COMPUTER-IMPLEMENTED METHOD FOR THE ANALYSIS OF MEASUREMENT DATA FROM A MEASUREMENT OF AN OBJECT

Description

[0001] The invention relates to a computer-implemented method and computer program product for analyzing measurement data from a measurement of an object, wherein the analysis evaluates whether the object corresponds to a target state.

[0002] Non-destructive testing can be used for quality control of objects to capture their structure, which is not visible from the outside. These objects can be, for example, components that are assembled into larger objects. If the non-destructive testing reveals defects in the component, such as pores, cavities, inclusions, areas of increased porosity, or structural loosening, etc., it must be assessed whether these defects impair the object's functionality. A conformity decision is then made, determining whether the object is acceptable or not, and thus whether it conforms to the requirements defined in the technical drawing, product specification, or elsewhere. This quality control is particularly relevant for objects manufactured using additive manufacturing or casting processes, such as injection molding, die casting, or die casting.However, the fundamental necessity of this quality control is independent of the manufacturing method of the object.

[0003] From WO 2018 / 217903 A1, a method is known in which a knowledgeable user provides training data for defect detection in an additive manufacturing process. A machine learning algorithm is trained using this training data.

[0004] There are automatable algorithms that assess whether an object is functioning correctly or not, based on a user-defined set of criteria. This predefined set of criteria generally infers the object's functionality from its geometric properties. However, this approach is only possible with a certain degree of imprecision, as further information about the defect is usually relevant for assessing the object's functionality, information which the algorithms do not consider.

[0005] For these reasons, the assessments of the automated algorithms are submitted to an expert for manual review if these algorithms cannot clearly determine whether relevant defects are present and / or if every detected defect should be reviewed by an expert as a precaution. If the expert review reveals that the algorithm's decision was incorrect, the expert corrects it. Similarly, cases in which the automated algorithm did not reach a clear result can be submitted to the expert. Both procedures, however, are time-consuming and require a relatively high level of personnel.

[0006] Against this background, the present invention is based on the objective technical problem of creating an improved method for analyzing measurement data from a measurement of an object.

[0007] The main features of the invention are specified in claims 1 and 14. Embodiments are the subject of claims 2 to 13.

[0008] In a first aspect, the invention relates to a computer-implemented method according to claim 1.

[0009] The computer-implemented method for analyzing measurement data from a measurement of one object automatically acquires measurement data from a large number of objects. Measurement data can be acquired based on similar measurement tasks, meaning that the objects are measured or analyzed, for example, at similar locations, in a similar manner, or to detect similar deviations. The acquisition of measurement data can be performed, for example, using imaging techniques such as computed tomography, although this does not exclude other measurement methods. The acquisition of measurement data can be performed "inline" or "atline," thus concurrently with the manufacturing process. Furthermore, the acquisition of measurement data can involve using measurement data, for example, from previously performed measurement series. The measurement data from these previously performed series can be loaded from a data carrier or other storage medium or location.

[0010] Analysis data sets are automatically generated from the measurement data of the objects. These analysis data sets can be generated, for example, using the machine learning algorithm that is to be trained. According to the invention, a different, non-machine learning, i.e., conventional, algorithm generates the analysis data sets. Each analysis data set is assigned to an object. Furthermore, the analysis data sets include at least one analysis result regarding the conformity of the actual state of the assigned measured object with the target state of the object. An analysis data set can comprise a multitude of analysis results from different analyses of an object. The algorithm used, for example, provides an assessment of whether a component, with regard to its geometry and material properties, is within the tolerances specified by the designer and can therefore fulfill its intended function.Furthermore, the analysis results can include information such as where a critical defect was identified and / or which regions were analyzed and how. Additionally, the analysis results can include a section of the (raw) measurement data or a visualization of the defect, which enables or simplifies verification for the user.

[0011] At least some of the analysis data sets are reviewed by a user. The user checks the individual analysis results within each data set. Only those analysis data sets whose results indicate a nonconformity of the object with the target state can be selected for review. Alternatively, only those analysis data sets can be selected for review where the analysis results do not provide a clear indication of the object's conformity. This might mean, for example, that the algorithm used cannot clearly determine from the analysis results whether the object is functional. Another alternative is for the user to review all analysis data sets. A nonconformity of the object with the target state refers to deviations from the target state, such as defects.

[0012] If individual analysis results are assessed differently by the user than they are stored in the verified analysis data set, the corresponding differently assessed analysis result will be changed in the corresponding verified analysis data set.

[0013] At least the modified analysis data sets resulting from user review are transmitted to the machine learning algorithm. The machine learning algorithm uses these modified data sets to train itself. In doing so, it modifies itself based on the modified data sets. Using these verified data sets, the machine learning algorithm learns to determine the functionality of objects—that is, their conformity with the target state—based on measurement data and to provide corresponding analysis results. The machine learning algorithm can then independently create analysis data sets from additional measurement data, which may have a lower user review rate than before the training process.

[0014] According to an exemplary embodiment, the preceding steps can be carried out sequentially, with measurement data for all objects being determined first, before an analysis takes place, and the verification being performed only after the analysis is complete. In an alternative example, provided the logical prerequisites are met, at least a partial temporal overlap of the steps can be provided, so that, for example, the determination of the analysis data sets for the measurements already performed can begin while the measurement data is being determined. Furthermore, if the corresponding prerequisites are met, the subsequent steps can, for example, be carried out while the aforementioned steps are being performed.The verification process can begin when the relevant analysis datasets are available during the dataset generation phase, and the modification of the machine learning algorithm can be performed as soon as the transmission of updated analysis datasets begins. Furthermore, iterative repetition of individual steps or sequences of steps is also conceivable.

[0015] The invention thus provides a computer-implemented method for analyzing measurement data, generating real-world training data for a machine learning algorithm to perform the analysis task. By training the machine learning algorithm with real-world training data from ongoing measurements, the trained algorithm can save time in the long run compared to manually monitoring conventional analysis algorithms, as fewer unclear cases are presented to the user. Furthermore, the reduced effort required by the user reduces the potential for errors.

[0016] The determination of analysis data sets from the measurement data for the objects is carried out using an evaluation algorithm different from the learning algorithm, the procedure additionally comprising the following step: replacing the evaluation algorithm with the learning algorithm after a predefined minimum number of adapted analysis data sets have been transmitted to the learning algorithm and / or after a predefined minimum number of analysis data sets have been determined.

[0017] This ensures that an evaluation algorithm first determines the analysis data sets. The machine learning algorithm receives training data from the evaluation algorithm's analysis data sets, which are reviewed by a user, allowing it to modify itself. Furthermore, the machine learning algorithm is only used to determine the analysis data sets once it has received a predefined minimum number of adapted analysis data sets as a basis for its own modifications. This predefined minimum number can be chosen by estimation such that, after training with this minimum number of analysis data sets, the machine learning algorithm produces fewer analysis results requiring user review or correction than the evaluation algorithm. Alternatively, the machine learning algorithm can replace the evaluation algorithm once a predefined minimum number of analysis data sets has been determined.The predefined minimum number of analysis data sets can be based on an estimate of the associated number of user reviews. Both of these factors increase the efficiency of training the machine learning algorithm and reduce the number of user reviews during the training process.

[0018] According to another example, the determination of analysis data sets from the measurement data for the objects can be carried out using an evaluation algorithm different from the machine learning algorithm, and the procedure, after the user has checked the analysis data sets, includes the following step: determining training analysis data sets from the measurement data for the objects using the machine learning algorithm; and comparing the adapted analysis data sets with the corresponding training analysis data sets before submitting the adapted analysis data sets to the machine learning algorithm; wherein the evaluation algorithm is replaced by the machine learning algorithm if at least some of the adapted analysis data sets match the corresponding training analysis data sets.

[0019] This allows the machine learning algorithm to determine training analysis datasets from the measurement data before or during its training using the verified and adapted analysis datasets. These training analysis datasets serve only for comparison with the adapted analysis datasets and are not used to assess the conformity of the objects with the target state. The machine learning algorithm only replaces the conventional evaluation algorithm in determining the analysis datasets when the comparison between the training analysis datasets and the analysis datasets of the evaluation algorithm reveals a sufficiently high degree of agreement. The agreement between the conventional evaluation algorithm and the training datasets can also be taken into account.can be used as a reference to determine a suitable time at which the learning algorithm replaces the conventional evaluation algorithm.

[0020] According to another exemplary embodiment, the method may additionally include the step of providing the adapted analysis data sets for the assigned objects via an output unit.

[0021] This allows the user-customized analysis datasets to be directly output as the final analysis result via an output unit. The customized analysis datasets are thus used both to improve the machine learning algorithm and as the final result of quality assurance.

[0022] It can also be advantageous if the procedure includes the following step before the analysis data sets are checked by a user: marking an analysis data set if at least one analysis result regarding the conformity of the assigned object with the target state is not clear; and using only the marked analysis data sets when checking the analysis data sets by the user.

[0023] The user is thus presented with only uncertain analysis results for review; that is, results that the process classifies as either not clearly acceptable or not clearly acceptable. Results classified as acceptable are not reviewed by the user. This reduces the effort and increases the speed of the process. The classification as an uncertain analysis result can be achieved, for example, by using a three-tiered tolerance range. Each tolerance range comprises analysis results classified as acceptable or unacceptable. Between these two tolerance ranges is a third tolerance range in which the process cannot make a clear assessment. These analysis data sets, containing an analysis result classified in the third tolerance range, are reviewed manually by the user.Alternatively, the procedure can provide an output variable used to check the conformity of the assigned object. This output variable can indicate how closely the examined structure resembles a target pattern, which might, for example, be a problematic defect. Depending on the definition, this output variable could be 1 if it is definitely a problematic defect. A value of 0 would indicate that it is not a problematic defect. A value of 0.5 would therefore indicate a high degree of uncertainty. In this way, a measure of uncertainty can be implicitly derived. Furthermore, the classification can alternatively be based on the output of an independent, dedicated uncertainty measure. This uncertainty measure could be generated, for example, by a conventional algorithm or by another machine learning algorithm.A simple example would be the specification of a signal-to-noise ratio. The noisier the data, the less certain the analysis results. The subsequent learning algorithm can be a separate algorithm for calculating the actual analysis result, or it can be a combined algorithm that determines both the analysis result and its associated uncertainty.

[0024] The measurement data can provide at least a partial representation of a volume arranged within an object.

[0025] This allows for the analysis of an object's interior. Deviations from the target state within the object that affect its functionality can be identified. The analysis can be performed based on volumetric data. Alternatively, other data depicting the object's interior, such as two-dimensional radiographs, can be analyzed.

[0026] The process of obtaining measurement data from a large number of objects may further involve the following step: providing volume data using a computed tomography measurement.

[0027] Computed tomography allows for the efficient provision of meaningful, high-resolution volumetric data. This data captures the object's structure in its entirety, enabling the detection of variations in both volume and surface texture.

[0028] Furthermore, it can be advantageous that the determination of measurement data for a large number of objects includes the following step: providing volume data as measurement data using process data from a measurement during the additive manufacturing of an object.

[0029] This allows the internal structure of the object to be determined directly from the process data of a measurement during its manufacture. The process data can be spatially resolved information about physical quantities, such as the proportion of laser power reflected by the object while a volume element of the object is being manufactured. Immediately after or even during the object's completion, this data can be analyzed for deviations from the target state; that is, the analysis data sets can be determined immediately. Because defects manifest themselves in this data in a comparatively complex way, only a few algorithms suitable for such automated analyses currently exist, and a large number of analysis data sets have typically had to be reviewed by a user.By training and using the adaptive algorithm, the number of analysis data sets that a user has to check can be reduced.

[0030] The process of deriving analysis data sets from the measurement data for the objects may include the following step: evaluating deviations inside the object.

[0031] If the measurement data indicates an internal deviation within the object, this deviation can be evaluated as part of the analysis to determine the object's functionality. Deviations or defects that do not affect the object's functionality can, for example, be assessed as irrelevant, resulting in a positive analysis result, i.e., indicating that the object is functioning correctly with respect to this analysis. Conversely, the defect can be assessed as relevant, and the analysis result will indicate that the object is not functioning correctly.

[0032] The defects or deviations from the target state of the object can be, for example, air inclusions, which can have a wide variety of shapes, sizes, positions and other properties and can negatively affect the functionality of the object.

[0033] Furthermore, the determination of analysis data sets from the measurement data for the objects, prior to evaluating deviations within the object, can include the following step: determining segmentation data based on the measurement data, where the segmentation data describes an internal composition of the object; where the evaluation of deviations within the object is carried out on the basis of the segmentation data using the machine learning algorithm.

[0034] This allows the geometry of individual defects or deviations to be measured. Therefore, the system not only checks for the presence of a defect but also determines its shape. This can be done at the voxel level or with sub-voxel precision. The machine learning algorithm thus receives training data with additional parameters related to the deviations, based on which an analysis result and a corresponding modification of the machine learning algorithm can be performed. The segmentation data, which can be determined, for example, by a separate segmentation algorithm, can describe the composition of the object by providing spatially resolved information on whether a region or volume element contains material, air, or defects. The segmentation algorithm itself can also be a machine learning algorithm, which, for example, has been trained using simulations.

[0035] Furthermore, the determination of analysis data sets from the measurement data for the objects, prior to evaluating deviations inside the object, can include the following step: determining a local wall thickness at a location of a deviation; whereby the evaluation of deviations inside the object is carried out based on the local wall thickness.

[0036] Deviations or defects within an object generally weaken its internal structure. Even with the same size defect, its impact on the object's functionality can be greater if it is located in an area of ​​thinner walls than if it is located in an area of ​​thicker walls. Therefore, the influence of a defect on an object's functionality can be determined, among other things, based on wall thickness.

[0037] Advantageously, transmitting at least the adapted analysis data sets to the learning algorithm, whereby the learning algorithm itself changes based on the adapted analysis data sets, includes the following step: transmitting simulated analysis data sets based on simulated measurement data to the learning algorithm, whereby the learning algorithm itself changes based on the simulated analysis data sets.

[0038] Simulation allows the entire measurement process, such as in computed tomography (CT), consisting of X-ray imaging of the object, reconstruction, and data analysis, to be realistically reproduced. By simulating analysis datasets, a large number of datasets can be provided to the machine learning algorithm, which can then adapt to these datasets. A key advantage is that the input parameters of the simulation, and thus the geometry of the object, are known. This allows for the automated derivation of ground truth for conformity decisions, eliminating the need for additional user input. The simulated analysis datasets can be used in addition to those generated from actual measurements.This eliminates the need to wait for a large number of measurements during the ongoing production process to train the machine learning algorithm. Especially with a small number of deviations from the target state in the manufacturing process, accumulating suitable analysis datasets, which the machine learning algorithm can use to modify itself, can take some time. Using these simulated analysis datasets, the machine learning algorithm can thus reach a state more quickly where it can detect deviations from the target state with greater certainty than a conventional, non-machine learning algorithm.

[0039] In an exemplary embodiment, the method may, prior to determining analysis data sets from the measurement data for the objects, include the following steps: Determining preliminary analysis data sets from the measurement data for the objects using a defect detection algorithm, wherein a preliminary analysis data set is assigned to one of the objects and contains at least one analysis result regarding the conformity of the assigned object with the target state; Determining, using the defect detection algorithm, whether the analysis result indicates a deviation of the conformity of the assigned object to the target state within a predefined range;Transmitting the measurement data of the objects whose preliminary analysis data sets show an analysis result indicating a deviation of the assigned object's conformity to the target state within the predefined range to the machine learning algorithm for determining analysis data sets from the measurement data for the objects.

[0040] In this example, the identification of analysis data sets from the measurement data for the objects is performed using the machine learning algorithm. Each analysis data set is assigned to one of the objects and contains at least one analysis result regarding the conformity of the assigned object with the target state. Using the defect detection algorithm, which in this example is not the machine learning algorithm but a conventional algorithm, measurement data that clearly indicate deviations or clearly indicate no deviations from the target state can be filtered before the machine learning algorithm checks the measurement data. This ensures that the machine learning algorithm only analyzes measurement data for objects that do not yield clear analysis results using the defect detection algorithm. Since the machine learning algorithm typically requires more computing power to execute than a conventional non-machine learning algorithm, i.e.,If the process is slower than the defect detection algorithm, the entire procedure can be accelerated if, in the case of unambiguous analysis results, analysis by the learning algorithm is omitted.

[0041] In an alternative or additional example, analyses that have already yielded a positive result can be repeated after the machine learning algorithm has modified itself based on the adjusted analysis datasets. In this way, if necessary, these analyses can be subsequently categorized as incorrect or ambiguous. This allows objects that were incorrectly categorized as correct to be identified as such or presented to the user for a decision. The initial analyses are thus considered provisional until the machine learning algorithm has modified itself sufficiently to achieve an acceptably low error rate. The acceptable error rates can be predefined by the user.As long as the analysis is considered provisional, the corresponding objects are not yet used further, e.g., delivered or processed further. The newly performed, second analysis by the machine learning algorithm is assigned to the corresponding analysis dataset or measurement data. Similarly, objects that were incorrectly categorized as defective can subsequently be identified as acceptable or as ambiguous and presented to the user for a decision. In this way, unnecessary waste can be minimized.

[0042] Alternatively or additionally, it can be stipulated that all analyses are considered preliminary. The measurement data are stored until the machine learning algorithm achieves an acceptably low error rate. Afterwards, all measurement data are re-analyzed for conformity, and only then, depending on the result, are the corresponding objects classified and, if necessary, used further.

[0043] In another example, the analysis data sets are generated by examining defects in the measurement data with regard to geometric properties such as size, shape, orientation, location in the component, but also the proximity to other defects, and deriving a statement about the conformity of the object from this.

[0044] The method can also be used to analyze more complex geometries, such as foam structures, for conformity. Furthermore, image data of objects can be examined in this way to determine the presence of structures or geometries. Examples include whether a required solder joint is missing or whether the assembly of components, such as a printed circuit board or a connector with corresponding connections, has been carried out correctly and in accordance with the desired target state.

[0045] Furthermore, in another example, the user can be presented with samples of the analysis data sets identified by the machine learning algorithm for evaluation. The sample can be randomly selected or specifically comprise analysis data sets containing relatively clear assessments. The analysis data sets reviewed by the user can then be presented to the machine learning algorithm, allowing the algorithm to modify itself based on these data sets. This minimizes the risk of the machine learning algorithm being incorrectly trained with regard to certain characteristics.

[0046] In another example, analysis datasets from different measurement systems can be merged, with the machine learning algorithm using these merged datasets to modify itself. This merging of the analysis datasets can be performed by various users, with the datasets being transmitted to a central location, for example, via a network application. Ideally, this involves investigating identical or similar measurement tasks with identical or similar acquisition parameters.

[0047] Furthermore, it may be provided, for example, that the learning algorithm receives separate analysis data sets for different areas of the object, in which, for example, different tolerances may be defined, in order to modify itself on the basis of these.

[0048] Alternatively or additionally, different machine learning algorithms can be used for the analysis in the different areas, i.e., the different machine learning algorithms can be specialized for a specific area.

[0049] In another aspect, the invention relates to a computer program product with instructions executable on a computer, which, when executed on a computer, cause the computer to carry out the method according to the preceding description.

[0050] The advantages and further development opportunities of the computer program product result from the preceding description.

[0051] Further features, details and advantages of the invention will become apparent from the wording of the claims and from the following description of exemplary embodiments with reference to the drawings. The drawings show: Fig. 1 a schematic representation of the computer-implemented method for analyzing measurement data from a measurement of an object, Fig. 2 a schematic representation of a prior art algorithm, Fig. 3 a schematic representation of deviations in measurement data of an object, Fig. 4a, b flowcharts of various embodiments of the method, Fig. 5 a flowchart of an embodiment of a method step, and Fig. 6 a partial flowchart of a further embodiment of the method.

[0052] In Fig. 1 An embodiment of the computer-implemented method according to the invention is presented. The procedure is explained in more detail using the example of measuring an object 30. However, the method is used to measure and analyze a large number of objects.

[0053] In this example, object 30 is automatically measured using an imaging technique. However, the measurement can also be performed manually, for example. The imaging technique yields measurement data 44 of object 30. An analysis data set is automatically generated from this measurement data 44 using an algorithm and assigned to object 30. This analysis data set includes analysis results 46 and 48 regarding the conformity of the assigned object 30 with its target state. That is, the analysis results 46 and 48 assess whether the corresponding measured values ​​evaluated in the respective analyses lie within or outside of predefined tolerance ranges.

[0054] Analysis results 46 indicate that the corresponding analyses at specific positions in object 30 are OK, meaning they are within the predefined tolerance ranges. Analysis result 48 is marked with an exclamation mark (!), indicating that it requires review. It may be specified that analysis results are reviewed if they show a result outside the predefined tolerance ranges, i.e., if an analysis at a specific position in object 30 reveals that object 30 is not OK at that position. Alternatively or additionally, it may be specified that analysis results are reviewed if they are marked as uncertain, meaning the algorithm cannot determine, or cannot determine with the required certainty, whether the analysis result is rated as OK or not OK.

[0055] The analysis data sets containing analysis results 48 that need to be verified are presented to a user 52 via a user interface 50. The user interface 50 can be, for example, a computer monitor or a touchscreen, but is not limited to these examples.

[0056] User 52 checks whether analysis result 48 is correct. If they are satisfied with analysis result 48, it is not changed. Furthermore, user 52 has the option to change analysis result 48 to an adjusted analysis result 54 if they conclude that analysis result 48 is incorrect.

[0057] The adapted analysis dataset with the adapted analysis result 54 is transmitted to the machine learning algorithm. The machine learning algorithm then changes its state from an initial state 60 to a modified state 62. The transition from the initial state 60 to the modified state 62 can, for example, be achieved by changing a step 64 of the algorithm's initial state 60 to a step 66 of the modified state 62. In the modified state 62, the machine learning algorithm is improved compared to the initial state 60 and, in future analyses based on a similar measurement task underlying the adapted analysis dataset, will be more likely to deliver an analysis result 54 that does not require user verification.

[0058] The identification of analysis data sets prior to user review can be performed using a conventional, state-of-the-art algorithm. Such a conventional algorithm evaluates whether an object is OK or not based on predefined decision criteria. An example of a decision tree underlying the decisions of a conventional algorithm is shown in Figure 2 depicted.

[0059] The decision tree indicates according to Figure 2Several conditions, 12 to 22, are linked together using either "AND" or "OR" operators. For example, condition 12 might require a wall thickness within a predefined tolerance. Furthermore, condition 14 might require a pore size in the object's material that is less than 0.3 mm. Conditions 12 and 14 must be met simultaneously. Alternatively, conditions 16 and 18 must be met simultaneously, with condition 16 requiring that the wall thickness be no more than 0.1 mm outside the tolerance and condition 18 requiring that no pores be present in the material.

[0060] Additionally, condition 20 must be met, which requires a deviation from a CAD model of less than 0.1 mm. Furthermore, all other measurement results must fall within the tolerances according to condition 22.

[0061] Based on this decision tree, result 24 indicates that object 30 is either OK or not OK. However, this does not necessarily guarantee the object's functionality, as the decision tree cannot perfectly represent the complex relationships that influence functionality. To avoid unnecessary scrapping of otherwise functional objects 30, a user should therefore perform an inspection if there are deviations from the object's target state, for example, if result 24 indicates that object 30 is not OK. Furthermore, the decision tree can also be structured so that result 24 indicates that a clear decision cannot be made and must be verified by the user. These decisions can apply to object 30 as a whole, as well as to individual defects or areas within it.to be analyzed at object 30.

[0062] An example of deviations from the target state of an object are defects in composite fiber materials, which are found in Figure 3 The following are shown. A cross-sectional image 32 from the volume of object 30 is shown, which may, for example, have been determined based on a computed tomography measurement. A fiber break is shown in area 34, where the fiber has been severed. A missing fiber fragment is shown in area 36.

[0063] The area where a fiber was previously present is now an empty space within the matrix material in which the fibers are embedded. These empty spaces may contain air inclusions. In area 38, the connection between the fiber and the matrix material has been lost, resulting in another vacancy within the matrix material. Area 40 shows fractures in the structure of the matrix material.

[0064] Areas 34, 36, 38, and 40 are connected by a subsequently occurring break 42, which runs across section 32. Break 42 can be a consequence of the defects in areas 34, 36, 38, and 40. Individually, each of the defects in areas 34, 36, 38, and 40 might not impair the usability of object 30. However, in this example, break 42, taken together, leads to a lack of functionality for object 30. Attempting to represent the lack of functionality of object 30 due to the potential occurrence of a break using a predefined decision tree and a conventional algorithm can lead to incorrect decisions if the defects in the analyzed areas deviate only slightly.

[0065] In an alternative or additional example, one can simply analyze whether voids or air inclusions in the material negatively affect the functionality of object 30. This simplifies and accelerates the analysis. The material of object 30 does not necessarily have to be a composite fiber material, but could be, for example, a plastic, a metal, or a ceramic, etc., that contains air inclusions.

[0066] The computer-implemented method 100 according to the invention, which supplies a learning algorithm with realistic, user-adapted analysis data sets to improve the learning algorithm, is described in the Figure 4a and 4b illustrated, which show exemplary different embodiments of method 100.

[0067] According to Figure 4aIn a first step, procedure 102 comprises the acquisition of measurement data for a large number of objects. The acquisition of this measurement data can be based on similar measurement tasks. A large number of objects are measured sequentially or simultaneously using the same or similar analyses and / or with the same or similar analytical objectives. The acquisition of measurement data can also involve loading previously performed measurements from a memory. Measurement tasks for composite materials, for example, might include searching for voids in the matrix material caused by fractures, fiber loss, or fiber detachment. For cast objects or objects manufactured using additive manufacturing processes, the measurement task might include, for example, detecting voids or foreign material inclusions. However, the measurement tasks can vary further and be individually tailored to a specific object or process.The manufacturing process of the object must be coordinated.

[0068] The measurement data can be, for example, volume data obtained using computed tomography. The computed tomography measurement takes place during step 102 in step 124 after the object has been manufactured.

[0069] Alternatively or additionally, in an additive manufacturing process, the process data determined during manufacturing can be provided as volume data in step 126 for step 102 of the object's production. This process data is directly spatially resolved and therefore available even during the object's manufacture. This allows analyses of the already completed parts of the object to be performed during its production.

[0070] In other examples of determining volume data not shown, measurement data from ultrasound procedures, magnetic resonance imaging and other imaging procedures can also be used.

[0071] However, the possible analyses are not limited to volumetric data. Deviations from the object's conformity with the target state can also be determined using two-dimensional radiographic images, which are provided, for example, by radiographic methods, or by optical inspection of objects using camera images.

[0072] In step 104, analysis data sets are determined from the measurement data for the objects. Each analysis data set is assigned to one of the objects. Furthermore, each analysis data set contains at least one analysis result regarding the conformity of the assigned object with the target state of the object. The determination of the analysis data sets is performed automatically using a computer-implemented algorithm. In a first exemplary embodiment, the computer-implemented algorithm can be a self-learning algorithm that is trained and improved in subsequent steps. In another exemplary embodiment, the determination of the analysis data sets can be carried out by a conventional algorithm according to the state of the art.

[0073] According to Figure 5Step 128 can be included in step 104, which evaluates deviations from the target state inside the object. These deviations could be defects, for example. In this case, the measurement data also shows the interior of an object and not just its surfaces.

[0074] Optionally, further options can be found according to Figure 5 While step 104, in a further step 130, provides for the determination of segmentation data based on the measurement data. This segmentation data describes the internal spatial composition of the object. Furthermore, the evaluation of the deviations within the object is carried out using the machine learning algorithm based on this segmentation data. Based on the segmentation data, the deviations can be expressed geometrically. That is, the geometric shape of the region in which the deviation is located can be determined.

[0075] Furthermore, according to Figure 5 Optionally, step 132 can be included in step 104. This involves determining a local wall thickness at a point of deviation. The evaluation of deviations within the object is then performed based on this determined local wall thickness. This evaluation can be conducted with regard to the object's functionality.

[0076] In Figure 4aStep 120 is further illustrated, in which an analysis record is marked if it contains at least one ambiguous analysis result regarding the conformity of the assigned object with the target state. Only the marked analysis records are used in a subsequent step 122, which are then presented to the user for review in step 106. The analysis records can be marked using standard data processing techniques. By marking the analysis records to be reviewed, the user is only presented with uncertain or ambiguous analysis results for review. Certain analysis results do not need to be reviewed by the user. Steps 120 and 122 are optional.

[0077] In step 106, the user reviews the analysis results of at least some of the analysis data sets. The user has the opportunity to examine the automatically generated analysis results. This allows the user to view the measurement data and assess the analysis results accordingly. Depending on the implementation, the user may be presented with only the marked analysis data sets or also with additional data sets, such as all analysis data sets with negative results, i.e., also the unambiguous results.

[0078] According to step 108, the user can adjust an analysis result in an analysis dataset based on the measurement data if they disagree with the analysis result, i.e., if the user determines an analysis result that differs from the original analysis result regarding the conformity of the assigned object with the target state. This then results in an adjusted analysis dataset.

[0079] The adjusted analysis datasets can be provided via an output unit as described in step 118, and thus directly output as the final result of the analysis. The adjusted analysis datasets are therefore used as the final result of quality assurance.

[0080] In step 110, the modified analysis datasets are transmitted to the machine learning algorithm. The machine learning algorithm then modifies itself based on these modified datasets. This means that the machine learning algorithm improves based on the modified datasets. The improved machine learning algorithm can then derive analysis datasets from additional measurement data of objects that require less user verification compared to the datasets before the modification.

[0081] If the machine learning algorithm was used in step 104 to determine the analysis data sets from the measurement data for the objects, the combination of steps 106 to 110 will directly improve the analysis results of the machine learning algorithm.

[0082] In an alternative embodiment, if a conventional algorithm determines the analysis data sets from the measurement data for the objects in step 104, the adaptive algorithm improves upon the conventional algorithm in step 110. In this case, the conventional algorithm can be an evaluation algorithm. The adaptive algorithm replaces the conventional algorithm in step 112 if a predefined minimum number of adapted analysis data sets have been submitted to the adaptive algorithm and / or after a predefined minimum number of analysis data sets have been determined by the conventional algorithm.

[0083] Another alternative embodiment of method 100 is described in Figure 4b As shown. Instead of step 112 after step 110, steps 114 and 116 can be placed before or after step 110. In Figure 4bThis only illustrates the execution of steps 114 and 116 before step 110, thus not excluding the alternative after step 110. The explanations above apply to the subsequent steps performed before steps 114 and 116.

[0084] In step 114, the machine learning algorithm determines training analysis datasets based on the measurement data. This determination of training analysis datasets is similar to the determination of analysis datasets according to step 104. However, the training analysis datasets are not reviewed by a user and are not used to determine the functionality of the object.

[0085] The adjusted analysis datasets, i.e., the analysis datasets reviewed and modified by the user, are compared in step 116 with the corresponding training analysis datasets, each associated with the same object, before the adjusted analysis datasets are submitted to the machine learning algorithm. The evaluation algorithm in step 104 is replaced as soon as the machine learning algorithm identifies at least some training analysis datasets that produce the same result as the adjusted analysis datasets. That is, as soon as the machine learning algorithm produces fewer analysis results requiring user review than the conventional algorithm, the conventional algorithm is replaced by the machine learning algorithm according to step 116.

[0086] Optionally, step 110 in all embodiments can further include step 134, in which simulated analysis data sets are transmitted to the machine learning algorithm. The machine learning algorithm then modifies itself based on the simulated analysis data sets. These simulated analysis data sets are based on simulated measurement data resulting from a realistic simulation.

[0087] In Figure 6 Another exemplary embodiment is described in which steps 136, 138 and 140 are performed between steps 102 and 104.

[0088] According to step 136, after the measurement data has been acquired, preliminary analysis data sets are generated from the measurement data for the objects using a defect detection algorithm. The defect detection algorithm can be a conventional algorithm. A preliminary analysis data set is assigned to one of the objects and contains at least one analysis result regarding the conformity of the assigned object with the target state.

[0089] In step 138, the preliminary analysis result is checked by the defect detection algorithm to determine whether it shows a deviation of the assigned object's conformity to the target state within a predefined range. This predefined range may include analysis results that do not allow for a clear assessment of the functionality of the measured assigned object.

[0090] Preliminary analysis results that are not assigned to the predefined range are output as final quality control analysis results. If a preliminary analysis result is assigned to the predefined range, the measurement data underlying the analysis result are transmitted to the machine learning algorithm according to step 140, which then repeats step 102 instead of the defect detection algorithm. The analysis data set resulting from the machine learning algorithm is then output as the quality control result if no user review is required.

[0091] The computer-implemented method 100 described above can further be carried out in any embodiment by a computer which, controlled by a computer program product, executes instructions that cause the computer to carry out the computer-implemented method 100.

[0092] The preceding steps can be carried out sequentially or with at least partial temporal overlap, provided that the respective logical prerequisites for carrying out the steps are met.

[0093] All features and advantages arising from the claims, the description and the drawing, including design details, spatial arrangements and process steps, can be essential to the invention both individually and in various combinations. Reference symbol list

[0094] 30 Object 32 Sectional view 34 Area 36 Area 38 Area 40 Area 42 Fracture 44 Measurement data 46 Analysis result 48 Analysis result 50 User interface 52 User 54 Adapted analysis result 60 Initial state 62 Change state 64 Step of an algorithm 66 Modified step of an algorithm

Claims

1. Computer-implemented method for analysing measurement data from a non-destructive measurement of an object for capturing a structure of the object that is not visible from outside, wherein the analysis assesses whether the object corresponds to a target state, wherein the method (100) comprises the following steps: determining (102) measurement data of a plurality of objects; determining (104) analysis data sets from the measurement data for the objects in an automated manner by means of a non-adaptive assessment algorithm, which is different from an adaptive algorithm, wherein an analysis data set is assigned to one of the objects and has at least one analysis result about the conformity of the assigned object to the target state; checking (106), by a user, the analysis results of at least some of the analysis data sets; adapting (108) an analysis result of a checked analysis data set if the checking by the user yields a deviating analysis result about the conformity of the assigned object to the target state; and communicating (110) at least the adapted analysis data sets to the adaptive algorithm, wherein the adaptive algorithm modifies itself on the basis of the adapted analysis data sets in order to determine analysis data sets from further measurement data of objects by means of the modified adaptive algorithm; wherein the steps are carried out successively or with at least partial temporal overlap, replacing (112) the assessment algorithm by the adaptive algorithm after a predefined minimum number of adapted analysis data sets have been communicated to the adaptive algorithm and / or after a predefined minimum number of analysis data sets have been determined.

2. Computer-implemented method according to Claim 1, characterized in that after the checking of the analysis data sets by the user the method comprises the following step: determining (114) training analysis data sets from the measurement data for the objects by means of the adaptive algorithm; and comparing (116) the adapted analysis data sets with the corresponding training analysis data sets before communicating the adapted analysis data sets to the adaptive algorithm; wherein the assessment algorithm is replaced by the adaptive algorithm if at least some of the adapted analysis data sets match the corresponding training analysis data sets.

3. Computer-implemented method according to either of Claims 1 and 2, characterized in that the method additionally comprises the following step: providing (118) the adapted analysis data sets for the assigned objects by way of an output unit.

4. Computer-implemented method according to any of Claims 1 to 3, characterized in that before the checking of the analysis data sets by a user the method comprises the following step: marking (120) an analysis data set if at least one analysis result about the conformity of the assigned object to the target state is not unambiguous; and using (122) only the marked analysis data sets during the checking of the analysis data sets by the user.

5. Computer-implemented method according to any of Claims 1 to 4, characterized in that the measurement data provide at least one partial representation of a volume arranged within an object.

6. Computer-implemented method according to any of Claims 1 to 5, characterized in that determining measurement data of a plurality of objects comprises the following step: providing (124) volume data as measurement data by means of a measurement by computed tomography.

7. Computer-implemented method according to any of Claims 1 to 5, characterized in that determining measurement data of a plurality of objects comprises the following step: providing (126) volume data as measurement data by means of process data of a measurement during additive manufacturing of an object.

8. Computer-implemented method according to any of Claims 1 to 7, characterized in that determining analysis data sets from the measurement data for the objects comprises the following step: assessing (128) deviations from the target state in the interior of the object.

9. Computer-implemented method according to Claim 8, characterized in that the deviations are air inclusions.

10. Computer-implemented method according to Claim 8 or 9, characterized in that determining analysis data sets from the measurement data for the objects, before assessing deviations in the interior of the object, comprises the following step: ascertaining (130) segmentation data on the basis of the measurement data, wherein the segmentation data describe an internal composition of the object; wherein assessing deviations in the interior of the object is carried out on the basis of the segmentation data by means of the adaptive algorithm.

11. Computer-implemented method according to any of Claims 8 to 10, characterized in that determining analysis data sets from the measurement data for the objects, before assessing deviations in the interior of the object, comprises the following step: determining (132) a local wall thickness at a position of a deviation; wherein assessing deviations in the interior of the object is carried out on the basis of the local wall thickness.

12. Computer-implemented method according to any of Claims 1 to 11, characterized in that communicating at least the adapted analysis data sets to the adaptive algorithm, wherein the adaptive algorithm modifies itself on the basis of the adapted analysis data sets, comprises the following step: communicating (134) simulated analysis data sets based on simulated measurement data to the adaptive algorithm, wherein the adaptive algorithm modifies itself on the basis of the simulated analysis data sets.

13. Computer-implemented method according to any of Claims 1 to 12, characterized in that before determining analysis data sets from the measurement data for the objects the method comprises the following steps: determining (136) provisional analysis data sets from the measurement data for the objects by means of a defect recognition algorithm, wherein a provisional analysis data set is assigned to one of the objects and has at least one analysis result about the conformity of the assigned object to the target state; determining (138), by means of the defect recognition algorithm, whether the analysis result has a deviation of the conformity of the assigned object to the target state within a predefined range; communicating (140) the measurement data of the objects whose provisional analysis data sets have an analysis result having a deviation of the conformity of the assigned object to the target state within the predefined range to the adaptive algorithm for determining analysis data sets from the measurement data for the objects.

14. Computer program product comprising instructions which are executable on a computer and, when executed on a computer, cause the computer to carry out the method according to any of the preceding claims.