Method for process monitoring of a laser welding process in an interaction zone of a workpiece, and process monitoring arrangement for implementing the method

Abstract

The invention relates to a method for process monitoring of a laser joining process at an interaction zone 6 of a workpiece 2, wherein the laser joining process is observed with an event camera 7 and event data is generated by the event camera 7, wherein the event data is transformed into transformation data, wherein the transformation data is evaluated via a machine learning module 12 and the machine learning module 12 assigns a label 13 - 17 from a label set with at least two labels 13 - 17 for a test section of the laser joining process, wherein process information is assigned to the labels 13 - 17.

Description

[0001] The invention relates to a method for process monitoring of a laser joining process and a process monitoring arrangement for implementing the method. State of the art

[0002] In laser material processing, the manufacturing process is often monitored using sensors to provide online indications of errors. With appropriate process monitoring systems, it is possible to monitor 100% of the production process and thus implement 100% quality control cost-effectively. Microphones are typically used to record acoustic process emissions, and optical sensors to record optical process emissions. Cameras have proven particularly effective as optical sensors, as they provide a very broad data set for process monitoring.

[0003] The publication DE 10 2020 205 637 A1 discloses a monitoring device in which a dynamic vision sensor is used as the sensor, i.e. an event-based camera. Disclosure of the invention

[0004] The invention relates to a method for process monitoring of a laser joining process with the features of claim 1 and a process monitoring arrangement with the features of claim 15.

[0005] Preferred or advantageous embodiments of the invention will become apparent from the dependent claims, the following description, and the accompanying figures.

[0006] The invention relates to a method for process monitoring of a laser joining process at an interaction zone of a workpiece. The laser joining process can, in principle, be implemented as a laser soldering process. Preferably, the laser joining process is implemented as a laser welding process. In particular, two metallic components are joined together as the workpiece within the laser welding process. Specifically, the laser welding process is implemented as a laser deep penetration welding process.

[0007] The laser welding process is observed with an event camera, which is directed towards the interaction zone. The event camera is particularly preferred and easy to integrate if its beam path is arranged coaxially to the processing laser beam of the laser welding process. Alternatively or additionally, the event camera is positioned such that the interaction zone remains at the same position in the event camera's field of view throughout the laser welding process. The event camera generates event data during the laser welding process.

[0008] The event camera is also known as an event-based camera (EBC). It is also referred to as a neuromorphic camera, silicon retina, or dynamic vision sensor. The event camera features an image sensor that responds to local changes in brightness. Unlike conventional cameras, the event camera does not use a shutter to capture images. Instead, each pixel of the image sensor operates independently and asynchronously, reporting brightness changes when they occur and remaining inactive otherwise. The pixels of the event camera react independently to changes in brightness. Each pixel stores a reference value for brightness and continuously compares it to the current brightness level.When the brightness difference exceeds a threshold, the pixel resets its reference value and generates an event: a discrete packet containing the pixel address [x,y] and the timestamp [t]. Depending on the image sensor setup, events may also include the polarity [p] (increase or decrease) of a brightness change or an instantaneous measurement of the illuminance. The entirety of these events is referred to as event data. The event camera outputs an asynchronous stream of events as event data, triggered by changes in brightness within its field of view.

[0009] The event data is intended to be transformed into transformation data. This corresponds in particular to preprocessing the event data in order to bring it into an analyzable and / or meaningful format.

[0010] According to the invention, the transformation data is evaluated using a machine learning module, which utilizes learning algorithms to perform the evaluation. Specifically, the machine learning module is trained in a first step, optionally validated in a subsequent step, and can then perform the evaluation.

[0011] The machine learning module is trained to assign a label from a set of labels to at least one inspection segment of the laser welding process. The laser welding process can be divided into multiple inspection segments, with each segment corresponding to a specific time period within the process. In this case, a label can be assigned to some or all of the inspection segments. Alternatively, the inspection segment may encompass the entire laser welding process, in which case the entire process is labeled.

[0012] The label set has at least two labels, with process information assigned to the labels.

[0013] One aspect of the invention is that, by using the machine learning module, the advantages of an event camera for process monitoring can be combined with the advantages of machine learning-based analysis. The particular advantage of this combination lies in the fact that event data, as such, is difficult or impossible for a human user to interpret. Through transformation into process data and subsequent analysis by the machine learning module, the information from the event data can be converted into easily understandable labels for the user, with the labels describing or being assigned to the process information.

[0014] In a preferred embodiment of the invention, at least one label is assigned to an iO (OK) process information. This label indicates that the laser welding process was carried out correctly and / or without errors, according to specifications and / or based on the training data. Furthermore, at least one label is assigned to an niO (not OK) process information. This label indicates that the laser welding process was not carried out correctly and / or without errors, according to specifications and / or based on the training data. The niO process information label can be global and / or unqualified and may not contain any detailed information about the cause of the error. Alternatively, several niO process information labels can be present, each describing a single cause of the error. The niO process information can be structured, for example, as follows: - Welding defect process information, such as an indication of ejection, a void, a welding interruption, etc., - Position error process information, such as mispositioning of the processing laser beam, misposition of the workpiece, excessively large joining gap, etc. - Contamination fault process information, such as contamination of the workpiece in the area of ​​the interaction zone by oil or other contaminants, etc. - Weld quality process information, such as the weld seam profile and, optionally, the weld seam tightness. Tightness can be inferred, for example, from a good weld seam profile.

[0015] In one possible implementation, the transformation data is represented as an event rate. The event rate indicates the number of events occurring within each time period. Thus, the event rate counts how many events occur within each time period. The time periods can be arranged sequentially at equidistant intervals. Alternatively, the event rates could be calculated using a moving average filter. With this filter, a window of fixed width is shifted across the data, and a new data point is calculated from the values ​​within that window.

[0016] In a preferred further development, a first event rate is generated for a first image sub-region, where the first sub-region includes or is formed by the interaction zone. In particular, the first image sub-region is formed by the so-called keyhole in laser beam welding. A second event rate is generated for a second, different image sub-region. Optionally, further event rates can be generated. For example, event rates could be generated for the leading heat-affected zone or for the trailing weld pool.

[0017] In a preferred training course, the machine learning module uses a decision tree for evaluation and label assignment. A decision tree is a model from data analysis and machine learning used to classify data. It is based on a tree structure where each branch represents a decision or condition, and each leaf of the tree represents a final decision or class to which the data belongs. Essentially, the decision tree serves to systematically sort event rates based on criteria. It consists of nodes that represent specific characteristics of the data. Each node checks a condition and then splits the data into two or more branches, depending on whether the condition is true or false.

[0018] The structure of a decision tree begins at a "root node," which represents the entire dataset. At this point, an initial decision is made about the data, based on a specific attribute (characteristic). The data is then divided into different subsets accordingly. This process is repeated for each subset until a final classification into labels is reached. This final decision is represented in the so-called "leaves" of the tree.

[0019] In particular, the decision tree is given a combined data set comprising the first, the second event rate (and possibly further event rates) as input; the output of the decision tree is a label from the label set.

[0020] In an alternative embodiment of the invention, the transformation data is configured as a series of data matrices. In particular, the data matrices are configured as rectangular or square matrices with data points. The data matrices each correspond to one and / or the same section of the field of view of the event camera. For example, during the joining process, the interaction zone is always located in the same area of ​​the field of view and thus, from a data perspective, in the same area of ​​the data matrix. The data matrix can have the same dimensions as the event camera; alternatively, the dimensions can be changed for data processing reasons.

[0021] The data matrices each represent a time range of the event data. A separate data matrix is ​​created for each time range. Within each data matrix, a characteristic value for the events in the assigned time range is entered at the event's position in the field of view and / or at the corresponding position in the data matrix.

[0022] In this way, the event data is transformed into a series of data matrices, similar to images, with the data matrices forming transformation data. By transforming the event data into data matrices, it can be subsequently evaluated more easily and, due to preprocessing, is also more meaningful compared to the event data itself in conjunction with the evaluation.

[0023] One possibility is to define the parameter as a binary value, which can therefore take on two different values. The binary value takes a first value, for example, 1, if at least one event has occurred at the position in the data matrix and / or in the field of view within the specified time period. The binary value takes a second value, for example, 0, if no event has occurred at the position in the data matrix and / or in the field of view within the specified time period. The transformation data generated by this transformation step yielded very good results in the evaluation.

[0024] A second possibility is that the parameter is expressed as a frequency, where the frequency represents the number of events that occurred at the position in the data matrix and / or in the field of view. Specifically, the frequency is expressed as a natural number and corresponds to the number of events.

[0025] A third possibility is that the parameter is represented as time information. This time information indicates the point in time at which an event or multiple events occurred. For example, the time information represents the time of the first or the last event at a given position in the field of view and / or in the data matrix within that time range.

[0026] Firstly, the time periods can be arranged at equal intervals. This simplifies the transformation step. Secondly, the time periods can be chosen such that a predefined number of events has occurred within each period. The same predefined number of events is used in each time period. This configuration results in time periods of varying lengths, but the consistent number of events makes them highly comparable. In particular, data matrices can only be evaluated if "enough has happened," thus leveraging the advantage of the event camera: data is only generated when something occurs.

[0027] In an alternative configuration, the transformation data is represented as a process data matrix. Within this matrix, each event is represented by a parameter correctly positioned within a workpiece coordinate system. This parameter represents the time of the event. The process data matrix thus depicts the path of the laser beam on the workpiece, with the events correctly positioned. The events are represented by a time parameter. The process data matrix can depict the entire laser welding process; alternatively, it can show sections of the process. Dividing the process data matrix into sections is particularly advantageous for longer laser welding processes.

[0028] Alternatively, the parameter can be expressed as a binary parameter or as a frequency, as previously described. The binary parameter indicates whether an event occurred at the position. The frequency indicates how many events occurred at the position over a predefined time period or over the entire laser welding process.

[0029] The machine learning module can use any type of neural network, especially a deep one. It particularly favors a convolutional neural network (CNN), as extensive experience with such networks already exists in image processing, and this powerful technology can be advantageously transferred to process monitoring.

[0030] A further aspect of the invention is a process monitoring arrangement suitable and / or configured for implementing the method described above. The process monitoring arrangement includes the event camera described above. Furthermore, the process monitoring arrangement comprises an evaluation unit with an input interface for receiving the event data. A transformation module, which transforms the event data into transformation data, is connected to the input interface as a component of the evaluation unit. The evaluation unit includes the machine learning module, which evaluates the transformation data and assigns it to the respective label. The evaluation unit includes an output interface for outputting the labels.Optionally, the process monitoring arrangement includes a laser processing device which comprises the laser source and the scanner device or alternatively another beam guidance and / or shaping device for guiding and shaping the processing laser beam.

[0031] Further features, advantages, and effects of the invention will become apparent from the following description of preferred embodiments of the invention and the accompanying figures. These show: Fig. 1 a schematic representation of a process monitoring arrangement as an embodiment of the invention; Fig. 2 a block diagram, an evaluation device of the process monitoring arrangement, and an explanation of an exemplary embodiment of the method; Fig. 3 representations of different data matrices as transformation data from the event data.

[0032] The Fig. Figure 1 shows a schematic diagram of a process monitoring arrangement 1 as an embodiment of the invention. The process monitoring arrangement 1 includes a workpiece 2 on which a laser joining process, in this embodiment a laser welding process, in particular a laser deep penetration welding process, is carried out. A laser source 3 is provided which supplies a processing laser beam 4 for the laser joining process. For beam guidance and shaping, the process monitoring arrangement 1 includes a laser scanner 5, which shapes the processing laser beam 4 and guides it over the workpiece 2. During the laser joining process, an interaction zone 6 is formed in which the processing laser beam 4 is coupled into the workpiece 2 to carry out the laser joining process. Other beam guidance and shaping devices can also be used.

[0033] The process monitoring arrangement 1 includes an event camera 7, which is optically coupled such that the workpiece 2 is observed coaxially with the processing laser beam 4 at the interaction zone 6. The event camera 6 (event-camera EC), inspired by biological mechanisms, measures brightness changes pixel by pixel and offers advantages such as high temporal resolution, high dynamic range, low power consumption (since no routine processes need to be recorded), and low latency.

[0034] Unlike standard cameras that capture images at a fixed rate (e.g., 10 fps), event-controlled cameras (ECs) respond to brightness changes asynchronously and independently for each pixel. Therefore, ECs stream a data sequence of events. Each event represents a brightness change of a predefined magnitude at a pixel at a specific time. When the change exceeds a threshold, the camera sends an event. The event includes the x,y position of the change in the EC observation file, where x ∈ {1,...,X} and y ∈ {1,...,Y}, the time t at which the event occurred, and the 1-bit polarity p ∈ {-1,1}. The polarity indicates the type of change, with 1 representing an increase in brightness and -1 a decrease.

[0035] An event e is represented by the 4-tuple e=(x,y,t,p).

[0036] Therefore, the data sequence E supplied by an EC consists of N events. E={ei}i=1N.

[0037] The EC encoding described here is inspired by the spiking nature of biological visual pathways. Compared to standard cameras, ECs offer numerous potential advantages: ECs have extremely high temporal resolution and low latency (both on the order of microseconds), enabling these sensors to respond quickly to visual stimuli. Furthermore, ECs have a very high dynamic range (120 dB compared to 60 dB for standard cameras) and low power consumption.

[0038] The process monitoring arrangement 1 includes an evaluation unit 8, which is designed as a digital data processing unit, such as a computer. The evaluation unit 8 receives data from the event camera 7 and optionally data from the laser scanner 5.

[0039] The Fig. Figure 2 shows a schematic block diagram of the evaluation unit 8 and also serves as an illustration for an embodiment of the method according to the invention.

[0040] The evaluation unit 8 has an input interface 10 for receiving the event data from the event camera 7. In step 100, the laser welding process is thus carried out and the event data is made available to the evaluation unit 8, in particular to the input interface 10.

[0041] The evaluation unit 8 has a transformation module 11, in which the event data is transformed into transformation data. The transformation of the event data to the transformation data thus takes place in step 200.

[0042] The evaluation unit 8 includes a machine learning module 12, which is trained to evaluate the transformation data for a test section of the laser welding process and to assign a label 13, 14, 15, 16, or 17 from a label set 18 to the test section. Each label 13, 14, 15, 16, or 17 is assigned process information about the laser welding process, in particular quality information.

[0043] In this embodiment, label 13 is assigned an OK process information, indicating that the laser joining process was carried out correctly according to specifications. Labels 14, 15, 16, and 17 each indicate qualified errors in the laser joining process. Label 14 indicates that workpiece 2 was contaminated; label 15 indicates that a joining partner of workpiece 2 was missing; label 16 indicates that the laser beam 4 was defocused during processing; and label 17 indicates that the joining gap between the joining partners of workpiece 2 was too large. Thus, labels 14, 15, 16, and 17 refer to processing errors and their causes. From a physical perspective, the processing errors and their causes can be derived from the optical emissions during the laser joining process.

[0044] The evaluation unit 8 has an output interface 19, via which the labels 13, 14, 15, 16, 17 for the test section are output. Thus, in step 300 the transformation data is evaluated and in step 400 the labels 13, 14, 15, 16, 17 are output.

[0045] Labels 13, 14, 15, 16, and 17 are linked to the corresponding workpiece 2 as process information, so that the workpiece can be automatically sorted out, assigned quality information, or sent to a rework process based on labels 13, 14, 15, 16, and 17. In a simpler configuration, output interface 19 can be connected to a signaling device, such as a traffic light, which indicates that workpiece 2 should be sorted out.

[0046] In a first implementation variant (event rate), the transformation data is formed as an event rate ε, where the event rate indicates the number of events occurring per time interval.

[0047] An event rate transforms a data sequence E of event data into a time series. More precisely, an event rate indicates the number of events that occur in equidistant time intervals, per unit length of the intervals, as inspection segments of the laser welding process. Mathematically expressed, this means: Suppose a laser process was observed during the time interval T = [ts, te], resulting in an EC data sequence E. T is divided into L equal subintervals of length Δt as inspection segments, i.e., L = (te - ts) / Δt.

[0048] Then lk=ts+(k−1)⋅Δt Let k ∈ {1,...,L} define the start times of the subintervals. Let ft(e) be the time of the event e. Then G ⊆ E is defined as G(tc)={e|c1∧c2} With c1: e ∈ E and c2: tc ≤ ft(e) < tc+Δt, this includes all events in an interval of length Δt starting at tc. Then the event rate is ε ∈ ℕ. 1×L given by ε=(|G(lk)|Δt)k=1L.

[0049] Since the laser process is observed in situ, the laser beam always appears at the same position on the EC x,y-plane. The x,y-plane therefore lies in a local coordinate system, at the center of which the laser beam intersects the workpiece. In this setup, the processing laser beam 4 intersects the workpiece 2 at the center of the x,y-plane. To obtain separate information near the laser beam, e.g., about the keyhole, and information further away from the laser beam, the x,y-plane is divided into two regions: The first image sub-region is a circle with radius r around the center in the x,y-plane, and the second image sub-region is the x,y-plane without the circle. Therefore, an inner event rate εi and an outer event rate εo are defined below: d defines the distance of the x,y-position of an event e from the center in the x,y-plane. Therefore, Gi ⊆ G is defined as Gi={e|c1∧c2∧d≤r} and includes all events whose x,y positions lie within a circle of radius r centered at an interval of length Δt. Then the rate of events in the interior εi ∈ N is εi=(|Gi(lk)|Δt)k=1L.

[0050] Go also includes G, defined as G0={e|c1∧c2∧d>r} all events whose x,y positions lie outside the circle.

[0051] Then the rate of events outside ε0 ∈ ℕ 1×L ε0=(|G0(lk)|Δt)k=1L.

[0052] In step 200, the event data is fed into a decision tree algorithm (DT). This algorithm is first trained and validated using comparable data. The DT 20 was implemented using the Scikit-learn library. This library uses an optimized version of the CART algorithm. To measure the quality of a split, Gini impurity was used as a criterion. A grid search was performed to find suitable hyperparameters, namely the maximum tree depth and the minimum number of blocks required to split an internal node. Projection onto the xy-plane

[0053] In a second implementation variant, in step 200, the transformation module 11 projects the data onto an xy-plane in a camera coordinate system, so that the transformation data is formed as a series of data matrices 23, each data matrix 23 corresponding to a section of the event camera's field of view of the workpiece. In particular, the interaction zone is always located at the same position in the projection and / or in the data matrix 23. The data matrices 23 each represent a time range of the event data as an inspection section, with each data matrix 23 containing a characteristic value for the events in the assigned time range at the event position in the field of view within the data matrix 23.

[0054] This event representation in the form of data matrices 23 projects events into the x,y-plane, resulting in images of dimension X × Y. Three different such x,y projections are presented below: binary, density or frequency, and temporal.

[0055] Binary quantity: The binary projection marks all x,y positions where at least one event occurred during a time interval or time range. It is assumed that a laser welding process was observed during the time interval T = [ts,te], resulting in an EC data sequence E. C={(fx(e),fy(e))|c1} Let P be the set of x,y coordinates at which at least one event occurs. Then the entries of the binary projection P ∈ {0, 1} are X×Y Pxy={1,if(x,y)∈C,0,otherwise.

[0056] P as data matrix 23 therefore has the entry 1 as a characteristic value if at least one event occurs at the position (x,y), and 0 as a characteristic value if no event occurs. Density / Frequency

[0057] In addition to providing information about whether at least one event occurred at a given x,y position, the density projection also provides information about how many events occurred at that position. Let fx(e) and fy(e) be the x and y entries of an event e. Then D(x,y)={e|c1∧c3∧c4} With c3: (x,y) ∈ C and c4: (fx(e), fy(e)) = (x,y), the set of events occurring at position (x,y) is given. Then the entries of the density projection are Qx,y={|D(x,y)|D^,if D≠00,otherwise, where D̂ = max({D(x,y)}). The higher an entry in Q is, the more events have occurred at its (x,y) coordinates. The frequency of events is thus entered into data matrix 23 as a key parameter. Time-wise:

[0058] In addition to whether at least one event occurred at an x,y position or not, the temporal projection retains the temporal information of the events. Therefore, let G(x,y)={ft(e)|c1∧c3} The set of times of events occurring at position (x,y). Then the entries of the time projection R1 ∈ [0, 1] X×Y as data matrix 23 Rxy,1={max(G(x,y))te−ts,if G≠00,otherwise.

[0059] Each x,y position where an event occurred is scaled according to the time at which the event occurred by including time information as a parameter in data matrix 23. If multiple events occurred at the position (x,y), the longest time is used.

[0060] Analogously, the minimum time leads to the time projection R2 ∈ [0,1] X×Y Rxy,2={min(G(x,y))te−ts,if G≠00,otherwise.

[0061] Sampling: The data matrices 23 are each assigned to a time range.

[0062] The time intervals can each be of equal length and arranged at equal time intervals. It is possible for the time intervals to be consecutive without gaps. Alternatively, they can be designed as sliding time intervals, so that the time intervals overlap in time, with each overlap being of the same duration.

[0063] Suppose a laser welding process was observed during the time interval T, resulting in an EC data sequence E as event data. T is divided into equal subintervals of length Δt. The constant-time sampling strategy then generates the event representations, i.e., the data matrices 23, for each subinterval of length Δt. A special feature of considering a constant-time sampling strategy is that even if few or no events occur—i.e., no or only minor changes occur in the laser welding process—a representation in the form of a data matrix 23 is still generated.

[0064] Alternatively, the time intervals are arranged and / or chosen such that a predefined number of events is arranged within each time interval. The number of events is the same for each time interval. Assume that a laser welding process was observed during the time interval T, resulting in an EC data sequence E as event data. E is divided into subsets with the same number of events |S| = Δe. Thus, each entry of the event rate ε has the value Δe. This constant event sampling strategy then generates the event representations in the form of data matrices 23 for each subset with Δe events.

[0065] Projection along the processing path: In another implementation variant, a process data matrix for the entire laser welding process is generated as a test section or sections thereof in step 200 by transformation module 11. In the process data matrix, a characteristic value for each event is entered correctly in a workpiece coordinate system, specifically in a top view of the workpiece. This characteristic value is represented as a time value, a binary quantity, or a frequency. The time value represents the point in time of the event.

[0066] This event representation additionally uses the recorded scanner data or the processing path during the laser welding process. This data includes the xs, ys position on the laser's processing plane and the corresponding time ts. Therefore, a scanner data input s is represented by s=(xs,ys,ts) Let ft(s) be the time of an entry s. Then S=(si)i=1M, with ft(si) ≤ ft(si+1) the sequence of M scanner entries with ascending time entries. N(i)={e:c1∧c5} with c5: ft(si) ≤ ft(e) < ft(si+1) are all events that occur between two consecutive time entries of the scanner.

[0067] Let fx(s) and fy(s) be the xs and ys entries of s. Then the midpoint (cx,cy) of the EC observation field at time ft(e) for an event e ∈ N(i) lies at Cx(e,i)=ft(e)ft(si+1)−ft(si)⋅(fx(si+1)−fx(si)), and cy(e,i) accordingly in the laser's processing plane. γ(·) converts a distance in the laser's processing plane into the corresponding number of pixels in the EC observation field.

[0068] Then the entries of the projection along the scanner are V ∈ {0,1} bx×b with b X =γ(max(f x (s))-min(f x (s)))+X and by accordingly Vjk={1,if j=γ(cx(e,i)−min(fx(s)))+fx(e)−X / 2 and k=γ(cy(e,i)−min(fy(s)))+fy(e)−Y / 2, where e∈N(i),i∈{1,2,…,M−1}0,otherwise.

[0069] In step 300 of the machine learning module 12, the data matrices 23 and / or the process matrix(s) are fed to a CNN algorithm 21 (Convolutional Neural Network) for evaluation. A CNN 21 was trained as grayscale images for the binary projection P, the density projection Q, and the time projection R1. This CNN then assigns the labels 13, 14, 15, 16, and 17, which are output by the output interface 19 in step 400.

[0070] The dataset consists of 249 welded metal plate pairs. The number of metal plate pairs for training, validation, and testing is determined by ntrain (60%), nval (20%), and ntest (20%). Data enrichment with adjustment factors was integrated for CNN 21.

[0071] To teach CNN 21 invariance and robustness properties, data magnification using rotation, reflection, translation, and shearing was employed. The image size was changed from 720 × 1280 pixels to 100 × 177 pixels. The models were developed and trained using TensorFlow. A variety of CNNs were tested: initially, the networks were kept shallow and later expanded to deeper networks. The kernel size was varied between (3,3) and (7,7). In addition to convolutional layers, max pooling, global average pooling, stack normalization, and dropout were also used. Various activation functions such as ReLU, LeakyReLU, and tanh were tested. The final layer used softmax activation. During the training of CNN 21, categorical cross-entropy was used as the loss function. Furthermore, an Adam optimizer with a learning rate Ir = 1 - 10⁻⁴ was employed.In this way, CNNs 21 could be generated with sufficient accuracy.

[0072] The Fig. Figure 3 shows different data matrices 23 generated by the binary transformation. The respective data matrices 23 are assigned different labels 13, 14, 15, 16, and 17, as previously described. Even without the machine learning module 12, differences in the laser welding process are already apparent from the representation. For the other data matrices 23, the displayed pixels are, for example, color-coded. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] DE 10 2020 205 637 A1

[0003]

Claims

[1] Method for process monitoring of a laser joining process at an interaction zone (6) of a workpiece (2), wherein the laser welding process is observed with an event camera (7) and event data is generated by the event camera (7), where the event data is transformed into transformation data, wherein the transformation data are evaluated via a machine learning module (12) and the machine learning module (12) assigns a label (13 - 17) from a label set with at least two labels (13 - 17) for a test section of the laser joining process, wherein process information is assigned to the labels (13 - 17). [2] Method according to claim 1, wherein at least one label (13) is assigned to an iO process information and at least one label (14 - 17) is assigned to an niO process information. [3] Method according to claim 1 or 2, wherein the at least one label (14 - 17) is assigned to a weld defect process information, a positional defect process information, a weld quality process information and / or a contamination defect process information. [4] Method according to any of the preceding claims, wherein the transformation data is configured as an event rate, the event rate indicating the number of events occurring per time interval. [5] Method according to claim 4, wherein a first event rate is formed for a first image part area of ​​the event camera (7) comprising the interaction zone (6) on the workpiece (2) during the laser joining process and a second event rate is formed for a second, different image part area, wherein the first and the second event rate form the transformation data. [6] Method according to claim 4 or 5, characterized by, that the machine learning module (12) uses a decision tree (20) to evaluate the event rates. [7] Method according to one of claims 1 to 3, wherein the transformation data are formed as a series of data matrices (23), wherein each data matrix (23) corresponds to a section of the field of view of the event camera (7), wherein the data matrices (23) represent a time range of the event data, wherein in each data matrix (23) for the assigned time range a characteristic value for the events in the time range is entered at the position of the events in the field of view. [8] Method according to claim 7, wherein the characteristic parameter is designed as a binary quantity, wherein the binary quantity takes a first value if at least one event has taken place at the position in the time domain and a second value is entered if no event has taken place at the position in the time domain. [9] Method according to claim 7, wherein the characteristic parameter is designed as a frequency, the frequency representing the number of events that have taken place at the position in the time domain. [10] Method according to claim 7, wherein the characteristic parameter is designed as a time information, the time information representing information about the time of the first or the last event at the position in the time domain. [11] Method according to any one of the preceding claims 7 to 10, wherein the time intervals are arranged at equal time intervals. [12] Method according to any one of the preceding claims 7 to 10, wherein the time intervals are arranged such that a predefinable number of events is arranged in each of the time intervals. [13] Method according to any one of the preceding claims 1 to 3, wherein the transformation data are formed as a process data matrix, wherein a characteristic value for the events is entered in the process data matrix in the correct orientation, wherein the characteristic value is formed as a time information, wherein the time information represents the time of the event, as a binary quantity or as a frequency of the events. [14] Method according to any one of the preceding claims 7 to 13, characterized by , that the machine learning module (12) uses a CNN algorithm (21) to evaluate the data matrices (23) and / or the process matrix or matrices. [15] Process monitoring arrangement (1) for implementing the method according to one of the preceding claims, comprising the event camera (7) and an evaluation unit (8), wherein the evaluation unit (8) comprises an input interface (10), a transformation module (11) for transforming the event data into transformation data, the machine learning module (12) and / or an output interface (19) for outputting the label (13 - 17).

Images