Method for the non-destructive testing of a connection between two components, method for producing a welded connection between at least two components, welding device, computer program product, computer-readable storage medium, method for training a machine learning model for predicting a welding quality, and machine learning model for predicting a welding quality
The method addresses the unreliability of existing weld quality prediction by using process data and AI models to provide non-destructive testing and adaptive welding adjustments, enhancing reliability and efficiency.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-09-08
- Publication Date
- 2026-04-02
AI Technical Summary
Existing methods for predicting the quality of welded joints, particularly ultrasonic welded joints, are unreliable and often destructive, leading to high costs and time consumption.
A method for non-destructive testing of welded joints using process data, including temperature profiles and other parameters, to determine an assessment value and an expected value, with confidence thresholds for quality prediction, and optional extended testing or adjustment of welding processes based on AI models.
Enables reliable and repeatable prediction of weld quality, reducing the need for destructive testing and improving the efficiency of welding processes through continuous learning and adaptive parameter adjustment.
Smart Images

Figure EP2025075454_02042026_PF_FP_ABST
Abstract
Description
[0001] PTESI OIWO / 04 . 09. 2025 1 2025094 075
[0002] Method for non-destructive testing of a connection between two components, method for producing a welded joint between at least two components, welding device, computer program product, computer-readable storage medium, method for training a machine learning model to predict weld quality, and machine learning model for predicting weld quality
[0003] The invention relates to a method for non-destructive testing of a connection between two components, a method for producing a welded connection between at least two components, a welding device, a computer program product, a computer-readable storage medium, a method for training a machine learning model to predict weld quality, and a machine learning model for predicting weld quality according to the features of the preamble of the independent claims.
[0004] In industrial joining processes such as welding, and especially ultrasonic welding, it is crucial to be able to reliably predict or determine the quality of the resulting joint. Determining the quality of the joint through mechanical tests is unsatisfactory for several reasons. Such tests are very costly and time-consuming. Furthermore, they lead to the destruction of the joint being tested.
[0005] Therefore, there is a need for methods and devices that can determine or predict the quality of a connection without causing damage.
[0006] Several solutions are already known, including those using AI-based systems. PTESIOIWO / 04.09.2025 2 2025094075
[0007] From DE 10 2021 111 962, for example, an assessment procedure for evaluating a machining process, a training procedure for training an AI system and an ultrasonic machining system are known, wherein the quality of the workpiece machined with a machining process can be determined using a coded image file.
[0008] From DE 10 2022 112 098 a method for determining process parameters for an ultrasonic machining process is known, in which an objective, reproducible and accurate process parameter finding should be possible.
[0009] US 2017 / 225275 discloses a device and a method for measuring the quality of an ultrasonic weld. A weld quality measurement unit processes a sensor signal. Based on the energy absorbed during welding, it is determined whether a weld defect might be present.
[0010] KR 1020130036104 shows a method for monitoring the welding status during ultrasonic welding in real time. A measurement step is performed simultaneously with the welding step.
[0011] It is known from WO 2017 / 129250 that the quality of an ultrasound treatment is to be assessed using a set of evaluation criteria. Factors such as welding time, power input, or energy input are taken into account.
[0012] In DE 34 29 776, US 2014 / 0203066 or US 10,466,204, a method for quality control in ultrasonic welding is shown, which is based on the acquisition of welding parameters. PTESI OIWO / 04 . 09 . 2025 3 2025094 075
[0013] EP 1 677 942 discloses a method for quality control of conductors welded in a compression chamber of an ultrasonic welding device. After welding, the compression chamber is depressurized, ultrasound is applied to the welded conductors, and a characteristic parameter is then measured. The weld inspection is carried out directly in the compression chamber without the need for additional equipment.
[0014] German patent application DE 10 2021 102 100 describes a method and a device for testing objects intended for welding. Depending on a set of test parameters, an object intended for testing is subjected to ultrasonic waves in such a way that it is not yet welded. During this exposure, a measured value and a test signature are determined, which can then be compared with a reference signature.
[0015] US 2019 / 271669 discloses a method and apparatus that enable the non-destructive determination of the weld quality of an ultrasonic weld. Vibration characteristics during the welding process are taken into account.
[0016] However, all these known solutions have various disadvantages. In particular, reliably predicting weld quality is not possible in every case.
[0017] The object of the present invention is therefore to avoid the disadvantages of the known method and, in particular, to provide methods and devices that allow a reliable and repeatable prediction of the quality of a connection between two components. This is intended to be, in particular, a PTESI OIWO / 04.09.2025 4 2025094 075
[0018] welded joints, and especially ultrasonic welded joints, are involved.
[0019] According to the invention, these problems are solved by a method for non-destructive testing of a connection between two components, a method for creating a connection between at least two components, a welding device, a computer program product, a computer-readable storage medium, a method for training a machine learning model to predict weld quality, and a machine learning model for predicting weld quality with the features of the independent claims.
[0020] A first aspect of the invention therefore relates to a method for the non-destructive testing of a connection between at least two components. This is in particular an ultrasonic welding process. The components are typically metallic components, for example, stranded wires welded to terminals or to other stranded wires. However, other components and other materials are also conceivable.
[0021] In a first step, process data is provided which represents at least one process parameter recorded during the welding process. Within the scope of the present invention, the term "process parameter" includes parameters that are directly related to the operation of the welding device, for example, data related to a generator or to the power system (press, sonotrode, anvil), but also parameters that are measured on or adjacent to the components to be welded. The process data can be a temperature of the at least one component and / or a tool, which is measured particularly during or after the welding process.
[0022] The process data, preferably measured by infrared measurement, can include a spatial and / or temporal temperature profile of the at least one component and / or tool. This can be measured, for example, by recording the spatial distribution of infrared radiation emitted by the component / tool, particularly at different times, preferably with infrared cameras.
[0023] An assessment value is determined based on the process data.
[0024] The assessment score is then checked against the fulfillment of target criteria. These target criteria are typically relevant to the quality characteristics of the resulting connection.
[0025] The quality of a weld depends on a multitude of different factors (e.g., process parameters, component properties, environmental parameters, condition of the welding equipment components). Determining evaluation values is therefore always associated with a certain degree of uncertainty. Consequently, in a further step, an expected value is determined, which is also based on the process data. The expected value characterizes the reliability of the predetermined evaluation value. The expected value can provide a measure of reliability for the evaluation value and, in particular, include, define, or be descriptive of a measure of dispersion, such as variance or standard error; an interval measure, especially a confidence or credibility interval; an information measure; or an ensemble uncertainty measure, such as the dispersion of several model instances. PTESI OIWO / 04.09.2025 6 2025094 075
[0026] The expected value allows us to assess whether the result of the evaluation of the assessment value with regard to the fulfillment of the target criteria can be trusted. The expected value can indicate how reliable the statement is as to whether the assessment value meets the target criteria or not. For example, the expected value can indicate a probability that the assessment value meets or fails to meet the target criterion, perhaps by comparing it to a confidence interval.
[0027] If the target criteria are not met, a direct notification is preferably issued indicating that the weld is not considered satisfactory. In this case, the expected value does not need to be specified. It is also conceivable to refrain from calculating an expected value altogether in this case.
[0028] If the target criteria are met, the assessment value and the expected value are preferably provided, especially for further processing of the data.
[0029] The present invention is therefore initially based on determining an evaluation value and an expected value using process data provided in situ during the welding process. No properties of the resulting joint are yet taken into account.
[0030] If the expected value reaches an initial confidence threshold, a quality prediction is preferentially issued in a further step. The quality prediction characterizes the expected quality of the weld joint. This can be, for example, quantitative data such as an expected pull-off force or qualitative data according to a specific gradation, for example, a binary gradation (good / poor) or a multi-level classification (e.g., very good / good / medium / poor).
[0031] If the expected value does not reach the first confidence threshold, it can be assumed that the assessment value does not provide any meaningful information about the weld quality. This does not necessarily mean that the weld quality is poor. Therefore, in this case, an extended inspection based on the properties of the welded component is preferably initiated. This extended inspection can be initiated by instructing the user. However, it is also conceivable to trigger further inspection steps directly and automatically. It is also conceivable to prompt the user for additional data or information via an input interface. Therefore, if necessary, the prediction of the quality is improved using values provided ex situ after the welding process. Properties of the resulting joint can also be taken into account.
[0032] The generated quality prediction can correspond to the previously determined assessment value. However, it is also conceivable to derive a quality prediction from the assessment value, particularly to simplify interpretation.
[0033] In addition to or as an alternative to reporting the expected weld quality or initiating extended testing, it is also conceivable to adjust future welding or testing processes based on the assessment value and the expected value. This is particularly relevant when a process yields inconsistent results, enabling self-learning improvement during operation. PTESI OIWO / 04.09.2025 8 2025094 075
[0034] Preferably, the extended test is initiated by issuing instructions to the user on an output device. This could, for example, be a request to subject the created connection to an individual test.
[0035] Alternatively, the output device can be instructed to read in measured properties of the welded components. Preferably, the properties of the welded component for the extended test are the results of non-destructive tests, in particular optical, electrical, or acoustic tests. The properties of the welded component can include physical and / or material-related parameters, in particular electrical conductivity, optical reflectivity, and / or a geometric property, preferably a geometric deviation from a predetermined shape. These tests can be performed in a manner known per se, in particular by taking and analyzing images of the weld, measuring the electrical conductivity, determining the response to sound stimulation, etc.The results of the corresponding test can then be fed into the user interface of the welding device and taken into account in a further process step.
[0036] However, it is also conceivable to request the user in certain cases to perform a destructive test and to read in the results of the test (for example, a pull test or the measurement of the adhesion number concentration) via the input interface.
[0037] Typically, a destructive test results in a pull-off force (that is, the force at which a weld breaks) or information regarding the PTESI OIWO / 04.09.2025 9 2025094 075
[0038] Part adhering components of the other part ("adherence number concentration").
[0039] Particularly if a final, non-destructive analysis within the extended testing process fails to yield reliable, usable results, a destructive test can be performed. The results of the destructive test can then be used to further improve the prediction, e.g., to further train a model.
[0040] If the extended expected value reaches a second confidence threshold, an extended quality prediction can be issued, which characterizes the expected quality of the weld joint. This extended quality prediction can be identical to the quality prediction determined in the previous step. In this case, the extended expected value has only strengthened the validity of the extended quality prediction. However, it is also conceivable to adjust the extended quality prediction based on the extended testing and the extended expected value.
[0041] Alternatively or additionally, the process data, the assessment value, the expected value, the extended expected value, the properties of the welded component can be used individually or in combination to train a computer model, to adjust operating parameters, or to adapt algorithms to determine future assessment values or expected values.
[0042] Operating parameters that need to be adjusted for future operation typically include parameters for operating an ultrasonic generator, such as amplitude, current, voltage, and / or frequency. However, other operating parameters may also be required, such as those listed in PTESI OIWO / 04.09.2025 10 2025094 075.
[0043] Phase, active power, apparent power, reactive power can be adjusted.
[0044] It is also conceivable to adjust additional parameters that are relevant to the welding process. For example, temperature controls (cooling, temperature control) can be activated in a manner known per se. It is also conceivable to adjust the welding process itself. For instance, insufficient quality may indicate worn tools. Depending on the situation, it may therefore be helpful to subject a sonotrode to an additional vibration after the actual welding process, while the welding press is released. Such additional vibrations can serve to shake off any components that may still be adhering to the sonotrode. Adjusting additional parameters can include, in particular, a "shaking impulse" and its improvement, as described in WO 2018 / 141660, or cooling, as described, for example, in WO 2018 / 145769.
[0045] It is also conceivable to adapt subprograms of the welding device depending on the information provided, in particular integrated subprograms that modify the welding process sequence, as well as subprograms that can communicate with other systems or subsystems via interfaces. These other systems or subsystems can be, for example, measuring devices or manufacturing execution systems (MES). The subprograms can be integrated based on the recipe, whereby the recipe can contain information about materials or welding parameters. They can also contain process-relevant information (e.g., with scripts / algorithms). Furthermore, additional models can be integrated into the subprograms that can generate predictions and / or evaluate the welding process.
[0046] If the expanded expected value does not reach the second confidence threshold, information is displayed indicating that no quality prediction is possible.
[0047] Additionally or alternatively, process data, assessment value, expected value, extended expected value and / or properties of the welded component can also be used for the purposes described above to train AI models or to adjust operating parameters or algorithms.
[0048] The assessment value is preferably determined directly based on operating parameters. If the operating parameters are within certain limits, a positive or negative assessment value is issued. These limits could, for example, be energy amounts entered within predefined times, travel distances, or forces acting during welding. However, other values to be considered (distance, force, power, or similar) depending on other predefined values (force or time) or combinations thereof are also conceivable.
[0049] A notification, in particular a status, service, error, or condition notification regarding process data and / or a property of the welded component, and / or a control instruction for adjusting the process data, can be issued. The service notification and / or the control instruction is configured for output to a user and / or for output to a welding device, preferably for controlling the welding device. The output of the notification and / or control instruction can optionally be automated. PTESI OIWO / 04.09.2025 12 2025094 075
[0050] This output can be made if the expected value or the extended expected value reaches or fails to reach the respective first or second confidence threshold.
[0051] These notifications and / or control instructions enable continuous monitoring / control, which can be performed either by a user or semi- or fully automatically, directly controlling the welding device or performed by the welding machine itself. Such a user / welding machine monitoring loop allows for flexible and rapid alerting to, and / or response to, changes in process data and / or properties of the welded component.
[0052] The AI model can consist of or include at least one decision tree, random forest, support vector machine, or neural network model, in particular a supervised neural network model, which preferably includes at least one convolutional layer and / or a dropout layer.
[0053] The expected value and / or the extended assessment are preferably carried out using a KL model. A separate KL model can be used for each of these steps, or a more complex model can be used to perform all assessment steps.
[0054] The AI model is particularly favored for its use of linear regression based on singular value decomposition (SVD). PTESI OIWO / 04.09.2025 13 2025094 075
[0055] Preferably, the AI model for determining the expected value and / or the extended expected value is based on a hybrid model. This hybrid model can be based on a first and a second model. The first model is trained with a first subset of training data, and the second model with a second subset of training data. The first subset is based on training data that exhibits a first category of properties. The second subset is based on training data that exhibits a second category of properties. The first category of properties could, for example, represent a certain number of the best welding results. The second category of properties could represent a number of the worst welding results. The best and worst welding results could, in particular, be the best and worst deduction values.
[0056] Preferably, the AI model is a combined model. This combined model is based on a plurality of process parameters and on a plurality of first and / or second models, each defined / trained for different process parameters. The process parameters include, in particular, power, force, and travel distance, which are plotted as a function of time, energy input, or other parameters. The combined model is preferably based on a plurality of mixed models, each generated from the plurality of first and second models.
[0057] Process parameters can include, in particular, generator data such as power, energy, amplitude, current and / or frequency of the generator for generating ultrasonic vibrations, or other operating parameters such as phase, active power, apparent power, or PTESI OIWO / 04 . 09 . 2025 14 2025094 075
[0058] Reactive power. It can also include press data, in particular the force, displacement, and / or travel speed of a press used to join the components to be joined. Process data can also include information on the components to be welded, in particular material data or dimensional properties of the components. Process data can also include measurement data based on measurements of the components before or during joining. Likewise, process data can include tool data, in particular data on the condition of a sonotrode and / or anvil for ultrasonic welding. Process data can also include values derived from the above data, such as mathematical derivatives, e.g., with respect to time. Process data can also include outputs from subprograms, in particular from the subprograms mentioned above.
[0059] According to a further preferred embodiment of the invention, machine limits are additionally determined, output, or adjusted in the method. These machine limits define permissible operating parameters for the specific welding situation based on process data. The machine limits are determined in such a way that the operating parameters ensure the Kl model can generate reliable information about the quality of the welded joint. This ensures that the claimed method operates in the most stable environment possible, thus guaranteeing high reliability of the quality prediction. The number of required extended tests or unusable predictions is thereby significantly reduced. It is also conceivable to automatically adjust the machine limits in a welding device so that correct operating parameters are set and maintained without operator intervention.However, it is also conceivable to output the machine limits via an output interface and to advise the user that work should be carried out within these limits.
[0060] It is also preferable to determine and output or adjust service parameters. These service parameters can include, in particular, information about the condition or replacement of parts such as tools. Sonotrodes and / or anvils, in particular, wear down over time. The available process data and the determined assessment and expected values can be used to draw conclusions about the condition of the tools. For this purpose, it is especially useful to store the determined assessment and expected values in a time series. If, for example, the assessment values worsen over time, or if the prediction becomes less reliable and the expected values decrease as a result, this can be interpreted as an indication of tool wear. Similarly, this can be an indication that maintenance of parts of the device is required.Accordingly, information regarding the timing or necessity of maintenance work can also be issued. It is also conceivable to transmit this information or these instructions to a supplier of spare parts and / or maintenance services via a communication interface.
[0061] The procedure may include obtaining or determining a large amount of training data, including training process data, which were used to train the AI model to determine the expected value and / or the extended expected value. Alternatively, the procedure may include obtaining or determining an initial statistical analysis based on the large amount of training data. It may determine a model fit based on the provided process data and the initial statistical analysis. PTESI OIWO / 04.09.2025 16 2025094 075
[0062] The model's quality can be determined through analysis. Model quality characterizes the reliability of a multitude of assessment values, expected values, and / or extended expected values. Optionally, detection limits, particularly predetermined detection limits, can be obtained, determined, or defined. These detection limits define the predictive accuracy of the model's quality for specific welding situations, ensuring that the multitude of assessment values, expected values, and / or extended expected values provide meaningful information. The model quality can be compared with the detection limits to determine whether the provided process data is suitable for the permissible or meaningful determination of the multitude of assessment values, expected values, and / or extended expected values.
[0063] This enables the determination and provision of a model fit as a measure of the reliability of the multitude of quality values, expected values / extended expected values, by comparing the process data with previously used training data / an initial statistical analysis. Thus, the reliability of determining these values based on the process data can be assessed using training data already used for training, i.e., known data. In other words, the model fit provides a measure of the overall reliability of the entire process as a function of the provided process data.
[0064] The model accuracy of a welding device can thus steadily decrease over time, as increasingly unreliable process data is obtained due to wear and tear on the sonotrode. Therefore, model accuracy does not provide information about individual values, but rather a trend over time. PTESI OIWO / 04.09.2025 17 2025094 075
[0065] A large number of assessment values, expected values, and / or extended expected values in view of changing process data.
[0066] The training data and / or initial statistical analysis can include a variety of different process parameters. The initial statistical analysis can include one or more, in particular multivariate linear, regressions of the training process data.
[0067] Model fit can be determined using one or more significance metrics, such as F-statistics, Bayesian model comparisons, and / or AI-supported significance analysis. Furthermore, multiple significance metrics can be compared to determine the model fit, particularly of different models.
[0068] The model fit can be determined based on at least a number of the assessment values, expected values, extended expected values, and / or properties of the welded component in the training data. Alternatively, the model fit can be determined based on a second statistical analysis using the number of assessment values, expected values, extended expected values, and / or properties of the training data.
[0069] This allows for a better prediction of model performance, as it does not only compare the new process data with the training process data. Instead, it also considers the associated values / properties or secondary statistical analyses of known training data to determine model performance. This is particularly informative for determining the model performance. PTESI OIWO / 04.09.2025 18 2025094 075
[0070] Quality is the multitude of judgment, expectation and / or extended expectation values.
[0071] The numerous assessment values, expected values, extended expected values, and / or properties of the training data can be at least partially ordered chronologically, allowing for a more meaningful model fit assessment based on comparison with the process data. Preferably, the training data or initial statistical analysis for determining the model fit can comprise a large number of time series of one or more process parameters. The time series can have been measured at a single welding machine or at a large number of different welding machines. These welding machines can be physically separated from one another.
[0072] Determining the expected value can be done based on the process data and the model accuracy. This allows the reliability of the assessment value to be determined even more precisely, especially when considering the model accuracy for a specific welding device.
[0073] Another aspect of the invention relates to a method for creating a connection between at least two components. This is in particular an ultrasonic welding connection. Within the framework of the method, at least one component is welded using the welding device. The at least one component can be a metal component or a plastic component, in particular a vehicle component, preferably an automotive component, a medical device component, and / or an electronic component. The at least one component can be a connecting element, an electrical conductor, in particular a cable, preferably a high-voltage cable, or a structured connecting element. A quality prediction is determined using a non-destructive testing method as described above.The determined assessment value and the expected value and / or the extended expected value and the quality predictions are provided for further use as described above and preferably output.
[0074] According to a further aspect of the invention, in a method for creating a connection between at least two components, machine limits are provided in a first step. The method is preferably an ultrasonic welding process. The machine limits define permissible operating parameters for a specific welding situation. The specific welding situation is defined in particular by the nature of the components to be welded (for example, materials and dimensions), but can also include further information such as tool temperatures or tool conditions. The operating parameters are set such that a computer model can generate reliable information about the quality of the welded connection. The welding is then carried out with operating parameters that are within the machine limits.Subsequently, the Kl model can be used to predict the quality of the weld joint, in particular as described above.
[0075] Another aspect of the invention relates to a method for the non-destructive testing of a connection between at least two components. Again, the method is preferably an ultrasonic welding process and the resulting connection an ultrasonic weld. First, process data is provided, representing at least one process parameter acquired during the welding process. The process data is provided in a computer arrangement. [PTESI OIWO / 04.09.2025 20 2025094 075]
[0076] The process data preferably consists of the process parameters described above.
[0077] Based on the process data, the computer system determines an evaluation value and checks it against the fulfillment of target criteria. The determination and verification are based on the process data using the computer system.
[0078] Furthermore, an expected value is determined, which characterizes the reliability of the assessment value. The expected value is also determined based on the process data using the computer configuration.
[0079] Furthermore, properties of the welded component are provided for extended testing in the computer setup.
[0080] Process data refers to data acquired in situ, meaning data obtained during or before the welding process in a welding fixture or on the components to be welded. Properties of the welded component for extended testing are ex situ data, meaning data acquired after the welding process. This data can be acquired within or outside the welding area.
[0081] The computer system then determines a quality prediction that characterizes the expected quality of the weld joint. The determination of the quality prediction is based on the assessment value, the properties of the welded component, and / or the expected value. The inventive method allows for non-destructive testing with particularly informative quality predictions. PTESI OIWO / 04.09.2025 21 2025094 075
[0082] A further aspect of the invention relates to a welding device with a computing arrangement configured to perform one of the methods described above. This preferably involves a known ultrasonic welding device comprising an ultrasonic generator, a converter, a sonotrode, a press, and an anvil, as well as means for determining process data. The process data may relate to parameters of the ultrasonic vibrations or to parameters resulting from the interaction of the sonotrode, anvil, and press.
[0083] Another aspect of the invention relates to a computer program product comprising instructions which, when the program is executed by a computer, cause it to carry out one of the aforementioned methods.
[0084] A further aspect of the invention relates to a computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out one of the methods described above.
[0085] According to another aspect of the invention, a machine learning model is provided for predicting the weld quality of a joint between two components. The machine learning model is based on a mixed model. The mixed model is based on a first model and a second model. The first model was trained with a first subset of training data. The second model was trained with a second subset of training data. The first subset is based on training data that exhibits a first category of properties. The second subset is based on training data that exhibits a second category of properties. Preferably, the first category of properties represents the n best PTESI OIWO / 04.09.2025 22 2025094 075
[0086] Welding results, especially pull-off values. The second category of properties represents, in particular, the m worst welding results, especially pull-off values.
[0087] Preferably, the machine learning model is a combined model. Starting from a plurality of process parameters, the model is built on a plurality of models, in particular first and / or second models. The plurality of process parameters are, in particular, power, force, and travel distance. The combined model is based, in particular, on a plurality of mixed models generated from the plurality of first and second models.
[0088] The machine learning model can be an incrementally and / or continuously learning online machine model. The machine learning model can be connected to, connectable to, and / or run on a computing unit of a cloud-based and / or edge computing platform.
[0089] This makes it possible to provide a machine learning model that learns dynamically and can be continuously improved. Process data from a large number of other test fixtures can thus be considered for training, and in addition, the specific process data of the particular test fixture can be incorporated into the machine learning model for training.
[0090] Another aspect of the invention relates to a computer-implemented method for processing training data for use in a machine learning model. The method comprises obtaining or determining training data and identifying properties in the training data that the PTESI OIWO / 04.09.2025 23 2025094 075
[0091] The method represents the quality of welding results, particularly pull-off values. It involves creating at least a first and a second subset of training data. The first subset is based on training data exhibiting a first category of properties. The second subset is based on training data exhibiting a second category of properties. The first property can represent the n best welding results, particularly pull-off values. The second property can represent the m worst welding results, particularly pull-off values.
[0092] Another aspect concerns a method for training a machine learning model to predict weld quality.
[0093] As a first step, properties in training data are identified that represent the quality of the welding results. These are primarily measured pull-off values.
[0094] Based on the identified properties, a first and a second subset of training data are created. The first subset of training data is based on training data exhibiting a first category of properties. The second subset is based on training data exhibiting a second category of properties. The first and second categories represent, in particular, the n best and m worst welding results, respectively. The welding results are specifically deduction values. PTESI OIWO / 04.09.2025 24 2025094 075
[0095] As an alternative to identifying the properties and forming the first and second subsets, an already formed first and second subset of training data can be obtained directly.
[0096] Based on the first and second subsets, a first and a second model are then trained. A mixed model is then created based on the first and second models.
[0097] Preferably, a plurality of first and / or a plurality of second models are trained, starting from a plurality of process parameters. These are typically process parameters such as power, force, and travel distance. From the plurality of first and second models, a plurality of mixed models are generated. The plurality of the mixed models can then be combined into a single combined model.
[0098] It goes without saying that the order can also be changed. For example, first and second combined models can be created first, which are then combined into a mixed model based on the combined models.
[0099] Another aspect of the invention relates to a training dataset for use in a method for training a machine learning model to predict weld quality, particularly as described above. The training dataset comprises a first and second subset of training data. The first subset is based on training data exhibiting a first category of properties. The second subset is based on training data exhibiting a second category of properties. The first property can represent the n best welding results, including pull-off values, and the second PTESI OIWO / 04.09.2025 25 2025094 075
[0100] Property representing the worst welding results, including deduction values.
[0101] The invention is explained in more detail below using exemplary embodiments and with reference to the drawings. The drawings show:
[0102] Figure 1: A schematic representation of a welding system according to the invention.
[0103] Figure 2: A schematic representation of the process according to the invention.
[0104] Figure 3a: A schematic representation of individual values determined within the scope of the present invention
[0105] Figure 3b: A schematic representation of the definition of a model based on the values from Figure 3a
[0106] Figures 4a to 4b: Representation of the definition of models based on welding force and the highest / lowest deduction values
[0107] Figures 5a to 5d: Representation of combined models based on power, travel distance, force and combinations thereof and
[0108] Figure 6: Schematic representation of actions based on determined values.
[0109] Figure 1 schematically shows an ultrasonic welding system 20. The ultrasonic welding system 20 has a sonotrode 21. PTESI OIWO / 04.09.2025 26 2025094 075
[0110] Generator 26 produces high-frequency oscillations which are fed to a converter 22, which converts the electrical oscillations into mechanical vibrations. The vibrations are then transferred from the converter 22 to the sonotrode 21 in a known manner.
[0111] Two components or workpieces 40 to be joined are arranged between the sonotrode 21 and an anvil 27. In the specific embodiment shown, the sonotrode 21 is depicted as a longitudinally oscillating sonotrode. However, other designs, and in particular torsional sonotrodes, are also conceivable.
[0112] The two components 40 are shown schematically and, after welding, together form a welded component 41. These are typically strands or terminals or other types of metallic or non-metallic components. A press 28 applies a force to the sonotrode 21 directed against the anvil 27, so that the components 40 can be pressed against each other between the sonotrode 21 and the anvil 27.
[0113] The ultrasonic welding system 20 has a computing unit 24. The computing unit 24 is used to control the generator 26. In particular, it is used to specify operating parameters of the generator 26, such as amplitude and frequency. It is also used to control the press 28 and, in particular, to monitor the pressing force and feed rate.
[0114] A measuring device 23 determines operating parameters from the generator 26 and transfers them to the computing unit 24.
[0115] Furthermore, a measuring sensor 25 in the welding area for the components 40 to be welded determines measured values from the ongoing operation of the PTESI OIWO / 04 . 09 . 2025 27 2025094 075 and transmits them to the processing unit 24. The measuring sensor 25 can, in particular, be a temperature sensor.
[0116] The computing unit 24 is equipped with or functionally connected to a storage medium 29. In particular, a Kl model is stored on the storage medium 29, which, based on data transmitted by the measuring device 23 and / or the measuring sensor 25, enables a prediction of the quality and thus a non-destructive testing of the connection between the components 40 / of the welded component 41 in the manner described below.
[0117] Information obtained through the AI model can be used to control the generator 26, the press 28, and, in particular, to output information to an output device 30. The output device 30 can, for example, be a monitor with a graphical user interface. A user can also provide instructions or further information via an input interface 31, which is received by the computer unit 24. The input interface 31 and the output device are shown schematically separately. It can also be a combined input / output device, in particular a computer with a screen, mouse and keyboard, or a touchscreen.
[0118] Furthermore, additional information can be provided by the computing unit via an output interface 32, for example for transmission to external computer systems.
[0119] Figure 2 schematically shows the sequence of a test method according to the invention. Starting from a welding process, process data P are provided. The process data P include, in particular, the frequency, amplitude, voltage, phase, active power, apparent power or reactive power and / or current of the generator 26. Typically, the process data P consists of values such as welding energy, travel distance or force as a function of other values such as time. In addition, they include information on the material and dimensions of the welded components 40. Furthermore, parameters of the press 28, such as applied force or travel distance, can be taken into account.
[0120] In a first step, an assessment value B is determined based on the process data P. The assessment value B characterizes the expected quality of the connection between the components 40. If the assessment value B does not meet the specified target criteria SK, an output is generated on the output device 30 indicating that the expected quality is insufficient.
[0121] If the determined assessment value B meets the target criteria SK, an expected value E is determined in a further step. The expected value E indicates how reliable the determined assessment value B is. If the determined expected value E is greater than a predetermined confidence threshold KE, an output is generated on the output device 30 indicating that the expected quality of the connection between the components 40 is sufficient.
[0122] Preferably, the assessment value is binary, meaning that the weld is considered either good or not good. This is achieved by assessing whether certain predefined process parameters lie within a specified process window. PTESI OIWO / 04.09.2025 29 2025094 075
[0123] For example, it is checked whether, for certain assessment pairings, the energy applied within a certain time period or for a certain travel path lies within a window for the energy.
[0124] If the determined expected value E does not reach the confidence threshold KE, an extended test is performed. In the extended test, an extended expected value EE is determined. For this extended test, in addition to the process data P, further data D are provided. These are supplied, for example, via the input interface 31 of a device according to the invention. These are, in particular, physical properties of the connected component 41, such as optical, electrical, magnetic, or mechanical properties.
[0125] It is also conceivable to input measured values that can provide information about the physical properties of the welded joint. For example, it is possible to measure the temperature of the welded joint at a specific time interval after the welding process has ended, for instance, using non-contact infrared measurements. Similarly, it is conceivable to determine the temperature in a joining zone of the welded component using thermocouples, as described, for example, in DE 10 2021 117 697. The data obtained in this way can also be provided and taken into account when determining the expected value.
[0126] If the calculated expanded expected value EE meets a second confidence threshold KE2, an output indicating that the expected quality is good is generated. If the expanded expected value is below the second confidence threshold KE2, a PTESI OIWO / 04.09.2025 30 2025094 075
[0127] Output that no prediction can be made regarding weld quality or that the weld quality is likely to be insufficient.
[0128] The invention has been explained above with a focus on the mechanical strength of a welded joint. However, the resulting joint can also be non-destructively tested with regard to other relevant properties. Alternatively or in parallel, for example, mechanical pull-off values, electrical conductivities, and optical properties (such as discoloration or particle adhesion) can be determined or predicted.
[0129] The extended testing may also include a destructive test step. This may, in particular, be a test to determine the pull-out force of the connection. The destructive test is optional and is therefore shown with a dashed line.
[0130] The determination of the individual values, in particular the expected value E and the extended expected value EE, is carried out by an AI model. To improve the AI model, it is conceivable to feed the determined assessment value B, the expected value E, the extended expected value EE, and / or values W from the destructive testing into the AI model to enable continuous training of the model. These optional steps are represented by dashed lines.
[0131] In the extended inspection, the user is specifically instructed to subject certain welded components to further inspection if the prediction of quality is not sufficiently clear. This can be done by issuing instructions to the user. However, it is also conceivable to initiate corresponding tests directly and automatically in the device according to the invention, for example, by means of built-in testing equipment such as cameras, measuring devices for electrical conductivity or magnetic properties.
[0132] Regardless of the test result and the determined values, it is conceivable to output information about the determined values via the further interface 31. In particular, if the quality deteriorates, this can indicate worn tool parts. Therefore, the determined values B, E, and EE are stored in a time series on memory 29 or in another suitable storage medium. Accordingly, service instructions S can be determined and transmitted by applying the AI model or by means of other computational models.
[0133] Typical parameters used in the prediction can be:
[0134] - Welding time, especially when welding is performed with respect to a certain energy or force or path, that is, the time that elapses until an energy is applied / a force is reached / a path is traveled.
[0135] - The initial and / or the final dimension (for example, height and / or width) of a knot welded from strands.
[0136] - Optical properties of the welded component.
[0137] The cross-sections of the welded components and their modification / compaction.
[0138] Withdrawal forces. PTESI OIWO / 04.09.2025 32 2025094 075
[0139] Working volume of the material in the connection area of the
[0140] Workpieces, e.g., based on several cross-sectional areas
[0141] ( "Weld nugget cross-section" )
[0142] - Electrical properties, in particular electrical conductivity or resistance or voltage drops.
[0143] - Damage to components adjacent to the welding point and connected to the workpieces to be joined, or to adjacent components of the welding system (both the ultrasonic stack and the unit generating a pressure force, such as a press).
[0144] - Fracture points on welded components such as terminals.
[0145] - Visual properties of insulation material.
[0146] - Condition of tools.
[0147] - Other criteria specified or relevant for welding, such as quality criteria according to USCAR38-2.
[0148] Typically, the welding time until a desired energy input is reached should remain approximately constant over time. Unusual changes in welding time may indicate that the process has changed.
[0149] The initial height of the welded joint can be measured directly in the welding fixture using sensors. The initial height can also be used, in particular, to ensure that the material used is within predefined tolerances. PTESI OIWO / 04.09.2025 33 2025094 075
[0150] Figures 3a and 3b schematically illustrate and explain how the AI model is structured and trained. To train the model, pull-off values of connected components 41 are measured. This means that the connection is subjected to a tensile force under standardized conditions until it breaks. The pull-off force required for breakage is recorded. The relevant operating parameters were also recorded over time.
[0151] Figure 3a schematically shows a table displaying the parameters derived from individual welds for the formation of a model. A total of 20 welds were performed. For each of these welds, process parameters from the welding equipment are stored in an individual file for each weld / component.
[0152] To define a model, all 20 welded components are subjected to a destructive test. The fourth column in Figure 5b lists the forces (pull values) required to break the joint. Using the four highest pull values, a first model optimized for high pull values is defined via SVD as described below. In this specific example, these are welds 4, 6, 7, and 11.
[0153] In a second step, a second model optimized for the four worst trigger values is defined. These are welds 2, 15, 18, and 19. PTESIOIWO / 04.09.2025 34 2025094075
[0154] The second and third columns show errors resulting from the use of a reduced matrix in the SVD. The error indicates the deviations arising from the reduced matrix.
[0155] The second column shows the error resulting from applying the model defined based on the four highest deduction values. The third column shows the error resulting from applying the model defined with the lowest deduction values.
[0156] Columns 5 to 8 show a classification: The high / low classification represents the probability that a weld can be classified as a high-tensile weld ("high") or a low-tensile weld ("low"). It goes without saying that the previously defined highest pull-off values result in a "high classification" of almost 100%, while the four lowest pull-off values result in a "low classification" of almost 100%.
[0157] The previously defined models are then applied to the 12 further welds that lie between the four highest and the four lowest deduction values.
[0158] This results in an error for each measurement for both models (high / low). A high / low classification is then calculated for each measurement. The classification is obtained by normalizing the errors to 100% (no deviation for the respective low or high errors).
[0159] The classification yields a reliability value (referred to as a score), which, in the eighth column, was normalized to 100% for the best value PTESI OIWO / 04.09.2025 35 2025094 075. This normalized score value indicates the expected weld quality, taking into account the high and low models.
[0160] Figure 3b shows graphically how the individual measurements can be represented according to the predicted quality ("scores") and the measured deduction values. The x-axis shows the score normalized to 100%, and the y-axis shows the measured deduction values.
[0161] Based on this set of points, a linear regression is applied (shown with a dashed line in Figure 3b) and a straight line with the same slope is plotted through the lowest deduction value.
[0162] For new measurements, the high and low models are applied to the measured values, and a score value is determined based on the models. Based on the score values and the linear relationship shown, it can be predicted that the deduction for a given score should lie above the solid line shown.
[0163] To determine the score value, the mean of the high classification and the low classification subtracted from the amount 1 is calculated.
[0164] The error results from the difference between the original curve (e.g., power) and the curve processed with the reduced matrix.
[0165] High score values in the high-model indicate a high probability of good welding, while in the low-model they indicate a high probability of poor welding. Low score values on the other side indicate a likely unreliable prediction.
[0166] The score determined in this way corresponds to the expected value.
[0167] The reliability of the AI models used depends, among other things, on the specific operating parameters. In this exemplary embodiment, machine limits are determined and provided within which the device 20 according to the invention is to be operated. If operation takes place within these machine limits, it is ensured that the AI model can generate the most reliable information possible regarding the quality of the welded joint.
[0168] Figures 4a and 4b show the definition of models based on measurement parameters derived from power. The generator's power is continuously measured, and welding continues until a predefined energy input is reached.
[0169] Figure 4a shows the definition of a "high-model", where in this case the five measurements with the highest deduction values are taken into account and therefore correspond to a score of 100%.
[0170] Figure 4b shows the definition of a "low-model", where the five values with the worst deduction values correspond to a score of 100%.
[0171] Figure 5a shows the definition of a mixed model, based on the previously described high and low models. This results in an optimized model that is optimized with respect to both poor and good trigger pull values. As described above in connection with Figure 3b, PTESI OIWO / 04.09.2025 37 2025094 075, a model is defined from the set of points by linear regression, which allows the assignment of future measurements.
[0172] The process steps described with reference to Figures 4a, 4b and Figure 5a are subsequently applied to process data representing the distance traveled by the sonotrode as a function of the applied energy (Figure 5b) and the force as a function of the applied energy (Figure 5c).
[0173] From these three models (power / distance / force) a combined model is generated by creating a weighted average of the three conditions (see Figure 5d).
[0174] Future measurements will be processed using this combined average model. This results in particularly meaningful and stable predictions.
[0175] It is also conceivable to define a medium model based on welds with medium pull-off forces. This would allow the model to be further improved.
[0176] The determined score values can also be used to decide whether further measurements are necessary.
[0177] Figure 6 schematically illustrates how further actions result from three models (high / mid / low) and the calculated scores. Measurements where the score value is high (1) in one model and low (0) in the other models lead to a clear classification. For example, the weld according to the second row is clearly identifiable as a poor weld (low), the weld according to the third row as a medium weld (mid), and the weld according to the sixth row as a good weld (high). PTESIOIWO / 04.09.2025 38 2025094075
[0178] If the scores lead to contradictory results (for example, rows six and eight), a destructive test is performed, and the values are used for further training of the model ("training run" according to the Action column). The same applies (see first row) if no score is obtained. Thus, one or more, in particular three, new AI models are trained based on this training data. These new models can determine better scores by taking the results of the destructive test into account. The AI models may be pre-trained, in which case they must be replaced with the new AI models. Alternatively, adaptive AI models can be used, which are further trained during operation.
[0179] However, if a medium score of 1 results in conjunction with a high (line 7) or low (see line 4) score, further non-destructive tests are performed. The results can then be made available for use in further models.
Claims
PTESIOIWO / 04.09.2025 39 2025094075 Patent claims 1. Method for non-destructive testing of a connection between at least two components (40) , in particular an ultrasonic welded joint, comprising the following steps - Providing process data (P) which represents at least one process parameter recorded during the welding process - Based on the process data, determining an assessment value (B) and checking the assessment value (B) with regard to the fulfillment of target criteria (SK) - Determining an expected value (E) based on the process data (P), where the expected value (E) characterizes the reliability of the assessment value (B), in particular further comprising the steps: o If the target criteria (SK) are not met: Output of a message that the weld joint is not considered "good" o If the target criteria (SK) are met: Provision of the assessment value (B) and the expected value (E).
2. The method of claim 1, comprising the further steps: - When the expected value (E) reaches a first confidence threshold: Output a quality prediction (Q) which characterizes the expected quality of the weld joint. - If the expected value (E) does not reach the first confidence threshold: Initiate an extended test based on properties of the welded component to achieve an extended expected value (EE). PTESIOIWO / 04.09.2025 40 2025094075 3. Method according to one of claims 1 or 2, wherein the initiation of the extended test is carried out by (i) issuing instructions to the user and / or by (ii) generating a request to read in properties of the welded components (41) .
4. A method according to claim 2 or 3, wherein the properties of the welded component (41) for the extended test are selected from the group of - non-destructive tests, in particular optical, magnetic, electrical or acoustic tests. - a destructive test, in particular by means of a tensile test or measurement of the adhesion number concentration.
5. A method according to any one of claims 1 to 4, comprising the further steps: - when the extended expected value (EE) reaches a second confidence threshold: o Output an extended quality prediction (QE) that characterizes the expected quality of the weld joint and / or o Provide at least one of the process data, assessment value, expected value, extended expected value, properties of the welded component for training a KL model or for adjusting operating parameters or algorithms for determining the assessment value or expected value. - If the extended expected value (EE) does not reach the second confidence threshold: o Output of information that no quality prediction is possible and / or PTESI OIWO / 04 . 09 . 2025 41 2025094 075 o Providing at least one of the process data, assessment value, expected value, extended Expected value and / or properties of the welded component for training a computer model or for adjusting operating parameters or algorithms to determine the assessment value or expected value.
6. A method according to any of the preceding claims, comprising the further steps: - Outputting at least one notification, in particular a condition, service, error, or status notification regarding the process data (P), and / or a property of the welded component (41), and / or - Outputting at least one control instruction to adjust the process data (P), wherein the service notification and / or control instruction is configured for output to a user and / or for output to a welding device, preferably for controlling the welding device, wherein the output is optionally automated, and wherein the output is particularly triggered when the expected value (E) or extended expected value (EE) reaches or fails to reach the respective first or second confidence threshold.
7. Method according to any one of claims 1 to 6, wherein the AI model consists of or comprises at least a decision tree, random for rest, support vector machine, or neural network model, in particular a supervised neural network model, which preferably comprises at least one convolution layer and / or dropout layer. PTESIOIWO / 04.09.2025 42 2025094075 8. Method according to any one of claims 1 to 7, wherein at least one of the steps is (i) determining an assessment value (B), (ii) determining an expected value (E) and (iii) extended testing by one or more Kl models.
9. Method according to claim 8, wherein the Kl model for determining the expected value (E) and / or (iii) determining the extended expected value (EE) is a Kl model based on a mixed model based on a first model and a second model, in particular according to claim 7, wherein the first model was trained with a first subset of training data and the second model with a second subset of training data, wherein the first subset is based on training data having a first category of properties, and the second subset is based on training data having a second category of properties, wherein in particular the first category of properties represents the n best welding results, in particular pull-off values, and the second category of properties represents the m worst welding results, in particular pull-off values.
10. Method according to claim 9, wherein the Kl model is a combined model which, starting from a plurality of process parameters, in particular from power, force and travel distance, is based on a plurality of first and / or a plurality of second models, wherein in particular the combined model is based on a plurality of mixed models which were generated starting from the plurality of first and second models. PTESIOIWO / 04.09.2025 43 2025094075 11. Method according to any one of claims 1 to 10, wherein the process data are selected from - Generator data, in particular energy, power, phase, amplitude, current and / or frequency of the generator (26) for generating ultrasonic vibrations - Press data, in particular force, stroke and / or travel speed of a press (28) for joining the components to be joined - Material data of the components to be welded (40) - Measurement data based on measurements of at least one of the components (40) before joining or of the components (40) during joining - Tool data, in particular data on the condition of a sonotrode (21) and / or an anvil (27) for ultrasonic welding - Outputs from subprograms, in particular recipe-related subprograms that modify the welding process, as well as subprograms that can communicate with other systems or subsystems via interfaces.
12. Method according to any one of claims 8 to 11, wherein machine limits are determined and output or adjusted, wherein the machine limits define permissible operating parameters for the specific welding situation, with which operating parameters the Kl model can generate reliable information on the quality of the welded joint.
13. Method according to any one of claims 1 to 12, wherein service parameters are determined and output or adjusted, wherein the service parameters are in particular selected from PTESIOIWO / 04.09.2025 44 2025094075 Information on the condition or replacement of parts, especially tools, and information on the timing or necessity of maintenance work.
14. Method according to one of the preceding claims comprising the additional steps: - Obtaining or determining (i) a variety of training data, including training process data, used to train the Kl model to determine the expected value (E) and / or the extended expected value (EE), or (ii) an initial statistical analysis based on the variety of training data, - Determining a model fit based on the provided process data (P) and (i) the multitude of training data or (ii) the first statistical analysis, wherein the model fit is characterized by the reliability of a multitude of assessment values, expected values (E), and / or extended expected values (EE), in particular further comprising the steps of: o Obtaining, determining, or ascertaining detection limits, in particular predetermined detection limits, wherein the detection limits define the predictive accuracy of the model fit for specific welding situations with which the multitude of assessment values (B), expected values (E), and / or extended expected values (EE) provide meaningful information, o Comparing the model fit with the detection limits to determine whether the provided process data (P) are suitable for the permissible or meaningful determination of the multitude of assessment values (B), expected values (E),and / or extended expected values (EE) are suitable. PTESIOIWO / 04.09.2025 45 2025094075 15. as described in claim 14, wherein the determination of the model quality is additionally based on - At least a large number of the assessment values, expected values, extended expected values, and / or properties of the welded component of the training data are performed, or - A second statistical analysis will be performed based on the multitude of assessment values, expected values, extended expected values, and / or properties of the training data.
16. Method for producing a connection between at least two components (40) using a welding device, in particular an ultrasonic welding connection using an ultrasonic welding device (20), comprising the following steps: - Performing a welding of at least one component (41) with the welding device (20), wherein the at least one component (41) is optionally a metal component or a plastic component, in particular a vehicle component, preferably an automotive component, a medical device component, and / or an electronic component, - Determining and providing an assessment value (B) and an expected value (E) using a method for non-destructive testing of a compound according to one of claims 1 to 15.
17. Method for producing a connection between at least two components (40) using a welding device, in particular an ultrasonic welding connection using an ultrasonic welding device (20), comprising the following steps: PTESI OIWO / 04 .
09. 2025 46 2025094 075 - Providing machine limits, whereby the machine limits define permissible operating parameters for the specific welding situation, with which operating parameters a KL model can generate reliable information about the quality of the welded joint. - Performing a weld with operating parameters that are within the machine limits - Determining a quality prediction for the welded joint using the Kl model, in particular according to one of claims 5 to 12.
18. Method for automatic non-destructive testing of a connection between at least two components ( 40 ), in particular an ultrasonic welded joint, comprising the following steps - Providing process data which represent at least one process parameter recorded during the welding process in a computer arrangement ( 24 ) - Determining an assessment value (B ) and checking the assessment value with regard to the fulfillment of target criteria, based on the process data, using the computer arrangement ( 24 ) - Determining an expected value (E) that characterizes the reliability of the assessment value (B) based on the process data, using the computer arrangement (24) - Providing properties of the welded component ( 41 ) for extended testing in the computer arrangement ( 24 ) PTESIOIWO / 04.09.2025 47 2025094075 - Determining a quality prediction (Q; QE) with the computer arrangement (24) which characterizes the expected quality of the welded joint, wherein the determination is based on at least one value selected from the o assessment value (B) o properties of the welded component (41) o expectation value (E;EE) .
19. Welding device (20) with a computer arrangement (24) wherein the computer arrangement (24) is configured to perform the method according to any one of claims 1 to 16 .
20. Welding device (20) according to claim 19, comprising an ultrasonic generator (26), a converter (22), a sonotrode (21), an anvil, and means (23, 25) for determining process data.
21. Computer program product comprising instructions which, when the program is executed by a computer, cause it to execute a method according to any one of claims 1 to 16.
22. Computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to execute a method according to any one of claims 1 to 16.
23. Machine learning model for predicting the weld quality of a joint between two components, wherein the machine learning model is based on a mixed model which is based on a first model and a second model, PTESI OIWO / 04.09.2025 48 2025094 075 wherein the first model was trained with a first subset of training data and the second model with a second subset of training data, wherein the first subset is based on training data having a first category of properties, and the second subset is based on training data having a second category of properties, wherein in particular the first category of properties represents the n best welding results, in particular pullback values, and the second category of properties represents the m worst welding results, in particular pullback values.
24. Machine learning model according to claim 23, wherein the machine learning model is a combined model which, starting from a plurality of process parameters, in particular from power, force and travel distance, is based on a plurality of first and / or a plurality of second models, wherein in particular the combined model is based on a plurality of mixed models which were generated starting from the plurality of first and second models.
25. Machine learning model according to one of claims 23 or 24, wherein the machine learning model is an incrementally and / or continuously online learning machine learning model, and in particular is connected or connectable to a computing unit of a cloud-based and / or edge computing platform and / or is executed on it.
26. Computer-implemented method for processing training data for use in a machine learning model to predict weld quality, comprising the steps: PTESI OIWO / 04 .
09. 2025 49 2025094 075 - Obtaining or determining training data, identifying properties in training data that represent the quality of welding results, especially pullback values, - Forming at least a first and a second subset of training data, wherein the first subset is based on training data having a first category of properties, and the second subset is based on training data having a second category of properties, wherein in particular the first property represents the n best welding results, especially pullback values, and the second property represents the m worst welding results, especially pullback values.
27. Methods for training a machine learning model to predict weld quality include the steps of: identifying properties in training data that represent the quality of the welding results, in particular pull-off values; and forming at least a first and a second subset of training data, wherein the first subset is based on training data exhibiting a first category of properties, and the second subset is based on training data exhibiting a second category of properties, wherein in particular the first property represents the n best welding results, in particular pull-off values, and the second property represents the m worst welding results, in particular pull-off values. PTESI OIWO / 04 . 09 . 2025 50 2025094 075 o Obtaining a first and second subset of training data according to claim 28 - Training a first and a second model based on the first and second subsets - Forming a mixed model based on the first and second models.
28. Method according to claim 27, wherein a plurality of first and / or a plurality of second models are trained, starting from a plurality of process parameters, in particular starting from power, force and travel distance, and wherein a plurality of mixed models are generated starting from the plurality of first and second models, and wherein the plurality of mixed models are combined into a combined model.
29. Training data set for use in a method for training a machine learning model to predict weld quality, in particular according to claim 27 or 28, comprising a first and second subset of training data, wherein the first subset is based on training data having a first category of properties, and the second subset is based on training data having a second category of properties, wherein in particular the first property represents the n best welding results, in particular pull-off values, and the second property represents the m worst welding results, in particular pull-off values.
Citation Information
Patent Citations
Method and apparatus for testing objects intended for ultrasonic welding
DE102021102100A1
Assessment procedures for evaluating a processing procedure, training procedures for training an AI system, and ultrasonic processing equipment
DE102021111962A1
Method for determining the temperature in a joining zone
DE102021117697B3
Methods for determining process parameters for an ultrasonic machining process and ultrasonic machining tools
DE102022112098A1
Method for quality control in ultrasonic welding and associated apparatus
DE3429776A1