Method for producing a motor vehicle
Patent Information
- Application Number
- PCT/EP2024/074770
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-09-22
- Filing Date
- 2024-09-05
- Publication Date
- 2025-03-27
Smart Images
Figure EP2024074770_27032025_PF_FP_ABST
Abstract
Description
[0001] METHOD FOR MANUFACTURING A MOTOR VEHICLE
[0002] The invention relates to a method for producing a motor vehicle, in which at least one quality parameter of at least one component of the motor vehicle is determined based on an image evaluation of at least one image of the motor vehicle captured by means of at least one optical capture device, in particular a camera.
[0003] Methods for manufacturing motor vehicles, in which at least one quality parameter is determined for at least one component of the motor vehicle under manufacture, are generally known from the prior art. In addition to the possibility of manually and / or visually checking the quality of the component, for example a part or assembly of the motor vehicle, for example a paint surface, an emblem, an add-on part and the like, it is also known to use an optical detection device, in particular a camera, to draw conclusions about the quality of the component by taking an image of the component, which image is then evaluated. In particular, the image evaluation is intended to determine whether the component in question meets the quality requirements that apply to the manufacture of the motor vehicle or whether reworking of the component is necessary.
[0004] For such evaluation tasks, the use of artificial intelligence is generally known, whereby a comparatively complex conditioning process, i.e., training of the artificial intelligence, is required. It must be ensured that the artificial intelligence is trained with a high variance of different images of components in different states in order to be able to reliably make a decision about the quality parameter of the component when executing the process. Since, particularly in the manufacture of motor vehicles, with regard to the multitude of equipment variants, i.e.Colors, structures, model-specific add-on parts, manufacturing conditions, location conditions, and the individual positioning of the component in the image result in significant differences in image analysis. Therefore, both the selection and the scope of the image data sets required for training the artificial intelligence are complex. Since the different combinations of equipment variants are distributed very differently during vehicle production, even if several thousand images are provided, it cannot be guaranteed that all equipment variants are adequately covered. Such provision also requires a great deal of effort, particularly for the so-called "labeling" of the data used for conditioning or training. For example, a statistical selection of images can be used for training, and the function can then be tested.However, such a trial-and-error approach requires multiple iterative improvement loops, which are time-consuming and costly.
[0005] The invention is based on the object of providing a method for producing a motor vehicle which is improved compared to the method in question, in which in particular the training of the artificial intelligence is improved.
[0006] The object is achieved by a method according to claim 1. The dependent claims relate to possible embodiments.
[0007] As described, the invention relates to a method for manufacturing a motor vehicle, in which at least one quality parameter of at least one component of the motor vehicle is determined based on an image analysis of at least one image of the motor vehicle captured by at least one optical capture device, in particular a camera. In other words, at least one image of the manufactured or to-be-manufactured motor vehicles is to be captured using the optical capture device, for example, a camera. The quality parameter of the component of the motor vehicle is subsequently determined by the image analysis.The manufacture of the motor vehicle can, in particular, involve the assembly of a component on an already partially manufactured motor vehicle, for example, the arrangement of an add-on part on a frame of the motor vehicle, or the processing of the component, for example, a painted or otherwise treated surface of the component. In principle, any property of the component, such as its mounting position, surface quality, identification and verification of the correct component, and the like, can be determined through image analysis. The method described herein comprises the following steps:
[0008] Training an artificial intelligence executed on at least one computing device based on an initial image set with a predetermined quality parameter;
[0009] Providing an image data set comprising a plurality of images of a plurality of different components and / or motor vehicles captured by the at least one optical capture device;
[0010] Extracting a first number of images through active learning as a first sub-dataset;
[0011] Extracting a second number of images based on at least one piece of product information of the motor vehicle having the at least one component as a second sub-data set;
[0012] Extracting a third number of images based on at least one piece of inspection information of the motor vehicle having the at least one component as a third sub-data set;
[0013] Merging the subsets into the training dataset;
[0014] Retraining the artificial intelligence based on the training dataset;
[0015] Determining the quality parameter of the at least one component by means of the trained artificial intelligence;
[0016] Releasing the at least one motor vehicle or reworking the motor vehicle based on the quality parameter determined for the motor vehicle.
[0017] In other words, the artificial intelligence, which can be executed as a program or as "software" on a computer, is first trained based on an initial image set with predetermined quality parameters. The initial image set can, for example, contain a defined number of images that were captured with the optical capture device and for which the quality parameter was determined. These can therefore be fed to the artificial intelligence in a "labeled" form, so that the artificial intelligence can essentially be trained based on the initial image set. In other words, the initial image set tells the artificial intelligence whether the components of the motor vehicle shown in the images are "OK" or "not OK," so that the artificial intelligence can determine the corresponding quality parameters.
[0018] Subsequently, an image dataset is provided which comprises a plurality of images of a plurality of different components and / or motor vehicles captured by the at least one optical capture device. For example, a set of images taken during a production period can be provided as the image dataset. The image dataset thus comprises the images of various motor vehicles or their components that were manufactured during the production period. The production period can be selected arbitrarily, for example over at least one day, at least one shift, one week, or any other period. Likewise, in principle, all available images of components or motor vehicles of the same type or of relevant types can be provided as the image dataset.This ensures that the image data set can in principle contain a very large number of different variants that are to be examined in the process.
[0019] To do this, an initial number of images is extracted as a first sub-dataset, namely using "active learning." The term "active learning" refers to a form of machine learning in which special learning algorithms can be used. Examples of such algorithms include so-called "variational adversarial active learning" or "learning loss for active learning." This way, an initial number of images are extracted from the image dataset and kept ready, for example, stored, as a first sub-dataset. Extracting the first sub-dataset from the entire image dataset can be performed, in particular, as the first extraction step.
[0020] Furthermore, a second sub-data set can be extracted from the image data set, in which a second number of images are extracted based on at least one piece of product information of the motor vehicle having the at least one component. Product information is generally understood to be information that describes a property of the component and / or the motor vehicle. The product information can be used, for example, via an identification number of the motor vehicle, in particular the so-called VIN (“vehicle information number”). Based on the product information, it can be determined which specific equipment or configuration variant the motor vehicle is equipped with, or the motor vehicle captured in the image can be identified as an individual. In particular, each image captured by the optical capture device is linked to such product information, in particular the identification number of the motor vehicle.
[0021] The method further involves extracting a third sub-data set from the entire image data set. For this purpose, a third number of images is extracted from the image data set based on at least one piece of test information from the motor vehicle having the at least one component. In the context of this application, test information is understood to mean any information that has already been collected by means of a previous test procedure. The test information can be linked to the previously described identification number, in particular the VIN. For example, at least within the image data set there can be a number of components or motor vehicles that have already been tested by means of a test procedure, for example carried out on a random basis. The test information indicates, for example, whether the component captured in the image was classified as OK or not OK or, in general, which quality parameter the component was assigned to.
[0022] The sub-datasets extracted from the image dataset are then combined into a training dataset. This offers the advantage that the generated training dataset can ultimately draw images for training the artificial intelligence from three different sources: the first sub-dataset selected using active learning, the second sub-dataset selected based on product information, and the third sub-dataset extracted from the image dataset using inspection information. The artificial intelligence is then retrained based on the training dataset, meaning its training process can be further improved.
[0023] Advantageously, it is therefore not necessary to label a significantly larger number of different images, for example, the entire image dataset, and use them as a training dataset, or to make a statistical selection from the image dataset and then manually check these for their relevance for reconditioning or retraining, possibly repeating this process several times. Instead, by extracting the individual sub-datasets, a training dataset can be automatically created, which can then be used for reconditioning the artificial intelligence.
[0024] The quality parameter of at least one component can then be determined in an improved manner using the reconditioned or newly trained artificial intelligence. Based on the quality parameter, the motor vehicle containing the component can then be released, or the motor vehicle can be sent for reprocessing, during which any identified quality defect can be remedied. For the purposes of this description, the term "training" can also be replaced by "conditioning" or "retraining" by "reconditioning," and vice versa.
[0025] The described method can be further developed by specifying a total number of images to be extracted from the image dataset. This makes it possible, for example, to determine at the outset the total number of images to be used for retraining the artificial intelligence. The total number can, in principle, be set arbitrarily and can, for example, be determined depending on the variance of the various components and / or vehicles, so that the total number of images extracted from the image dataset allows for reliable retraining of the artificial intelligence.
[0026] The method can further provide for a distribution, in particular a percentage distribution, of the first number, the second number, and the third number to be determined. This can, for example, be used to control the proportion that the individual sub-data sets contribute to the re-learning process. In principle, the individual proportions can be set or specified as desired. Purely as an example, the first sub-data set can contribute 30% to the total number of extracted images, with the second sub-data set accounting for 50% and the third data set accounting for 20% of the total extracted images. Depending on the desired focus or the resources available for extracting the images, the proportions can be changed as desired, as described.By specifying the percentage distribution, it can be ensured in particular that a defined number is available in the individual sub-datasets, so that the different advantages achieved by using the different sub-datasets are sufficiently taken into account.
[0027] As already described, the images in the first sub-dataset are extracted using active learning. In one embodiment of the method, active learning can be used to extract images into the first sub-dataset that exhibit a defined, in particular maximum, deviation from the initial image set. In other words, active learning is used to extract those images from the entire image dataset that are as different as possible from the images used for the initial training of the artificial intelligence.
[0028] This ensures that the artificial intelligence training can be carried out as efficiently and effectively as possible, preventing identical or excessively similar images that were already used for training using the initial image set from also being included in the first sub-dataset. For example, it can be determined which of the images in the image dataset, which can be obtained from production as described above, are as different as possible from the images in the initial image set. To do this, the deviations between the images can be determined, and those images that are at least above a minimum deviation can be selected.
[0029] It is also possible to sort the images in the image dataset according to their deviation or degree of deviation from the images in the initial image set and to select the images with the greatest deviation as the first set of images. The deviations can be of various types, for example with regard to the color of the component, the position of the component, the type of component, and the like. This advantageously results in a much greater diversity than with a purely statistical selection. Advantageously, any weaknesses that may arise when extracting the images into the first sub-dataset, for example a small area-related proportion of the identified features in the image, e.g. an emblem on a car hood, identified deviations that have no relation to the inspection task, and the comparatively high computational proportion or computational effort, are compensated for by the two other sub-datasets.
[0030] As already described, the second sub-data set is created based on the product information of the component or the motor vehicle in the images of the image data set, or images are extracted from the image data set into the sub-data set. The product information can comprise or describe at least one equipment variant and / or a configuration of the motor vehicle. As is generally known, a large number of individual configurations, i.e. combinations of paint finishes, component colors, designs of add-on parts, implementation of special equipment, and the like, are produced during the manufacture of motor vehicles. The individually selected equipment variant or the individual configuration can be assigned to the individual images in the image data set using the product information. This allows different components orMotor vehicles can be selected by specifically selecting individual equipment variants or configurations from the multitude of images in the image dataset. This can generally be performed using a so-called principal component analysis (PCA) based on the image dataset.
[0031] For example, it is possible to assign the images of the image dataset to different groups, which can also be referred to as clusters, based on the product information, with at least one image per group being extracted into the second sub-dataset. In other words, the individual vehicles can be assigned to different groups that match in at least one parameter of the product information. For example, groups can be formed according to type, color, special equipment, left-hand drive or right-hand drive, add-on parts, and the like. This results in individual groups of motor vehicles that match in at least one such parameter, for example, the group of black vehicles, the group of red vehicles, the group of vehicles with sunroofs, and the like.
[0032] In the described embodiment, at least one image from each of these groups is selected to ensure that an individual from each of these groups is represented in the second sub-dataset. In particular, it can be ensured that a comparison is made with the images extracted from the first sub-dataset. In other words, a potential candidate for the second sub-dataset can be compared with the first sub-dataset. If the image is already contained in the first sub-dataset, it is not included in the second sub-dataset. Instead, for example, another individual from the same group is extracted into the second sub-dataset. This individual may advantageously differ from the previously "drawn" individual in at least one other parameter, thus further increasing diversity.
[0033] The third sub-data set is formed, as previously described, by extracting images from the image data set for which specific test information is available. According to one embodiment of the method, the at least one piece of test information can describe or include at least one test result of a component captured in the at least one image. In principle, any test data available for a specific motor vehicle or component can be used as test information. As already described, this data can be linked to the identification number, in particular the VIN.
[0034] Test data is understood to mean, for example, a test report or the result of a previous inspection for at least one component. By selecting the images of the third sub-data set based on the test information, it is advantageously possible to specifically adjust the proportion of images that have already been checked as "OK" or "not OK". The proportion of images in the third sub-data set that contain components found to be not OK can be specifically increased. For example, 80% of the images contained in the third sub-data set can be selected as "not OK", since in reality this usually represents a very small proportion. This can specifically ensure that the artificial intelligence can subsequently reliably detect component errors that lead to a classification as "not OK" or "not OK".Furthermore, this advantageously allows for targeted use of the already available knowledge to train the artificial intelligence. As already described with regard to product information, this also ensures that images already extracted into the first or second sub-dataset cannot be incorporated into the third sub-dataset, but instead, a different image with the same test result or the same or similar test information is selected.
[0035] Based on the test information and / or the product information, a further embodiment of the method allows images to be extracted depending on at least one time parameter and / or one location parameter. As described, environmental conditions also generally influence the capture of images, for example, the prevailing lighting conditions (e.g., artificial or natural light), the location within a facility, or the facility or plant in which the image was captured. For this purpose, time parameters can be used, for example, summer time, winter time, a time of day, a day, a shift, a calendar week, or a specific date.
[0036] In particular, a location within a plant can be used as a location parameter, for example in the case of flexible location changes to individual production lines or in the case of parallel provision of several similar production lines. It is also possible to identify in which plant the image was optically captured. This ensures that a diverse selection of images is made with regard to the time parameter or the location parameter, which can then be extracted into the second sub-data set and / or the third sub-data set. This allows environmental influences to be taken into account more effectively, so that, for example, an artificial intelligence trained in the summer achieves good results in winter and the training and thus the resulting determination result, i.e. the specific quality parameter, is not dependent on the time of day, the lighting or the location.Instead, a robust determination is made possible.
[0037] The described method can further be developed such that an image extracted into the first sub-dataset is removed from the set of images from which images can be extracted into the second sub-dataset and / or the third sub-dataset and / or that an image extracted into the second sub-dataset is removed from the set of images from which images can be extracted into the third sub-dataset and / or that an image extracted into the third sub-dataset is removed from the set of images from which images can be extracted into the second sub-dataset. In principle, it is therefore possible to prevent so-called "double drawings". In other words, when extracting images from the image dataset, a drawing can be performed without replacement, so that an image extracted into one of the sub-datasets is no longer available for the other sub-datasets.A specific order can be specified when dragging, so that, for example, the images in the first sub-dataset are extracted first, making them unavailable for the other two sub-datasets. In principle, the second sub-dataset can then be dragged first, followed by the third sub-dataset, or vice versa.
[0038] In addition to the method, the invention relates to a plant for producing motor vehicles, comprising at least one assembly device designed to assemble at least one component for a motor vehicle and / or at least one processing device designed to process at least one component and a computing device designed to carry out the method described above.
[0039] Furthermore, the invention relates to a method for generating a training data set, in which at least one quality parameter of at least one component of a motor vehicle is determined based on an image evaluation of at least one image of the motor vehicle captured by means of at least one optical detection device, in particular a camera, comprising the steps:
[0040] Training an artificial intelligence executed on at least one computing device based on an initial image set with a predetermined quality parameter;
[0041] Providing an image data set comprising a plurality of images of a plurality of different components and / or motor vehicles captured by the at least one optical capture device;
[0042] Extracting a first number of images through active learning as a first sub-dataset;
[0043] Extracting a second number of images based on at least one piece of product information of the motor vehicle having the at least one component as a second sub-data set; Extracting a third number of images based on at least one piece of inspection information of the motor vehicle having the at least one component as a third sub-data set;
[0044] Merge the subsets into the training dataset; output the training dataset.
[0045] All advantages, details, designs, and / or features described with regard to the method for manufacturing the motor vehicle are fully applicable to the system and the method for generating a training data set. The method for generating the training data set can be used, in particular, to generate and output the training data set, which can subsequently be used to retrain the artificial intelligence within the manufacturing process.
[0046] The invention is explained using exemplary embodiments with reference to the figures. The figures are schematic representations and show:
[0047] Fig. 1 is a schematic diagram of a plant for manufacturing motor vehicles according to an embodiment; and
[0048] Fig. 2 is a schematic diagram of a flow chart of a method for producing a motor vehicle according to an embodiment.
[0049] Fig. 1 shows a system 1 which is designed to manufacture a motor vehicle. The system 1 can, for example, have a production line 2 in which various components are assembled on the motor vehicle or a plurality of motor vehicles, or components installed on the motor vehicle are processed, or components which are intended for assembly on the motor vehicle are processed. The system 1 can thus be regarded as a production system or processing system, wherein at least one production line 2 can be designed as an assembly line or the like. At least one optical detection device 3 is assigned to the system 1, in particular to the at least one production line 2, and is designed to detect the at least one component of the motor vehicle or at least similar components of the motor vehicles manufactured on the production line 2, for example as a camera by capturing images.For example, the optical detection device 3 can be used to detect a section of the motor vehicle on the assembly line. In principle, the system 1 can have any number of production lines 2 and any number of optical detection devices 3. It is also possible to provide multiple optical detection devices 3, for example, in different orientations or at different positions, to detect the corresponding component or other components.
[0050] The optical detection device 3 is particularly intended to determine a quality parameter of the component, in particular to determine whether the component and thus the motor vehicle is to be classified as "OK" or "not OK". For this purpose, the system 1 has a computing device 4 on which artificial intelligence is implemented or which provides artificial intelligence. In other words, the images of the individual motor vehicles produced on the production line 2 are fed to the computing device 4 and thus to the artificial intelligence, so that the artificial intelligence can determine by means of image evaluation whether the component of the motor vehicle depicted in the image meets quality requirements. If the quality requirement is met, i.e.If the result of the image evaluation classifies the component as being OK, the motor vehicle can be released and, for example, fed to another station in production line 2 or output at the end of production line 2. If the evaluation shows that the component is not OK, the corresponding motor vehicle can be fed to a corresponding post-processing station in production line 2 or externally to production line 2 for reworking of the component, so that a new evaluation can be carried out after reworking if necessary.
[0051] The functioning of the artificial intelligence and its training and retraining are explained below by way of example with reference to Fig. 2. Fig. 2 schematically shows a method for producing the motor vehicle, as this can be carried out, for example, on the system 1 shown in Fig. 1. The method starts schematically in a block 5, in which the artificial intelligence executed on the computing device 4 is conditioned or trained with an initial image set. The conditioning or training with the initial image set can be understood as a basic "first training" of the artificial intelligence. Here, a certain number of images are supplied as the initial image set, which images comprise a predetermined quality parameter. In particular, the images of the initial image set are thus "labeled" so that the artificial intelligence can learn from them which images orwhich components are to be described or assessed as OK and which as not OK.
[0052] In order to further train the artificial intelligence or to further improve the basic training process, a method for generating training data is shown in a dashed block in Fig. 2. The dashed block thus delimits the method for generating training data, which is or can be carried out as part of the method for manufacturing the motor vehicle, so that this method can also be considered or carried out independently, for example on the computing device 4 or another computing device implemented separately from the computing device 4. The result or the renewed training of the artificial intelligence can be carried out based on the training data obtained by the corresponding method.
[0053] In block 6, an image data set is provided comprising a plurality of images captured by the capture device 3, which images show a plurality of motor vehicles and components. For example, the image data set can be collected during operation of the system 1, with the capture device 3 capturing an image of each manufactured motor vehicle or each corresponding component in the production line 2. In principle, the image data set can comprise any number of images, in particular a significantly higher number of images compared to the initial image set. In particular, the image data set can comprise several thousand images, whereby the initial image set can comprise, for example, only 100 images. The numbers are arbitrary and can be changed depending on the embodiment or specific application and are provided herein for illustrative purposes only.
[0054] For example, all images captured by the capture device 3 can also be used on the production line 2, for example all those that show the same motor vehicle type or that are relevant for the subsequent inspection task because they show the same type of motor vehicles and components that are subsequently to be inspected using artificial intelligence. From the image data set provided in block 6, a first sub-data set is extracted in block 7, a second sub-data set in block 8, and a third sub-data set in block 9. Although the order is fundamentally arbitrary, it may be preferable to extract the first sub-data set from the image data set first. For this purpose, a total number of images to be extracted into the individual sub-data sets can be specified, for example, in block 6 or in blocks 7-9 respectively.For example, a total number of images to be extracted from the image dataset can be specified, for example, 100 images. As already described, the exact numbers are arbitrary and serve only as an explanation. For example, in each of blocks 7-9, the percentage or total number to be extracted into the respective sub-dataset is specified. Also purely as an example, 30 of the 100 images are to be extracted into the first sub-dataset, 50 of the 100 images into the second sub-dataset, and 20 of the 100 images into the third sub-dataset.
[0055] In Block 7, the first set of images, i.e., the 30 images, are extracted from the entire image dataset using active learning. The initial image set is used as a reference, and the 30 images that exhibit the greatest deviations from the initial image set are extracted from the entire image dataset. These deviations can be performed, for example, as part of image analysis or pattern recognition, so that images that are as different as possible from the initial image set on which the artificial intelligence was initially trained can be found.
[0056] In block 8, the second number of images to be extracted into the second sub-data set is extracted based on the product information of the images in the image data set. The product information is evaluated, for example, using the identification number assigned to and linked to each image captured by the capture device 3. The product information describes the specific equipment or configuration of the individual motor vehicle shown in the individual images. For this purpose, the entirety of the images in the image data set can be evaluated using a principal component analysis, and the images can be assigned to individual groups or so-called "clusters" according to their equipment variants. Subsequently, at least one image from each group is included in the second sub-data set.For example, ten groups can be formed that differ in their product information, with the individual individuals within the groups having the same product information, for example a first group of red motor vehicles, a second group of motor vehicles with a sunroof, a third group of motor vehicles with a black interior, etc.
[0057] To extract the second number of images for the second sub-dataset—for example, the 50 images—five images are extracted from each of the ten groups. In conjunction with Block 7, this ensures that the images already extracted from the image dataset for the first sub-dataset are no longer available for extraction into the second sub-dataset. This prevents duplicate drawings and instead increases the diversity of the training data generated at the end.
[0058] In block 9, the third number of images for the third sub-data set is extracted based on test information. The test information describes a test result already available for the images, for example, as a test report or other test data. In particular, the test information for an image can indicate whether the motor vehicle or component shown in it was found to be OK or not. The third sub-data set thus extracts images for which a test result is already available so that they can be specifically fed into the artificial intelligence. In particular, a targeted distribution of the test results can be set in block 9. The focus can be placed in particular on components or their images that led to a negative test result.For example, it can be specified that 80% of the images extracted into the third sub-dataset show components that were found to be “not OK” during previous testing, so that the artificial intelligence can be trained specifically for such error cases.
[0059] As already described, it is advantageously also ensured in block 9 that the images extracted from the image data set into the first sub-data set or the second sub-data set do not end up in the third sub-data set or are no longer available for it. In principle, the extraction in block 7 can be carried out first, so that the greatest diversity in terms of deviations from the initial image set is present there. Subsequently, the images can be extracted first into the second sub-data set or the third sub-data set as desired, but it is always ensured that no double extractions occur. In blocks 8 and 9, a selection of the images can also be made based on a time parameter and / or a location parameter. For example, groups can be formed based on the test reports or the product information, which are extracted at different times, i.e.Times of day, seasons, shifts, and / or at different locations or different places within a facility. This also ensures that diverse image extraction is performed across different time and location parameters.
[0060] Subsequently, in block 10, the three sub-datasets are combined to form the training dataset. The training dataset refers to the previously described 100 images extracted from the entire image dataset into the first sub-dataset, the second sub-dataset, and the third sub-dataset. The training dataset can then be used in block 11 to retrain the artificial intelligence running on the computing device 4. Each image in the training dataset can be pre-labeled to provide the relevant information for training the artificial intelligence.
[0061] The newly trained artificial intelligence can then be used in block 12 to determine the quality parameters of components in system 1, namely based on images captured by the capture device 3. Accordingly, in block 13, based on the determined quality parameters, it can be determined for each component or for each motor vehicle whether it is OK or not, and thus whether it can be released or whether it should at least be sent to a post-processing station.
[0062] As described, the methods described with reference to Fig. 2 can be executed as desired on the system 1, or the computing device 4 executes the methods described in Fig. 2. All advantages, details, and features described with reference to the methods and the system can therefore be combined with one another as desired, are interchangeable, and are transferable to one another.
[0063] List of reference symbols
[0064] 1 system
[0065] 2 Production line 3 Recording device
[0066] 4 Computing device
[0067] 5-13 blocks
Claims
CLAIMS 1. A method for producing a motor vehicle, in which at least one quality parameter of at least one component of the motor vehicle is determined based on an image evaluation of at least one image of the motor vehicle captured by means of at least one optical capture device (3), in particular a camera, comprising the steps: Training an artificial intelligence executed on at least one computing device (4) based on an initial image set with predetermined quality parameters; Providing an image data set comprising a plurality of images of a plurality of different components and / or motor vehicles captured by means of the at least one optical capture device (3); Extracting a first number of images through active learning as a first sub-dataset; Extracting a second number of images based on at least one piece of product information of the motor vehicle having the at least one component as a second sub-data set; Extracting a third number of images based on at least one piece of inspection information of the motor vehicle having the at least one component as a third sub-data set; Merging the sub-datasets into the training dataset; Retraining the artificial intelligence based on the training dataset; Determining the quality parameter of the at least one component by means of the re-trained artificial intelligence; Releasing the at least one motor vehicle or reworking the motor vehicle based on the quality parameter determined for the motor vehicle.
2. Method according to claim 1, characterized in that a total number of images to be extracted from the image data set is determined.
3. Method according to claim 2, characterized in that a distribution, in particular a percentage distribution, of the first number and the second number and the third number is determined.
4. Method according to one of the preceding claims, characterized in that images are extracted into the first sub-data set by means of active learning, which have a defined, in particular the maximum, deviation from the initial image set.
5. Method according to one of the preceding claims, characterized in that the product information comprises or describes at least one equipment variant and / or a configuration of the motor vehicle.
6. Method according to one of the preceding claims, characterized in that the images of the image data set are assigned to different groups based on the product information, wherein at least one image per group is extracted into the second sub-data set.
7. Method according to one of the preceding claims, characterized in that the at least one piece of test information describes or comprises at least one test result of a component captured in the at least one image.
8. Method according to one of the preceding claims, characterized in that based on the test information and / or the product information, images are extracted as a function of at least one time parameter and / or one location parameter.
9. Method according to one of the preceding claims, characterized in that an image extracted into the first sub-data set is removed from the set of images from which images can be extracted into the second sub-data set and / or the third sub-data set and / or that an image extracted into the second sub-data set is removed from the set of images from which images can be extracted into the third sub-data set and / or that an image extracted into the the image extracted from the third sub-dataset is removed from the set of images from which images can be extracted into the second sub-dataset.
10. Plant (1) for the production of motor vehicles, comprising at least one assembly device designed for assembling at least one component for a motor vehicle and / or at least one processing device designed for processing at least one component and a computing device (4) designed to carry out the method according to one of the preceding claims.
11. A method for generating a training data set, in which at least one quality parameter of at least one component of a motor vehicle is determined based on an image evaluation of at least one image of the motor vehicle captured by means of at least one optical capture device (3), in particular a camera, comprising the steps: Training an artificial intelligence executed on at least one computing device (4) based on an initial image set with predetermined quality parameters; Providing an image data set comprising a plurality of images of a plurality of different components and / or motor vehicles captured by means of the at least one optical capture device (3); Extracting a first number of images through active learning as a first sub-dataset; Extracting a second number of images based on at least one piece of product information of the motor vehicle having the at least one component as a second sub-data set; Extracting a third number of images based on at least one piece of inspection information of the motor vehicle having the at least one component as a third sub-data set; Merging the subsets into the training dataset; Output of the training dataset.