Processor circuit and method for monitoring a joining process of workpieces in an automatic joining machine, and automatic joining machine

Description

[0001] The invention relates to a method and a processor circuit for monitoring a joining process of workpieces in an automatic joining machine in order to detect faulty joined workpieces. The invention also includes an automatic joining machine, in particular a welding machine (e.g., a welding robot), with the said processor circuit.

[0002] Today, quality control during joining processes in automotive body construction, such as resistance spot welding, is primarily performed using ultrasonic testing. The inspection is performed manually with the assistance of an ultrasonic device. The quality results are determined by visual inspection of the test signal, i.e., the ultrasonic image, and the measurements of the testing device. The inspection is performed via random sampling, in which a component is randomly removed and its quality checked.

[0003] If defects are detected in the joined workpieces of a component, countermeasures must be taken in the form of parameter adjustments in the welding control and / or rework of the components, depending on the degree of the defect.

[0004] However, quality control is not "comprehensive," meaning not all car bodies / weld points are consistently inspected. Therefore, the inspection depends on how representative the sample is of the overall population. However, welding is performed using adaptive welding control, meaning each weld is individually controlled, which is why consistent inspection is important.

[0005] With a sample size of 3 components and a total population of 600 vehicles, assuming a confidence level of 80%, due to the high process variance, conclusions can only be drawn from the sample to the total population with a probability of 38%.

[0006] US 2003 / 0234239 A1 discloses the use of ultrasound to inspect the quality of spot welds in a spot welding process. The resulting inspection signals are fed as input to an artificial neural network trained to evaluate the quality of spot welds based on inspection signals. This process also requires each spot weld to be inspected to be examined using ultrasound.

[0007] WO 2020 / 106725 A1 discloses generating control signals for a joining process using a machine learning model. To train the model, simulation data on the joining process can be generated, which can then be used for training. The model can also be used to predict poor-quality joints. The training is specifically carried out based on a large number of different workpieces with different shapes and / or properties. This makes the acquisition of training data particularly complex.

[0008] DE 10 2019 209100 A1 describes a method for monitoring a laser beam welding process. The welding process is observed using an event-based sensor divided into pixels, which outputs an event associated with a pixel whenever the light intensity incident on that pixel changes by at least a specified percentage. From the total number of events output by the sensor, subsets of events are determined, each resulting from different splashes of molten material released during the welding process. A quality assessment for the welding process and / or the workpiece is evaluated from the determined subsets of events. The events output by the sensor can be fed into an artificial neural network (ANN).The events can also be fed directly to the ANN, allowing the ANN to take over the task of identifying subsets of events resulting from different splashes. The ANN can support the determination of the quality assessment or take over the entire process.

[0009] EP 3 812 105 A1 describes a system and method for controlling a welding robot using various machine learning methods. The method includes obtaining operating data of the welding robot during welding and obtaining image data of a welded workpiece. The image data is processed to determine parameters of the welded workpiece, and the parameters are processed to determine a quality of the welded workpiece by evaluating the parameters using an evaluation function. The results of the evaluation function are input into a machine learning model. The results of the evaluation function are applied to the obtained operating data using the machine learning model to modify the operating data to improve welding. The modified operating data is used to control the welding robot.

[0010] The invention is based on the object of monitoring an automatic joining machine in such a way that inferior joined workpieces are detected and sorted out.

[0011] The problem is solved by the subject matter of the independent patent claims. Advantageous developments of the invention result from the features of the dependent patent claims, the following description, and the figure.

[0012] As one solution, the invention comprises a method for monitoring a joining process of workpieces in an automatic joining machine. A joined workpiece can be created by joining at least two parts to be joined. Machine learning is used to determine the quality of the joints. In a training phase for a machine learning estimation model, precursor workpieces are joined (permanently connected) in the automatic joining machine and / or in at least one other automatic joining machine using the joining process (from precursor parts to be joined). In this process, a digital process image of the joining process is recorded for the respective precursor workpiece, formed from at least one predetermined process parameter and at least one measured and / or calculated process variable. The precursor workpieces are of the same workpiece type as the workpieces themselves.In other words, the joining process for workpieces of the workpiece type to be monitored is observed (these workpieces are referred to here as precursor workpieces). The workpieces can be, for example, frames or car bodies made from sheet metal or profile parts. The joining can involve, for example, welding and / or soldering and / or gluing. The process parameters available for the joining process (specified variables for the joining process) and process variables (as they arise dynamically during the course of the joining process, i.e. a respective process signal) are summarized or stored in the digital process image. The respective process image thus represents a data set that describes a joining run of specific workpieces, i.e. a single run of the joining process. Corresponding data can be read out or recorded, for example, from a fieldbus and / or from sensors.The resulting quality must be assigned to this respective process image for each joining pass.

[0013] For this purpose, a test signal that correlates with or is dependent on the joining quality of the joined workpieces (i.e. the manufactured component) is determined on the joined precursor workpieces using a predetermined test method, and a quality indication that signals the determined joining quality is generated using an evaluation method. Firstly, a test signal, for example a measurement signal, can be generated using a conventional test method, for example by manual testing using a testing device. A person skilled in the art can determine which test method is suitable based on the joining process to be monitored. Test signals that correlate with the joining quality to be monitored, for example of joints, are known per se from the prior art. A test signal can be, for example, a received ultrasound signal.The evaluation method specifies an assignment of a test signal to a quality specification (for example, a binary quality specification "sufficient" and "unsatisfactory").

[0014] The assignment of process image to the resulting quality specification must now be automated. Using a predetermined training method, a machine learning estimation model is generated. This model receives the respective digital process image as input and outputs an estimated quality specification for the input, which is trained according to the signaled quality specification. For example, an artificial neural network and / or a decision tree model can be used as the estimation model. However, these are only examples, and a particularly advantageous embodiment of the estimation model will be described later in the description. The training method can be based on the well-known supervised learning and / or unsupervised learning.

[0015] Once the estimation model has been created or trained, it can be used during operation of the joining machine when joining the workpieces to be monitored. During this operating phase, in which the joining machine joins the workpieces to be automatically monitored using the joining process, the respective digital process image of the joining process for the respective workpiece is also determined and fed into the estimation model as input. The estimated quality information for the joined workpiece is generated using the estimation model. The actual testing method can therefore be omitted or omitted here. For the workpieces joined by the joining machine, a rejection measure is then triggered to sort out those workpieces for which the quality information estimated by the estimation model meets a predetermined rejection criterion.

[0016] As soon as the estimation model is available, an immediate, automated assignment of the process image to an estimated quality specification can be performed. A test using the test method is no longer necessary at this stage. The rejection measure can be triggered whenever it is determined that the estimation model estimates a quality specification for joined workpieces that corresponds to or fulfills the rejection criterion. By defining the rejection criterion, the specialist can decide or control which joined workpieces should be subjected to the rejection measure, i.e., which should be classified or assessed as defective or of inadequate quality.

[0017] The invention offers the advantage that precursor workpieces can be manufactured using the joining machine in a conventional joining process, and the training material or training data for the estimation model can be determined (incidentally or incidentally). Once the training or calculation of the estimation model is completed, it is possible to switch to automated quality monitoring using the estimation model, i.e., the operating phase for the estimation model can be started. In this case, the testing method can be advantageously dispensed with.

[0018] The invention also includes further developments which result in additional advantages.

[0019] A further development comprises the training method comprising a machine learning algorithm and / or the estimation model comprising an artificial neural network and / or a decision tree and / or at least one of the following estimation algorithms: Random Forest, Support Vector Machine, SVM, and XGBoost, to name a few examples. In addition to or as an alternative to the known machine learning algorithms, such as artificial neural networks and / or decision trees, the estimation model can thus also be based on an estimation algorithm or several combined estimation algorithms, such as those mentioned. Machine learning or machine learning refers to adapting the configuration parameters of these estimation algorithms to the training data as determined in the training phase (specified or labeled assignment of digital process images to quality specifications), which can also be done automatically.The methods of supervised and / or unsupervised learning described above can be used.

[0020] A further development involves combining several or all of the estimation algorithms in the estimation model according to the majority voting principle, known as ensemble learning. This has proven to be a particularly advantageous variant of an estimation model. In particular, two or three of the aforementioned estimation algorithms (Random Forest, SVM, XGBoost) can be run, i.e., they can be supplied with the respective process image as input or input data, after which an estimated quality indication or a binary intermediate indication with regard to the rejection criterion (met or not met) is available for each estimation algorithm. Then, for example, according to the majority voting principle (more than 50 percent or more than X percent with X greater than 50), a single final statement or a final signal can be determined or summarized as to whether the rejection criterion is met overall.It has been shown that this can compensate for weaknesses or inadequacies of individual estimation algorithms.

[0021] A further development comprises scaling or stretching / compressing a respective time signal of at least one process variable to a predetermined signal length using dynamic time warping to generate the inputs (for the estimation model), and in particular comparing the signal length with a reference signal by the estimation model. This allows time signals of different lengths to be processed uniformly. Time signals of different lengths can arise, for example, if, in a controlled joining process, the control allows the joining process to last for different lengths of time for each workpiece to be joined (e.g., with a tolerance or fluctuation in duration in a range of 1 second to 5 minutes). Nevertheless, a uniform predetermined signal length can be set using dynamic time warping, so that workpiece joining processes of different lengths can also be evaluated or checked using the same estimation model.The reference signal can specify a target curve that must result if the joining process runs smoothly.

[0022] A further development comprises the process parameter providing for or taking into account at least one of the following process parameters: a sheet thickness of at least one of the workpieces, a geometry of the workpieces. The process parameters therefore represent static or predetermined sizes or quantities, such as can be taken from design data, e.g. CAD data (CAD - Computer Added Design). This advantageously makes it possible to interpret or evaluate a process variable depending on a process parameter. For example, different values ​​of the sheet thickness can result in different time signals and / or different values ​​for process variables and yet a sufficiently high quality can always be achieved (separation criterion not met).This can be recognized or correctly interpreted by always considering or observing a process parameter, such as sheet thickness, and interpreting or evaluating at least one process variable differently depending on the current value of at least one process parameter. Conversely, one and the same value for a process variable can indicate a fulfilled rejection criterion (poor quality) and another an unfulfilled rejection criterion (adequate quality), depending on the value of at least one process parameter, e.g., depending on the sheet thickness.

[0023] A further development includes the at least one process variable taking into account or providing for at least one of the following process variables: an intervention signal from a process control system in the joining process; a current intensity, a voltage signal, a voltage intensity, an electrical resistance, a transport speed of the sheet metal pieces and / or a joining material. These process variables have proven particularly meaningful in determining a quality specification.

[0024] A further development involves monitoring more than 80% or all of the joined workpieces using the estimation model during the operational phase. In other words, not only is a random inspection of individual joined workpieces and / or their joints necessary or possible, but more than 80 percent of all joined workpieces or all joined workpieces can be inspected.

[0025] A further development involves applying the testing method during the operational phase to only some of the joined workpieces or to none of the joined workpieces. In particular, it is possible to dispense with the testing method, for example, manual testing using an ultrasonic device, or to monitor only selected joined workpieces and / or a predetermined sample using the testing method.

[0026] A further development includes the testing method comprising an ultrasonic test, and the test signal comprising an ultrasonic image of a joint created by the joining process. Ultrasonic testing has proven particularly suitable as a testing method for training an estimation model, as its test signal has proven particularly suitable for use in the field of machine learning. Alternative or additional testing methods include at least one of the following: X-ray analysis, load testing.

[0027] A further development includes the evaluation method providing for the test signal to be presented to an operator at an operator interface, and for user input from the operator to be received via the operator interface. The user input can contain or specify the quality information. The evaluation method can therefore also provide user feedback from an operator, i.e. a user input, so that the operator's expertise can be incorporated or used when generating training data. The operator can, for example, have the test signal displayed on a screen, such as the described ultrasound image, and the user input regarding the quality information can be received via an input element, such as a keyboard and / or a mouse and / or a GOI (Graphical User Interface). The user input can then be used as a so-called label for the training phase in machine learning.

[0028] A further development stipulates that the separation measure includes examining the separated joined workpieces using the test method and / or scrapping them. A separated joined workpiece can therefore be re-examined using the test method. This avoids unnecessary scrap. Scrapping can be provided as an additional or alternative separation measure.

[0029] A further development includes the joining process being a welding process, in particular resistance spot welding. In connection with a welding process, in particular resistance spot welding, a particularly high correlation has been shown between a digital process image, such as can be generated by an automatic joining machine during operation while performing the joining process, and the estimated quality specification. This means that an incorrect estimate is particularly low for a welding process, in particular resistance spot welding, which can provide or offer one or more electrical variables as a process image.

[0030] As a further solution, the invention comprises the process solution with a processor circuit for monitoring an automatic joining machine, wherein the processor circuit is configured to determine process parameters and process variables of a workpiece joining process during an operating phase of the automatic joining machine and to feed them as input to an estimation model. Using the estimation model, it generates a respective estimated quality indication regarding the joined workpieces and triggers rejection measures for rejecting those workpieces for which the estimated quality indication meets a predetermined rejection criterion. The processor circuit can have a data processing device or a processor device configured to carry out the described method steps.For this purpose, the processor circuit can have at least one microprocessor and / or at least one microcontroller and / or at least one FPGA (Field Programmable Gate Array) and / or at least one DSP (Digital Signal Processor). Furthermore, the processor circuit can have program code that is configured to carry out the method steps when executed by the processor device. The program code can be stored in a data memory of the processor circuit. The processor circuit can be provided, for example, as a component of a control circuit or controller of the automatic joining machine. The processor circuit can also be provided separately from the automatic joining machine, for example in a control center or in a central computer in a factory shop. The processor circuit can be implemented as a network of multiple control circuits, i.e. as a distributed solution.The processor circuit can be implemented entirely or partially as an Internet server, for example, as a cloud server. The method can be implemented or executed in a virtual machine. The described data, e.g., the digital process image, can be transmitted from the joining machine to the processor circuit, for example, via a communication connection, e.g., an IP connection (Internet connection).

[0031] As a further solution, the invention comprises an automatic joining machine, in particular an automatic welding machine (e.g., a welding robot), with an embodiment of the processor circuit according to the invention. The automatic joining machine can, for example, provide a robot for the joining or the joining process. The processor circuit can be coupled to the robot. The processor circuit can be coupled to at least one sensor of the automatic joining machine and / or at least one controller of the automatic joining machine in order to determine the described digital process image.

[0032] The invention also encompasses combinations of the features of the described embodiments. The invention therefore also encompasses implementations that each have a combination of the features of several of the described embodiments, unless the embodiments are described as mutually exclusive.

[0033] Exemplary embodiments of the invention are described below. Shown are: Fig. 1 is a schematic representation of an embodiment of the joining machine according to the invention; Fig. 2 is a flow chart illustrating an embodiment of the method according to the invention.

[0034] The exemplary embodiments explained below are preferred embodiments of the invention. In the exemplary embodiments, the described components of the embodiments each represent individual features of the invention that can be considered independently of one another, each of which also develops the invention independently of one another. Therefore, the disclosure is intended to encompass combinations of the features of the embodiments other than those shown. Furthermore, the described embodiments can also be supplemented by further features of the invention already described.

[0035] In the figures, the same reference symbols denote elements with the same function.

[0036] Fig. 1 shows an automatic joining machine 10, which can be, for example, an automatic welding machine, in particular for resistance spot welding. By means of the automatic joining machine 10, workpieces 11 can be connected or joined to one another at joining points 13 in a joining process 12. A joined workpiece 11 can be produced from two or more than two joining parts, e.g., metal sheets, by the automatic joining machine 10. This can involve the production of motor vehicle bodies. In the automatic joining machine 10, a processor circuit 14 can estimate or determine whether the quality of the joining points 13 on the workpieces 11 is sufficient or whether (conversely) the joined workpieces 11 meet a predetermined rejection criterion 15. According to the check by means of the rejection criterion 15, the joined workpieces 11 can be deemed to be in order for further use 16 (e.g.,into a motor vehicle) or as not in order and thus subjected to a predetermined separation measure 17, which in . Fig. 1 symbolized by a switch. Using the processor circuit 14, it is not necessary to subject the joined workpieces 11 themselves to a separate or dedicated testing method, for example, using ultrasound. Furthermore, all workpieces 11 joined in the joining process 12 can be examined, meaning that there is no need to limit sampling to less than 100 percent, in particular less than 80 percent.

[0037] This can be achieved by the processor circuit 14 determining a digital process image 18 from the joining machine 10 and / or from predetermined design data, e.g., CAD data, during the execution of the joining process 12. The process image 18 can receive at least one process parameter 19, for example, a geometry of the processed or joined parts, and at least one process variable 20, for example, a controller signal and / or a current signal and / or a voltage signal, as they were set or adjusted, for example, by at least one controller of the joining machine 10 during the execution of the joining process 12, for the workpieces 11 to be joined. The digital process image 18 can be fed to an estimation model 21, which can be operated by the processor circuit 14.The estimation model 21 can, for example, be based on software that can be executed or operated by at least one microprocessor of the processor circuit 11. The estimation model 21 can, in particular, be based on a plurality of estimation algorithms 22, all of which can be supplied or operated individually or independently of one another with at least part or all of the data of the process image 18. In . Fig. 1 Three estimation algorithms 22 are shown as examples. Ellipsis 23 is intended to indicate that more or fewer than three estimation algorithms 22 can be used. Examples of the estimation algorithms have already been mentioned. A respective estimation output 24 of each of the estimation algorithms 22 can be summarized by means of a majority voting principle 25 such that, if a predetermined majority of the estimation algorithms 22 (e.g., more than 50%) signals that a quality is less than a predetermined threshold value or that a respective algorithm-specific rejection criterion is met, the rejection criterion 15 is considered to be met overall. Otherwise, if the majority signals that the quality of the joints 13 is sufficient, the rejection criterion 15 is not met.

[0038] To generate the estimation model 21, the joining process 12 can be carried out in advance or in historical joining runs on precursor workpieces 30. For this purpose, the joining machine 10 and / or at least one other, similar joining machine (not shown) can be used. The joined precursor workpieces 30 can be analyzed or tested using a predetermined test method 31, for example, an ultrasonic test, resulting in a test signal 32 in addition to the respective process image 18 for the joined precursor workpieces 30. The test signal 32 can be presented to an operator 33, for example via a user interface UI, which can comprise a screen, for example, and the operator 33 can receive a user input 34 as a predetermined or ground truth or quality indication 35, for example via a keyboard or generally a GUI.Using the quality information 35 and the respective process image 18 for the joined precursor workpieces 30, the estimation model 21 can be calculated, generated, or determined in a training phase 37 using a training algorithm or training method 36. This trained estimation model 21 can then be transferred to the processor circuit 14, where it can then be used in an operating phase 38 in the manner described.

[0039] The training phase 37, in particular the training method 36, can be carried out in a data processing system 39, which can provide a computer or a computer network to carry out the training method 36.

[0040] Fig. 2 illustrates the training phase 37 once again. It shows how production data from the joining process 12 for the precursor workpieces 30 can be combined into the digital process image 18, for example, time signals 40, which can be adapted to a predetermined signal length 42 using dynamic time warping 41, for example. The estimation model 21 can then be generated using the training method 36.

[0041] The quality information 35 obtained using the precursor workpieces 30 and the associated process images 18 represent historical data 43 that can be determined when executing the joining process 12. The historical data 43 can be divided into training data 44 and test data 45. The training data 44 can be used to operate the training method 36, which can include, for example, supervised learning 46 and / or unsupervised learning 47. The estimation model 21 resulting from the training method 36 can then be operated on the basis of the test data 45 using a test method 48, as is known per se from the prior art, in order to determine a test metric 49 that signals whether the historical data 43 is correctly recognized, classified, or categorized using the estimation model 21.If this is detected by means of the test metric 49, the actual operating phase 38 can take place, in which current or real-time process data of the process image 18 for the workpieces 11 can be checked by the estimation model 21 in the manner described in order to control a control signal 50 depending on the check of the rejection criterion 15 for feeding the joined workpieces 11 to the further use 16 or the rejection measure 17, as already described.

[0042] A machine learning (AI) algorithm enables the automated analysis of large amounts of data, such as those generated by digital process images. They can identify predetermined patterns and anomalies in the collected data from digital process images.

[0043] Such a method is therefore used for quality inspection. An algorithm is trained on the quality results (e.g., "thin spot"). Process variables and process parameters serve as input for training the algorithm. The algorithm recognizes patterns in these and links them to the output variables in the form of quality results. Using the digital process image of the welding control system, the algorithm can perform a 100% inspection of all weld spots in real time.

[0044] The algorithm automatically checks all weld spots for quality. This eliminates the need for a non-representative sampling procedure. As a result, product quality can be improved and process efficiency can be increased.

[0045] In the Fig. 2An implementation of the method is shown. Several algorithms are used to monitor quality results. These are prepared for pattern recognition using a historical training dataset. They link input variables with a quality result. Once training is complete, the algorithm is validated using a historical test dataset. If the algorithm passes the test, it is used in the quality process.

[0046] Data from the welding process is streamed or transmitted to the algorithm. This algorithm determines the learned quality results, preferably in real time during the joining process, and predicts them based on the detected patterns.

[0047] Overall, the examples demonstrate how a machine learning algorithm can be deployed to predict ultrasonic test results.

Claims

1. Method for monitoring a joining process (12) of workpieces (11) in an automatic joining machine (10), wherein in a training phase (37): precursor workpieces (30) are joined in the automatic joining machine (10) and / or in at least one other automatic joining machine (10) by means of the joining process (12) and hereby a digital process image (18) of the joining process (12), which is formed from at least one predetermined process parameter (19) and at least one measured and / or calculated process variable (20), is recorded for the respective precursor workpiece (30), wherein the precursor workpieces (30) are the same type of workpiece type as the workpieces (11), and a test signal (32), which correlates with a joining quality of the joined workpieces (11), is determined on the joined precursor workpieces (30) by means of a predetermined testing method (31), and a quality indication (35) signaling the determined joining quality is generated by means of an evaluation method, and an estimation model (21) of the machine learning is generated by means of a predetermined training method (36), which receives the respective digital process image (18) as input and assigns an estimated quality indication (35) to the input as output, which is trained in accordance with the signaled quality indication (35), and in an operating phase (38), in which the automatic joining machine (10) joins the workpieces (11) by means of the joining process (12): the digital process image (18) of the joining process (12) of the respective workpiece (11) is determined and supplied to the estimation model (21) as input and the estimated quality indication (35) relating to the respective joined workpiece (11) is generated by means of the estimation model (21), and for the workpieces (11) joined by the automatic joining machine (10), a rejection measure (17) is triggered for rejecting those workpieces (11) for which the estimated quality indication (35) meets a predetermined rejection criterion (15).

2. Method according to claim 1, wherein the training method (36) comprises a machine learning algorithm and / or the estimation model (21) comprises an artificial neural network and / or a decision tree and / or at least one of the following estimation algorithms (22): - Random Forest, - Support Vector Machine, SVM, - XGBoost.

3. Method according to claim 2, wherein in the estimation model (21) multiple or all of the estimation algorithms (22) are operated in combination according to the majority voting principle (25).

4. Method according to any one of the preceding claims, wherein, for generating the inputs, a respective time signal (40) of at least one process variable (20) is scaled to a predetermined signal length (42) by means of a dynamic time warping (41) and is compared with a reference signal in the estimation model (21).

5. Method according to any one of the preceding claims, wherein the at least one process parameter (19) comprises at least one of the following process parameters: a sheet thickness of at least one of the workpieces (11), a geometry of the workpieces (11).

6. Method according to any one of the preceding claims, wherein the at least one process variable (20) comprises at least one of the following process variables: an intervention signal of a process controller in the joining process (12); a current strength, a voltage signal of a voltage strength, an electrical resistance, a transport speed of the sheet metal pieces and / or of a joining material.

7. Method according to any one of the preceding claims, wherein in the operating phase (38) more than 80% or all of the joined workpieces (11) are monitored by means of the estimation model (21).

8. Method according to any one of the preceding claims, wherein in the operating phase (38) the testing method (31) is applied to only some of the joined workpieces (11) or to none of the joined workpieces (11).

9. Method according to any one of the preceding claims, wherein the testing method (31) comprises an ultrasound test and the test signal (32) comprises an ultrasound image of a joint (13) generated by the joining process (12).

10. Method according to any one of the preceding claims, wherein the evaluation method comprises presenting the test signal (32) to an operator (33) at an operator interface and receiving a user input (34) from the operator (33) via the operator interface, which user input shows the quality indication (35).

11. Method according to any one of the preceding claims, wherein the rejection measure (17) comprises that the respective rejected workpiece (11) is inspected by means of the testing method (31) and / or is scrapped.

12. Method according to any one of the preceding claims, wherein the joining process (12) is a welding process, in particular resistance spot welding.

13. Processor circuit (14) for monitoring a joining process (12) of workpieces (11) in an automatic joining machine (10) by a method according to claim 1, wherein the processor circuit (14) is configured, in an operating phase (38) of the automatic joining machine (10), to determine process parameters (19) and process variables (20) of the joining process (12) of workpieces (11) and to supply them as input to an estimation model (21) and, by means of the estimation model (21) to generate a respective estimated quality indication (35) relating to the joined workpieces (11) and to trigger a rejection measure (17) for rejecting those workpieces (11) for which the estimated quality indication (35) meets a predetermined rejection criterion (15).

14. Automatic joining machine (10), in particular automatic welding machine, with a processor circuit (14) according to claim 13.