can therefore serve as a basis for identifying and classifying processing errors. Processing errors can be classified in particular as follows: gap, offset, missed weld penetration, missed weld penetration, spattering, pore formation. Hyperspectral images can also be used to identify deviations or anomalies in the laser processing process. Here, anomalies are, for example, deviations from a weld that was previously marked or classified as "good".

[0038] In order to learn a supervised machine learning method, such as a neural network, a large number of error-free processing steps (such as welding) and a large number of error-prone processing steps are performed and hyperspectral images are recorded in each case. The training data can be used for anomaly recognition, especially when welding with a small number of errors can be produced. The training of the neural network can be performed using standard methods.

[0039] The regulation or optimization of the processing steps can be achieved with the help of the so-called "reinforcement learning" of the neural network. In the reinforcement learning method, the control actions of the processing steps (for example, changes in laser power or changes in welding speed) are evaluated in terms of whether they lead to a process or process flow that is considered optimal, such as a process without processing errors. The actions that lead to the optimal process are evaluated as high. Here, a large number of processing steps are performed using corresponding detection of hyperspectral images, wherein the hyperspectral images present the state of the processing steps. Based on this, a neural network, especially a deep neural network or a deep convolutional neural network, can be trained in the following aspects: each possible control action is associated with the state of the processing step and the control action is classified according to the extent to which the control action leads to the optimal state of the processing step. After the method is completed, the neural network can optimize the processing step in each state through the corresponding control action and keep it in the optimized state.

[0040] The neural network can be adapted to the changed laser processing steps by so-called transfer learning based on training data. The training data can include: a plurality of hyperspectral images of the changed laser processing steps for determining the corresponding input tensors, and predetermined output tensors corresponding to the corresponding hyperspectral images, which contain the corresponding predetermined information about the changed laser processing steps.

[0041] Instead of using a neural network, significant spectral bands in the form of individual frames can be selected from the hyperspectral image, and the intensity distribution of the individual frames can then be evaluated. Thus, for example, the geometry of the keyhole and / or the weld pool can be evaluated. The evaluated geometry can be compared with those of a "good" or "bad" machining process (in particular a good or bad welding process). Deviations in the geometry can be used for identification and classification.

[0042] The laser processing system can further comprise at least one imaging optical device for imaging the beam emitted by the region of the workpiece onto the at least one hyperspectral sensor. The at least one imaging optical device can be arranged in the beam path of the laser beam and / or in front of the hyperspectral sensor along the propagation direction of the beam emitted by the region of the workpiece, for example between the beam splitter and the at least one hyperspectral sensor.

[0043] The method described above can be performed during a laser processing step. In this case, the processing result can be a current processing result. In particular, the processing result can be obtained in real time, or the method can be performed after the laser processing step has ended. In this case, the output tensor can contain information about processing errors. Depending on whether the detected hyperspectral image contains workpiece areas in the preceding part of the laser beam, in the current processing area and / or in the following part of the laser beam, pre-process, in-process and / or post-process monitoring can be performed using the described method. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In the following, the present invention is described in detail with the aid of the accompanying drawings. In the accompanying drawings:

[0045] Figure 1A A schematic diagram showing a hyperspectral sensor according to an embodiment of the present disclosure;

[0046] Figure 1B A schematic diagram showing a hyperspectral image that has been detected by a hyperspectral sensor according to an embodiment of the present disclosure;

[0047] Figure 1C Schematic diagram showing a region and a hyperspectral image of a workpiece according to an embodiment of the present disclosure;

[0048] Figure 2A schematic diagram showing a laser processing system according to a first embodiment of the present disclosure;

[0049] Figure 3 A schematic diagram showing a laser processing system according to a second embodiment of the present disclosure;

[0050] Figure 4 A schematic diagram showing a laser processing system according to a third embodiment of the present disclosure;

[0051] Figure 5 A schematic diagram showing a laser processing system according to a fourth embodiment of the present disclosure;

[0052] Figure 6 A method for monitoring a machining process on a workpiece using a laser beam is shown according to specific embodiments of the disclosure. DETAILED DESCRIPTION

[0053] Unless otherwise indicated, the same reference numerals are used below for identical or identically acting elements.

[0054] Figure 1A A schematic diagram of a hyperspectral sensor according to an embodiment of the present disclosure is shown. Figure 1B A schematic diagram showing a hyperspectral image that has been detected by a hyperspectral sensor according to an embodiment of the present disclosure is shown. Figure 1A The hyperspectral sensor shown in can be combined with other embodiments of the present disclosure.

[0055] The hyperspectral sensor 100 has a plurality of pixels 1001, which are arranged along a first direction (x direction or column direction) and a second direction (y direction or row direction) perpendicular to the first direction. These two directions correspond to the two spatial dimensions of the hyperspectral image detected by the hyperspectral sensor 100. All pixels 1001 of the hyperspectral sensor are capable of detecting or detecting beams within a certain wavelength range. A 5×5 optical mosaic filter 1002 is applied to the pixel 1001. The 5×5 mosaic filter 1002 includes 25 different single filters 1003, which are arranged in a unit grid, here for example in a square. The unit grid is repeated in two directions, and a single filter 1003 of the mosaic filter 1002 is assigned to each pixel 1001 of the hyperspectral sensor. Figure 1A Only a single area (einzelnen) of 5×5 consecutive pixels 1001 is shown. ) on a mosaic filter 1002. However, a single filter 1003 of the mosaic filter 1002 is arranged on each pixel 1001. Each single filter 1003 of the mosaic filter 1002 represents an optical bandpass filter and has a passband region, which corresponds to a spectral band within a wavelength range. The 25 single filters 1003 of the mosaic filter 1002 respectively have different passband regions or corresponding spectral bands. The spectral bands can be continuous and non-overlapping, so that they completely cover the wavelength range.

[0056] With the help of the hyperspectral sensor 100, it is possible to detect or probe the beam over a large wavelength range in a spectrally resolved manner. The beam is filtered by the single filter 1003 of the mosaic filter 1002 before being detected by the corresponding pixel 1001 according to the corresponding spectral band. Therefore, each pixel 1001 of the hyperspectral sensor 100 can detect or probe the intensity value of the beam in the spectral band corresponding to the passband area of ​​the single filter 1003. Thus, each pixel can correspond to an intensity value after bandpass filtering. As a result, 25 single frames 301 with the intensity values ​​are generated in the spectral band determined by the single filter 1003.

[0057] Figure 1C A schematic diagram of a region and a hyperspectral image of a workpiece according to an embodiment of the present disclosure is shown. A directional row hyperspectral sensor (i.e., a hyperspectral sensor with a row-based bandpass filter) according to an embodiment of the present disclosure has a plurality of pixels, which are arranged along a first direction (i.e., in the x-direction or column direction) and a second direction (i.e., in the y-direction or row direction). All pixels of the hyperspectral sensor are capable of detecting or detecting beams within a certain wavelength range. The hyperspectral sensor has a row-based filter with L different single optical bandpass filters. The L different optical bandpass filters are arranged on at least one row of pixels of the hyperspectral sensor, i.e., arranged in a directional row. In other words, the L different optical bandpass filters extend on at least one row of sensor pixels in a first direction. The optical bandpass filter can change every row or every n rows in the second direction. In other words, a single optical bandpass filter covers one or more rows of pixels. Each bandpass filter has a passband region, which corresponds to a spectral band within the wavelength range. Therefore, each row of the hyperspectral sensor detects the intensity of the emitted beam filtered by the corresponding bandpass filter. Different bandpass filters of the filter have different passband regions or corresponding spectral bands. The spectral bands can be continuous and non-overlapping, so that they completely cover the wavelength range.

[0058] With the help of a hyperspectral sensor, it is possible to detect or detect a beam over a large wavelength range in a spectrally resolved manner. Through a bandpass filter, the beam is filtered before being detected by the corresponding pixel according to the corresponding spectral band. Therefore, each pixel of the hyperspectral sensor can detect or detect the intensity value of the beam in the spectral band corresponding to the passband area of ​​the bandpass filter. Therefore, the intensity value after bandpass filtering can be assigned to each pixel. In order to obtain a hyperspectral image of the area of ​​the workpiece 2, the row 21 of the workpiece area is scanned by the hyperspectral sensor in the second direction (i.e., in the y direction or row direction). Therefore, each row 21 of the area of ​​the workpiece 2 is recorded multiple times (i.e., using each bandpass filter in the bandpass filter of the hyperspectral sensor). In the case of a hyperspectral sensor with a row-based filter with L different bandpass filters, L bandpass filtered single frames 301 with the intensity value are generated in the spectral band determined by the corresponding bandpass filter to form the hyperspectral image. This method is also called the "Push Broom" method. This method can be particularly applied in laser processing systems with high-precision material transportation. The scanning or synchronization of the row reading can be run in the direction of relative movement between the laser processing head and the workpiece or in the direction of material conveying and can be matched with the material conveying speed. In other words, a single frame of the hyperspectral image is recorded at a time, and the single frame corresponds to another wavelength range in each pixel row. Subsequently, the rows are scanned in another wavelength range respectively by scanning movement or relative movement of each row. This method allows the number of bandpass filters to be increased. The scanning movement or relative movement is preferably performed vertically on the pixel rows of the hyperspectral sensor, that is, the scanning movement or relative movement can be performed in the y direction or the row direction. Therefore, according to the resolution of the hyperspectral sensor, more wavelength bands can be achieved.

[0059] Therefore, the hyperspectral sensor 100 is configured to detect the radiation and output it as 25 bandpass filtered individual frames 301. The ensemble of these 25 frames is referred to as a hyperspectral image 300, a hyperspectral cube or a hyperspectral data cube.

[0060] The hyperspectral sensor according to an embodiment of the present disclosure has pixels having a spectral sensitivity between 400 nm and 1800 nm. When a 5×5 mosaic filter is used, a hyperspectral image or a hyperspectral cube can be detected, which has 25 single frames. Therefore, each of the 25 single frames covers a spectral band with a width of 56 nm. In other words, each single frame corresponds to an intensity distribution in a spectral band of 56 nm. In other words, the single frames have a spacing of 56 nm in the spectral dimension.

[0061] According to other embodiments of the present disclosure of a hyperspectral sensor (not shown), the hyperspectral sensor can have a 4×4 mosaic filter with 16 single filters. Thus, the hyperspectral sensor is configured to output 16 single frames as a hyperspectral image. For example, the hyperspectral sensor can have 2048×1088 pixels and be configured as a CMOS sensor. The hyperspectral sensor has a 4×4 mosaic filter. As a result, the local or spatial resolution is reduced by a factor of 4. In other words, a hyperspectral image detected by such a hyperspectral sensor has 16 single frames, each of which has 512×272 pixels.

[0062] Thus, a hyperspectral sensor can substantially simplify the image recording in different spectral bands and separate the snapshot recording into consecutive spectral bands.

[0063] The hyperspectral sensor 100 can be based on graphene or InGaAs and have a sensitivity between 400nm and 1800nm, for example. According to other embodiments, the hyperspectral sensor 100 can be based on InGaAs with a spectral sensitivity between 900nm and 1800nm. According to other embodiments, the hyperspectral sensor can be a near infrared enhanced CMOS sensor. The graphene-based hyperspectral sensor has a constant high sensitivity between 400nm and 1800nm.

[0064] Figure 2 A schematic diagram of a laser processing system for performing a processing operation (eg, welding or cutting) on ​​a metal workpiece by means of a laser beam according to a first embodiment of the present disclosure is shown.

[0065] The laser processing system 1 comprises a laser processing head 20 and a sensor unit 10 for monitoring a processing process. The laser processing head 20 can be a laser cutting head, a laser soldering head or a laser melting welding head.

[0066] The laser processing system 1 includes a laser device (not shown) for generating a laser beam 201 (also referred to as a “processing beam” or a “processing laser beam”).

[0067] According to an embodiment, the laser processing system 1 or a part thereof (e.g., the processing head 20) can be movable along a processing direction. The processing direction can be a cutting direction, a brazing direction, or a welding direction and / or a movement direction of the laser processing system 1 (e.g., the processing head 20) relative to the workpiece 2. In particular, the processing direction can be a horizontal direction. The processing direction can also be referred to as a "feed direction".

[0068] The laser processing system 1 is controlled by a control unit (not shown) which is configured to control the processing head 20, the sensor unit 10 and / or the laser device. The laser processing system 1 can also include a computing unit (not shown) which is configured to determine information about the processing sequence. According to one embodiment, the computing unit is combined with the control unit (not shown). In other words, the functionality of the computing unit can be combined with the functionality of the control unit in a common processing unit.

[0069] The laser device 1 can have collimation optics 202 for collimating the laser beam 201 and focusing optics 203 which are provided to focus the laser beam 201 onto the workpiece 2. The laser device 1 can also include a beam splitter 204 which is provided to deflect the laser beam 201 onto the workpiece 2.

[0070] During the machining process or when machining the workpiece 2 with the aid of the laser beam 201, a beam 205 is generated which is emitted by the workpiece 2. The beam 205 can be light of the laser beam 201 which is reflected or backscattered by the surface of the workpiece 2, a plasma beam, a temperature beam or visible light. The beam 205 can also include light which is radiated by an irradiation source (not shown) onto the workpiece 2 and reflected back.

[0071] The sensor unit 10 comprises a hyperspectral sensor 100 and an imaging optical system 101 . Depending on the embodiment, the imaging optical system 101 can also be omitted. The imaging optical system 101 is provided for imaging or focusing the beam 205 onto the hyperspectral sensor 100 .

[0072] According to Figure 2 In the embodiment shown in FIG. 1 , the sensor unit 10 comprises a hyperspectral sensor 100 having a spectral sensitivity of 400 nm to 1800 nm. According to other embodiments, the sensor unit 10 comprises a hyperspectral sensor having a spectral sensitivity of 400 to 950 nm or to 1000 nm. According to embodiments, the hyperspectral sensor 100 can be a graphene-based or InGaAs-based hyperspectral sensor with a mosaic filter, such as a combination thereof. Figure 1A and 1B According to other embodiments, the hyperspectral sensor 100 can be a near-infrared enhanced hyperspectral sensor with a mosaic filter, such as a combination of Figure 1A and 1B Just as described.

[0073] The sensor unit 10 or the hyperspectral sensor 100 is configured to detect the beam 205 emitted by the workpiece 2 and output it as a hyperspectral image. The computing unit is configured to determine information about the processing step, in particular information about the processing result, such as processing errors, based on the hyperspectral image. This determination can be achieved with the aid of a deep neural network, in particular a deep convolutional neural network.

[0074] According to Figure 2 In the embodiment shown in FIG. 2 , the sensor unit 10 is mounted on a laser processing head 20 or is coupled thereto.

[0075] According to Figure 2 , the sensor unit 10 is arranged such that the beam 205 emitted by the workpiece is guided to the sensor unit 10 coaxially with the laser beam 201 in the laser processing head 20. Thus, the beam 205 emitted by the workpiece enters the laser processing head 20 in the opposite direction of the propagation direction of the laser beam 201 and impinges on the sensor unit 10. Here, the beam 205 passes through the focusing optical device 203 and the beam splitter 204 before it impinges on the sensor unit 10. In other words, the beam 205 and the laser beam 201 overlap in the laser processing head 20.

[0076] The laser processing system can also include a computing unit that determines a multi-channel input tensor based on the hyperspectral image detected by the sensor unit and determines an output tensor based on the input tensor with the help of a transfer function, the output tensor containing information about the processing process. The transfer function can be formed by a neural network, preferably a deep neural network or a deep convolutional neural network. The computing unit outputs adjustment data or control data to a control unit of the laser processing system based on the output tensor. Anomaly recognition can be performed based on the hyperspectral image or hyperspectral cube, in which the hyperspectral image of the error-free process is used as a database. Anomaly recognition can be performed using standard methods, building a model from training data and calculating the deviation of features from the model in the derivation.

[0077] The classification of typical error categories (such as gaps, offsets, leaky penetration and leaky welds, spattering, and pore formation) requires the generation of a large amount of training data containing these typical errors. This means that a weld must be generated for each error type. The training data generated in this way consists of a hyperspectral image and an added error description. These training data sets can be used to use monitored machine learning methods for classification. For example, the generated hyperspectral image can be used as a multi-channel input tensor of a deep neural network (typically a convolutional neural network). The generalized transfer function generated after the training process maps the hyperspectral image to an output vector in the derivation, which provides a prediction about the error classification. The learned generalized transfer function formed by the model of the deep convolutional neural network contains a feature vector in the last fully connected layer, which describes the process.

[0078] Treating the hyperspectral cube as a multi-channel input tensor for deep neural networks (especially deep convolutional neural networks) allows training these networks and, in this context, forming a generalized mapping function that maps the input tensor to a classification result. The use of hyperspectral images for mapping the process can be used in the "Reinforcement Deep Q Learning" method: the hyperspectral images, which are considered as the states of the process, allow the effects of actions on the process to be learned with the help of deep convolutional neural networks. After the learning is complete, the program can perform the optimal change of the process for each state.

[0079] By evaluating the actions acting on the process (for example, changes in laser power or changes in welding speed), the optimization of the laser welding process is achieved by means of a reinforcement learning method. The action that leads to the optimal process is evaluated as high. Since there is professional knowledge about the laser process, the number of actions in the process can be limited. Therefore, through a large number of welding experiments, a deep neural network (typically a convolutional neural network) can be used to train the assignment of the state presented by the hyperspectral cube in terms of the evaluation of each possible or allowed action. Then, the trained network allows the optimal action to be implemented for the process based on any state in order to optimize the process and keep it in the optimal state.

[0080] Figure 3 FIG. 1 is a schematic diagram of a laser processing system according to a second embodiment of the present disclosure. Figure 3 The laser processing system shown in Figure 2 The laser processing system is shown in FIG. 1 , except for the differences described below.

[0081] According to Figure 3, the beam path of the beam 205 emitted from the workpiece region and directed to the sensor unit 10 extends outside the laser processing head 20. The beam 205 thus does not enter the laser processing head 20 before it impinges on the sensor unit 10. In other words, the beam 205 and the laser beam 201 do not overlap in the laser processing head 20. Any filtering effects of the focusing optics of the laser processing head 20 are eliminated.

[0082] According to Figure 3 In the embodiment shown in FIG. 1 , the sensor unit 10 can be mounted on a side surface of the laser processing head. This arrangement of the sensor unit is also referred to as an “off-axial arrangement”.

[0083] Figure 4 FIG. 1 is a schematic diagram showing a laser processing system according to a third embodiment of the present disclosure. Figure 4 The laser processing system shown in Figure 2 The laser processing system is shown in FIG. 1 , except for the differences described below.

[0084] The sensor device 10 comprises a first hyperspectral sensor 100a, a second hyperspectral sensor 100b and a beam splitter 102. Instead of a sensor having broadband sensitivity, as in Figure 2 and Figure 3 In the embodiment of FIG. 1 , the corresponding wavelength range is covered by means of two hyperspectral sensors 100a and 100b and a beam splitter 102. Hyperspectral cubes of individual hyperspectral sensors (eg, an InGaAs image sensor and an NIR enhanced CMOS image sensor) can form a hyperspectral cube.

[0085] The beam splitter 102 is configured to split the beam 205 emitted by the workpiece and redirect it to the first hyperspectral sensor 100a and the second hyperspectral sensor 100b. The sensor device 10 can also include a first imaging optical device 101a and a second imaging optical device 101b arranged in front of the first and second hyperspectral sensors 100a and 100b, respectively, so as to image a portion of the beam 205 thereon.

[0086] According to Figure 4 In the embodiment shown in FIG. 1 , the first hyperspectral sensor 100a has a spectral sensitivity of 400 nm to 950 nm or 400 nm to 1000 nm, and the second hyperspectral sensor 100b has a spectral sensitivity of 1000 nm to 1700 nm or 900 to 1800 nm. According to an embodiment, the first hyperspectral sensor 100a can be a near-infrared enhanced CMOS sensor with a mosaic filter, such as a combination thereof. Figure 1A and Figure 1BThe second hyperspectral sensor 100b can also be an InGaAs-based hyperspectral sensor with a mosaic filter, such as a combination thereof. Figure 1A and Figure 1B Just as described.

[0087] according to Figure 4 The sensor unit 10 is configured to combine the hyperspectral images detected by the first hyperspectral sensor 100a and the second hyperspectral sensor 100b into a single hyperspectral image and output it. In this case, the combined hyperspectral image can include 50 single frames. The first hyperspectral sensor 100a and the second hyperspectral sensor 100b can have different resolutions, i.e. different numbers of pixels. Therefore, the sensor unit 10 can be configured to adapt the resolutions of the corresponding hyperspectral images obtained from the two hyperspectral sensors to each other, so that the single frames in the hyperspectral cube have the same spatial dimensions.

[0088] Figure 5 FIG. 1 is a schematic diagram showing a laser processing system according to a fourth embodiment of the present disclosure. Figure 5 The laser processing system shown in Figure 4 The laser processing system is shown in FIG. 1 , except for the differences described below.

[0089] According to Figure 5 , the beam path of the beam 205 which is directed to the sensor unit 10 and which originates from the workpiece region extends outside the laser processing head 20. The beam 205 thus does not enter the laser processing head 20 before it strikes the sensor unit 10. In other words, the beam 205 and the laser beam 201 do not overlap in the laser processing head 20.

[0090] According to Figure 5 In the embodiment shown in FIG. 1 , the sensor unit 10 can be mounted on a side surface of the laser processing head. This arrangement of the sensor unit is also referred to as an “off-axis arrangement”.

[0091] Figure 6 A method for monitoring a machining process on a workpiece using a laser beam is described.

[0092] The method can be implemented by a laser processing system according to an embodiment of the present disclosure.

[0093] The first step 601 comprises causing a laser beam to be incident on a processing area on a workpiece. In a second step 602, a hyperspectral image having N×M pixels of the area of ​​the workpiece is detected. Each pixel comprises L values, wherein the hyperspectral image has two spatial dimensions x and y and one spectral dimension λ, wherein N specifies the number of pixels in the first spatial dimension x, M specifies the number of pixels in the second spatial dimension y, and L specifies the number of spectral bands in the spectral dimension λ of the hyperspectral image, wherein M, N and L are all natural numbers.

[0094] The method can also include a third step 603, in which an input tensor is determined based on the hyperspectral image, and a fourth step 604, in which an output tensor is determined based on the input tensor and with the aid of a transfer function, the output tensor containing information about the machining result of the laser machining process. The transfer function between the input tensor and the output tensor can be formed by a learned neural network.

[0095] According to one embodiment, steps 602 to 604 are performed in parallel or simultaneously with step 601. According to another embodiment, steps 602 to 604 are performed after step 601 is completed.

[0096] According to an embodiment of the present disclosure, in a laser processing process in which a laser beam is incident on a processing area of ​​a workpiece, a hyperspectral image of an area of ​​the workpiece is detected in order to monitor the laser processing process. The detected area may include a processing area and / or an area in a preceding or following portion of the laser beam. The hyperspectral image has N×M pixels in two spatial dimensions. Each pixel has L values ​​in a spectral dimension. In other words, the hyperspectral image has L single frames of N×M pixels. L gives the number of spectral bands within the detected wavelength range. The L spectral bands can be adjacent to each other. A hyperspectral image or a hyperspectral cube over the entire wavelength range of the beam emitted by the processing process completely maps the process because all wavelengths are provided in a spatially resolved manner. The hyperspectral image contains information about the state of the laser processing process and can therefore be used as a basis for monitoring and regulating the laser processing process. In addition, information about the processing results of the laser processing process can be obtained with the help of the hyperspectral image, which includes information about processing errors. The state of the laser processing process or the information about the processing results can be determined with the help of a (deep) neural network, wherein the detected hyperspectral image is used as the basis of the input tensor of the deep neural network.

Claims

1. A laser processing system (1) for performing a welding process or a cutting process on a metal workpiece (2) by means of a laser beam (201), the laser processing system (1) comprising: A laser processing head (20), the laser processing head being used to cause a laser beam (201) to be incident on a processing area on the workpiece (2); and A sensor unit (10) for monitoring the welding process or the cutting process using at least one hyperspectral sensor (100, 100a, 100b), wherein the hyperspectral sensor (100, 100a, 100b) comprises a hyperspectral camera, wherein the sensor unit (10) is configured to continuously detect a hyperspectral image (300) having N×M pixels of an area of ​​the workpiece (2), wherein each pixel comprises L values, The hyperspectral image (300) has two spatial dimensions x and y and one spectral dimension λ, wherein N indicates the number of pixels in the first spatial dimension x, M indicates the number of pixels in the second spatial dimension y, and L indicates the number of spectral bands in the spectral dimension λ of the hyperspectral image, wherein M, N and L are all natural numbers greater than zero, wherein the L spectral bands are adjacent to each other, wherein the at least one hyperspectral sensor (100, 100a, 100b) has a spectral sensitivity range from 400 nm to 1800 nm, a computing unit configured to determine an input tensor based on the hyperspectral image (300) and to continuously determine an output tensor based on the input tensor by means of a transfer function, the output tensor containing information about the welding process or the cutting process, wherein the transfer function between the input tensor and the output tensor is formed by a deep neural network, The information about the welding process or the cutting process includes the existence of at least one processing error and the type of the processing error.

2. The laser processing system (1) according to claim 1, wherein: The sensor unit (10) is provided to detect a beam (205) emitted by a detected region of the workpiece (2) and / or a reflected laser beam and to output it as a hyperspectral image (300).

3. The laser processing system (1) according to claim 1 or 2, wherein: The sensor unit (10) is configured to detect at least one of the following beam types: a temperature beam, a beam in the infrared range of light, a beam in the near infrared range of light, a beam in the visible range of light, a plasma beam, reflected light of a laser beam, backscattered light of a laser beam, and light incident and reflected by an irradiation source.

4. The laser processing system (1) according to claim 1 or 2, wherein: The area of ​​the workpiece (2) detected in the hyperspectral image (300) includes at least one of the following areas on the workpiece: the processing area, the area in the preceding part of the laser beam (201), the area in the succeeding part of the laser beam (201), the area still to be processed, and the processed area.

5. The laser processing system (1) according to claim 1 or 2, wherein: L is equal to or greater than 16.

6. The laser processing system (1) according to claim 1 or 2, wherein: L is equal to or greater than 20.

7. The laser processing system (1) according to claim 1 or 2, wherein: L is equal to or greater than 25.

8. The laser processing system (1) according to claim 1 or 2, wherein: L is equal to or greater than 100.

9. The laser processing system (1) according to claim 1 or 2, wherein: The spectral bands are of equal size and / or consecutive to each other.

10. The laser processing system (1) according to claim 1 or 2, wherein: The sensor unit (10) is designed to detect a hyperspectral image (300) at predetermined time intervals.

11. The laser processing system (1) according to claim 1 or 2, wherein: The sensor unit (10) is configured to simultaneously detect all L values ​​for all N×M pixels of the hyperspectral image (300), or the sensor unit (10) is configured to simultaneously detect all pixel rows of the hyperspectral image (300) but in different spectral bands and to sequentially detect L values ​​of different spectral bands for each pixel row or for all n pixel rows.

12. The laser processing system (1) according to claim 1 or 2, wherein: The hyperspectral sensor (100, 100a, 100b) comprises a mosaic filter (1002) having N×M single filters (1003) with L different passband regions, or The hyperspectral sensor (100, 100a, 100b) has a plurality of pixel rows and the pixels in one pixel row are arranged in a first direction (x), and the hyperspectral sensor (100, 100a, 100b) comprises a row filter, in which each of a plurality of single filters (1003) extends on at least one pixel row of the hyperspectral sensor (100, 100a, 100b) and the plurality of single filters (1003) with L different passband regions are arranged in a second direction (y) perpendicular to the first direction (x).

13. The laser processing system (1) according to claim 1 or 2, wherein: The at least one hyperspectral sensor (100, 100a, 100b) has a spectral sensitivity range from 400 nm to 950 nm and / or from 400 nm to 1000 nm and / or from 1000 nm to 1700 nm and / or from 950 to 1800 nm and / or from 1200 nm to 2000 nm.

14. The laser processing system (1) according to claim 1 or 2, wherein: The at least one hyperspectral sensor (100, 100a, 100b) comprises a CMOS camera, an infrared enhanced CMOS sensor, a near infrared (NIR) enhanced CMOS sensor, an InGaAs based sensor, a graphene based sensor, a sensor array and / or a diode array.

15. The laser processing system (1) according to claim 1 or 2, wherein: The computing unit is configured to form the output tensor in real time and, based thereon, output control data to a control unit of the laser processing system (1).

16. The laser processing system (1) according to claim 1 or 2, wherein: The output tensor further comprises one of the following information: information about the status of the welding process or the cutting process, information about machining errors, the location of machining errors on the workpiece, the probability of a specific type of machining error and the spatial and / or planar extension of the machining errors.

17. A method for monitoring a welding process or a cutting process on a metal workpiece (2) by means of a laser beam (201), the method comprising the following steps: - causing a laser beam (201) to be incident on a processing area on a workpiece (2); and - continuously detecting a hyperspectral image (300) having N×M pixels from the area of ​​the workpiece (2) by means of a hyperspectral sensor (100, 100a, 100b), wherein the hyperspectral sensor (100, 100a, 100b) comprises a hyperspectral camera, wherein each pixel comprises L values, The hyperspectral image (300) has two spatial dimensions x and y and one spectral dimension λ, wherein N specifies the number of pixels in the first spatial dimension x, M specifies the number of pixels in the second spatial dimension y, and L specifies the number of spectral bands in the spectral dimension λ of the hyperspectral image, wherein M, N and L are natural numbers, wherein L spectral bands are adjacent to each other, wherein the hyperspectral sensor (100, 100a, 100b) has a spectral sensitivity range from 400 nm to 1800 nm, - determining an input tensor based on said hyperspectral image (300), and - based on the input tensor and with the aid of a transfer function, continuously determining an output tensor, which contains information about the welding process or the cutting process, wherein the transfer function between the input tensor and the output tensor is formed by a deep neural network, The information about the welding process or the cutting process includes the existence of at least one processing error and the type of the processing error.

Images