System and procedure for quality control in the manufacture of individual parts

Abstract

System for quality control in the manufacture of individual parts, the system comprising (1) a calculation unit (2) for creating images (8) of the individual parts to be produced, the images (8) forming training data for a neural network (3), b. the neural network (3) for classifying individual parts into fault-free and faulty parts on the basis of the training data, wherein the calculation unit (2) is in signal connection with the neural network (3) in order to make the training data available to the neural network (3), wherein the neural network (3) is self-learning and comprises artificial intelligence such that assignment parameters are iteratively improved during the classification of the training data or a comparison of the training data with real images.

Description

[0001] The invention relates to a system and a method for quality control during the production of individual parts.

[0002] DE 10 2014 218 132 A1 discloses a method and a device for determining the quality of a component.

[0003] DE 195 18 804 A1 discloses a method for monitoring a production process. DE 10 2005 038 889 A1 discloses a method for inspecting a manufactured product. DE 10 2014 218 096 A1 discloses a method and a device for monitoring a manufacturing and / or assembly process. DE 10 2016 012 451 A1 discloses a method for monitoring, analyzing, and operating at least one production facility. DE 10 2016 118 825 A1 discloses an image acquisition system for an automated production line.

[0004] During the industrial production of individual parts, especially in metalworking, defects can occur. During a quality inspection, manufactured parts are analyzed to distinguish defective, i.e., unusable, parts from flawless ones. The preparation and implementation of the quality inspection is significant.

[0005] The invention is based on the object of reducing the effort required for quality control during the production of individual parts.

[0006] The object is achieved by the features of claims 1 and 7. The core of the invention consists in creating images of the individual parts to be manufactured by means of a calculation unit. The images can be stored in a data memory and serve as the basis for classifying the images into defect-free and defective individual parts by means of a neural network. For classification with the neural network, a large number of images is required, in particular at least ten, in particular at least 50, in particular at least 100, in particular at least 500, in particular at least 1,000, in particular at least 5,000, in particular at least 10,000 and in particular at most 1,000,000 images. The images form so-called training data for the neural network.The neural network is particularly self-learning and includes artificial intelligence such that assignment parameters are improved iteratively, i.e. with increasing numbers of training data and / or real images, when classifying the training data or comparing the training data with real images. Because the images are created, i.e. calculated, in the calculation unit, synthetically created images of the individual parts can be used as training data. The effort required for the actual production of samples and their manual quality assessment for the provision of training data is reduced and, in particular, avoided. In particular, there is no longer any need to produce a large number of individual parts in order to provide training data for the neural network. The system can be used in the production of individual parts, in particular from the first individual part produced.For the system according to the invention, the training of new workpieces and / or processes is unnecessary. A start-up phase, during which the manufacturing process is adjusted and / or adapted to ensure error-free production of the individual parts, is also unnecessary. In particular, the conversion effort is reduced when parts are changed during the production of the individual parts, i.e., when the geometry and / or material changes. The new system can be deployed faster and more flexibly.

[0007] A calculation unit according to claim 2 enables a realistic simulation of the manufacturing process of the individual parts. The quality of the training data is improved. The calculation unit is designed, in particular, as a numerical simulation program, in particular as a finite element calculation program.

[0008] Alternatively, the calculation unit can be designed in such a way that real images of the individual parts are digitized. The digitized real images can then be synthesized, for example, by integrating error representations into the synthesized real images. The synthesized real images then form the training data. The training data generated in this way can also be generated in large quantities, thus quickly and at reduced cost. Providing the training data is straightforward.

[0009] A calculation program according to claim 3 can realistically represent different materials and / or different target geometries of the individual parts. In particular, the images generated by the calculation program can depict defects in and / or on the individual part. Such defects in sheet metal forming include, for example, necking, cracks, and / or undesired hardening. The training data is particularly suitable for depicting locally limited defects and / or weak points of the individual part.

[0010] A calibration unit according to claim 4 serves for an initial calibration of the images for the neural network.

[0011] With an image capture unit according to claim 5, actual manufactured individual parts can be optically captured. Additionally or alternatively, the image capture unit can comprise a laser triangulation device, an X-ray imaging device, and / or an ultrasound imaging device. It is essential that the image capture unit is capable of generating a real image of the individual part.

[0012] A manufacturing unit according to claim 6 serves to manufacture the individual parts. The individual parts are, in particular, metallic individual parts that have been manufactured, for example, using a forming process, in particular a sheet metal forming process or a bulk forming process.

[0013] The manufacturing unit can be a press for drawing sheet metal or a rolling device for thread rolling screws. The manufacturing unit can also be a primary forming device for casting individual parts from metallic or plastic materials. The manufacturing unit can also be a device for machining workpieces, for example, a lathe, a milling machine, a machine tool, or a machining center.

[0014] A method for quality control, in particular for automated quality control in the production of individual parts, essentially has the advantages of the system according to the invention, to which reference is hereby made.

[0015] The evaluation of the quality of the actually manufactured individual parts according to claim 12 enables fast and immediate feedback on the manufacturing process, i.e., the production of the individual parts. The evaluation of the actual images, in particular images, i.e., photographs, generated by the image acquisition unit is carried out by means of the neural network, in which the images are classified into faultless and faulty individual parts. During the evaluation, the actual images are classified by the trained neural network. A comparison of the actual images with the training data is unnecessary. The evaluation requires comparatively little memory capacity and processing power of a processor. The evaluation can be carried out quickly and easily. This makes it possible to reliably classify the actually manufactured individual parts as faultless or faulty.

[0016] In particular, monitoring of all manufactured individual parts is possible. Because quality inspection and thus the evaluation of the individual parts can be carried out automatically, the effort required is reduced. In particular, delays in the manufacturing process are avoided through automated quality inspection. The quality result is improved. Instead of selective sample inspections, a comprehensive, i.e., seamless, inspection of the manufactured individual parts is carried out.

[0017] Controlling the manufacturing process for the individual parts according to claim 13 enables continuous quality improvement in the manufacturing process. The result from the evaluation of the individual parts can serve as a control parameter in the manufacture of the individual parts. This enables direct feedback for the manufacturing process from the quality inspection. In particular, undesirable deviation trends in the manufacture of the individual parts can be identified if, for example, the actual dimension for the individual parts shifts from a permissible tolerance range towards an upper or lower threshold value during the manufacturing process. An undesirable deviation trend also exists if defective individual parts occur repeatedly in the course of a manufacturing process, in particular with the same or a similar defect pattern.Feedback allows this trend to be detected in a control unit and compensated for by taking adjustable machine parameters into account in the production unit. The feedback can also be used to identify defective materials that, for example, do not meet the required material quality and / or initial dimensions, such as sheet thickness or bar diameter.

[0018] Both the features specified in the patent claims and the features specified in the following exemplary embodiment of the system according to the invention are suitable, either individually or in combination with one another, for further developing the subject matter of the invention. The respective combinations of features do not represent any limitation with regard to further developments of the subject matter of the invention, but are essentially merely exemplary in nature.

[0019] Further features, advantages, and details of the invention will become apparent from the following description of the exemplary embodiment with reference to the drawings. They show: Fig. 1 a schematic representation of a system according to the invention for training a neural network, Fig. 2 a flow chart illustrating the individual steps according to the method in Fig. 1, Fig. 3 a schematic representation of the system in Fig. 1 for quality control of manufactured individual parts, Fig. 4 real images of manufactured individual parts generated by an image capture unit, Fig. 5 illustrations for classifying parts into faultless and faulty parts.

[0020] Corresponding components are shown in the Fig. 1 to 5 are provided with the same reference numerals. Details of the exemplary embodiment explained in more detail below may also constitute an invention in themselves or be part of a subject matter of the invention.

[0021] A system, designated as a whole by 1, is used for automated quality control during the manufacture of individual parts. The system 1 comprises a calculation unit 2 in the form of a computer, in particular a computer, which is used to execute a calculation program. The calculation program is in particular a numerical simulation program based on the finite element method. The system 1 further comprises a neural network 3 with an input 4 and an output 5. The neural network 3 is signal-connected to the calculation unit 2. The neural network 3 is further signal-connected to an image capture unit 6, in particular a camera. The neural network 3 is connected, for example, via the image capture unit 6 to a production unit 7. The neural network 3 can also be directly signal-connected to the production unit 7.According to the embodiment shown, the manufacturing unit 7 is a system for producing screws with external threads.

[0022] The procedure for automated quality control is explained in more detail below.

[0023] With the calculation program of the calculation unit 2, various Fig. created, i.e. calculated and made available to the neural network 3. The Fig. are fed to neural network 3. Neural network 3 is trained based on the images. The images are training data.

[0024] Creating the Fig. and providing the Fig. for the neural network 3 is based on Fig. 2 explained in detail.

[0025] In a calculation step 9, a calculation result is generated using the calculation program of the calculation unit 2. From the calculation result, in a data export step 10, the data required to create the Fig. The required data of the calculation result, in particular data for generating a three-dimensional model of the manufactured component, is exported. The data export step is also referred to as 3D export. The exported data for a three-dimensional representation of the calculation result are converted in a transformation step 11 into a target view and / or target orientation for the Fig. Transformation step 11 improves the comparability of the calculation results with the optical capture of manufactured individual parts. Transformation step 11 is carried out, in particular, automatically using a calculation program.

[0026] In a rendering step 12, the transformed and / or aligned 3D data model is rendered, in particular rendered in multiple layers and blended. The data model thus generated is calibrated by means of a calibration unit in a calibration step 13. For this purpose, optical images of real individual parts are generated in a reference step 14 and used to calibrate the Fig. used. Results of recordings generated in the reference step 14 are in Fig. 4 shown.

[0027] Fig. Figure 4 shows four different examples of real images in the form of optical images, i.e., photographs, of manufactured screws. The left column shows the raw data of the optical images 16. The corresponding right image 17 shows the images after image processing. In the illustrated embodiment, the image processing essentially involves binarization with a threshold value. Additionally, a mirroring of the image around the vertical axis can be provided to increase the number of possible images.

[0028] Fig. Figure 5 shows the result of calibration step 13, in which the images 17a to 17c from reference step 14 are classified. Accordingly, a defect-free individual part 17a or a defective part due to a faulty external thread 17b and / or a broken shaft 17c can be classified. The images generated in calibration step 13 and reference step 14 are stored in a data storage device in a storage step 15. The data storage device is connected to the neural network 3, in particular to the input 4 of the neural network 3. The data storage device can also be integrated into the neural network 3.

[0029] The calculation unit makes it possible to provide a large amount of input data for the neural network 3 quickly and easily. The provision of training data for the neural network 3 for the Fig. The training step of system 1 shown in Figure 1 is straightforward and inexpensive. In the neural network 3, the Fig. evaluated and classified. The output 5 of the neural network 3 can be used to decide whether the Fig. whether it is a faultless or faulty part.

[0030] During the training phase of the neural network 3, which is Fig. 1, the training data 8 are fed to the neural network 3. For each individual Fig. It is known a priori whether the individual part is faultless or faulty. During the training phase, the information about the class membership of the Fig. via an information bypass 20 past the neural network 3 and compared with the decision result of the neural network 3 via the output 5. Through this comparison, the parameters of the neural network 3 can be varied and, in particular, specifically adjusted.

[0031] In one of the training phases according to Fig. 1 downstream evaluation step, the quality of the actually manufactured individual parts is evaluated by means of the neural network 3. The individual parts manufactured in the production unit 7 are optically recorded by an image capture unit 6 in the form of a camera. The optical images of the individual parts can, as described above with reference to the Fig. 4, are digitally processed for the comparability of the data. The optical image 16, 17 of the individual part thus generated is fed to the input 4 of the neural network 3 and compared with the stored training data, i.e. the Fig. , compared and then evaluated.

[0032] The evaluation result, whether the individual part is faulty or defect-free, can be displayed on a display unit 18. The display unit 18 can serve as a visual and / or acoustic display and is, for example, a monitor with loudspeakers. The display unit 18 can also have a flashing light arranged in the immediate vicinity of the production unit. The display unit 18 signals, in particular, when defective individual parts are detected. This gives a machine operator the opportunity to intervene in the production process, for example, to interrupt the production process and / or to make changes to the production parameters, in particular to the parameters of the production unit 7.

[0033] Additionally or alternatively, it is conceivable to feed the result data provided at output 5 of the neural network 3 directly to a control unit 19. A control signal can be generated in the control unit 19 and transmitted to the manufacturing unit 7. This makes it possible to provide online monitoring of the manufacturing process and, in particular, to react directly to the manufacturing process. It is possible to provide online quality assurance and to control the manufacturing unit 7 via the feedback of information from the manufacturing process. With the method shown, quality control is possible continuously during the process.

[0034] The main advantage of System 1 and the method is that the training data is exclusively based on calculated Fig. consist.

Claims

[1] System for quality control in the manufacture of individual parts, the system comprising (1) a calculation unit (2) for creating images (8) of the individual parts to be produced, the images (8) forming training data for a neural network (3), b. the neural network (3) for classifying individual parts into fault-free and faulty parts on the basis of the training data, wherein the calculation unit (2) is in signal connection with the neural network (3) in order to make the training data available to the neural network (3), wherein the neural network (3) is self-learning and comprises artificial intelligence such that assignment parameters are iteratively improved during the classification of the training data or a comparison of the training data with real images. [2] System according to claim 1, characterized bythat the calculation unit (2) comprises a calculation program which is designed as a simulation program for simulating the production of the individual parts. [3] System according to claim 2, characterized by that the calculation program includes model data of different materials and / or different target geometries for the individual parts. [4] System according to one of the preceding claims, characterized by a calibration unit for calibrating the images (8) for the neural network (3). [5] System according to one of the preceding claims, characterized by an image acquisition unit (6) for optically capturing manufactured individual parts. [6] System according to one of the preceding claims, characterized by a manufacturing unit (7) for producing the individual parts. [7] Procedure for quality control in the production of individual parts with the process steps - creating images (8) of the individual parts to be manufactured by means of a calculation unit (2), the images (8) forming training data for a neural network (3), - Saving the training data in a data storage, - Classifying into error-free and defective individual parts by means of the neural network (3) on the basis of the training data), wherein the calculation unit (2) is in signal connection with the neural network (3) in order to make the training data available to the neural network (3), wherein the neural network (3) is designed to be self-learning and comprises artificial intelligence such that assignment parameters are iteratively improved during the classification of the training data or a comparison of the training data with real images. [8] Method according to claim 7, characterized by Simulating the production of individual parts using a simulation program. [9] Method according to claim 7 or 8, characterized by Calibrating the recordings for the neural network (3) using a calibration unit. [10] Method according to one of claims 7 to 9, characterized by Manufacturing the individual parts using a manufacturing unit (7). [11] Method according to one of claims 7 to 10, characterized by optical recording of manufactured individual parts by means of an image capture unit (6). [12] Method according to one of claims 7 to 11, characterized by Evaluating the quality of the manufactured individual parts, in particular all individual parts, by means of the neural network (3), wherein the neural network (3) has previously been trained by means of the images (8). [13] Method according to claim 12, characterized by Feeding back the quality result as a control parameter for the manufacturing unit (7).

Images