Calibration of a quality estimator for laser cutting processes.

The calibration of a real-time quality estimator using sensor data and offline measurements with machine learning algorithms addresses the unreliability of existing methods, ensuring accurate and adaptive quality assessment in laser cutting processes.

JP7592213B2Active Publication Date: 2024-11-29BYSTRONIC LASER AG
View PDF 10 Cites 0 Cited by

Patent Information

Application Number
JP2024501103
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2021-07-12
Filing Date
2022-05-30
Publication Date
2024-11-29
Estimated Expiration
2042-05-30

AI Technical Summary

Technical Problem

Existing quality estimation methods for laser cutting processes are susceptible to interference and drift due to aging and wear, and are not individually tailored to each laser cutting machine, workpiece, or cutting environment, leading to unreliable quality assessment.

Method used

A method and system for calibrating a real-time quality estimator using a calibration module that combines sensor data from a first sensor system during cutting with offline quality measurements from a second sensor system, employing machine learning algorithms and neural networks to dynamically adjust the estimator based on workpiece material, thickness, and cutting conditions.

Benefits of technology

Enables reliable, real-time quality estimation that adapts to changing parameters and environmental factors, improving cutting quality by continuously recalibrating the estimator to ensure accurate assessment of cut quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007592213000001
    Figure 0007592213000001
  • Figure 0007592213000002
    Figure 0007592213000002
  • Figure 0007592213000003
    Figure 0007592213000003
Patent Text Reader

Abstract

In one aspect, the invention relates to a calibration module KM for calibrating a real-time estimator ES intended for estimating the quality of a cutting method using a laser cutting machine L, the calibration module KM comprising: - a load interface E for loading a quality estimation result q' of the real-time estimator ES; - a first processor P1 intended for providing a quality measurement result q of a cutting edge of a workpiece, in particular the quality measurement result q can be provided by detecting S21 a measurement signal of the cutting edge of a finished cut workpiece by a measuring device K; - a second processor P2 intended for comparing the loaded quality estimation result q' with the quality measurement result q and based on that result for calculating a calibration data set k for calibrating the real-time estimator ES, the calibration module KM comprising: - an output interface A intended for outputting the calculated calibration data set k.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to the field of industrial laser processing, in particular to the cutting of workpieces made of metal or other materials, for example supplied as metal sheets on the work table of a laser cutting machine.The present invention relates to the calibration of a quality estimator for estimating the cutting quality of a laser cutting method. [Background technology]

[0002] Important aspects of a laser cutting machine are its cutting quality, cutting speed and machine autonomy. If the cutting process can be effectively monitored, the machine can be improved in all three essential criteria. Therefore, in recent years, it has become increasingly important to monitor the cutting process. Today's machines are increasingly equipped with photodiodes, and sometimes even cameras, which allow real-time monitoring of the machining process.

[0003] By monitoring the cutting process, it is also possible to detect or prevent splits, i.e. failure to cut the material. For this purpose, a photodiode that monitors the process is usually attached to the cutting head. However, a reliable detection of the cut quality remains a challenge. With the methods known from the prior art, it is not possible to distinguish, for example, whether the cut parts are cut too roughly with burrs and / or slag attached.

[0004] Subsequent measurements of the cutting edge of the cut parts are also established and currently mainly focus on the cutting quality. For this purpose, imaging devices such as cameras are often used. Using suitable methods and algorithms, the cutting quality can be extracted from the captured image of the cutting edge based on a number of quality parameters such as roughness, the presence or absence of burrs or slag, and other quality criteria. This information can also be used to optimally set the cutting parameters of future cutting operations.

[0005] DE 102019209088 A1 also proposes using mobile phone photos to assess the cutting edge.

[0006] US Patent No. 9529343 (B2) also refers to the currently established prior art, where it is proposed to collect both sensor data generated during the cut and image data of the cutting edge recorded after the cut, combining these to jointly derive a determination of the cutting performance and, if necessary, make recommendations to improve said cutting performance.

[0007] The applicant is also known from WO2020127004(A1), which proposes optimizing the cutting process itself using a deterministic cutting model in a calibration process. To adjust / adapt the cutting model, images of the cutting edge from a camera can be used. Furthermore, the quality of the cutting edge determined based on the model can be compared with the quality of the cutting edge determined using a camera.

[0008] WO2020069889A1 also proposes something similar. A method parameter algorithm is presented that provides optimized method parameters for the cutting process based on material and machine parameters and the desired cutting edge quality. In addition, the achieved edge quality can be recorded by a camera and made available by the algorithm to further optimize the method parameters.

[0009] WO 2021073946(A1) shows, for example, that the monitoring conditions can be defined initially in a configuration phase, so that, so to speak, a processing step is first carried out for setting the monitoring conditions, then a processing step is carried out with monitoring activated, and then a setting step is again entered for determining new monitoring conditions via the process signal.

[0010] Finally, mention should be made of EP Patent No. 3159093(B1), which describes the arrangement of a camera on the cutting head in order to make it possible to observe the process in real time, coaxially along the processing laser during cutting, in order to make it possible to interrupt the cutting process with an error message or to continue it in a modified way (with changed cutting parameters) or to start post-processing of defects.

[0011] The above mentioned documents disclose methods for improving the quality by adjusting the cutting parameters, in this case for example an estimation method may be used to estimate the quality of the cutting process prior to the cutting process in order to make it possible to adapt the cutting parameters if the estimation results output too low quality.

[0012] However, the estimation methods for estimating quality known from the prior art are prone to interferences, which reduces the reliability of the estimation result and thus has a negative impact on the subsequent cutting process and its quality. In particular, the estimation methods are prone to drifts over time in their long-term behavior, possibly due to aging, wear and cracks in sensors and optical components. Thus, the estimation methods have proven to be particularly flawed, and their data base is static. A further drawback is that the quality estimation is not individually adjusted to the respective laser cutting machine, the workpiece to be cut and / or the cutting environment. [Prior art documents] [Patent documents]

[0013] [Patent Document 1] German Patent Application Publication No. 102019209088(A1) [Patent Document 2] US Patent No. 9529343 (B2) [Patent Document 3] International Publication No. 2020127004(A1) [Patent Document 4] International Publication No. 2020069889(A1) [Patent Document 5] International Publication No. 2021073946(A1) [Patent Document 6] European Patent No. 3159093 (B1) [Patent Document 7] European Patent Application Publication No. 2357057(A1)

Non-licensed documents

[0014]

Non-patent document 1

Non-patent document 2

Non-patent document 3

[0015] Based on the above mentioned prior art, the objective of the present invention is to improve the (online) quality estimation, in particular to make the quality estimation more robust against interferences and dynamically changing parameters. In particular, the quality estimation should be configurable based on the currently cut workpiece material, its alloy and / or surface, and / or its thickness. [Means for solving the problem]

[0016] The above object is achieved by a method, a calibration module, a system and a computer program as set forth in the appended claims. Advantageous embodiments, alternative features and their associated advantages can be found in the dependent claims.

[0017] According to a first aspect, the above object is achieved by a method for calibrating a real-time estimator intended for quality estimation of a cutting process using a laser cutting machine, the method comprising the following method steps: - loading a quality estimation result of a first sensor system, in particular of a real-time estimator based on (online) sensors, - providing a quality measurement of the cut edge of the workpiece, which may in particular be provided by detecting a measurement signal of the finished cut workpiece by means of a second sensor system, in particular a measuring device (e.g. an offline camera); - comparing the loaded quality estimation result with the provided quality measurement result, and based on the result, - calculating a calibration data set for calibrating the real-time estimator; may include:

[0018] The terms used in this application are defined as follows:

[0019] Calibration in this context specifically means adjusting the estimator periodically or iteratively according to a preconfigurable scheme with feedback of actual measured quality values. Calibration therefore does not refer to the cutting process, but only to the real-time estimator. Calibration can also be understood as regulation.

[0020] The quality estimation is preferably performed in real time, i.e. during the cutting process, therefore online. The quality estimation is carried out, so to speak, in parallel with the movement of the laser cutting head on the sheet metal or material to be cut. In a variant, however, the quality can also be estimated in a test cut and thus not on parts of a production batch. However, one advantage of the quality estimation proposed here is that it provides a quality estimation continuously in an ongoing cutting process, without calibration or test cuts. The quality measurement is performed offline, i.e. usually after cutting. By "quality measurement" in this document is meant the calculation of the quality of the real, current cutting edge on the basis of measured signals, in particular optical signals (e.g. camera images).

[0021] In a preferred embodiment, the real-time estimator is designed to execute an estimation algorithm, which at least calculates a quality estimation result from the detected sensor signals of at least one sensor. The estimation algorithm itself is known from the prior art. For example, reference is made here to EP 2357057 A1, which describes a method for monitoring the quality of a laser machining process by detecting a signal via an optical sensor such as a photodiode. To estimate the quality, characteristic parameters are calculated from the signal using image processing methods and a reference to a standard cut quality is calculated for the corresponding parameters. Another method is outlined in Santolini, G. et al., Cut Quality Estimation in Industrial Laser Cutting Machines: A Machine Learning Approach, CVPR 2019, where sensors and DNNs (Deep Neural Networks) are used to draw conclusions about the cut quality.

[0022] The estimation algorithm may also be designed according to the specifications set out in European Patent Application No. 20165589.1, the contents of which are incorporated herein by reference.

[0023] The estimation algorithm of the real-time estimator can be designed using machine learning methods. In particular, the real-time estimator can include a model, preferably a convolutional network (CNN), with several hidden layers as network layers, in particular a convolutional layer (CONV) followed by an activation layer (ACT) followed by a pooling layer (POOL). This sequence (CONV, ACT, POOL) can be cascaded several times before one or more fully connected layers and an output layer are interconnected.

[0024] The real-time estimator may preferably also have a temporal architecture in order to process and extract the information content in the time dimension of the process signal, thereby allowing statements to be made about the signal profile over time. In this case, so-called gated recurrent units (GRUs) or long short-term memory networks (LSTMs) may be used in combination with CNNs to map the signal sequences in the DNN and thus allow learning time-dependent features.

[0025] In principle, there are many network structures, among which Convolutional Neural Networks (CNN) and Residual Neural Networks (RNN) are preferably used for the present task with camera images. For image sequences or time signals, temporal relations or temporal features can be learned. Gated Recurrent Units (GRU) and Long Short-Term Memory (LSTM) networks are suitable for this. It is advantageous to use composite networks that combine the above network structures in different layers.

[0026] The method according to the invention, in particular the calibration method including the calculation of the calibration data set, can use machine learning models. For example, a provision can be made for adapting the calibration model and / or the real-time estimator using a behavioral reward learning algorithm, in particular a reinforcement learning algorithm. A particularly preferred alternative embodiment is based on the fact that the calibration model and / or the real-time estimator is adapted to a new material (e.g. a new alloy or other material thickness) by a transfer learning method for transferring a real-time estimator and / or model trained for a material of type A to a material of type B. Depending on which layer of the DNN the real-time estimator retrains the network, different abstract features can be reused. In a further development, the transfer learning method can also be used to transfer the calculated or learned knowledge to other types of processes, in particular cutting process types such as laser melt cutting, laser flame cutting, laser-induced plasma cutting, plasma-assisted melt cutting or plasma-assisted flame cutting.

[0027] As an advantageous extension, the model can be supported by feature values ​​obtained directly from the sensor data using unsupervised learning methods. No quality measurements are required to calculate these supporting parameters. As a result, important process variations in the sensor signal or data can be recognized even when only a few quality measurements are available. Methods such as clustering, autoencoders, dimension reduction techniques such as k-NN / k-nearest neighbors, PCA / principal component analysis can be used.

[0028] The sensor data is preferably optical sensor data, recorded during cutting (online), i.e. during online process monitoring, by a suitable first sensor system, hereinafter referred to as "sensor" (e.g. photodiode, camera, OCT, acoustic sensor). This sensor data is transferred as input data to the estimation algorithm. The term "sensor" as used herein is understood to be a sensor unit and may include multiple separate sensors of the same or different type. The sensor data (digital) or sensor signals (analog) recorded in this way can be combined using sensor fusion techniques. Optional input data (if available) are material type, thickness, cutting process, laser power, cutting gas pressure, gas type, nozzle type / size, feed rate, nozzle spacing, focus position.

[0029] In a preferred embodiment of the invention, the detected sensor data and / or measurement signals are assigned location identification data (location identifiers) in order to make it possible to uniquely address and identify the data set and to provide spatial resolution. This is important for later carrying out a comparison on the data set acquired at the same position of the part, in particular the cutting edge. This is therefore important in order to make it possible to match (pair or assign) measurements and estimates at the same position for the comparison step.

[0030] In a preferred embodiment of the invention, the first sensor system is implemented as an optical sensor, in particular as a camera arranged coaxially in the cutting head. Alternatively or cumulatively, the first sensor system or sensor can also be a sensor for detecting other signal types, such as an acoustic sensor or a thermal sensor system with a pyrometric camera. The first sensor system or sensor transmits a sensor signal.

[0031] In a preferred embodiment of the invention, the second sensor system is implemented as a measuring device, in particular as an optical measuring device (e.g. as a camera) and can be used to record an offline measurement signal at the cutting edge of the workpiece. "Offline" in this specification means a sensory detection after the workpiece is cut or after a partial cutting process, the cutting process being interrupted for the measurement and allowing an optical detection of the cutting edge. The second sensor system or measuring device can be located in a downstream system of the laser cutting machine, such as a sorting system, a conveying system and / or a bending system. However, the measuring device can also be housed in the laser cutting machine itself, for example to measure the cutting edge, preferably intermittently, for cutting the production parts. The measuring device can also be designed for direct measurement of quality signals, as will be explained later (e.g. profilometer). The measuring device can be a multi-module device and can comprise a plurality of separate sensors, detectors, measuring modules of the same or different types. The second sensor system or measuring device delivers a measurement signal.

[0032] The output data of the estimation algorithm contains information about the cut quality currently achieved in the cutting process. The cut quality includes the following characteristics (not exhaustive): roughness, craters, burrs, slag residues, welds, some characteristics are also called quality defects. The features are preferably output in quality classes. For example, three to six quality classes may be output: "very good", "good", "medium", "fair", "poor". The type and number of classes may be configured in the configuration phase. Thus, the output data may be categorical. However, to provide a more detailed estimation result, alternatively or cumulatively, it is possible to specify both the burr height and the roughness using numerical values, with consecutive estimation numbers between 0 and 1. However, currently the output of quality classes is preferred.

[0033] Alternatively or cumulatively, it is possible for the estimation algorithm of the quality estimator to provide direct quantitative characteristics, i.e. burr height in units of measurement, e.g. μm, and / or roughness in units of measurement, which simplifies the subsequent digital comparison with the quality measurement.

[0034] Alternatively or cumulatively, it is possible to provide quality measurement results of the cutting edge of the workpiece after the quality measurement, each quality measurement result being subject to an additional method step, namely a binning method for classifying the quality measurement results into appropriate classes according to the qualitative properties of the estimator, which simplifies the comparison between the measurement and the estimation for categorical values.

[0035] The steps of the method of loading, providing and comparing are preferably performed in the order described. It is particularly preferred that the loading by the sensor and the sensor signal detection are performed online, i.e. during the productive cutting process or part of the production batch, while the provision and detection of the measurement signal, in particular the cutting edge image of the finished cut part, is performed at a later time step, in particular after the completion of the cutting process, by means of the measuring device. Alternatively, the detection of the measurement signal can also be integrated into the cutting process as explained above (interruption of the cutting process when the cutting edge is partially exposed).

[0036] The estimation algorithm is based on input data, preferably continuously provided by the sensor, and the provision of the quality measurement result is based on other input data provided by the measurement device.

[0037] Activation of the measurement device may be triggered automatically by an activation signal from the calibration module, or by manual intervention.

[0038] The comparison of the estimated values ​​and the measured values ​​(both related to quality) and / or the calculation of the calibration data set is preferably event-based. In a first variant, the comparison and / or the calculation (of the calibration data set) is performed continuously. In a second variant, the comparison / calculation is not performed all the time, but only when necessary, in particular when a preconfigured event occurs. The event can be time-based or situation-based. In particular, a threshold value can be configured. As soon as the comparison result shows a deviation above a configured threshold value, the calculation of the calibration data set is preferably triggered immediately and also the calibration of the real-time estimator. However, it can also be configured that the calculation is paused during a certain time phase, i.e. preferably when the deviation falls below the threshold value, and thus when the deviation indicates that the estimated values ​​and the measured values ​​comply to a tolerance range that can also be preconfigured and no adjustment of the estimator is necessary.

[0039] Calculation of the calibration data set may be initiated or triggered by various events, such as the provision of new measurements, the completion of the cutting process, after a timed event, or after a predefined event (e.g., when a particular area of ​​the workpiece sheet is cut or when a deviation exceeds a threshold, or other configurable event). This may be implemented in a trigger component. This saves computational power since the comparison is performed only when necessary. In one embodiment of the present invention, calculation of the calibration data set may be performed only if the real-time estimator requests a calibration.

[0040] Calibration of the real-time estimator may also be event-based and / or triggered when a change in the calibration data occurs.

[0041] The detection of the measurement signal of the cutting edge of the finished cut workpiece may be configurable. Thus, the capture of images may be initiated according to a time or event pattern, such as after a specific time interval. Alternatively or cumulatively, the reading of the sensor (photodiode or camera) may be configured and performed according to a specific time schedule.

[0042] The calculated calibration data set is preferably transferred directly to the real-time estimator and / or to a separate real-time module connected downstream of the real-time estimator in order to recalibrate the real-time estimator and / or the quality estimation result. In an advantageous embodiment of the invention, the calibration data set comprises an offset to be applied to the quality estimation result of the real-time estimator. In the simplest case, the calibration data set comprises new setting values ​​for adapting the real-time estimator. It is thus possible to subject the real-time estimator to continuous readjustment dynamically and also during the cutting process.

[0043] In a further preferred embodiment of the present invention, the calibration data set comprises trained parameters of a calibration model that has been trained using the mutually assigned quality estimates and quality measures.

[0044] The quality measurement results are preferably evaluated by capturing an image of the cutting edge of the part after cutting by the measuring device. The assessment can be performed automatically using image processing methods. Alternatively or cumulatively, machine learning methods can be applied to the recorded images of the cutting edge. Alternatively or cumulatively, the captured images can also be evaluated manually with respect to configurable quality criteria, in that the cutting edge is examined by a user. The operator can then provide his assessment results to the digital processing unit via a user interface. For this purpose, the user can be provided with a specific button. In this case, the quality measurement results can be provided by an input data set via a user interface. This can be implemented in particular in case the image detection and the automatic algorithmic image assessment fail due to an error.

[0045] Alternatively or cumulatively, quality measurement results can likewise be provided without imaging by using measurements in an automated manner by measuring the cut edge to directly measure optically detectable quality parameters such as groove depth, groove frequency, roughness and / or burr height. For this purpose, optical and / or mechanical measuring devices can be used to measure roughness (profilometer) and / or burr (gradiometer). In particular, optical 3D measuring systems can be used that function in particular according to the principle of the shifting focus method. The shifting focus method is an area-based method for optical high-resolution 3D surface measurements in the micro- and nano-range. The technique combines the functions of a roughness measuring device with those of a coordinate measuring system. Roughness is measured according to the DIN EN ISO4287 standard. Different roughness parameters are distinguished. For laser-cut surfaces, the Ra average profile roughness, the Rz average profile roughness and the Sa average surface roughness are suitable.

[0046] In one variant, the method or calibration module is designed to detect discoloration of the cutting edge (eg, an oxide layer), which is also detected by optical image detection.

[0047] The laser cutting machine is preferably a system used in the field of industrial production. The laser cutting machine may, for example, include a fiber laser or a CO2 laser with a suitable wavelength. Fiber lasers usually operate at near-infrared wavelengths, typically 1070 nm. Their current power is at least 2 kW. Currently, the maximum power used in industrial applications is about 30 kW, and this trend is increasing. The beam parameter product (BPP) is typically about 3-4 mm mrad, and at maximum power it is slightly higher. CO2 lasers can operate at a wavelength of 10.6 μm and a beam parameter product of about 10 mm mrad with a power of 2-10 kW.

[0048] In a preferred embodiment, the operation of the sensor and the operation of the measuring device are not performed synchronously, in particular according to different schemes. In particular, the sensor is continuously operating, and the measuring device can be switched on and / or operated sporadically only when necessary. Furthermore, the two image detection devices (sensor, measuring device) are preferably separate devices. The mentioned image capture devices can be located in different positions and / or on different devices or components. The sensor is typically located in the laser machine, preferably in the laser cutting head, and the measuring device is located separately, preferably in a mobile terminal device and / or on a handling system such as a sorting system. Both image detection devices are designed to have suitable data interfaces to the calibration module and to the real-time estimator.

[0049] In a preferred embodiment, the sensor may be designed as an optical sensor, in particular as an in-process camera and / or as a photodiode unit and / or as an acoustic sensor. The sensor is the basis and provides an input signal to a real-time estimator (hereinafter also called "estimator"), for example in the form of a photodiode signal.

[0050] The measuring device may include one or more sensors, preferably optical sensors. The measuring device may be at least one camera, for example a camera that is attached or assigned to a mobile terminal device and / or an automation system (sorting system). The measuring device is a device that is used for quality measurement and serves to provide a quality measurement result.

[0051] As explained above, the sensor and the measuring device are preferably two separate and distinguishable units. In particular, the measuring device may be an optical detection device in a mobile device, such as a camera-equipped smartphone, carried by a user. The smartphone or equivalent mobile terminal (e.g., a tablet) may have an app installed that automatically performs a quality assessment based on captured images.

[0052] In an alternative embodiment, the measurement signals of the measuring device may also be recorded intermittently for the cutting process. For this purpose, after the cutting process or after a partial cutting process that at least partially exposes the cutting edge, a (preferably automatic) nozzle change may be performed, which includes a deflection mirror arranged on the nozzle, in order to capture the exposed cutting edge from the side when viewed from above. In this case, the measurement signals are recorded using a measuring device configured on the laser cutting machine, in particular on the cutting head. In the latter case, the device (sensor) that observes the processing zone during cutting and detects the sensor signal may be the same as the device (measuring device) that subsequently detects the cutting edge of the cut workpiece. The sensor is preferably intended for image signal detection. For this purpose, the image signal may be captured either in the productive cut or in the calibration cut.

[0053] In a further preferred embodiment, the detection of the measurement signal, i.e. the quality measurement signal of the cutting edge, can be configured with respect to execution time and / or execution type. In particular, the time, frequency or number of recordings by the mobile terminal (e.g. the camera of a mobile phone) can be configured according to the respective application. For example, if an error-prone component needs to be cut, it may be necessary to measure and record more frequently and vice versa. Furthermore, if necessary, the detection parameters such as exposure time, focus, etc. can also be changed. For example, the detection of the signal from the at least one image and / or the provision of the quality measurement result can be controlled sporadically, at configurable intervals or according to a configurable pattern or based on an event (e.g. from the status of a sorting or mechanical handling system).

[0054] In another preferred embodiment, the calibration data set is calculated only when a deviation above a preconfigurable threshold is detected, which has the advantage that no calibration data has to be transmitted unnecessarily and no calibration procedure has to be performed on the estimator.

[0055] In another advantageous embodiment, metadata is assigned to the detected sensor signals and / or measurement signals and / or stored in an assigned manner. The metadata may relate, for example, to the detection time, part number, but also to the workpiece material (alloy, thickness, etc.). Useful information may be derived from the metadata of the recorded sensor signals. For example, patterns may be recognized and compared to the cutting plan. For example, when estimating, it may be automatically taken into account that the quality changes on the trajectory without changing the cutting parameters, e.g. a narrow radius results in a lower quality than a straight cut. This allows, for example, to operate the sensor more frequently in these areas. Alternatively or cumulatively, when calculating the calibration data set, metadata related to the recorded measurement signals of the measuring device (e.g. part, workpiece identification, and / or time information, etc.) may also be taken into account.

[0056] In principle, metadata can also be loaded from other sources (e.g. from the cutting plan), made available to the calibration module as input data and included in the calculation of the calibration data set. In the simplest case, a table can be created with all calibration data sets calculated for different material alloys. If the same material alloy is cut again later, the corresponding calibration data set can be loaded before the start of the cut, which means that the real-time estimator is calibrated as best as possible right from the start of the cut.

[0057] In further developments and advantageous embodiments, the calibration data set is calculated specifically for the material and / or cutting process. The material-specific information may include information on the type of material (e.g. alloys such as St37 or S235 as cheap steel sheets or Ruukki laser as high-quality laser sheets), the thickness of the material, and / or the surface properties (e.g. uncoated or colored, or sandblasted or scaled). The cutting process-specific information may include information on the cutting process, such as melt cutting, flame cutting, laser-induced plasma cutting, plasma-assisted melt cutting or plasma-assisted flame cutting. However, cutting process-specific also means that the important process parameters of the cutting machine are included in the calculation of the calibration data set. The inclusion of the process parameters is particularly advantageous when a calibration model is used, in particular when they can be used directly as additional input both when training the calibration model and when training its real-time application downstream of the real-time quality estimator. The important process parameters are the laser power, the feed rate, the focus position, the cutting gas pressure, and the nozzle spacing.

[0058] In another advantageous embodiment, a calibration model is used to calculate the calibration data set, and in the simplest case, the calibration model is preferably trained using quality estimation results and quality measurement results assigned to each other to calculate a calibrated quality estimation result for a current quality estimation result.

[0059] As a model trained using a supervised learning method (herein "calibration model"), a model common in artificial intelligence or machine learning is preferably used. The calibration model can be trained by additionally using as training data the stored intermediate results of the real-time estimator, in particular the intermediate results of the real-time estimator in the form of activation values ​​of neurons of a convolutional neural network (CNN) arranged before the output layer.

[0060] The models may include artificial neural networks (ANN), MLP (multi-layer perceptron), regression models, regression or classification trees, support vector machines (SVM), random forests, gradient boosted trees, or may be combined from the aforementioned models (using ensemble learning methods). The input data and training data are not image data, but two independent variables, namely estimates and measurements. The output data include calibrated estimates and / or calibration data sets. Basically, there are various implementation options. MLP (fully connected) may be used equivalently to regression models, but the latter has an advantage from an implementation point of view, since it is simpler. It should be noted that variants of the calibration model may also be used. In variants, intermediate results of the real-time estimator (activation values ​​by the CNN layer) are also used. Herein, MLP and / or flexible models (SVM, random forest, etc.) should be used.

[0061] The model can include a regression model, preferably a quadratic linear regression model, which is very easy to use with a small number of input variables (up to about 10) and can also map nonlinearities and interactions between two parameters. Here, we use the ordinary least squares (OLS) method, which does not require iteration, so that training requires little computational power.

[0062] A flexible calibration model such as a support vector machine (SVM) is preferred. Alternatively, models fitted using Random Forests (multiple decision trees), or XGboost (gradient boosted trees) can also be used.

[0063] Preferably, ensemble learning techniques may be implemented, in which several models are trained and evaluated simultaneously at run time by averaging output values ​​(in the case of regression) or by taking a majority vote (in the case of classification) to achieve increased accuracy or improved prediction.

[0064] The model may also be in the form of or include a statistical model.

[0065] The calibration model is preferably trained to calculate a calibrated quality estimate from the quality estimate initially provided by the real-time estimator. The calibration model may be trained by storing intermediate results of the real-time estimator together with the assigned quality measurement result data when the training data or input data is used. The intermediate results of the real-time estimator include in particular the features of the CNN extracted by the learned, i.e. parameterized, filter (kernel), i.e. the activation values ​​of the neurons from the convolutional neural network layer (CONV layer) onwards before the layer of the classifier (fully connected output layer) is placed. The filter kernel is a parameterized kernel of the CNN and is usually immutable during run-time or online operation, i.e. during disconnected operation while the first sensor technology is continuously supplying real-time data and the real-time estimator estimates the quality from this real-time data, while the activation values ​​(of the corresponding CONV layer) represent values ​​calculated in online mode from the current real-time image data with the filter kernel (of the corresponding CONV layer). However, the filter kernel may be adapted by retraining the real-time estimator, for example by transfer learning or by action-reward learning, in particular by reinforcement learning. With new materials, sheet thicknesses, and / or new cutting processes, more complex transfer learning methods allow learning of new correlations between process sensor signals and production-critical quality attributes during ongoing production, thus improving the cognition and performance of the online quality estimator.

[0066] The input data of the real-time estimator are sensor signals, in particular process sensors, i.e. mainly in-process cameras and / or photodiodes. Alternatively or cumulatively, other sensors such as light field cameras and / or acoustic sensors may be used. The output of the real-time estimator is the calculated quality estimation result.

[0067] In principle, one advantage of the method presented herein is that the method can be used without performing a calibration cut. A calibration cut can be optional and is not required. The real-time estimator and the method steps, in particular the loading, providing, and comparison, can be optionally performed after the calibration segment is performed. For example, the detection of the signal of the cutting edge can be performed directly on the calibration cut or without the calibration cut as part of the execution of the cutting plan.

[0068] The detection of the measurement signal of the image generator, preferably an optical measurement signal, can also take place intermittently via a deflection mirror and / or in particular relative to the cutting process (online). The detection of the measurement signal can also preferably take place intermittently. This has the advantage that it is not necessary to measure the edge outside the cutting machine, since the edge is already measured intermittently during the cutting process. For this purpose, for example, a calibration cut can be made at the edge of the metal sheet, or an image of the cut edge can be made directly by a sensor (first sensor system) in the area where the inner part has fallen through the cutting grid, using a deflection mirror.

[0069] In the above, the achievement of the object has been described in terms of a method. The features, advantages or alternative embodiments so described may also be applied to other claimed subject matter, and vice versa. In other words, the present claims (e.g. directed to a system, or a calibration module, or a computer program) may be further developed by using the features described and / or claimed in relation to a method. Thus, corresponding functional features of a method (e.g. a load) are formed by corresponding modules, in particular hardware modules or microprocessor modules, of a system or product (e.g. a load interface), and vice versa.

[0070] According to a further aspect, the invention relates to a calibration module for calibrating a real-time estimator intended for estimating the quality of a cutting method using a laser cutting machine, the calibration module comprising: - a load interface for loading the quality estimation results of the real-time estimator; - a first processor intended to provide a quality measurement result of a cutting edge of a workpiece, in particular the quality measurement result can be provided by detecting a measurement signal of the cutting edge of a finished cut workpiece by a measuring device; Equipped with - a second processor is intended to compare the loaded quality estimation results with the quality measurement results and, based on the results, in particular on the detection of deviations, A method for calculating a calibration data set for calibrating a real-time estimator, The calibration module is - an output interface (A) intended to output the calculated calibration data set (k); Equipped with.

[0071] In a further aspect, the invention relates to a system with the above-described real-time estimator and calibration module, used in a laser cutting machine. The real-time estimator and the calibration module can be designed as a distributed system and do not have to be implemented in the same structural unit. This allows for a great deal of flexibility, and computationally intensive processes such as the calculation of the calibration data set and / or the calibration of the real-time estimator can be outsourced to a powerful server. In particular, the first processor, which distributes the quality measurement results, as well as the second processor, which compares and calculates the calibration data set, can run locally on a computer on the machine, or these processors can also run as network services, in particular as cloud-based services. This means that computationally intensive steps, such as training of large calibration models or a complete retraining of the online quality estimator, do not have to be performed locally on the machine. Another advantage of a cloud-based solution is that the training data is collected centrally, allowing the production systems to potentially mutually benefit through a larger variety of materials to be cut.

[0072] In a preferred embodiment, the system comprises or is in data communication with a sensor, which may preferably be designed as an in-process sensor, for example a camera.

[0073] The calibration data set is used to calibrate or recalibrate the real-time estimator. A simple recalibration is simply to add an appropriate offset to the real-time quality estimate. A difference between the measured or actually determined value and the estimated value is formed for each quality parameter, such as roughness or burr. If several cutting edges are photographed, all difference values ​​of the same characteristic value are averaged. For the next cutting job, the averaged difference is added as an offset to the characteristic value of the real-time quality estimate. This corrects previous deviations and the cutting quality can be estimated more accurately in real time. The advantage of this simple method is that only the characteristic value of the real-time estimate needs to be stored for each photographed cutting edge and no complex calculations need to be performed to correct the real-time estimate.

[0074] Another method for realignment realizes that, overall, for each photographed cut edge, all determined quality parameters or measured quality measurement results are used as labels (Y values) and all estimated values ​​or previously stored quality estimation results of real-time quality estimation are used as explanatory values ​​(X values) as input values ​​for training a calibration model. Thus, a calibration model is trained with the above data for realignment. The result of the training is a parameter representation of the calibration model optimally adapted to the cut sheet metal.

[0075] For the next cutting job, the real-time quality estimates of the sensor are corrected in real-time by the already trained calibration model. The advantage of this further development is that the non-linear relationship between the original real-time estimates is mapped to the calibration model, making the realignment more accurate.

[0076] A further method for retuning consists in further using the stored intermediate results of the real-time quality estimation as input variables for training the calibration model. In case of using a deep neural network (DNN) as a real-time estimator of the cutting quality, the interim results of the real-time quality estimation are for example the activation values ​​of the neurons after a layer of a convolutional neural network (CNN), which is in front of the available output layer / classifier. These intermediate results contain very strongly compressed information on the quality parameters compared to the original sensor data. Possible classifiers include conventional neural networks composed of layers that are all connected (fully connected). The advantage of this method is that the input data for training the calibration model are high-dimensional and the retuning can be performed more precisely. Moreover, having access to the intermediate results of the real-time quality estimation makes it possible to achieve a better generalization of the retuning even in the case of small changes in the material composition and the sheet thickness.

[0077] An alternative method for post-adjustment is that the real-time quality estimator is directly post-trained without using a separate calibration model. For this purpose, the quality characteristic values ​​identified in the field measurements, i.e. the quality measurement results, are used as labels, and the assigned and stored original sensor data are used entirely as input data for post-training. As a result, good calibration can be achieved even when using DNNs as real-time quality estimators with relatively little new data.

[0078] Another advantage of the last two methods is that they can be used to implement transfer learning. In artificial intelligence, transfer learning refers to the task of transferring learned knowledge from a specific domain to another. The learned knowledge is initially available in the real-time quality estimator, which may already be specialized by previous retraining, and therefore can be applied to new materials and sheet thicknesses, as well as types of cutting processes (laser melt cutting, laser flame cutting, laser-induced plasma cutting, plasma-assisted melt cutting, plasma-assisted flame cutting) with the same method if sufficient new data is available.

[0079] It is particularly advantageous if the use of new materials allows the on-site quality measurement (using a measuring device) to be re-adjusted by the operator depending on the situation, so that the quality characteristics desired by the operator can be determined in the desired format.

[0080] A further development of the above-mentioned method for readjustment is that for each cutting material and sheet thickness, and optionally other distinguishing features (material composition or alloy, surface, cutting process settings), correction values, or parameterized representations of the calibration model, or retrained real-time quality estimators are saved. For the next cutting job, depending on the currently loaded material, sheet thickness, etc., the appropriate correction values, the appropriate parameterized representations of the calibration model, or the appropriate post-trained real-time quality estimators are reloaded. As a result, the real-time quality estimation can be performed very precisely for a specific material, and any loss in quality is reliably detected from the start of a new cutting job.

[0081] A further development of the above method for recalibrating the real-time cutting quality estimation consists in implementing the recalibration process as a continuous action-reward process (action-reward). This means that instead of completely retraining the entire real-time quality estimator or the calibration model, a model adaptation of the real-time quality estimator is performed by reinforcement learning for each in-situ measurement of the cutting quality. In this further development, this model adaptation advantageously directly influences the parameters of the DNN of the real-time estimator. The reinforcement learning algorithm can then adjust many parameters of the DNN to allow a better estimation. This makes repeated retraining unnecessary, which is advantageous for in-situ measurements at short intervals, without having to give up more complex methods.

[0082] Thus, using the described method, it is possible to measure the cut quality online or in real time with reliable accuracy even under adverse production conditions with variable influencing factors. With new materials, sheet thicknesses and cutting processes, more complex transfer learning methods allow learning of new correlations between process sensor signals and production-critical quality characteristics even during ongoing production, thus improving the perception and performance of real-time quality estimation.

[0083] If the measuring device is not mounted or is placed in an unfavourable position, the operator can also manually hold the cut part in front of the camera or even in front of the sensor.

[0084] If the determined cut quality is not satisfactory, the cutting parameters (laser power, laser light focal position, cutting gas pressure, etc.) can be modified to improve the cut quality. If real-time quality estimation is implemented, the change in cut quality due to the change in cutting parameters can be determined immediately. If real-time quality estimation is not implemented, the change in cut quality due to the change in cutting parameters can only be seen in the next cutting plan. The control (in "closed loop") of the aforementioned cutting parameters using the determined cut quality does not require human intervention.

[0085] In principle, the term "offline" or offline quality determination should be understood to be a quality measurement or determination that is not performed during cutting but is performed on the finished cut part, as opposed to an online quality estimation that occurs during the cutting process.

[0086] In a variant, the offline quality determination can also be performed on the cutting table instead of using the handling system, as follows: A calibration cut (rectangular hole) is made and the cut quality is determined using the cutting head camera, i.e. the in-process camera, using a special nozzle with a 45° mirror (similar to a dentist's mirror) in conjunction with a nozzle changer. Burrs in the area of ​​the lower edge can also be measured with the mirror. The advantage of this variant is that this can be done at the beginning and / or during the cutting process, thus optimizing the part quality immediately. Under certain circumstances, a rectangular hole needs to be removed from the sheet metal using a device or waste crushing is used, so that the notch is actually always free for inserting the mirror.

[0087] Another variant, which allows to carry out offline quality determination (measuring device) directly on the cutting table of the laser cutting machine instead of using a removal robot or a sorting system, can be implemented to create a calibrated cut at the edge of the sheet. The cut-out area falls off. The exposed cutting edge is photographed from the side with a measuring device, for example a camera integrated in the machine. This makes it possible to determine the cutting quality from the image immediately before cutting or several times during cutting. The determined offline cutting quality can then be used to update a real-time estimator of the cutting quality or to directly adapt the cutting parameters, i.e. to optimize the cutting quality. The advantage of this variant is that the image quality is always the same and no special nozzles and nozzle changers are required. Furthermore, calibration can be carried out at the beginning and / or during the cutting process and thus the part quality can be directly optimized.

[0088] If certain material properties (such as chemical composition, rolling direction, cold / hot rolling, surface structure and texture, etc.) are found to lead to characteristic real-time estimates and offline quality measurements, this correlation can be used to estimate said material properties. These can also be passed to a downstream system, such as a bending machine. With this information, the bending machine can, for example, calculate the elastic restoring force of the material and correctly bend the first part (by correctly selecting the required overbend).

[0089] Another way of achieving this object is to provide a computer program having a computer program code for performing all the method steps of the methods described in more detail above, when the computer program is run on a computer or loaded into a processor of a computer. It is also possible for the computer program to be stored on a computer readable medium. In a preferred embodiment of the invention, the computer program performs the method described above and can optionally perform the further developments mentioned above.

[0090] In the following detailed description of the drawings, non-limiting exemplary embodiments, together with their features and advantages, are described with reference to the drawings. [Brief description of the drawings]

[0091] [Figure 1] FIG. 1 shows an exemplary sketch of a cutting system with an assigned handling system. [Diagram 2] FIG. 13 is a further view of the cutting system with an integrated camera system acting as a measuring device. [Diagram 3] FIG. 1 shows photographs of two cut samples with different quality cut edges. [Figure 4] Further views of cut samples of different quality. [Diagram 5]FIG. 2 is a schematic diagram of a laser processing head having a process monitoring unit. [Figure 6] 1 is a schematic diagram of the component architecture according to a preferred embodiment of the calibration system of the present invention; [Figure 7] FIG. 1 is a schematic diagram of a simple calibration system. [Figure 8] 2 is a flow diagram according to a preferred embodiment of the present invention. [Figure 9] FIG. 2 is a schematic diagram of a system according to a further preferred embodiment of the present invention. [Figure 10] FIG. 1 is a schematic diagram of the interaction of a calibration model and a real-time estimator. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0092] The invention relates to a calibration tool or a calibration module and a calibration method for calibrating a real-time estimator ES for estimating the quality of the cutting result of a laser cutting machine L.

[0093] FIG. 1 shows a schematic representation of a laser cutting system with a shuttle table W intended to feed and remove workpieces into and out of the laser cutting machine L. The system also includes a removal robot R and a parts sorter T. The parts sorter T already has a camera with which the cutting quality can be estimated in addition to the parts recognition. The system can be equipped with two conveyor belts 23, 24. The shuttle table W is used to hold the workpieces. The system may also include an unloading pallet 27. Reference 21 denotes a sorting bridge with at least one conveyor belt. Reference 22 refers to a control panel containing a user interface. An unloading robot 20 is responsible for unloading the parts from the shuttle table W after they have been cut.

[0094] 2 shows a similar system with a removal robot R including a camera system K that is also used to capture cut surface images. The unloading robot 20 may be equipped with a gripping arm, in particular a suction cup, to pick up the cut parts 33 to be measured and sorted. The system may include a parts chute 31 that may be located on the sorting bridge.

[0095] A typical process is described below.

[0096] The cutting plan for the metal sheet is loaded. The machine control system identifies a suitable first straight cut from the in-situ camera recording of the cutting edge (measurement signals from the measuring device). A straight cut of a few cm in length is suitable so that the quality parameters can be properly averaged. The cutting process should also be stationary, i.e. the cutting head should not accelerate along the selected route. When this straight cut is performed, the process monitoring signals, which also include the camera recording of the processing zone by the sensor D, and the cut quality q' estimated therefrom (in particular the burr height and roughness) are stored. Meta information on the sensor signals, i.e. the coordinates and time of the selected cut and the consecutive part numbers, as well as the provisional results of the real-time cut quality estimation described below, are also saved.

[0097] Once the cutting plan is loaded, it can be determined which parts or which cutting edges are suitable for offline measurement. The cutting edges should ideally meet the following criteria and the selection parameters are set accordingly: - Cutting edges of parts that are not too heavy and easily replaceable - Straight cut edges of at least 1 cm length so that the quality parameters can be properly averaged - cutting edge where a steady cut is made, i.e. no acceleration of the cutting head occurs

[0098] For each cutting plan that meets the above criteria, the following parameters may be selected from all cutting edge zones: - Select in such a way that offline measurements (measurements using a measuring device2) do not cause time delays / loss of productivity - Selection in such a way that offline measurement leads to a maximum productivity loss of xy%, whereby the xy value can be set

[0099] Depending on the selected xy% values, the selected cutting edge areas are evenly distributed throughout the cutting plan.

[0100] In a preferred embodiment of the invention, poor measurements from a sensor (online) can also be compensated for using measurements from a measuring device or camera (offline). If the quality estimate of measurement 1 (using sensor D) indicates poor quality, this measurement 1 is saved and the part / cutting is then measured using sensor 2 (measuring device K). This gives data from or according to poor cuts in measurement 1, and the calibration model can be trained over a wider quality range, which improves the accuracy of the calibration. As a result, the rate of false positives (false positives: measurement 1 shows bad, measurement 2 shows good) can be reduced.

[0101] Once all parts have been cut across the metal sheet, the cut metal sheet is transported to the machine and transported to a sorting table using an automatic shuttle table. At the sorting table, a robotic arm and a camera K are present. The robotic arm sorts the cut sheet metal and stacks the different pieces in appropriate bins (or pallets). When a suitable cut edge is identified for on-site measurement and a part whose coordinates are known is grabbed by the robotic arm, the selected cut edge is fed to the camera K before stacking. The cut edge is then photographed in a suitable manner, for example from various directions. The cut quality can be determined or measured from the photograph in a known manner.

[0102] Now, a real-time quality estimate based on the detected sensor signal can be compared with an (in situ) quality measurement based on the measurement signal. If the two quality values ​​deviate from each other only within an acceptable tolerance, typically <10% (threshold value can be parameterized), the real-time estimate does not need to be readjusted. If the deviation is larger, the real-time estimate should be calibrated in a suitable way.

[0103] Figures 3 and 4 show cutting edges with different cutting qualities. The respective qualities can be clearly distinguished. Thus, in Figure 3, the left side is clearly less rough than the right side. Figure 4 shows pictures of several cutting samples with different cutting qualities. In particular, on the left side a strong formation of ridges is evident, while in the central image only slight ridges are present. The right image shows the cutting edge from below, whereby the formation of burrs is again clearly visible. Images from the side (left and central) as well as from below (right) can be used for the offline quality measurement. Other perspectives are also possible. Samples of known quality (e.g. the left and / or right edge) can also be used to calibrate the offline quality measurement based on the camera images.

[0104] 5 shows an exemplary cutting head 1 with an exemplary sensor system 5 in the form of a coaxial camera for performing first measurements leading to a quality estimation result of the real-time estimator ES or for acquiring sensor signals. The exemplary cutting head also shows a focusing lens 2 for a working laser beam 6, which is led to the cutting head via a conveying fiber 7, a dichroic deflection mirror 3 for separating a process observation beam path 4 for process observation via a process observation unit / camera 5, diaphragms and / or filters for process observation optimization, as well as a unit 9, by means of which the cutting parameters can be adapted, shown as an example using an automatic focusing lens adjustment. The workpiece is given the reference number 10.

[0105] A calibration system according to a preferred embodiment is shown in figure 6. The calibration module KM is the focus of the invention and is used to calibrate the real-time estimator ES.

[0106] As shown in Figure 6, there are two monitoring processes. 1. Online process during disconnection: In this process, the first sensor system with the sensor or detector D is active (left side of Fig. 6). 2. Offline process after disconnection: In this process the second sensor system with the measuring device, in particular the camera K, is active (right side of FIG. 6).

[0107] Sensor signals from a sensor (online) are fed to a real-time estimator ES so that the real-time estimator ES can output a quality estimation result. Measurement signals from a measuring device, in particular a camera K, are recorded and calculated by an algorithm to calculate a quality measurement result q.

[0108] In particular, the calibration model is evaluated or applied or assessed online in the calculation cycle of the real-time estimator ES using the quality estimation result q' and the calibration dataset k as inputs (and any intermediate results). The calibration model is trained in the calibration module, which is marked in Fig. 6 by the box "Calculation of the calibration dataset".

[0109] As shown in FIG. 6, an optical sensor system (e.g., a camera) does not necessarily have to be used as the sensor (online, first sensor system) and / or the measuring device (offline, second sensor system). Alternatively or cumulatively, other sensor types can be used, such as acoustic sensors, profilometers, etc. Another advantage is seen in that the first and second sensors do not necessarily have to be identical. It is therefore possible for the first sensor to be a module made up of various light detectors (cameras, photodiodes), optimized for use as a fast online sensor. On the other hand, the second sensor with the measuring device is optimized for recording the offline cutting quality, for example using a high-resolution camera. The advantage of a camera as the second sensor is that the cutting surface can be recorded quickly, which is a requirement when unloading, for example by a robot. Nevertheless, an optical measuring device for detecting the surface topography or a mechanical measuring device for measuring roughness (profilometer) or another mechanical measuring system can also be used as the second sensor system. Both detected signal types are converted by an algorithm into results, i.e. quality estimation results and quality measurement results, which can be compared again.

[0110] As shown in Fig. 6, the calculated calibration data set k may also be used to evaluate the calibration model, which serves to provide a calibrated quality estimation result q''.

[0111] Fig. 7 again shows diagrammatically the interaction between quality estimation results and quality measurements. The calibration module KM calculates a calibration data set k from the loaded quality estimation results and quality measurements. In the simplest case, this can be an offset applied to the estimate as a calibration. At the next cut, the calibrated estimator can already be applied to provide a more accurate estimate.

[0112] Fig. 8 shows a flow diagram according to a preferred embodiment of the invention for calibrating a real-time estimator ES according to a preferred embodiment of the invention. After the start of the method, in step S1, a quality estimation result q' of the real-time estimator ES is loaded, which has been calculated and stored during the cutting from a sensor signal of a sensor or detector D of a particular cutting section, typically from video data of an in-process camera. For this purpose, the sensor signal is recorded, and in step S11, a quality estimation result is calculated and stored therefrom using an estimation algorithm. In step S2, a quality measurement result q is provided. For this purpose, in step S21, in response to a trigger signal that may be output by the calibration module, a measurement signal is recorded (by the camera K), for example an image or an image sequence of the cutting edge of the workpiece, which provides the basis for the calculation of the quality measurement result q. Alternatively, the recording by the camera can also be omitted if the user manually enters his input on the user interface 22, as will be further explained below in connection with Fig. 9. In step S3, the two instances of the quality data set, i.e. the quality estimation result q' and the quality measurement result q, can be compared for consistency. If there is a deviation or if the deviation exceeds a certain tolerance range, a calibration data set k can be calculated and output in step S4 in order to calibrate or readjust the real-time estimator ES in step S5. This can be done by using appropriate setting parameters of the estimation algorithm. The method can be applied intermittently throughout the cutting process, as shown in FIG. 7, with an arrow branching back from step S5 to step S1. However, it is preferable that a pre-configuration is possible to define the duration and / or frequency of the calibration procedure. The trigger signal or trigger can also be parameterized. It is also possible that the estimation for a certain zone of the cutting process (e.g. the workpiece) remains constant (so that an estimation result for this zone needs to be provided only once), but several measurement signals, for example for different quality parameters, are provided, with a branching back from step S5 to step S2.

[0113] 9 shows a schematic diagram according to an embodiment variant of another possible implementation of the calibration module KM for a real-time estimator ES, which is designed for online, i.e. during the cutting process or in real time, preferably ongoing, continuous quality estimation of the signal b D Preferably, an optical signal is continuously fed to the real-time estimator ES from a sensor D, in particular an in-process camera, which may be mounted, for example, on the side of, coaxial with or within the cutting head and / or a photodiode. D are processed into a quality estimation result q', which, as the name suggests, is intended to estimate the expected quality of the current cutting process. However, due to changing laser conditions (wear, aging, etc.) and different workpiece materials as well as other influencing factors, the quality may be inaccurately estimated, which may lead to unfavourable settings and thus to an insufficient quality in practice. This can be avoided by the present invention, where the real-time estimator ES is subject to calibration or readjustment based on the actually measured cutting quality. For this purpose, a photograph b of the selected cut part is taken. K In the configuration phase, it can be set and configured, for example, how many parts and which parts should serve as the basis for recording, depending on the cutting plan. The measuring device, or in particular the camera K, can be located on the laser cutting machine L, separately from the laser cutting machine, or on a follow-up system (a sorting robot as shown in FIG. 2). The camera K takes pictures b according to preconfigurable events. KThis may occur sporadically. The camera K may also be arranged on a mobile device such as a smartphone or tablet. In FIG. 9, a load interface E of the calibration module KM for loading the quality estimation result q' of the real-time estimator ES is depicted. A deflection mirror, which may optionally be arranged on the laser head and is used for online process monitoring, is identified by the reference U. For this purpose, after the partial cutting process, which at least partially exposes the cutting edge, a deflection mirror, preferably arranged on the nozzle, may be used to detect the exposed cutting edge from the side while looking at it from above. In this case, the measurement signal is recorded using a measuring device arranged on the laser cutting machine, in particular on the cutting head. In the latter case, the detector D (sensor), which observes the processing zone during cutting and detects the sensor signal, may be the same component (measuring device) that later detects the cutting edge of the cut workpiece. The sensor is preferably intended for image signal detection. For this purpose, the image signal may be captured either in the productive cut or in the calibration cut.

[0114] As shown in Fig. 9, the camera K is essentially separate from the sensor D. Thus, two separate optical detection devices are provided in different components and / or locations of the system. In a variant of the invention, the functions can also be combined in one device (in particular the sensor D) so that a separate camera is not required.

[0115] The camera K is designed to transmit the captured image data to a first processor P1 of the calibration module KM via a wired or wireless connection, denoted in Fig. 9 as interface ss. The first processor P1 is intended to provide a quality measurement result q of the cutting edge of the finished cut workpiece, which is calculated mainly on the basis of the detection of the measurement signal of the cutting edge of the respective finished cut workpiece by the camera K and optionally on the basis of a subsequent image analysis by suitable image processing algorithms by means of an artificial neural network (ANN). The term "quality measurement result" is used because the quality is directly or indirectly (through the camera image b KThis is because measurements are "measured" (indirectly through) and calculations are based on actual recorded or measured values. The use of ANNs can be seen in the following publications: (1) Tatzel L., Leon F.: Image-based roughness estimation of laser cutting edges with a convolutional neural network, Lane 2020; (2) Giusti A. et al.: Image-based Measurement of Material Roughness using Machine Learning Techniques, ISEM 2020; (3) De Mitri O. et al.: Image acquisition, evaluation and segmentation of thermal cutting edges using a mobile device, SPIE Photonics 2019; (4) Bebis G. et al. (eds.) Advances in Visual Computing. ISVC 2014. Lecture Notes in Computer Science, vol 8887. Springer, Cham, Soukup D, Huber-Mork R (2014) Convolutional Neural Networks for Steel Surface See Defect Detection from Photometric Stereo Images.

[0116] In a further development of the invention, the calculation of the quality measurement result can take into account other measurement data in addition to or as an alternative to the captured camera image data, such as data from other measuring devices and / or metadata captured together with the measurement data captured by the sensor and / or machine data characterizing the laser cutting process, such as laser type, laser power, speed, etc. The first processor P1 derives from the provided data, in particular image data B of the cutting edge. KThe aim is to calculate the quality measurement result q from

[0117] The second processor P2 of the calibration module KM then aims to compare the loaded quality estimation result q' provided by the real-time estimator ES with the actual or real quality measurement result q provided by the first processor P1. In this regard, in FIG. 9, in order to clarify that the calibration data set k results from the deviation between the estimate q' and the measurement q or the manual input q*, the formula "k:=q'-q / q*" According to a preconfigurable event pattern, the second processor P2 is then intended to calculate a calibration data set k for calibrating the real-time estimator ES and to transfer the result together with the calibration data set k to an output interface A, which is intended to output the calculated calibration data set k.

[0118] In the first embodiment, the quality measurement q is calculated based on the quality of the image b captured by the camera K. K is calculated automatically by an algorithm on the first processor P1 based on

[0119] In a second embodiment, a fully or partially manual approach may be implemented, with the user manually inspecting and evaluating the cutting edge of the cut workpiece. For this purpose, suitable input fields may be made available to the user on the user interface UI, into which the user may enter specific values ​​w for selected and / or preconfigured quality parameters, such as roughness, burr formation, etc. The entered values ​​w are then fed to the first processor P1, which then executes a simple routine to determine a quality measurement result q* from the entered values. For this routine, the method described in WO2020069889A1 may be selected. Thus, partially manual inputs may also be used, with mainly qualitative values, such as "good", "medium", etc. being entered. If such qualitative values ​​can also be assigned online quality estimates (as suggested above), the values ​​from the online and offline quality may be easily compared or adjusted with each other.

[0120] In further embodiments, a combination of the two above-mentioned alternatives can also be used. This has the advantage that the automatically calculated quality measure q can be subject to verification. For this purpose, the automatically calculated quality measure q is compared with the quality measure q* determined on the basis of a manual input by the user. In case of a deviation (e.g. exceeding a threshold value), the check routine is triggered again and the user can be prompted, for example, to make another entry. If the deviation persists, this indicates an error in the calculation of the quality measure and an error message can be output, for example, to initiate an improvement of the calculation function executed on the first processor P1 to calculate the quality measure q.

[0121] Fig. 10 shows a schematic diagram of the calibration module KM in data exchange with the real-time estimator ES. Both the quality measurement result q and the quality estimation result q' are supplied to the calibration module KM in a very abstract way. Based on these data sets a calibration data set k can then be calculated, which of course relate to the same cutting edge. The calibration data set k is then forwarded to the real-time estimator ES for calibration.

[0122] Various methods can be used to derive the cut quality from the photographed cut edge. For example, comparative images of high quality cuts can be stored and compared to the current image to assess the current cut quality. For further details, see the publication "Soukup D, Huber-Mork R (2014) Convolutional Neural Networks for Steel Surface Defect Detection from Photometric Stereo Images." in Bebis G. et al. (eds.) Advances in Visual Computing. ISVC 2014. Lecture Notes in Computer Science, vol 8887. Springer, Cham. This publication covers the feasibility of characterizing metal surfaces using CNNs and images from different viewpoints, which can be found at https: / / pdfs.semanticscholar.org / b2b8 / ab163fb0183325dd3458e3cbaad2f8bf265e.pdf.

[0123] For more information on laser cutting roughness estimation using deep learning, please refer to the following publication: Tatzel L., Leon F.: Image-based roughness estimation of laser cutting edges with a convolutional neural network, Lane 2020.

[0124] Offline quality measurements of selected cut edges of selected parts for the purpose of calibrating the online quality estimates are especially useful for laser machines that have or are connected to a machine handling system. The cut parts are removed from the cutting table by a robot (or similar), removed, stacked and / or re-stacked. Systems that already have a camera connection during removal or sorting are especially suitable. It is easy for the sorting or removal robot to present the cut edges of the selected parts to the camera. A separate camera may also be beneficial. If an image of the cut edge is available, the real cut quality (roughness, burrs, etc.) can be determined and fed to the machine again. Cut quality includes the following features (not exhaustive): roughness, craters, burrs, slag residues, welds, some features are also called quality defects. The machine compares (i.e., calibrates) the measured quality of the cutting edge (e.g., sporadically via a second camera) with the quality estimated online via a first sensor (e.g., a process signal such as a process light) and can recalibrate or adjust the estimated quality if necessary, especially but not necessarily, if there is a deviation beyond a certain threshold.

[0125] The presented solution allows for a fully automatic cutting system, which can be periodically checked (calibrated) to maintain the required / desired cutting quality.

[0126] Even in the absence of the handling system (robot) described above, this idea can be incorporated into any automation system, at least online, with a first sensor in / on the cutting head for monitoring the process signal, and at least a second sensor or camera in a position where the quality of the cut edge can be determined.

[0127] Finally, it should be noted that the description and exemplary embodiments of the present invention should not be understood as being limiting with respect to a specific physical realization of the present invention. All features described and shown in connection with individual embodiments of the present invention can also be provided in different combinations in the subject matter according to the present invention to simultaneously realize their advantageous effects, even if the combination is not explicitly mentioned.

[0128] The scope of protection of the invention is given by the claims and is not limited by the features exemplified in the description or shown in the drawings.

[0129] It is particularly clear to those skilled in the art that the present invention can be used not only in sheet metal cutting systems but also in other systems involving lasers that require high quality. Furthermore, the components of the calibration module or the components of the system can be distributed over several physical products (distributed system) with suitable data connections.

Claims

1. A method for calibrating a real-time estimator (ES) using a calibration module, said real-time estimator (ES) being configured for the purpose of quality estimation for a cutting method using a laser cutting machine (L), a step (S1) of loading a quality estimation result (q′) of said real-time estimator (ES), said quality estimation result (q′) being determined on the basis of at least one sensor signal from at least one sensor (D); - providing a quality measurement (q) of the cutting edge of the workpiece (S2), said quality measurement (q) being provided by detecting (S21) a measurement signal of the cutting edge of the finished cut workpiece by means of a measuring device (K); a step (S3) of comparing the loaded quality estimation result (q') with the provided quality measurement result (q), in which location identifiers are assigned and / or stored for the detected sensor signal and the detected measurement signal, the detected sensor signal and the detected measurement signal being unambiguously identifiable via the location identifiers and assigned to each other for the comparison step (S3), and only if, based on the result, a determined deviation exceeds a preconfigurable threshold, - a step (S4) of calculating a calibration data set (k) for calibrating (S5) said real-time estimator (ES), a step of calibrating (S5) or retuning said real-time estimator (ES) using said calibration data set (k).

2. 2. The method according to claim 1, characterized in that the real-time estimator (ES) is designed to execute an estimation algorithm (S22) which calculates the quality estimation result (q') from at least a sensor signal detected from at least one sensor (D).

3. 3. The method according to claim 2, characterized in that the estimation algorithm (S22) of the real-time estimator (ES) comprises a machine learning method, in particular a deep convolutional neural network (CNN).

4. 4. The method according to claim 1 , characterized in that the at least one sensor (D) is designed as an optical sensor, in particular as an in-process camera and / or a photodiode unit, and / or the measuring device (K) is a camera, in particular a camera in a mobile terminal and / or a camera attached to or assigned to an automation system, and / or the measuring device (K) is designed as an optical measuring device, in particular as a post-process camera.

5. 2. The method according to claim 1, characterized in that the detection of the sensor signal and / or the measurement signal is configurable, in particular with regard to execution time and / or execution type.

6. 2. The method of claim 1, characterized in that metadata are assigned and / or stored to the detected sensor signal and / or the detected measurement signal.

7. 2. The method according to claim 1, characterized in that said calibration data set (k) comprises an offset applied to said quality estimation result (q') of said real-time estimator (ES).

8. 2. The method according to claim 1, characterized in that the calibration data set (k) is calculated in particular with regard to a material and / or a cutting process.

9. 2. The method of claim 1, wherein the calibration data set (k) includes trained parameters of a calibration model trained using quality estimates (q') and quality measures (q) that are assigned to each other.

10. 10. The method of claim 9, wherein the calibration model is selected from or combined with a regression model, a support vector machine, a random forest, a gradient boosted tree, an artificial neural network, and / or an ensemble learning approach combining different models is used.

11. 10. The method according to claim 9, characterized in that the calibration model is trained by using as training data additional stored intermediate results of the real-time estimator, in particular intermediate results of the real-time estimator in the form of activation values ​​of neurons of a convolutional neural network (CNN) arranged before an output layer.

12. 12. The method of claim 11, wherein the calibration model is trained by using stored input data and assigned quality measures (q) of the real-time estimator as training data.

13. 10. The method according to claim 9, characterized in that the calibration model and / or the real-time estimator (ES) are adapted using an action-reward learning algorithm, in particular using a reinforcement learning algorithm.

14. 10. The method according to claim 9, characterized in that the calibration model and / or the real-time estimator (ES) are adapted to new materials by a transfer learning method.

15. 2. A method according to claim 1, characterized in that the step (S21) of detecting a sensor signal of a cutting edge is performed on a calibrated cut or without a calibrated cut.

16. 2. A method according to claim 1, characterized in that the step (S21) of detecting the measurement signal by the sensor (D) is performed via a deflection mirror (U).

17. 2. The method of claim 1, wherein said quality measurement (q) is provided by an input data set via a user interface (UI).

18. 2. The method of claim 1, wherein the loading step (S1), the providing step (S2) and the comparing step (S3) of the steps of calibrating the real-time estimator (ES) are performed after a calibration disconnection has been performed.

19. A calibration module (KM) for calibrating a real-time estimator (ES) intended for quality estimation for a cutting method using a laser cutting machine (L), said calibration module (KM) comprising: a load interface (E) for loading the quality estimation results (q') of said real-time estimator (ES); a first processor (P1) intended to provide a quality measurement result (q) of a cutting edge of a workpiece, said quality measurement result (q) being provided by detecting (S21) a measurement signal of the cutting edge of a finished cut workpiece by a measuring device (K); Equipped with a step (S3) of a second processor (P2) intended to compare the loaded quality estimation result (q') with the quality measurement result (q), in which location identifiers are assigned and / or stored for the detected measurement signals, the detected measurement signals being unambiguously identifiable via said location identifiers and assigned to one another for the comparison step (S3), only if, based on the result, the determined deviation exceeds a preconfigurable threshold, configured to calculate a calibration data set (k) for calibrating the real-time estimator (ES), The calibration module (KM) a calibration module (KM), characterized in that it comprises an output interface (A) configured to output said calculated calibration data set (k) and to calibrate or retune said real-time estimator (ES) using said calibration data set (k).

20. A system for use in a laser cutting machine (L), characterized in that it comprises a real-time estimator (ES) and a calibration module (KM) according to claim 19.

21. 21. The system of claim 20, wherein the system comprises or is in data communication with a sensor (D).

22. 22. A computer program, the computer program being capable of being loaded into a memory unit of a processing unit and comprising program code segments for causing the processing unit to carry out a calibration method according to any one of claims 1 to 21 when the computer program is executed in the processing unit.

Citation Information

Patent Citations

  • Method for evaluating a laser-cut edge, mobile device and system

    DE102019209088A1

  • Method for monitoring the quality of laser-machining processes and corresponding system

    EP2357057A1

  • Method of controlling a laser cuting process in a high energy zone with interruption of the cutting process ; corresponding device and computer program

    EP3159093B1

  • Processing condition analyzer, laser processing device, laser processing system and processing condition analysis method

    JP2020121338A

  • Real-time adaptive control of additive manufacturing processes using machine learning

    JP2021008113A