Method and system for analyzing laser processing processes based on spectrograms

Abstract

This invention provides a method and system for analyzing laser processing processes based on spectrograms. [Solution] Disclosed is a method for analyzing a laser processing process, the method comprising: acquiring multiple spectra of process emissions at successive time points; creating at least one spectrogram based on the acquired spectra; and using a trained neural network to verify a predicted value of at least one physical quantity and / or verify at least one classification of a laser processing process, wherein the neural network receives the spectrogram as an input tensor and outputs the physical quantity and / or the classification of the laser processing process as an output tensor.

Description

Technical Field

[0001] The present disclosure relates to a method for analyzing a laser processing process, and in particular to a system and a laser processing system for analyzing a laser processing process based on a spectrogram.

Background Art

[0002] In a laser processing process, a workpiece, particularly a metal workpiece, is processed using a processing laser. The processing includes, for example, laser cutting, laser soldering, and / or laser welding. A laser processing system can include, for example, a laser processing head.

[0003] The laser processing process is often subject to quality control. In particular, when a workpiece is laser welded or laser soldered, the quality of the obtained joint is confirmed. Current monitoring systems for process monitoring and quality assessment during laser welding, laser soldering, or laser cutting are usually based on preprocess, inprocess, and / or postprocess monitoring systems. The preprocess monitoring system usually has the task of detecting or measuring the joint gap before the laser processing process, guiding the laser beam to an appropriate position, and determining the offset of the joint partner. In most cases, a triangulation system is used for this purpose.

[0004] Inprocess and postprocess monitoring systems are regularly used to monitor the laser processing process and manage and guarantee the quality of the obtained joint. As a result of the laser processing process, for example, a completed and cooled weld seam can be inspected and measured according to applicable standards (e.g., SEL100), so postprocess monitoring is particularly well used for quality monitoring. Postprocess monitoring or postinspection requires a great deal of effort in plant engineering. In many cases, a separate measurement cell must be installed for postprocess monitoring.

[0005] In-process monitoring systems (also called in-line or online process monitoring systems) are typically designed to detect at least some of the radiation emitted by laser processing processes. In many cases, it is not possible to record and process all signals with frequency and spatial resolution using in-process monitoring systems. Therefore, achieving quality monitoring that can be classified into error classes based on such monitoring systems is difficult. In laser processing processes, radiation is typically emitted from the molten pool in the visible range of approximately 400 nm to 850 nm, from the plasma in the range of approximately 400 nm to 1100 nm, from backscattered light from the processing laser in the range of approximately 900 nm to 1100 nm, and thermal radiation above 1000 nm. In other words, laser processing processes emit radiation over a wide range of approximately 400 nm to 1800 nm. This radiation is also called process emission or process radiation.

[0006] Depending on the application, for example, if testing in the application shows that a certain spectral emission range, such as thermal emission, does not contain information about the quality characteristics of interest, spectral detection can be limited to a specific wavelength range. In the case of joining different metals, obtaining frequency-related or frequency-dependent intensity of process emissions is particularly advantageous, as it allows for spectral evidence, for example, in overlap welding, of whether spectral lines of both joining partners are present in the spectrum. The process emission spectrum may also indicate changes in the alloy. Such changes can be caused, for example, by using materials from different manufacturers. Changes in the alloy can affect joining, cutting, and laser printing processes. During laser removal, the spectrum can indicate whether the coating is actually being removed, or whether the material beneath the coating is being heated by the laser or converted into plasma.

[0007] In existing in-process monitoring systems, diodes are typically used to analyze process radiation, which is detected within a narrow bandwidth. In principle, photodiodes with varying sensitivities are used. For example, a Si diode can detect the range from approximately 400 nm to 800 nm, an InGaAs diode from approximately 800 nm to 1200 nm, and another InGaAs or Ger diode from approximately 1200 nm to 2000 nm. Depending on the process, appropriate optical filters can be used to isolate regions from these wavelength ranges. For example, depending on the processing laser, the region of backscattered laser radiation can be reduced to approximately 1020 nm to 1090 nm. Wavelengths outside these detection ranges of the diodes and optical filters are not detected. The intensity components in the narrow wavelength range are no longer visible in the integrated intensity across the broad wavelength range.

[0008] Intensity curves recorded using such diodes in this manner are typically filtered to check whether they exceed a calculated or specified threshold. The filter parameters and thresholds are set individually for each signal, i.e., for each wavelength range. Therefore, the observation and evaluation of a single diode is equivalent to an individual sensor system.

[0009] For quality monitoring, reference curves can also be formed from many recorded signal curves, and so-called envelope curves can be arranged around these reference curves. The envelope curves represent thresholds at each point in time during welding. If the signal exceeds or falls below the values ​​of the envelope curves during the laser processing process, an error message is displayed or output using a predetermined error criterion. The criterion can be, for example, the integral value of the signal on the envelope curve, or the value of the signal that exceeds the envelope curve. An example of such a system is Precitec's LWM product.

[0010] The evaluation of spectra, and classification or regression based on spectra, cannot be satisfactorily achieved with feature-based methods. In particular, certain types of errors cannot be reliably concluded.

[0011] Another solution can be based on image sensors, and image processing can be used, for example, to analyze and / or measure molten pools and keyholes, allowing quality descriptions to be made with the measurement data obtained in this way.

[0012] When using an InGaAs sensor, for example, as disclosed in Patent Document 1, typically only the thermal profile in the range above approximately 1200 nm is evaluated, and Patent Document 1 describes a non-destructive quality determination method and apparatus for welding seams and welding equipment.

[0013] Furthermore, brightness profiles can be recorded using a CMOS sensor in the range of approximately 450 nm to 800 nm, and these brightness profiles are compared with a hypothetical model to recognize quality characteristics. Patent document 2 describes such a method for recognizing errors during a laser processing process.

[0014] Recording intensity versus frequency and evaluating spectra is described, for example, in Patent Document 3. In either case, the evaluation refers to data from a single spectrum and uses various spectral features to assess the laser process. In Patent Document 4, the frequency components important to the laser processing process are selected by the user. Using a spectrometer requires extensive integration work due to its structural size and the required precision, making it unsuitable for mass production. [Prior art documents] [Patent Documents]

[0015] [Patent Document 1] German Patent Application Publication No. 102008058187 [Patent Document 2] German Patent No. 102011078276 Specification [Patent Document 3] U.S. Patent Application Publication No. 2014149075 [Patent Document 4] German Patent Application Publication No. 102008043820 Specification [Patent Document 5] German Patent Application Publication No. 102020112116 [Patent Document 6] Chinese Patent Application Publication No. 111426639 [Patent Document 7] Chinese Patent Application Publication No. 103878479 [Patent Document 8] German Patent Application Publication No. 102018129441 [Summary of the Invention] [Problems to be Solved by the Invention]

[0016] Therefore, an object of the present invention is to provide a method and a system for analyzing a laser processing process in order to reliably evaluate the quality of laser processing. Furthermore, it is an object of the present invention to reliably and quickly detect processing errors without going through a complicated parameterization process.

[0017] Furthermore, an object of the present invention is to automate the quality evaluation of laser processing and the detection of processing errors, and thus enable process monitoring, particularly online process monitoring.

[0018] Adjusting the conditions or parameters of the laser processing process based on predicted values and / or classifications is a further object of the present invention. [Means for Solving the Problems]

[0019] One or more of these problems are solved by the features of the independent claims.

[0020] According to one aspect, a method for analyzing a laser processing process includes detecting a plurality of spectra of process emissions at successive time points or periods, generating at least one spectrogram based on the detected spectra, and determining at least one value or predicted value of a physical quantity or physical property and / or determining at least one classification of the laser processing process by a trained neural network, wherein the neural network receives the spectrogram as an input tensor and outputs the classification of the physical quantity and / or the laser processing process as an output tensor. The laser processing process can be, for example, a laser cutting, laser welding, laser soldering, or laser ablation process.

[0021] In other words, at least one value and / or at least one classification of the laser processing process can be determined based on at least one spectrogram by a transfer function formed by a trained neural network.

[0022] The neural network can be trained by an error feedback or error backpropagation learning method. The neural network can be a convolutional neural network and / or a deep neural network, such as a deep convolutional neural network or a convolutional network. The convolutional network may have at least one so-called "fully connected" layer.

[0023] The method according to the present invention enables efficient conclusions about specific types of errors or specific classifications without the need to process detected data and analyze it individually with computer assistance using conventional feature analysis. By using neural networks, particularly convolutional neural networks, it is possible to analyze a spectrogram with respect to a specific error and classify and / or map physical quantities or physical properties without knowing or extracting the spectrogram's features. Such errors may include, for example, poor welding when welding different materials, or the inclusion of coatings in the weld. Thus, machining errors can be reliably and quickly identified without complex parameterization processes. In other words, a particularly efficient, automated, and simple quality assurance process can be performed for each machined workpiece. Furthermore, neural networks are particularly sensitive and, with training, can identify slight changes or hard-to-recognize defects or machining errors.

[0024] At least one spectrogram serves as the input dataset or input tensor for a trained neural network. Multiple detected or recorded spectra can be combined as a spectrogram to form the input tensor for the neural network. Preferably, each generated spectrogram can form a separate input tensor. The spectrogram can represent the composition of process emissions from individual frequencies over time. Thus, the spectrogram can represent the time variation of the frequency distribution of process emissions, for example, using a short-time Fourier transform.

[0025] A trained neural network outputs a value for at least one physical property or classification as an output tensor. A neural network can also simultaneously determine values ​​for multiple physical properties and / or classifications and output them as an output tensor. Simultaneously quantifying multiple physical properties and / or classifications of the processing result means that the laser processing process can be monitored more reliably and accurately.

[0026] In particular, the quality of a laser processing process can be assessed by evaluating the workpiece immediately after processing. From this, it is possible to infer whether the processing work meets the desired standards for the processing process or the workpiece, or whether processing errors occur. The determination of at least one value or predictive value, in particular the detection of processing errors, can be automated, thus enabling process monitoring, especially online process monitoring.

[0027] The laser processing process can be performed on at least one workpiece, particularly a metal workpiece made of, for example, a pure metal and / or alloy. In particular, one workpiece may include multiple composite initial workpieces. In other words, two metal parts or initial workpieces made of the same or different materials can be joined by the laser processing process, for example, by welding.

[0028] Process emissions are generated on the workpiece, for example, during laser processing, i.e., during or immediately after laser processing, and are at least partially detected by a detector or sensor, particularly a spectrometer detector.

[0029] At least one spectrogram can be detected during the laser processing process in progress. Similarly, the values ​​of physical properties can be determined during the execution of the laser processing process or after its completion. Thus, the method according to the present invention can be formed particularly as an in-process method. Therefore, the determination of at least one physical property and / or classification of the laser processing process can be performed particularly in real time.

[0030] The method according to the present invention can be carried out continuously and / or repeatedly during a laser processing process. In other words, spectra and / or spectrograms can be detected continuously and / or repeatedly, and values ​​for at least one physical property and / or classification can be determined each time.

[0031] The generation of a spectrogram may have a time-series assembly of at least some or some of the detected spectra, where each spectrum is assigned, or may be assigned, to a point in time or time interval in which it was detected. The detected spectra are recorded at consecutive points in time. For example, recording may take from a few nanoseconds to a few milliseconds, so the point in time when the spectrum is recorded may actually correspond to a short period of time. However, some points in time can be approximated. For example, a spectrum may be assigned to the point in time when the detection or recording of the spectrum begins.

[0032] Therefore, the present invention is based on the idea of ​​using a neural network to specify the values ​​of physical properties and / or the classification of processing results, i.e., to quantify physical properties and / or classify laser processing processes or processing results, the neural network using at least one spectrogram recorded for a laser processing process as an input dataset. Thus, using the method according to the present invention, it is possible to determine the values ​​of physical properties and / or the classification of processing results of a laser processing process in a non-destructive manner. Physical properties can be quality features of the processing result and may be specified in standards or specifications, for example, regarding material properties. Therefore, the quality of processing results, such as weld seams or solder seams and cut edges, can be quantified or described quantitatively and evaluated based on determined values ​​of physical properties to specify fine-grained evaluation metrics for analyzing weld seams or solder seams and cut edges and the corresponding laser processing processes.

[0033] In summary, the method according to the present invention enables monitoring, particularly in real time, of laser processing processes, especially laser cutting, laser welding, laser soldering, or laser removal processes, from spectrograms generated by machine learning methods.

[0034] Each detection in the spectrum may involve detecting the intensity of process emissions as a function of wavelength at each point in time. Instead of process emission intensity, absorption, such as infrared absorption, may also be detected. The spectrum typically includes a plot or assignment of intensity to wavelength, with wavelengths usually specified in nanometers. Alternatively, the plot or assignment of intensity to wavelength may be replaced by a plot or assignment of intensity to frequency, with frequency specified in Hz.

[0035] The detected intensity can be raw data. Therefore, quality assessment of the laser processing process and detection of processing errors in the processed workpiece, particularly on the workpiece surface, can be performed based on the recorded raw data. This is called "end-to-end" processing or analysis. Analyzing raw data can reduce the number of steps required for analyzing the laser processing process, providing a particularly efficient method. As a result, complex preprocessing or data preparation can be omitted. In particular, there is no need for the program or user to perform mathematical calculations to analyze the detected intensity and detected spectrum. This makes the process particularly simple and efficient in terms of both time and cost. Raw data includes, for example, intensity determined based on electrical signals from sensors or detectors. Raw data can be data that has not undergone further mathematical processing or calculations after detection, such as filtering, smoothing, or normalization.

[0036] In particular, in-process monitoring systems can analyze laser processing processes using raw data, enabling rapid and efficient real-time analysis or evaluation.

[0037] The intensity of a single detected spectrum can essentially be recorded simultaneously. By simultaneously detecting the intensity of different wavelengths in the spectrum, all intensities for a particular state of the workpiece can be detected at a given time, thereby allowing for a complete and reliable mapping of that state. Furthermore, simultaneously detecting spectral intensities is particularly efficient.

[0038] The detection of each spectrum may involve local spectral splitting at the detector. Spectral splitting at or immediately before the detector can be performed, for example, by diffraction gratings and / or prisms. However, spectral splitting is preferably performed by reflection. Diffraction gratings may be preferred because they function "by reflection" and do not absorb any portion of the spectral emission through the material, and therefore do not distort the signal or spectrum. In particular, spectral splitting allows for the simultaneous recording of all or at least some of the intensity of the spectrum.

[0039] Process emissions include, for example, thermal radiation, plasma radiation, and / or laser radiation reflected from the workpiece surface. Predicted values ​​of physical quantities and / or classification of laser processing processes can be derived from the aforementioned process emissions, particularly characteristic spectral or wavelength ranges. In particular, these process emissions provide an indicator of processing error.

[0040] In particular, generating at least one spectrogram includes generating a first spectrogram for a first time interval and generating a second spectrogram for a second time interval, the second time interval may overlap with and / or immediately follow the first time interval.

[0041] When spectrograms overlap in time, for example, a first spectrogram may be generated from the spectrum from approximately 0 ms to 500 ms, and a second spectrogram from the spectrum from approximately 400 ms to 900 ms. This overlap may depend on the magnitude of the error and the processing speed, i.e., the time it takes for important features to form across the spectrum. In particular, the time intervals between two or more generated spectrograms are the same length and / or have the same number of spectra, regardless of whether the spectrograms overlap in time. Alternatively, the length of the time intervals may differ.

[0042] In particular, the physical quantities may include at least one of the following: tensile strength, compressive strength, electrical conductivity, keyhole depth, weld depth, gap size between two workpieces joined by a laser processing process, roughness of the cut edge of a workpiece cut by a laser processing process, burr on the cut edge of a workpiece cut by a laser processing process, burr height on the cut edge of a workpiece cut by a laser processing process, steepness of the cutting surface, and perpendicularity of the cut edge of a workpiece cut by a laser processing process. The classification of the workpiece may include classification into error classes, in particular into at least one of gap, offset, penetration and / or weld deficiency, removal defects, cutting quality, and alloy quality.

[0043] Furthermore, classification can also be determined based on these physical quantities. In particular, the classification can indicate whether the workpiece corresponds to a good weld or a poor weld, where a good weld is a welding result that meets a predetermined standard, and a poor weld is a welding result that does not meet a predetermined standard.

[0044] The spectrum can be detected, for example, at a sampling rate between approximately 100 Hz and 100 kHz, preferably between approximately 800 Hz and 10 kHz, and particularly between approximately 900 Hz and 2 kHz. A higher sampling rate leads to a higher temporal resolution of the spectrogram.

[0045] In particular, spectra can be detected in a wavelength range between approximately 100 nm and 1500 nm, preferably between approximately 130 nm and 1300 nm, especially preferably between approximately 150 nm and 1050 nm, and particularly between approximately 340 nm and 850 nm. Most of the physical values ​​or classifications of laser processing processes can be derived from the spectra in that wavelength range.

[0046] The spectrum can be detected with a spectral resolution between approximately 0.1 nm and 1 nm, preferably between approximately 0.2 nm and 0.8 nm, and particularly preferably between approximately 0.4 nm and 0.6 nm. The spectral resolution is primarily due to spectral resolving at or immediately before the detector. The stronger the resolving at the diffraction grating or other dispersive optical element, the higher the spectral resolution that can be achieved, provided the detector allows it.

[0047] The values ​​or predicted values ​​and / or classifications of physical quantities can be determined in real time. Based on this, closed-loop control data and / or control data can be output to the laser processing system that performs the laser processing process. Therefore, the values ​​and / or classifications of physical properties can be used to control the laser processing process in a closed loop, especially if each value is determined during the execution of the laser processing process. For example, the laser processing process can be closed-loop controlled so that the difference between the determined or measured values ​​and target values ​​of the physical properties of the current or subsequent processing results decreases. For example, if the physical property is the weld depth of a workpiece, and the determined value of the weld depth deviates from the target value of the weld depth, the laser processing process can be adapted so that the difference between the determined value of the weld depth and the current target value decreases in subsequent laser processing processes. Closed-loop control of the laser processing process may include adjustments to the focal position, the focal diameter of the laser beam, the laser power, and / or the distance of the laser processing head.

[0048] In particular, the laser processing process can be automatically controlled and / or closed-loop controlled based on predicted values ​​and / or classifications. For example, predicted values ​​of physical quantities can form the basis for controlling or closed-loop controlling the laser output of the processing laser. In particular, this can prevent the occurrence of processing errors. Furthermore, processing errors in the workpiece can also be corrected by closed-loop control and / or control of the laser processing process. Moreover, such methods can produce workpieces with desired physical properties generated by the laser processing process, thereby meeting very high quality standards. In particular, the conditions of the laser processing process can be adjusted based on predicted values ​​and / or classifications.

[0049] As a result, defective workpieces due to processing errors can be reduced or avoided, which is particularly important for very high-quality workpieces from an economic and environmental standpoint. Furthermore, workpieces can be improved by the optimized laser processing process compared to other workpieces manufactured by known laser processing processes; that is, the physical properties of workpieces processed according to the present invention can meet particularly high quality requirements. For example, electrical contacts can be virtually defect-free, have a particularly long service life, and possess particularly excellent conductivity through the optimized laser processing process.

[0050] The data detected on the workpiece, particularly the spectrogram and the predicted values ​​and / or classifications determined therefrom, can be recorded, stored, and attached to the product data sheet, for example, for assurance purposes. This is of particular interest to workpieces of very high quality or those that must meet high safety standards.

[0051] A trained neural network can be adapted to training data through transfer learning. In particular, a neural network can adapt to changes in process conditions, for example, due to a new batch of workpieces with slightly different material compositions. Adaptable neural networks that can be (re)trained with training data through transfer learning are especially flexible, versatile, and user-friendly.

[0052] The requirements for laser processing processes and the resulting workpieces can be diverse and varied. For example, a user might be interested in very specific physical parameters of a workpiece, as this would allow them to map very specific characteristics of the workpiece that they are interested in.

[0053] Transfer learning can be used, for example, to retrain a neural network if it produces incorrect predictions or classifications. In transfer learning, training data can be created from workpieces that have been specially processed for training. Spectra are detected for workpieces processed for training purposes, spectrograms are generated, and these, along with the respective measurements and / or expert classifications (so-called "ground truth" values) of the physical quantities, are used as training data for the neural network. The measurements or expert classifications of the physical quantities are based on the processed workpieces, and destructive methods may also be used as needed.

[0054] Transfer learning can be used to adapt a neural network to a modified situation or a modified laser processing process. Modified situations may include, for example, the workpiece to be processed having a different material, degree of contamination, and / or thickness, or the parameters of the laser processing being changed. In transfer learning, new examples may be added to the training dataset used to train or educate the neural network. Therefore, using a pre-trained neural network configured for transfer learning has the advantage that the system can be quickly adapted to modified situations, particularly modified laser processing processes.

[0055] A neural network can be a CNN, which may include fully connected layers, LSTM ("long short-term memory") layers, and / or at least one GRU ("gated recurrent units") layer. This can improve the performance of the neural network.

[0056] In a further embodiment, a system for analyzing a laser processing process is disclosed. This system can be adapted to carry out the method, and in particular embodiments thereof. The system comprises: at least one sensor or sensor unit, preferably having a sensor field, arranged to detect multiple spectra of process emissions at successive time points; at least one computing unit (also called a controller) arranged to generate at least one spectrogram as an input tensor based on the detected spectra; and a neural network arranged to calculate at least one value or predicted value of a physical quantity or characteristic and / or a classification of the laser processing process as an output tensor based on the input tensor. The neural network may be built into or implemented in the computing unit. Alternatively, the neural network may be provided on a server or in the cloud. In this case, the neural network may be wirelessly connected to the computing unit for data exchange.

[0057] A system for analyzing the laser processing process realizes all the advantages applicable to this method, and in particular to one or more embodiments of this method.

[0058] The process emission spectrum can be detected at least partially coaxially with the beam path of the processing laser used for the laser processing process. This allows for a particularly space-saving and efficient design of the system. As a result, optical elements can be utilized efficiently, and additional optical elements required for non-coaxial beam paths can be omitted.

[0059] The computing unit can be adapted to output closed-loop control data to a laser processing system that determines values ​​in real time and / or performs the laser processing process.

[0060] A sensor or sensor unit may have at least one spectrometer, in particular a MEMS spectrometer. MEMS (Micro-Electro-Mechanical Systems) are particularly cost-effective and space-saving, making them easy to integrate or mount on a laser processing head. MEMS devices can be used as "on-chip" spectrometers. This allows these miniaturized spectrometers to be integrated into the laser material processing head. In particular, a sensor or sensor unit may have two, three, four, five, six, seven, eight, nine, ten, or more MEMS spectrometers.

[0061] The computing unit can be adapted to perform closed-loop control and / or control of the laser processing process based on closed-loop control data and / or control data.

[0062] Such a system can autonomously produce high-quality workpieces, that is, without user intervention. Therefore, it can correct or stop laser processing processes that have resulted in processing errors, ensuring that the workpieces meet the user's individual quality requirements. In particular, it can control various parameters of the laser processing process and / or implement closed-loop control to ensure that the workpieces are inherently free from processing errors.

[0063] In a further embodiment, a laser processing system for processing a workpiece with a processing laser beam comprises a laser processing head for irradiating the workpiece with a processing laser beam, and a system for analyzing the laser processing process according to the present disclosure, particularly embodiments thereof.

[0064] A laser processing system for processing a workpiece also realizes all the advantages applicable to this method, and in particular to one of the embodiments of this method.

[0065] Embodiments of the present invention will be described in detail below with reference to the drawings. While the drawings depict various features of the embodiments, the features are not limited to these embodiments. Rather, all features that are not mutually exclusive can be combined with each other, or features can be omitted from the embodiments unless they are essential for carrying out the present invention. [Brief explanation of the drawing]

[0066] [Figure 1] This figure shows an example spectrum during a laser processing process at a certain time point t. [Figure 2] This figure shows an illustrative spectrogram of a laser processing process. [Figure 3] This is a schematic diagram illustrating the functional principle of a spectrometer according to one embodiment. [Figure 4] This is a schematic diagram of a laser processing head equipped with a spectrometer according to one embodiment. [Figure 5] This is a schematic diagram of an analysis method for a laser processing process according to one embodiment. [Modes for carrying out the invention]

[0067] Figure 1 shows an exemplary spectrum 3 of the laser welding process of aluminum to steel at a certain time t. Intensity I is plotted on the vertical axis, and wavelength λ is plotted on the horizontal axis. The result is the spectrum 3 shown.

[0068] Spectrum 3 was detected by a spectrometer. The spectrometer makes it possible to determine the intensity of light, particularly the intensity of process emissions, as a function of wavelength or frequency. Characteristic lines of process emissions, which are important for metal workpieces and their coatings, as well as for laser processing processes, are not masked by emissions that do not affect quality or regression assignment, which can occur when selectively viewing only a portion of the process emission spectrum. Therefore, this allows for a complete picture and more reliable quality analysis.

[0069] The main portion of spectrum 3 in Figure 1, i.e., the peak intensity of the spectrum, is in the visible spectral wavelength range 2 of light, i.e., between approximately 400 nm and approximately 730 nm.

[0070] Figure 2 is an exemplary spectrogram 5 of a laser processing process over approximately 0.2 seconds. Spectrogram 5 is assembled from a number of individual spectra 3 and plotted against time along the time axis. In other words, individual spectra 3 are placed in spectrogram 5 and can therefore be visualized over time. The formation of an image by combining spectra is achieved by the fact that each line in the image consists of the intensity of a spectrum.

[0071] The spectrum can be recorded, for example, at a rate of approximately 1 kHz per second and arranged in time to generate a spectrogram or image of 1000 lines. The length of the lines is determined by the resolution of the spectrometer. For example, in a wavelength range of approximately 150 nm to 1050 nm, a resolution of approximately 0.5 nm results in a line length of 1800 intensity values. The resolution of individual data points is, for example, 16 bits.

[0072] Figure 3 is a schematic diagram illustrating the operating principle of the spectrometer. A portion of the process emission passes through the gap 9 or slit as incident light 11 and is incident on the reflective concave diffraction grating 8 of the diffraction grating tip 7. The light is spectrally separated by the diffraction grating 8 and irradiated onto the sensor 10, particularly the sensor field.

[0073] The spectrometer 6 shown in Figure 3 can be a MEMS spectrometer, which is particularly cost-effective and easy to integrate into a laser processing head. The development of an "on-chip" spectrometer based on the MEMS spectrometer makes it possible to integrate this technology into a laser material processing head. In particular, a line sensor can be used as a detector or sensor 10, on which spectrally separated light is incident or imaged. Data can be read out from sensor 10 with, for example, a 16-bit resolution. Data rates can be achieved in the range of kHz per spectrum.

[0074] A Hamamatsu MEMS spectrometer (C12666MA) with a spectral resolution of approximately 15 nm in the range of approximately 340 nm to 850 nm is given as an example of spectrometer 6. The use of multiple MEMS spectrometers corresponding to different wavelength ranges can be easily carried out by combining the individually detected spectra and arranging them in a spectrogram. Preferably, the wavelength ranges are directly adjacent to each other or overlap in small regions.

[0075] Next, data or spectra, particularly wavelength-dependent intensities, are read out and combined to form a spectrogram 5, as shown as an example in Figure 2. For example, the first spectrogram 5 can be generated from a spectrum 3 from approximately 0 ms to 500 ms, and the second spectrogram 5 can be generated from a spectrum 3 from approximately 400 ms to 900 ms, and so on. In this case, there is overlap depending on the magnitude of the error and the welding speed, i.e., the time over which important features are formed across the entire spectrum.

[0076] The number of spectral 3s in the spectrogram 5 can be selected depending on the application, for example, the length or duration of a processing process such as welding, and the sufficient exposure of the line sensor within the MEMS device.

[0077] The use of neural networks allows for the classification and / or mapping of the spectrogram 5 according to its error, without the need to know or extract features within the spectrogram 5. For this purpose, the error is not defined for individual spectral 3 or spectrogram 5 data, which represents the so-called ground truth, but the welded workpiece is used to determine the ground truth.

[0078] For the purpose of classifying or regressing input data to a quality class, or mapping it to the value of a physical quantity such as strength or conductivity, the results of processing such as welded joints are subject to measurements such as force measurements or conductivity measurements. This can be used to physically determine the force that causes the weld seam to tear, or the conductivity generated between the joined materials.

[0079] These values ​​can be used as ground truth for training or tuning neural networks, particularly convolutional neural networks. Therefore, the user does not need to know the intensity distribution in individual spectra and its progression over time.

[0080] Spectrogram 5 forms the input tensor of a neural network, particularly a deep neural network, typically a convolutional neural network, and the neural network classifies the spectrogram 5, for example, by the type of error. The neural network may have, for example, an XCeption architecture network.

[0081] To classify welds into typical error classes such as gaps, offsets, insufficient penetration or weld, and poor removal, it is necessary to generate a large amount of training data with these typical errors. Therefore, welds must be generated for each type of error. Strength, conductivity, or other physical quantities must be determined for each weld, depending on which class to classify it into or which physical quantities to map. This provides a clear assignment of data to ground truth. Thus, the complexity of determining key features from spectrograms can be avoided.

[0082] When welding different materials, spectrogram 5 changes depending on the proportion of joining partners in the joining process. Regression to the weld depth value for overlapping joins of different materials is possible by using spectrogram 5 as the input tensor.

[0083] In particular, differences in cutting quality caused by alloy changes in the material can be recognized based on spectrogram 5. Classification by cutting quality is possible by using spectrogram 5 as the input tensor.

[0084] Figure 4 is a schematic diagram of a laser welding head 24 equipped with spectrometers 22 and 23 as an example of a laser processing head according to one embodiment. The beam path of the processing laser beam 16 is irradiated onto the workpiece 12 being processed via a collimating optical system 17, a beam splitter 15, and a focusing optical system 14. During processing, process emissions 13 are generated, such as thermal radiation, plasma radiation, and / or laser radiation reflected from the surface of the workpiece. Some of the process emissions 13 are guided from the surface of the workpiece to one of the two spectrometers 22 and 23, at least partially coaxially or parallel to the beam path of the processing beam 16. The process emissions 13 are split by the beam splitter 20. Some of the process emissions 13 are reflected by the beam splitter 20 to the spectrometer 23 via the imaging optical system 19, which can detect wavelengths between approximately 300 nm and 1050 nm. Another portion of the process emission 13 is transmitted through the imaging optical system 21 and the beam splitter 20 to the spectrometer 22, which can detect wavelengths between approximately 1000 nm and 1400 nm.

[0085] The laser welding head 24 is connected to a computing unit 18, which performs calculation operations, including assembling the spectrum 3 into a spectrogram 5, determining physical quantities and / or classifications, and employing or using a neural network to identify processing errors.

[0086] Figure 5 shows a schematic diagram of a method for analyzing a laser processing process according to one embodiment. In step 110, multiple spectra of process emissions are detected by a spectrometer at consecutive points in time or over a period of time. In other words, the spectrometer detects intensities at different wavelengths at determined points in time or over a determined period of time, where the dataset of intensities and associated wavelengths represents the spectrum. In step 120, a spectrogram is generated based on the detected spectra. The spectrogram is generated by plotting the multiple spectra along a time axis against the time t at which they were detected. In step 130, a neural network is used to determine the values ​​of physical quantities and / or the classification of the laser processing process. The neural network takes the spectrogram as an input tensor and generates the values ​​of physical quantities and / or the classification of the laser processing process as output tensors.

[0087] The neural network is specifically a convolutional neural network, which can consist of, for example, an XCeption network. The input layer can be tuned to match the dimensions of the spectrogram (e.g., 2315 × 500 × 1). The latter dimensions are due to the predicted probabilities for the class being classified, such as gaps, insufficient penetration or welding, and offsets.

[0088] The laser processing processes described herein may include joining or connecting workpieces, or separating or removing materials. Laser processing processes may be, or include, laser cutting processes, laser removal processes, laser welding processes, or laser soldering processes. The results of laser processing processes may include cut, joined, or connected workpieces, i.e., welded, soldered, or cut workpieces. In particular, the results in this case may represent welded or soldered joints between joined workpieces. Welded or soldered joints may be formed by weld seams. In other words, the results in this case may specify weld seams or soldered seams. The results may also represent a portion or area of ​​a welded joint or welded seam. Gaps may occur between workpieces joined by a laser processing process, which affect the welding result. Gaps may be described as the space between two opposing surfaces of workpieces to be joined in the case of butt welding, or as the space between workpieces to be joined in the case of overlap welding. The distance between opposing surfaces of joined workpieces can be called the gap size. An excessively large gap may indicate a processing error in the laser cutting process. In the case of butt welding, the gap is determined pre-processing, i.e., before welding; in the case of overlap welding, the gap is determined using clamping techniques.

[0089] The processing results of a laser processing process may include intermediate results of the laser processing process, i.e., features that exist (or are unique to) the execution of the laser processing process. In particular, the processing results may include a vapor capillary and / or molten pool, also known as a "keyhole." The keyhole depth may be defined as the distance between the bottom of the vapor capillary and the workpiece surface irradiated by the laser beam. The weld depth may be estimated from the keyhole depth.

[0090] The physical property values ​​of the processed material can correspond to the predicted measured values ​​of the physical property. In other words, determining the values ​​of the physical property can be considered a prediction of the measured values ​​of the physical property.

[0091] At least one physical property of the processing result may include at least one of the following: the strength of the welded or soldered joint produced by the laser processing process, particularly the tensile strength, compressive strength, and / or shear strength; the conductivity of the welded or soldered joint produced by the laser processing process; the keyhole depth; the weld depth of the workpiece; the gap size between two workpieces joined by the laser processing process; the roughness of the cut edge of the workpiece cut by the laser processing process; the burr or burr height of the cut edge of the workpiece cut by the laser processing process; the steepness of the cut front; and the perpendicularity of the cut edge of the workpiece cut by the laser processing process. If the keyhole depth or the steepness of the cut front is determined or predicted using a method according to the present invention, another measuring device, such as an optical coherence tomography (OCT) scanner, may be omitted in the laser processing system for performing the laser processing process. On the other hand, determining the value of the tensile strength is particularly relevant to butt-jointed workpieces. The processing results during laser cutting can be described by physical properties such as the roughness of the cut edge, the burr or burr height of the cut edge, or the perpendicularity of the cut edge.

[0092] The values ​​of physical properties can be determined in physical units, for example, in "SI units" (International System of Units). For example, intensity can be expressed in Newtons (N) or Newtons per unit area (N / m²). 2 The weld depth can be determined in μm, the gap size in μm, and the conductivity in Siemens (S). The roughness of the cut edge can be determined, for example, in μm units.

[0093] All discrete values ​​presented herein may deviate from their specifications by up to approximately 10%, and particularly by up to approximately 5-7%. Therefore, these specifications should be understood as approximate values ​​only. [Explanation of symbols]

[0094] 3. Spectrum 5. Spectrogram 6. Sensor or spectrometer device 7 Diffraction grating chips 8 Reflective concave diffraction grating 9 gaps 10 Sensor Field 11 Incident light 12 Workpiece 13 Process Emission 14. Focusing Optical System 15 Beam Splitter 16 Processing beam 17 Collimated Optics 18 computing units 19. Imaging Optical System 20 Beam Splitter 21 Imaging Optical System 22 Spectrometer 23 Spectrometer 24 laser welding heads 110 Detecting multiple spectra with a spectrometer 120 Based on the detected spectrum, generate at least one spectrogram. 130 A trained neural network determines the predicted value of at least one physical quantity and / or the classification of at least one laser processing process.

Claims

1. A method for analyzing a laser processing process, wherein the method is - A step (110) of detecting multiple spectra (3) of process emission (13) at consecutive time points, - A step (120) of generating at least one spectrogram (5) based on the plurality of spectra (3), - A step (130) in which a neural network determines the value of at least one physical quantity and / or determines the classification of at least one of the laser processing processes, Includes, A method characterized in that the neural network receives the spectrogram (5) as an input tensor and outputs the classification of the physical quantity and / or the laser processing process as an output tensor.

2. The method according to claim 1, characterized in that the step of detecting each of the spectra (3) includes the step of detecting the intensity of the process emission (13) as a function of wavelength at each of the time points.

3. The method according to claim 2, characterized in that the detected intensity is raw data.

4. The method according to claim 2 or 3, characterized in that the intensity of each spectrum (3) is detected simultaneously, and / or the detection of each spectrum (3) includes local spectral splitting in the detector.

5. The method according to any one of claims 1 to 3, characterized in that the process emission (13) includes thermal radiation, plasma radiation, and / or laser radiation reflected from the surface of the workpiece (12).

6. The step of generating the at least one spectrogram (5) is, A step of generating a first spectrogram (5) for a first time interval and a second spectrogram (5) for a second time interval, wherein the second time interval overlaps with the first time interval and / or follows immediately after the first time interval, and / or Steps include assembling the aforementioned multiple spectra in a time series, The method according to any one of claims 1 to 3, characterized by including

7. The physical quantity includes at least one of the following: tensile strength, compressive strength, electrical conductivity, keyhole depth, weld depth, gap size between two workpieces (12) joined by the laser processing process, roughness of the cut edge of the workpiece (12) cut by the laser processing process, burr on the cut edge of the workpiece (12) cut by the laser processing process, burr height on the cut edge of the workpiece (12) cut by the laser processing process, steepness of the cut surface, and perpendicularity of the cut edge of the workpiece (12) cut by the laser processing process, and / or The method according to any one of claims 1 to 3, characterized in that the classification of the workpiece (12) corresponds to a classification into error classes and includes at least one of gap, offset, penetration and / or weld deficiency, removal defects, cutting quality and alloy quality.

8. The method according to any one of claims 1 to 3, characterized in that the spectrum (3) is detected at a sampling rate between approximately 100 Hz and 100 kHz, or between approximately 800 Hz and 10 kHz, or between approximately 900 Hz and 2 kHz.

9. The spectrum (3) is detected in a wavelength range between approximately 100 nm and approximately 1500 nm, or between approximately 130 nm and 1300 nm, or between approximately 150 nm and 1050 nm, or between approximately 340 nm and 850 nm, and / or The method according to any one of claims 1 to 3, characterized in that the spectrum (3) is detected with a spectral resolution between approximately 0.1 nm and 1 nm, or between approximately 0.2 nm and 0.8 nm, or between approximately 0.4 nm and 0.6 nm.

10. The method according to any one of claims 1 to 3, characterized in that the predicted value of the physical quantity and / or the classification is determined in real time, and control data is output to the laser processing system that performs the laser processing process based on thereon.

11. The method according to any one of claims 1 to 3, characterized in that the neural network is a pre-trained neural network that can be adapted by training data obtained by transfer learning.

12. A system for analyzing a laser processing process, wherein the system - At least one sensor unit arranged to detect multiple spectra (3) of process emissions (13) at consecutive time points, - At least one computing unit (18) is configured to generate at least one spectrogram (5) as an input tensor based on the plurality of spectra (3), - A neural network configured to output an output tensor of at least one value of a physical quantity and / or a classification of the laser processing process based on the input tensor, A system characterized by comprising the following features.

13. The system according to claim 12, characterized in that the sensor unit has at least one spectrometer or MEMS spectrometer.

14. The system according to claim 12 or 13, characterized in that the at least one computing unit (18) is designed to control the laser processing process based on control data.

15. A laser processing system for processing a workpiece (12) with a processing laser beam, wherein the laser processing system is - A laser processing head (24) for irradiating the workpiece (12) with a processing laser beam (16), - The system according to claim 12 or 13, A laser processing system equipped with the following features.

Images