Process signal reconstruction and anomaly detection in laser processing processes

Description

[0001] The present invention relates to a method for monitoring a laser processing process on a workpiece, in particular on a metallic workpiece, a laser processing method, a system for monitoring a laser processing process, and a laser processing system for processing a workpiece by means of a laser beam with such a system. Background and state of the art

[0002] In a laser processing process, a workpiece, in particular a metallic workpiece, is processed using a processing laser. The processing may include, for example, laser cutting, soldering, welding, and / or ablation. The laser processing system may, for example, comprise a laser processing head.

[0003] Laser processing processes are often subject to quality control. In particular, during laser welding or soldering of a workpiece, the quality of the processing result, i.e., the resulting joint, is checked. Current solutions for process monitoring and quality assessment in laser welding usually include a pre-, in-, and post-process monitoring system. The pre-process system is responsible for detecting the joining gap in order to guide the laser beam to the appropriate position and determine the offset of the joining partners. In most cases, triangulation systems are used for this purpose. The in- and post-process systems are used to monitor the process and ensure the quality of the resulting joint. In-process systems are also used to control or regulate the laser process.

[0004] Post-process monitoring is often used for quality control, as it allows the laser processing result, for example, a finished and cooled weld seam, to be examined and measured according to applicable standards (e.g., SEL100). Post-process monitoring or post-inspection requires significant equipment complexity. A separate measuring cell often has to be set up for post-process monitoring. In-process monitoring systems (also called inline or online process monitoring systems) are typically designed to detect at least some of the radiation emitted by the laser processing process.

[0005] In-process processing refers to the processing of signals that arise directly when the laser strikes the workpiece to be processed. The signals can be recorded using photodiodes, line sensors, image sensors, or multi- or hyperspectral sensors. In each case, electromagnetic radiation emitted from the laser process zone is used as the basis for data processing. The data can be integrated over a specific wavelength range using image sensors with spatial resolution, and / or integrated with diodes with non-spatial resolution, and / or acquired with temporal frequency resolution using spectrometers. Other data, such as the laser power in the processing head, can also be used for data processing. The goal of such processing is typically to evaluate the process.

[0006] Typical anomaly detection methods based on radiation intensity curves create reference curves from many recorded signal curves. Envelopes are placed around the reference curves. If a signal exceeds or falls below an envelope during the process, an error is triggered based on predefined error criteria. Criteria can include, for example, the integral of the signal over the envelope or the frequency of the signal exceeding the envelope. This approach is based on the assumption that the process always remains constant. A laser material processing process under changing conditions is difficult to monitor this way. Furthermore, it is complex to parameterize signal changes that are not responsible for quality.

[0007] In a downstream step, quality can be determined by a worker through visual inspections or random sampling using other measuring methods. Since materials scientists place demands on a weld in the form of measured values, a worker tests the weld on the workpiece by determining the required physical parameters, such as strength. For example, the welded workpiece is subjected to a tensile test, and the tensile force in Newtons at which the weld fails is determined. When welding contacts, for example, the conductivity is determined in Siemens, since a specific conductivity of the weld structure is required. This downstream step increases cycle times and is generally more expensive and complex than automated in-process monitoring and should therefore be avoided if possible.The quality characteristics determined by the operator through visual inspections or other measurement methods on a random basis can be documented using discretized or continuous values.

[0008] A further method for in-process quality monitoring was proposed, based on these documented values ​​and on learning methods with deep neural networks. These methods learn the most generalized relationship possible between the incoming signal curves, images, and / or control signals and a given value (a readily interpretable quality characteristic). A regressor is used for continuous values, and a classifier for discretized values. The learned relationship is then predicted by the in-process monitoring system. This allows even a process with changing conditions to be monitored, since the generalized signal curve has been assigned to a quality characteristic. Furthermore, it is easier to filter out signal changes that are not responsible for quality.

[0009] EP 3 736 645 A1 describes a method for the automated control of material or laser processing processes. The method is controlled in a closed-loop system by a control unit that calculates correction output signals and controls an energy generation unit, an energy output unit, an energy output measurement unit, and a material-energy interaction measurement unit for measuring actual processing results. The actual processing results are fed to the control unit. Generated correction output signals, which lead to desired processing results, are fed to the energy generation unit and the energy output unit via a control unit output. The calculation of the correction output signals is performed using machine learning methods.The method comprises storing time series of measurements, desired machining results, and correction output signals in a memory of the control unit, analyzing by comparing the time series of the correction output signals and the time series of measurements in an observation subunit based on a machine learning technique, and sending the correction output signals in the form of production conditions to the energy output control subunit and / or the second energy output control subunit.

[0010] DE 10 2018 129 441 A1 relates to the monitoring of a laser processing process using deep convolutional neural networks and describes a system for monitoring a laser processing process for processing a workpiece, comprising: a computing unit which is configured to determine an input tensor based on current data of the laser processing process and to determine an output tensor based on the input tensor by means of a transfer function, which output tensor contains information about a current processing result, wherein the transfer function between the input tensor and the output tensor is formed by a trained neural network.

[0011] DE 10 2018 129 425 A1 describes a system for detecting a machining error for a laser machining system for machining a workpiece, the system comprising: a detection unit for detecting image data and height data of a machined workpiece surface; and a computing unit, the computing unit being configured to form an input tensor based on the detected image data and height data, and to determine, based on the input tensor, by means of a transfer function, an output tensor containing information about a machining error.

[0012] Johannes Günther et al., "Intelligent laser welding through representation, prediction, and control learning: An architecture with deep neural networks and reinforcement learning," Mechatronics, Vol. 34, December 31, 2016, pages 1–11, XP029535343, ISSN: 0957-4158, DOI: 10.1016 / J.MECHATRONICS.2015.09.004, describes the use of integrated machine intelligence in the control of a laser welding process. Summary of the invention

[0013] In quality monitoring based on learning methods with deep neural networks, it is generally not possible to train all possible defect types for a process. Furthermore, the user typically wants to monitor quality with as few defect examples as possible. The behavior of a classifier or regressor with unknown defect types or extrapolations is often unpredictable. The results can deviate significantly from the expected prediction and lead to misinterpretations. For example, large extrapolations in a regression can lead to extreme outliers (processes / products that would actually be "not OK") being rated with a similar value to actual good welds (processes / products that are "OK"). Furthermore, the result of a regression or classification using deep neural networks can only be explained with additional methods, which are generally complex to evaluate.Without the additional methods, the user cannot understand how the algorithm made its decision.

[0014] It is an object of the invention to provide a method for monitoring a laser processing process with a neural network, which allows detection when an anomaly of the laser processing process occurs and in which as few examples of faulty processing as possible are required for training the neural network.

[0015] It is a further object of the invention to provide a method for monitoring a laser processing process using a neural network, which allows a quality characteristic of the laser processing process to be determined and, in addition, the plausibility of the determined quality characteristic to be checked. It is desirable to achieve additional certainty for the correctness of the determination of the quality characteristic through the plausibility check, even if faulty processing operations or error types occur that were not taken into account during training of the neural network. In particular, it is desirable that as few examples of faulty processing operations as possible are required to train the neural network.

[0016] One or more of these objects is achieved by the features of the independent claims. According to one aspect of the present disclosure, a method for monitoring a laser processing process according to claim 1 is provided.

[0017] This aspect may include one or more of the features described below.

[0018] The laser processing process can include, for example, laser cutting, soldering, welding, and / or ablation. The autoencoder is formed, in particular, by a trained deep neural network. The neural network can be trained using error feedback or backpropagation. Here and below, a neural network is understood to be an artificial neural network.

[0019] The process signal data set can also be referred to as the sensor data set. The reconstructed process signal data set can also be referred to as the reconstructed sensor data set. The process signal data set input to the autoencoder can be referred to as the process signal input data set. The reconstructed process signal data set generated by the autoencoder can be referred to as the process signal output data set.

[0020] The generation of a reconstructed process signal data set by the autoencoder can also be referred to as reconstructing a process signal data set by the autoencoder. That is, the method comprises: reconstructing the (input) process signal data set by generating a reconstructed process signal data set by the autoencoder. The method comprises the step: generating an encoding by the autoencoder based on the input at least one process signal data set of the laser processing process. The generation of a reconstructed process signal data set by the autoencoder is carried out based on the generated encoding. The autoencoder comprises an encoder and a decoder. The at least one process signal data set of the laser processing process is input to the encoder. The encoder generates an encoding based on the input at least one process signal data set of the laser processing process.

[0021] The reconstructed process signal dataset is generated by the decoder. The decoder generates the reconstructed process signal dataset from the generated encoding (reconstructing a process signal dataset). The encoding can also be referred to as latent feature space, compressed mapping, dimensionality reduction, or bottleneck.

[0022] The incoming process signal data set and the reconstructed process signal data set can have the same dimension and size (number of components per dimension). For example, they can have a dimension N with kn elements each, where N, kn are natural numbers with n = 1, ..., N. In particular, the process signal data set input to the autoencoder (or process signal input data set) can be input to the autoencoder in the form of a tensor, and the reconstructed process signal data set generated by the autoencoder (process signal output data set) can be generated in the form of a tensor, with the input process signal data set and the reconstructed process signal data set being tensors of identical dimension and size. Thus, although dimensionality reduction occurs within the autoencoder, the reconstructed process signal data set again has the same dimension and size as the process signal data set input to the autoencoder.The term tensor encompasses a vector (1-dimensional tensor). The respective tensor space in which the process signal input data set and the process signal output data set are defined can be identical. Generating a reconstructed process signal data set by the autoencoder can thus, in particular, comprise: mapping the process signal input data set as a reconstructed process signal output data set in the form of a tensor with identical dimension and size (identical tensor space). This can be referred to as mapping "to itself." The autoencoder can be referred to as a machine learning algorithm autoencoder.

[0023] Using dimensionality reduction via a deep autoencoder, the signals or the process signal dataset can be reconstructed. The reconstruction error can determine whether an anomaly is present or whether a regression and / or classification result is valid. A special feature is that a process anomaly can be detected during and / or after the laser processing process being monitored based on the determined reconstruction error.

[0024] The processing of process signals (or the process signal dataset) using an autoencoder for anomaly detection can be described as follows. In one embodiment, the autoencoder performs the following processing: The encoder, consisting of a deep neural network, extracts features from incoming signals for encoding. This encoding simultaneously represents a dimensionality reduction of the incoming process signals. This enables the description of the process signals in a few dimensions. The decoder, consisting of a deep neural network, reconstructs the process signals.

[0025] The present invention therefore deals with anomaly detection using deep neural networks, as well as a plausibility check or verification of a regression and / or classification result in the monitoring of a laser material processing process, as well as the quality assessment of the products. The invention is based on the idea of ​​obtaining a reconstructed process signal data set from the processing of a process signal data set by an autoencoder and investigating whether the reconstructed process signal data set matches the original process signal data set input into the autoencoder. If this is the case, it can be assumed that the autoencoder is processing the key features of the process signal data set as expected, and the laser processing process thus corresponds to what was considered during the creation or training of the autoencoder.However, if the reconstruction error suggests that the reconstructed process signal dataset does not match the original process signal dataset input to the autoencoder, an anomaly in the laser processing process can be detected. It can then be concluded that the laser processing process does not correspond to what was considered during the creation or training of the autoencoder.

[0026] If the data is sufficiently compressed, at least one process signal data set (i.e. the input signal) can only be correctly reconstructed if the autoencoder has been trained in advance with this data. When training the data, the weights of the autoencoder can be adjusted using backpropagation in several epochs until the reconstruction error is minimal. The reconstruction error can be the averaged absolute or squared deviation of the process signal data set (i.e. the input signal) from the reconstructed process signal data set (i.e. reconstructed signal). The reconstruction error can be the signed, absolute or squared deviation summed or integrated along the time axis. For example, when trained with multiple data sets, the autoencoder learns to encode recurring features and patterns and to decode them again.For a known laser processing process, the measured process signals can be perfectly reconstructed with a well-trained autoencoder, with the exception of signal noise, even if fluctuations occur in the process signal. Only those fluctuations in the process signal that also appeared in the trained signals are perfectly reconstructed. The structure of the autoencoder determines how well the features and patterns can be abstracted and generalized. If anomalies occur, for example, in the form of gross differences in the signal curve compared to the previously trained signal curves, these will not be correctly reconstructed by the autoencoder. Therefore, the reconstruction error for unknown laser processing processes, such as those that occur in the case of quality defects, is significantly greater than for known laser processing processes.

[0027] The method according to the invention can be carried out continuously and / or repeatedly during the laser processing process.

[0028] A process signal data set of the laser processing process is understood to be a data set of at least one process signal of the laser processing process. Process signals can be measured or recorded for the laser processing process. A process signal data set can, for example, comprise a time series of a process signal, for example, measured intensities over a time interval, and / or can comprise a measured process parameter of the laser processing process. Particularly in in-process monitoring systems, a process signal data set can comprise raw data in order to enable particularly fast and efficient real-time analysis or evaluation.

[0029] Process signals can be recorded or measured, for example, with one or more sensors, for example with at least one of at least one photodiode, at least one line sensor, at least one image sensor, at least one camera, at least one triangulation sensor, at least one sensor for optical coherence tomography (OCT), at least one acoustic sensor and / or at least one multi- or hyperspectral sensor.A process signal from the laser processing process can, for example, include: process radiation / emission from the laser processing process, such as thermal radiation and / or plasma radiation, an image of the workpiece surface of a workpiece, radiation intensity from the laser processing process reflected by the workpiece, a process signal from the laser processing head, such as a scattered radiation intensity from scattered radiation within the laser processing head or a temperature of an element within the laser processing head. A process signal can, for example, be measured in optical coherence tomography by irradiating measuring light and detecting the reflected measuring light. Electromagnetic radiation emitted from the laser processing zone can also be measured as a process signal and serve as a process signal data set as the basis for monitoring a laser processing process.The electromagnetic radiation can, for example, include process emissions such as thermal radiation, plasma radiation, and / or reflected laser radiation from a workpiece surface. However, the process signals can also include acoustic signals.

[0030] Process signals may also include control signals used by a controller for the laser processing process, e.g., laser power, processing speed, etc.

[0031] The reconstruction error preferably describes a deviation of the at least one reconstructed process signal data set from the at least one process signal data set input to the autoencoder. The reconstruction error can, for example, comprise an averaged absolute or squared deviation. The reconstruction error can be calculated using different criteria. For example, in the case of sequential time series, the reconstruction error can only be used for certain signal ranges. It is also possible to consider the reconstruction error separately for individual dimensions (if these are not uniformly standardized, for example) and, for example, to estimate a covariance matrix and a mean vector. The reconstruction error can, for example, be determined based on a metric. Metrics such as the Mahalanobis distance can be used as a measure of the degree of abnormality (or as a measure of the reconstruction error).In addition, a reconstruction error can also be calculated using FFT or wavelets (continuous CWT and / or discrete DWT) for different frequency ranges.

[0032] The method may include determining a degree of abnormality, wherein the detection of an anomaly in the laser processing process is based on the determined degree of abnormality. The determination of a degree of abnormality may, in particular, be based on a weighted summation or on the Mahalanobis distance relative to individual characteristic values ​​of the reconstruction error.

[0033] The method may include: determining a degree of abnormality, in particular based on the reconstruction error; and detecting an anomaly of the laser processing process based on the determined degree of abnormality. In particular, the step of detecting an anomaly of the laser processing process based on the determined reconstruction error may include: determining a degree of abnormality based on the reconstruction error; and detecting an anomaly of the laser processing process based on the determined degree of abnormality.

[0034] The step of determining a reconstruction error may comprise: determining a Mahalanobis distance, in particular based on the at least one process signal data set and the at least one reconstructed process signal data set.

[0035] Determining a degree of abnormality may include: determining a degree of abnormality based on the Mahalanobis distance and / or a weighted sum of individual reconstruction error metrics.

[0036] Determining a degree of abnormality may comprise: determining a Mahalanobis distance, in particular based on the at least one process signal data set and the at least one reconstructed process signal data set.

[0037] The step of determining the Mahalanobis distance can be done with respect to: the deviation of the at least one process signal data set (sensor data set) from the at least one reconstructed process signal data set (sensor data set); and / or individual characteristic values ​​for the reconstruction error; and / or the coding of a process signal data set (sensor data set).

[0038] Reconstruction error metrics may include and / or be referred to as reconstruction error metrics, reconstruction error metrics, reconstruction error parameters, a reconstruction error feature or features, and / or a reconstruction error metric.

[0039] The method may comprise or the step of determining a reconstruction error may comprise: determining the parameters mean vector and covariance matrix of the Mahalanobis distance, in particular determining the parameters mean vector and covariance matrix of the Mahalanobis distance using defect-free or labeled data sets.

[0040] The method may include determining an abnormality level based on the weighted summation and / or the Mahalanobis distance relative to individual characteristic values, in particular relative to the individual characteristic values ​​for the reconstruction error. The weighted summation may be the signed, absolute, or squared deviation summed or integrated along the time axis.

[0041] The method may include setting a threshold value that distinguishes between normal and anomalous based on the degree of abnormality. For example, in the step of detecting an anomaly, the detection of an anomaly in the laser processing process may be based on whether the degree of abnormality reaches and / or exceeds the threshold value.

[0042] Detecting an anomaly in the laser processing process based on the determined reconstruction error can also be referred to as detecting whether an anomaly in the laser processing process exists based on the determined reconstruction error. Detecting an anomaly in the laser processing process based on the determined reconstruction error can mean or include determining, based on the determined reconstruction error, that either an anomaly in the laser processing process exists or that an anomaly in the laser processing process does not exist. For example, a criterion used for the reconstruction error or its measure (judgment criterion) can determine or define whether an anomaly in the laser processing process exists. The reconstruction error decides whether it is an anomaly.For example, detecting an anomaly of the laser processing process may include: detecting an anomaly of the laser processing process based on a measure of the determined reconstruction error.

[0043] Detecting an anomaly in the laser processing process can, for example, include: detecting / recognizing that an anomaly in the laser processing process exists if the determined reconstruction error meets an error criterion, e.g., the reconstruction error exceeds a threshold or a measure of the reconstruction error exceeds a threshold. Assuming that information about the laser processing process is contained in the process signal data set, a small reconstruction error indicates that the autoencoder has encoded relevant features of the laser processing process. However, if the reconstruction error is relatively large, this in turn indicates an anomaly and thus that features of the laser processing process are not sufficiently captured by the autoencoder.

[0044] The method may include classifying the at least one process signal data set as an anomaly or as a normal case based on the determined reconstruction error. For example, the at least one process signal data set may be classified as an anomaly if the presence of an anomaly in the laser processing process is detected based on the determined reconstruction error.

[0045] The method may include classifying the laser processing process or (in particular) a processing result of the laser processing process as "OK" or "not OK" based at least on whether an anomaly of the laser processing process is detected in the step of detecting an anomaly of the laser processing process. For example, if an anomaly of the laser processing process is detected, the laser processing process or a product of the laser processing process may be classified as "not OK." The processing result may, for example, be a product of the laser processing process. Classifying the laser processing process may, in particular, include classifying a processing result of the laser processing process.

[0046] If the method is used for quality control in the laser processing process, anomalies indicate a quality defect or a faulty process, and the product is classified as "not OK" if the reconstruction error exceeds a threshold. The autoencoder will then typically be trained exclusively with signals (process signal data sets) from a laser processing process whose product or process has been classified as "OK."

[0047] Preferably, the method comprises: measuring at least a portion of the process signals of the process signal data set, and / or transmitting (or transmitting) at least a portion of the process signals of the process signal data set from a controller.

[0048] At least some of the process signals can be recorded or measured, for example, using one or more sensors, as described above. The controller can, in particular, be a controller that controls the laser processing process. The controller can, for example, be a controller of a laser processing system for processing a workpiece using a laser beam. Process signals transmitted by a controller can, for example, include control signals, e.g., a specified laser power, a target focus position, a processing speed, etc.

[0049] The method may include inputting one or more process conditions of the laser machining process into the autoencoder. Process conditions may include, for example, a workpiece material, a machining configuration, a welding geometry (e.g., lap welding, butt welding, etc.).

[0050] The method for monitoring a laser processing process can, for example, include a pre-process monitoring method, an in-process monitoring method, and / or a post-process monitoring method. An in-process monitoring method can enable monitoring while the workpiece is being processed.

[0051] Especially when laser welding or soldering a workpiece, it is important to check, assess, and / or ensure the quality of the resulting joint. One goal is to reliably locate and detect all defects.

[0052] In embodiments, the method further comprises the steps of: determining a quality characteristic of the laser processing process; and evaluating the quality characteristic as valid if no anomaly is detected in the step of detecting an anomaly of the laser processing process; and evaluating the determined quality characteristic as invalid if an anomaly is detected in the step of detecting an anomaly of the laser processing process.

[0053] The method can therefore provide an application of an anomaly detector, which preferably comprises or consists of an autoencoder with a deep neural network and which reconstructs process signals (a process signal data set), wherein a reconstruction error is used for evaluation. Thus, a validation of the quality characteristic takes place depending on whether no anomaly is detected in the step of detecting an anomaly in the laser processing process. The method thus allows a quality characteristic of the laser processing process to be determined and, in addition, the plausibility of the determined quality characteristic to be checked. In particular, errors can be detected not only on the basis of the quality characteristic, but also, for example,Not only can a laser processing result be classified as "not OK," but it can also intercept cases in which an anomaly is detected and the quality characteristic is therefore considered invalid. Especially in the field of laser material processing, sufficient data is generally not available for quality monitoring using deep neural networks. Especially in laser material processing, it is desirable to ensure process monitoring under changing process conditions. It is therefore particularly advantageous for the user that quality monitoring can be implemented in this way with as few examples of defectively processed materials as possible.

[0054] The quality characteristic can, in particular, be a predictive value. Determining a quality characteristic of the laser processing process can be a prediction of the quality characteristic of the laser processing process. The quality characteristic can describe a prediction for a quality of the laser processing process or (in particular) the processing result of the laser processing process. For example, it can characterize a processing quality or any processing errors. The term "prediction for a quality of the laser processing process" can, in particular, include a prediction for a quality of the processing result of the laser processing process.

[0055] Two cases in particular are worth mentioning here: The quality characteristic can be determined from discrete values ​​or from continuous values. This means that the quality characteristic can be determined from a range of values ​​that includes discrete (or discretized) values ​​(e.g., discrete classification values), or it can be determined from a range of values ​​that includes a continuous range of values.

[0056] The quality characteristic can comprise a physical quality characteristic, for example, a prediction for a physical quantity (physical property, physical value) of the laser processing process or the processing result of the laser processing process. For this purpose, the quality characteristic can be determined, in particular, from continuous values. In a laser welding process, the physical quantity can be, for example, the strength / tensile strength of a weld seam, the electrical conductivity of the weld structure, or the smoothness of the weld seam. For example, the determined quality characteristic can be used to influence the control parameters of the process during the ongoing laser processing process. This allows the laser processing process to be controlled. The quality characteristic can, for example, characterize the processed workpiece.

[0057] The quality characteristic can, for example, classify the quality of the laser processing process or the processing result of the laser processing process. From this, it can be deduced whether the processing operation leads to the desired criteria of the processing process or the workpiece, or whether processing errors occur. For this purpose, the quality characteristic can be determined, in particular, from discrete values. For example, a classification can be made as "OK" / "defect-free" or "not OK" / "faulty" (i.e., a defect is present). The workpiece can, for example, be marked or classified as "OK" / "good" (i.e., suitable for further processing or sale) or as "not OK" / "poor" (i.e., as scrap). The quality characteristic can, for example, include a prediction that a defect is present.In a laser welding process, for example, the quality characteristic can include a prediction that a gap is present. The determination of at least one quality characteristic and, in particular, the detection of processing errors can be automated, thus enabling process monitoring, particularly online process monitoring.

[0058] A quality characteristic of the laser processing process can be determined based on the at least one process signal data set of the laser processing process. From the at least one process signal data set, significant features are extracted or calculated, for example, as quality characteristics for the quality of the laser processing process. These can, for example, represent or describe the quality of the weld and solder seams in a laser welding or laser soldering process.

[0059] Determining a quality characteristic of the laser processing process can comprise determining a quality characteristic of the laser processing process by a regressor or a classifier which is formed by a (trained) neural network. Preferably, a regressor is used for continuous values ​​and a classifier is used for discrete (e.g. discretized) values. The autoencoder and the regressor or classifier can be designed as separate algorithms. The autoencoder and the regressor or classifier can be arranged in parallel. The autoencoder and the regressor or classifier can receive the at least one process signal data set via a common input layer. Alternatively, the regressor or classifier can be arranged in parallel to a decoder of the autoencoder. In this case, the regressor orClassifier can be arranged after an encoder of the autoencoder and receive the coding from it (so-called extended autoencoder).

[0060] The quality characteristic can be determined in real time. Based on this, control data and / or regulation data can be output to a laser processing system performing the laser processing process. The value of a physical property and / or a classification can thus be used to regulate the laser processing process, particularly if the respective value is determined during the laser processing process. Control of the laser processing process can include adjusting a focus position, a focus diameter of the laser beam, a laser power, and / or a distance of a laser processing head.

[0061] In embodiments of the method, in the step of determining a quality characteristic of the laser processing process by a regressor formed by a neural network, a value for the quality characteristic is determined; and / or in the step of determining a quality characteristic of the laser processing process by a classifier formed by a neural network, a classification value for the quality characteristic is determined.

[0062] The quality characteristic determined by the regressor or classifier is thus assessed as valid or invalid, depending on whether an anomaly is detected during the anomaly detection step of the laser processing process. This allows a plausibility check of the regressor or classifier to be performed. The regressor or classifier can, in particular, be formed by a deep neural network. A classification value can also be determined based on a physical quantity.

[0063] Using the reconstructed signals, a plausibility check and verification of the regressor or classifier can be performed. The validity of the prediction (i.e., the quality characteristic) can be checked. The reconstruction error determines whether an anomaly is present or whether the regression and / or classification result is valid. Thus, regression and / or classification to determine a quality characteristic of a laser processing process is combined with a plausibility check and anomaly detection using a deep neural network. Especially in laser material processing, there is usually not enough data available for quality monitoring using deep neural networks for regression and classification. It is therefore particularly advantageous for the user that quality monitoring can be implemented in this way with as few examples of faulty processed materials as possible.The behavior of a classifier or regressor with unknown error types or extrapolations is often unpredictable. Such cases can be addressed by rating the specific quality attribute as "not valid."

[0064] Anomaly detection using reconstruction errors can mask the predicted value of the regressor or classifier (the quality characteristic). A product can be classified as "not OK" regardless of the predicted value if an anomaly is detected, for example, if the reconstruction error exceeds a threshold. If the reconstruction error is below the threshold, the result of the predicted value of the regressor or classifier can be assumed to be valid. In this way, the autoencoder can also be trained with signals from a laser processing process whose process / product has been classified as "not OK." Furthermore, the regressor enables assignment to a physical quality characteristic, while the classifier enables assignment to a discrete defect pattern.

[0065] Preferably, the autoencoder is trained together with the regressor and / or classifier using known data. The autoencoder can be trained either as a separate algorithm running parallel to the regressor or classifier, or as an extended autoencoder that incorporates the regressor or classifier, forming a unified algorithm. The autoencoder and the regressor or classifier can be trained as follows: The data is labeled with a discrete or continuous value. This value represents a quality characteristic. With an extended autoencoder, a joint code can then be created that represents recurring features and patterns.For quality monitoring of laser processing, the quality characteristic can then be predicted from this coding using the regressor or classifier, for example. The signals (at least one process signal data set) can be reconstructed in parallel or simultaneously using the decoder. Using the reconstructed signals, a plausibility check and verification of the regressor or classifier can then be performed.

[0066] Assuming that the information is contained in the signal trace, a small reconstruction error indicates that the predicted quality attribute is valid. However, if the reconstruction error is relatively large, this may indicate an anomaly and thus an invalid predicted quality attribute, as well as a quality defect or faulty process. The user can see when certain signals or signal regions are not predicted well and can initiate a review or retraining of the algorithm.

[0067] An anomaly can, in turn, be verified using additional methods, either through visual inspection by the operator or additional measurement methods, and assigned to a quality characteristic. In the case of a regressor or classifier plausibility check, an extended autoencoder, or a regressor or classifier with the autoencoder, can be additionally trained using this newly labeled data set.

[0068] In embodiments, the autoencoder and at least one of the regressor and the classifier have a common encoder; and / or the regressor and / or the classifier determines the quality characteristic based on data from an encoder of the autoencoder.

[0069] The determination of the quality characteristic is thus based on processing the at least one process signal data set by the encoder of the autoencoder or the shared encoder. At least one of the regressor and the classifier can determine the quality characteristic based on a code generated by the encoder or data from the encoder. This establishes a close connection between the determined quality characteristic and the reconstructed process signal data set. Therefore, the detection of an anomaly can determine with high certainty whether the determined quality characteristic is valid. The encoder can thus be a shared part of a neural network comprising the autoencoder and at least one of the regressor and the classifier. It is also advantageous that the regressor or the classifier can process dimensionally reduced data from the autoencoder.If the deep autoencoder and at least one of the regressor and the classifier share a common encoder, the autoencoder can also be an extended autoencoder comprising at least one of the regressor and the classifier, in particular an extended deep autoencoder. The extended autoencoder can be trained as a unified algorithm comprising the regressor and / or classifier.

[0070] In other embodiments, the autoencoder and at least one of the regressor and the classifier are parallel to each other; and / or the regressor and / or the classifier determines the quality characteristic based on the at least one process signal data set; and / or the autoencoder and at least one of the regressor and the classifier have a common input layer.

[0071] In this case, the autoencoder and at least one of the regressor and the classifier form separate algorithms. In particular, they can comprise separate neural networks, wherein the neural network of the autoencoder and the neural network of at least one of the regressor and the classifier can have a common input layer. The autoencoder can be trained as a separate algorithm in parallel to the regressor and / or the classifier.

[0072] Preferably, the autoencoder and at least one of the regressor and classifier are trained with the same data. This creates a close relationship between the determined quality characteristic and the reconstructed process signal data set. Thus, the validation of the quality characteristic can be improved by the anomaly detection step.

[0073] The step of determining a reconstruction error may comprise: determining a deviation of the at least one process signal data set from the at least one reconstructed process signal data set. The step of determining a reconstruction error may comprise: determining an averaged absolute or squared deviation of the at least one process signal data set from the at least one reconstructed process signal data set, in particular determining a signed, absolute, or squared deviation summed along the time axis.

[0074] The step of determining a reconstruction error may comprise: filtering at least a portion of the process signal data set and / or the reconstructed process signal data set; and, based thereon, determining the reconstruction error.

[0075] Filtering may, for example, include filtering to reduce noise, such as filtering with a noise filter. To minimize the influence of unsystematic signal noise, the signals may be filtered to generate the reconstruction error.

[0076] The reconstruction error can be determined, for example, as the determined deviation or based on the determined deviation. The deviation can be averaged, subjected to the calculation of the absolute value, and / or squared. The reconstruction error can be determined, for example, as an averaged absolute or squared deviation.

[0077] The method may include normalizing the reconstruction error with respect to the process signal data set. Normalizing the reconstruction error using the input signal level serves to improve comparability of the reconstruction error and can therefore simplify the detection of an anomaly.

[0078] In both the case of anomaly detection and the case of the regressor or classifier's plausibility check, an autoencoder reconstructs the incoming data, i.e., at least one process signal data set, and a reconstruction error is determined from the difference between the input and the reconstruction signal. If current measurements are not trained, the reconstruction error becomes large. A large reconstruction error indicates an anomaly, which in turn can represent a quality defect. In the case of pure anomaly detection, the autoencoder can be trained exclusively with training data in the form of signals (or corresponding process signal data sets) for laser processing processes whose product or process has been classified as "OK" or "good."In the case of a plausibility check of the regressor or classifier, the autoencoder can be trained together with these and with all input signals of the quality monitoring system of the laser processing processes that were also trained by them. These are preferably essentially verified examples with a quality attribute. In the case of a plausibility check of the regressor or classifier, training data for laser processing processes whose product or process was classified as "OK" or "good" can be used for training, as can training data for laser processing processes whose product or process was classified as "not OK" or "bad." In the case of a plausibility check of the regressor or classifier, the autoencoder can be trained either as a separate algorithm in parallel or as an extended autoencoder as a combined algorithm.In the case of a plausibility check of a regressor or classifier, for example, an anomaly can be checked, assigned to a quality characteristic and additionally learned by the regressor or classifier as well as the autoencoder or extended autoencoder.

[0079] In embodiments, the neural network of the autoencoder and / or the regressor and / or the classifier is adaptable. This enables transfer learning. The trained neural network can be adapted using training data through transfer learning. In particular, the neural network can be adapted to new process signal data sets or new laser processing processes that have been newly labeled with a quality characteristic, for example. An adaptable neural network that can be (re)trained using training data through transfer learning is particularly flexible, versatile, and user-friendly.

[0080] Using transfer learning, a neural network can be retrained, for example, when an anomaly is detected. Measuring a physical quantity or classifying it by experts to specify a quality characteristic can be done based on the machined workpiece and, if necessary, using destructive techniques.

[0081] The use of a trained neural network configured for transfer learning therefore has the advantage that the method can be quickly adapted to changing situations, especially to a modified laser processing process. This can reduce the occurrence of anomalies. For example, a valid quality characteristic can be determined for multiple or different laser processing processes.

[0082] According to a further aspect of the present disclosure, a laser processing method is provided, the laser processing method comprising the steps of: processing a workpiece using a laser beam; and monitoring the laser processing process according to the method for monitoring a laser processing process.

[0083] Processing a workpiece using the laser beam can also be referred to as performing the laser processing process or be encompassed thereby. Processing a workpiece using the laser beam can include, for example, laser cutting, soldering, welding, and / or ablation.

[0084] The laser processing method may further comprise: measuring process signals of the laser processing process, and / or transmitting process signals from a controller; wherein monitoring the laser processing process comprises: inputting at least one process signal data set of the laser processing process into the autoencoder, wherein the process signal data set comprises the measured process signals and / or the process signals transmitted from a controller.

[0085] According to a further aspect of the present disclosure, a system for monitoring a laser processing process is provided, the system comprising: at least one sensor unit configured to measure process signals of the laser processing process; at least one autoencoder formed by a deep neural network; and at least one computing unit configured to execute the method for monitoring the laser processing process. At least one process signal data set of the process signals of the laser processing process is input into the autoencoder.

[0086] According to a further aspect of the present disclosure, a laser processing system for processing a workpiece using a processing laser beam is provided, the laser processing system comprising: a laser processing head for irradiating the processing laser beam onto the workpiece; and the system for monitoring a laser processing process. The laser processing system can be configured to perform the described laser processing method. The computing unit can be configured to regulate and / or control the laser processing process based on control data and / or control data.

[0087] The system and the laser processing system for processing a workpiece can also realize all the advantages that also apply to the method and in particular to one of the embodiments of the method. Brief description of the drawings

[0088] Embodiments of the present invention are described in detail below with reference to figures. The figures depict various features of embodiments, whereby the features are not limited to the embodiments alone. Rather, all features that are not mutually exclusive can also be combined with one another, or features can be removed from embodiments, provided they are not essential to the implementation of the invention. Fig. 1 shows a schematic representation of a laser processing system for processing a workpiece using a laser beam and a system for monitoring a laser processing process according to an embodiment; Fig. 2 shows a schematic representation of a deep autoencoder for processing process signal data sets of the laser machining process according to an embodiment; Fig. 3shows schematically a process signal data set and a reconstructed process signal data set; Fig. 4 schematically shows a process signal data set and a reconstructed process signal data set in case of an anomaly of the laser processing process; Fig. 5 shows a schematic representation of a deep autoencoder in parallel with a regressor or classifier according to one embodiment; Fig. 6 shows a schematic representation of an extended deep autoencoder together with a regressor or classifier according to one embodiment; and Fig. 7 shows a schematic representation of a method for monitoring a laser processing process according to an embodiment. Detailed description of implementation examples

[0089] In the following, unless otherwise stated, the same reference symbols are used for identical and equivalent elements.

[0090] Figure 1shows a schematic representation of a laser processing system 100 for processing a workpiece using a laser beam 10 according to embodiments of the present disclosure. The laser processing system 100 comprises a laser processing head 101, in particular a laser cutting, laser soldering, or laser welding head, and a system 300 for monitoring the laser processing process with a sensor unit 310 as an example of a laser processing head according to one embodiment. The system 300 for monitoring the laser processing process further comprises a computing unit 320.

[0091] The laser beam 10 is supplied, for example, via an optical fiber 104 from a laser source (not shown). The beam path of the (processing) laser beam 10 runs via a collimation optics 122 and a focusing optics 124 to a workpiece 18 being processed.

[0092] According to embodiments, the laser processing system 100 or parts thereof, such as the processing head 101, can be movable relative to the workpiece 18 along a processing direction 20. The processing direction 20 can be a cutting, soldering, or welding direction and / or a direction of movement of the laser processing system 100, such as the processing head 101, with respect to the workpiece 18. In particular, the processing direction 20 can be a horizontal direction.

[0093] The laser processing system 100 is controlled by a control unit 140 which is configured to control the processing head 101.

[0094] During processing, process emissions 13 are generated, for example, thermal radiation, plasma radiation, and / or laser radiation reflected from a surface of a workpiece. A portion of the process emissions 13 is guided at least partially coaxially or collinearly to the beam path of the processing beam 10 from the surface of the workpiece via a beam splitter 115 and a collimator optics 312 to the sensor unit 310.

[0095] The computing unit 320 comprises a deep neural network in the form of a deep autoencoder 400 and receives process signals 11 from the sensor unit 310, which are input into the neural network 400 in the form of process signal data sets 30.

[0096] Fig. 2shows a schematic representation of the deep autoencoder 400 for processing process signal data sets 31 of the process signals 11 of the laser processing process. The autoencoder 400 comprises an input layer 1, an encoder 2, a coding layer 3, a decoder 4, and an output layer 5. The coding layer 3 is a common layer of the encoder 2 and the decoder 4.

[0097] Autoencoder 400 performs the following processing: At least one process signal data record 31 of a process signal 11 is input into input layer 1. Encoder 2, consisting of a deep neural network, extracts features of incoming signals 11 for encoding in coding layer 3. This encoding simultaneously represents a dimensional reduction of the incoming process signals 11. This enables the description of process signals 11 in a few dimensions. Decoder 4, consisting of a deep neural network, reconstructs the process signals. A reconstructed process signal data record 51 of the reconstructed signal is output at output layer 5.

[0098] Fig. 3shows a schematic representation of a process signal data set 31 and a corresponding reconstructed process signal data set 51, exemplified as a curve of a signal intensity I over time t. Except for signal noise, fluctuations in the process signal 11 are reconstructed virtually perfectly, which also occurred in the signals taught into the autoencoder 400.

[0099] Fig. 4 schematically shows a process signal data set 31A and a corresponding reconstructed process signal data set 51A in the event of an anomaly 40 occurring in the laser processing process. If anomalies 40 occur in the form of gross differences in the signal curve compared to the previously learned signal curves, they are not correctly reconstructed by the autoencoder 400.

[0100] The arithmetic unit 320 determines a reconstruction error as the averaged absolute or squared deviation of the input signal 11 or process signal data set 31 / 31A from the reconstructed signal or reconstructed process signal data set 51 / 51A.

[0101] If the reconstruction error exceeds a threshold value, the computing unit 320 detects the presence of an anomaly 40. If the method is used directly for quality control in the laser processing process, the anomalies 40 indicate a quality defect or faulty process, and the product is classified as "not OK" by the computing unit 320 if an anomaly 40 has been detected, ie, if the reconstruction error exceeds a threshold value.

[0102] A portion of the process signals of process signal data set 31 can also be transmitted by control unit 140. Process signals transmitted by control unit 140 can, for example, include control signals, e.g., a specified laser power, a desired focus position, a processing speed, etc. Corresponding reconstructed process signals are then part of the reconstructed process signals of reconstructed process signal data set 51.

[0103] Fig. 5 and Fig. 6show schematic representations of the deep autoencoder 400 together with a regressor 600 or classifier 600 for predicting a quality characteristic x of the laser processing process. During quality monitoring of the laser processing process, the quality characteristic x of the laser processing process is predicted. The quality characteristic x is evaluated as valid if no anomaly 40 is detected, and the quality characteristic x is evaluated as invalid if an anomaly 40 is detected.

[0104] The autoencoder 400 is trained with known data together with the regressor 600 or classifier 600. For this purpose, the data is labeled with a discrete or continuous value representing the quality characteristic x. The regressor 600 or classifier 600 is implemented as a deep neural network 6 with an output layer 7 in the computing unit 320.

[0105] Fig. 5shows a schematic representation of a deep autoencoder 400 in parallel with a regressor 600 or classifier 600.

[0106] For the quality monitoring of laser processing processes, at least one process data set 31 is input into the common input layer 1 of the autoencoder 400 and the regressor 600 or classifier 600. From this, the quality characteristic x is predicted using the deep neural network 6 of the regressor 600 or classifier 600, and simultaneously, the process signals are reconstructed as process signal data set 51 via the autoencoder 400.

[0107] Fig. 6 shows a schematic representation of an extended deep autoencoder 400 comprising a regressor or classifier 600.

[0108] The extended autoencoder 400 creates a common coding in coding layer 3 that represents recurring features and patterns. For quality monitoring of laser processing, the quality feature x is predicted from the coding of coding layer 3 using the deep neural network 6 of the regressor 600 or classifier 600. At the same time, the process signals are reconstructed from the coding of coding layer 3 as a process signal data set 51 via the decoder 4.

[0109] Fig. 7 shows a schematic representation of an exemplary method for monitoring a laser processing process, such as that implemented by the system 300 according to Fig. 4 or Fig. 5 can be executed.

[0110] The procedure includes the following steps: Step S10: Measuring at least some of the process signals of the process signal data set 31, and / or transmitting at least some of the process signals of the process signal data set 31 from the controller 140; Step S20: Inputting the process signal data set 31 of the laser processing process into the autoencoder 400; Step S30: Generating an encoding by the encoder 2; Step S40: Generating a reconstructed process signal data set 51 by the autoencoder 400; Step S50: Determining a reconstruction error by the computing unit 320; Step S60: Detecting an anomaly of the laser processing process based on the determined reconstruction error by the computing unit 320; Step S70: Determining a value of a quality characteristic of the laser processing process, for example, by the regressor 600 or classifier 600; Step S80: Evaluate the quality characteristic as valid if no anomaly is detected in step S60 (case N);and step S90: Evaluating the quality characteristic as invalid if an anomaly is detected in step S60 (case A). ;

[0111] As described in the exemplary embodiments, a plausibility check and anomaly detection in combination with deep neural networks for regression and classification in laser material processing are enabled by an extended autoencoder. Based on the autoencoder, the signals can be reconstructed and a reconstruction error can be calculated. The reconstruction error can be used both to evaluate the quality of the laser processing process and to verify the validity of the regressor's or classifier's prediction.

[0112] The initial effort required to generate defect examples in laser material processing for training quality monitoring systems based on a regressor or classifier can thus be significantly reduced, as their available space can be limited according to the known defect examples. The user can monitor quality with a few defect examples and a few good examples. The prediction accuracy for the quality of the process and the product can be increased during the ongoing production process by having the operator analyze anomalies, assign them a quality attribute, and retrain them with the proposed algorithm or method—that is, the autoencoder and, if applicable, the regressor and / or classifier.

Claims

1. A method for monitoring a laser machining process, said method comprising the steps of: - inputting (S20) at least one process signal data set (31) of the laser machining process into an autoencoder (400); - reconstructing the input process signal data set (31) by means of said autoencoder (400), comprising: generating an encoding (3) by means of an encoder (2) of the autoencoder (500) based on the at least one process signal data set (31), wherein generating the encoding comprises a dimension reduction of the at least one process signal data set (31), generating a reconstructed process signal data set (51, 5) by a decoder (4) of the autoencoder (500) based on the generated encoding, wherein the encoder (2) and the decoder (4) consist of a deep neural network, respectively; - determining (S50) a reconstruction error based on the at least one process signal data set (31) and the at least one reconstructed process signal data set (51) comprising determining a deviation of the at least one process signal data set (31) from the at least one reconstructed process signal data set (51); and - detecting (S60) an anomaly of the laser machining process based on the determined reconstruction error.

2. The method according to claim 1, further comprising: measuring (S10) at least some of the process signals of the process signal data set (31), and / or transmitting at least some of the process signals of the process signal data set (31) from a control (140).

3. The method according to one of the preceding claims, said method further comprising the steps: - determining (S70) a quality feature (x) of the laser machining process; and - evaluating (S80) the quality feature (x) as valid when no anomaly is detected in the step of detecting (S60) an anomaly of the laser machining process; and - evaluating (S90) the determined quality feature (x) as not valid when an anomaly of the laser machining process is detected in the step of detecting (S60) an anomaly.

4. The method according to claim 3, wherein, in the step of determining (S70) a quality feature (x) of the laser machining process by means of a regressor (600) formed by a neural network, a value for the quality feature (x) is determined, and / or wherein, in the step of determining (S70) a quality feature (x) of the laser machining process by means of a classifier (600) formed by a neural network, a classification value for the quality feature (x) is determined.

5. The method according to claim 4, wherein said autoencoder (400) and at least one of said regressor (600) and said classifier (600) have a common encoder (2), and / or wherein said regressor (600) and / or said classifier (600) determines (S70) the quality feature (x) based on data from an encoder (2) of said autoencoder (400).

6. The method according to claim 4, wherein said autoencoder (400) and at least one of said regressor (600) and said classifier (600) are parallel to each other, and / or wherein said regressor (600) and / or said classifier (600) determines (S70) the quality feature (x) based on the at least one process signal data set (31), and / or wherein said autoencoder (400) and at least one of said regressor (600) and said classifier (600) have a common input layer (1).

7. The method according to one of claims 4 to 6, wherein said autoencoder (400) and at least one of said regressor (600) and said classifier (600) are trained with the same data.

8. The method according to one of the preceding claims, wherein the step of determining (S50) a reconstruction error comprises: - determining a mean absolute or squared deviation of the at least one process signal data set (31) from the at least one reconstructed process signal data set (51); and / or - determining a signed, absolute or squared deviation summed up along the time axis; and / or - determining the Mahalanobis distance, in particular determining the Mahalanobis distance with respect to: - the deviation of the at least one process signal data set (31) from the at least one reconstructed process signal data set (51); and / or - individual characteristic values of the reconstruction error; and / or - the encoding of a process signal data set (31).

9. The method of claim 8, wherein said method comprises: - determining the parameters mean vector and covariance matrix of the Mahalanobis distance using defect-free or labeled data sets.

10. The method according to one of the preceding claims, wherein the reconstruction error for individual dimensions is determined (S50) separately and / or based on a metric and / or by means of a fast Fourier transformation and / or a wavelet transformation.

11. The method according to one of the preceding claims, wherein said method comprises: normalizing the reconstruction error with respect to the process signal data set (31) and / or wherein the step of determining (S50) a reconstruction error comprises: - filtering at least part of the process signal data set (31) and / or the reconstructed process signal data set (51); and - based thereon, determining the reconstruction error.

12. The method according to one of the preceding claims, wherein said method comprises: - determining a degree of abnormality, in particular determining a degree of abnormality based on a weighted summation or on the Mahalanobis distance with respect to individual characteristic values for the reconstruction error, wherein detecting (S60) an anomaly of the laser machining process is based on the determined degree of abnormality.

13. A laser machining method, comprising the steps of: - machining a workpiece by means of a laser beam (10); and - monitoring the laser machining process according to the method according to one of the preceding claims.

14. A system (300) for monitoring a laser machining process, said system (300) comprising: - at least one sensor assembly (310) configured to sense process signals (11) of the laser machining process; - at least one autoencoder (400); and - at least one processor (320) configured to carry out the method for monitoring the laser machining process according to one of claims 1 to 13.

15. A laser machining system (100) for machining a workpiece (18) by means of a machining laser beam (10), said laser machining system (100) comprising: - a laser machining head (101) for radiating the machining laser beam (10) onto said workpiece (18); and - a system (300) according to claim 14.