Detection of technical states or technical qualities
Patent Information
- Authority / Receiving Office
- EP · EP
- Patent Type
- Applications
- Current Assignee / Owner
- SIEMENS AG
- Filing Date
- 2024-07-23
- Publication Date
- 2026-04-29
AI Technical Summary
Existing AI-based detection methods for anomalies, errors, and damage in technical devices are prone to incorrect results due to improper training on flawed data, leading to imprecise or incorrect detection, especially in visual quality control applications.
A redundant inspection process using two AI models in a data processing system, where the first model detects anomalies through image-based visual quality control and the second model verifies the results, allowing for automated inspection and error detection, thereby enhancing security and efficiency.
This approach ensures high-security automated inspection by comparing detection results from both models, reducing complexity and resource usage, and enabling effective quality assurance across various industrial applications.
Smart Images
Figure EP2024070814_30012025_PF_FP_ABST
Abstract
Description
[0001] Detection of technical conditions or technical properties
[0002] The invention relates to a computer-implemented method for detecting technical conditions or technical properties and to a data processing system, wherein at least one first artificial intelligence model implemented in a first computing unit of the data processing system and a second artificial intelligence model implemented in a second computing unit of the data processing system that is connected to the first computing unit in a signal-transmitting manner are trained on the basis of training data representing technical conditions or technical properties, wherein in the data processing system a federal third artificial intelligence model is formed at least from first training results relating to the first artificial intelligence model and from second training results relating to the second artificial intelligence model.
[0003] In numerous technical applications, such as visual quality control, the reliable detection of anomalies, errors, and / or damage is of great importance. If methods based on artificial intelligence models are used for such applications, faulty detection and / or classification results can lead to imprecise or erroneous results being generated by these models if they are trained on the basis of these faulty detection and / or classification results and this remains undetected.
[0004] WO 2022 / 225506 A1, for example, is known from the prior art, which discloses a method for detecting chassis faults, wherein an autoencoder is trained with first acceleration data of a first chassis of a first rail vehicle. Second acceleration data are acquired during operation of a second chassis of a second rail vehicle and received by the autoencoder. Faults in the second chassis can be detected using the autoencoder trained on the basis of the first acceleration data and the second acceleration data received by the autoencoder.
[0005] Furthermore, WO 2023 / 117775 A1 describes a method and a device for the automated inspection of bogies for rail vehicles. A test section is defined on a bogie, which is captured by a camera. An actual camera image content relating to the test section is compared with a target camera image content relating to the test section. A warning is generated if the actual camera image content deviates from the target camera image content, and this is evaluated as an error using a machine learning method.
[0006] The invention is based on the object of specifying a redundant inspection method which is further developed compared to the prior art.
[0007] According to the invention, this object is achieved with a method according to claim 1, in which at least one technical first state or at least one technical first quality is detected at least by means of the first artificial intelligence model and by means of the third artificial intelligence model, wherein at least the technical first state or at least the technical first quality relates to a mechanical, electrical and / or electronic quality property of at least one technical device, wherein for visual quality control for image-based detection of anomalies, errors and / or damages of the at least one technical device, at least one first detection result is formed by means of the first artificial intelligence model and at least one first control detection result is formed by means of the third artificial intelligence model, and wherein a first detection error is then detected,if at least the first detection result and the first control detection result differ from each other.
[0008] This measure enables automated inspection processes by means of the first computing unit and the second computing unit, whereby the first detection result is checked by means of the first control detection result.
[0009] This ensures a high level of reliability when detecting anomalies, errors, and / or damage. Complex monitoring of a detection step (e.g., using a two-person inspection principle) can be avoided. The method according to the invention can be used, for example, for the automated inspection of technical devices (e.g., machines, vehicles, aircraft, machine elements, devices, apparatus, etc.), whereby the technical devices can be inspected, for example, by evaluating images of the technical devices using image processing. It is also possible, using the method according to the invention, to automatically inspect medical images (e.g., X-ray images) and thus detect anomalies, etc.
[0010] The process also enables automated quality assurance with high effectiveness and efficiency.
[0011] Further advantageous embodiments of the method according to the invention emerge from the subclaims.
[0012] Large-scale industrial (e.g., mechanical engineering and / or electrical engineering) application areas, etc., can be opened up with the method according to the invention if the at least one technical device is inspected automatically to form at least the first detection result and the first control detection result. If, for example, facts to be detected are processed in image data, it can be helpful for automated inspection if data relating to at least the first technical state or at least the first technical quality are encoded in at least one signal, which represents at least one image and which is evaluated to form at least the first detection result and the first control detection result.
[0013] For an automated inspection based on images, it may also be advantageous if at least the technical first state or at least the technical first quality is at least one state or at least one quality of at least one signal which represents at least one image and which is evaluated to form at least the first detection result and the first control detection result.
[0014] This measure can be used, for example, to detect differences in images based on differing signal parameters. For example, by evaluating signal parameters that represent light-dark differences in images, the contents of medical X-ray images can be compared and anomalies detected.
[0015] A designation of the first detection result is achieved when the first detection result is assigned as a label to a data set which concerns the first detection result.
[0016] This measure simplifies automated processing of the first detection result. A repeated, and thus resource-intensive, generation of the first detection result is avoided.
[0017] Detection results can be compared with each other, for example, based on the labels assigned to them. Incorrect detection results can be identified based on labels and corrected by changing the labels. This also simplifies the processing of detection results, for example, in databases.
[0018] A reduction in the complexity of an artificial neural network by compressing input information is possible if at least the first model of artificial intelligence is an autoencoder .
[0019] An automated evaluation of the technical first state or the technical first quality is made possible if at least the technical first state or at least the technical first quality is classified by means of at least one first classification model implemented in the first computing unit.
[0020] An advantageous solution is achieved if a federal third classification model is formed in the data processing system on the basis of at least the first classification model and a second classification model implemented in the second computing unit, of which a first classification model instance is formed in the first computing unit and a second classification model instance is formed in the second computing unit.
[0021] This measure avoids the merging of local data sets and their transfer to a central processing unit. The quality of assessment results regarding the initial technical status or initial technical quality is improved through federalization.
[0022] Training of the first classification model on the basis of the first detection result and a second detection result formed by means of a fourth artificial intelligence model is made possible if the fourth artificial intelligence model implemented in the first computing unit and a fifth artificial intelligence model implemented in the second computing unit are trained on the basis of training data representing technical conditions or technical properties, wherein the federal third artificial intelligence model is formed from the first training results and the second training results as well as from third training results relating to the fourth artificial intelligence model and fourth training results relating to the fifth artificial intelligence model,wherein at least one technical second state or at least one technical second condition is detected at least by means of the fourth artificial intelligence model and by means of the third artificial intelligence model, wherein at least the second detection result is formed by means of the fourth artificial intelligence model and at least one second control detection result is formed by means of the third artificial intelligence model, and wherein a second detection error is detected when at least the second detection result and the second control detection result deviate from one another.
[0023] With regard to a timely detection of a misguided training process of the first classification model, it is helpful if at least the first classification model is trained to distinguish at least the technical first state from the technical second state or at least the technical first condition from the technical second condition on the basis of at least the first detection result and the second detection result, wherein a first training process of the first classification model for detecting anomalous or erroneous first training inputs into the first classification model concerning at least the first detection result and the second detection result is monitored by means of the third artificial intelligence model on the basis of at least the first control detection result and the second control detection result.
[0024] This measure allows for the registration, deletion and / or correction of anomalous or erroneous initial training entries.
[0025] It is advantageous if a first model instance of the third artificial intelligence model is formed in the first computing unit and a second model instance of the third artificial intelligence model is formed in the second computing unit, wherein the second training results are transferred from the second computing unit to the first computing unit.
[0026] This avoids the merging of local data sets and their transfer to a central processing unit. To form the federal third model, the first processing unit and the second processing unit can coordinate themselves.
[0027] A promising field of application for the method according to the invention is opened up with a data processing system comprising means for carrying out the method according to the invention.
[0028] The means comprise the first computing unit and the second computing unit.
[0029] The invention is explained in more detail below using exemplary embodiments.
[0030] Examples include:
[0031] Fig. 1: A flowchart of an exemplary embodiment of a computer-implemented method according to the invention for detecting technical conditions or technical properties,
[0032] Fig. 2: A schematic representation of an exemplary first embodiment of a data processing system according to the invention with a first computing unit and a second computing unit, and
[0033] Fig. 3: A schematic representation of an exemplary second embodiment of a data processing system according to the invention with a first computing unit, a second computing unit and a third computing unit.
[0034] Fig. 1 shows a flow chart for an exemplary embodiment of a computer-implemented method according to the invention for detecting technical states or technical properties, wherein a first artificial intelligence model 4 implemented in a first computing unit 1 of a data processing system, as shown by way of example in Fig. 2 and Fig. 3, and a second artificial intelligence model 5 implemented in a second computing unit 2 of the data processing system, which is connected to the first computing unit 1 in a signal-transmitting manner via a first signal line 9, are trained on the basis of training data representing technical properties (first model training step 14), which are federated to form a third artificial intelligence model 6 (first federalization step 16).Furthermore, in the first model training step 14, a fourth artificial intelligence model 7 implemented in the first computing unit 1 and a fifth artificial intelligence model 8 implemented in the second computing unit 2 are trained on the basis of the training data representing technical properties, wherein the fourth model 7 and the fifth model 8 are also included in the first federalization step 16.
[0035] The first model 4, the second model 5, the third model 6, the fourth model 7 and the fifth model 8 are shown as examples in Fig. 2 and Fig. 3.
[0036] The training data is generated by carrying out a digital image analysis, whereby digital images which show technical devices in the form of vehicle bodies, with and without damage, are automatically analyzed in the image processing system. The images are provided with digital labels which indicate defective and error-free technical properties. On the basis of these images, the first model 4, the second model 5, the third model 6, the fourth model 7 and the fifth model 8 learn to recognize vehicle bodies as defective. The digital images used in the first model training step 14 show various defects which can be differentiated between using a classification step 18. The first model 4 and the second model 5 are trained on the basis of first images which show scratches in vehicle bodies.The fourth model 7 and the fifth model 8 are trained based on second images showing cracks in vehicle bodies. Using classification step 18, scratches can be distinguished from cracks in vehicle bodies.
[0037] According to the invention, it is also possible to form the training data on the basis of measurements on technical devices which make anomalous or faulty and fault-free technical states of the devices recognizable (e.g. a faulty state of a friction brake in which a maximum permissible temperature of the friction brake is exceeded, etc.), whereby technical states can be detected.
[0038] The digital images for forming the training data are fed from external data sources into the first computing unit 1 via a second signal line 10, as shown by way of example in Fig. 2 and Fig. 3, and into the second computing unit 2 via a third signal line 11, as shown by way of example in Fig. 2 and Fig. 3.
[0039] In the data processing system, the federal third model 6 is then formed in the first federalization step 16 from first training results relating to the first model 4, from second training results relating to the second model 5, from third training results relating to the fourth model 7 and fourth training results relating to the fifth model 8. In this case, a first model instance 19 of the third model 6, as shown by way of example in Fig. 2, is formed in the first computing unit 1 and a second model instance 20 of the third model 6, as shown by way of example in Fig. 2, is formed in the second computing unit 2, the first training results and the third training results being transferred from the first computing unit 1 to the second computing unit 2 and the second training results and the fourth training results being transferred from the second computing unit 2 to the first computing unit 1.
[0040] The first training results and the second training results relate to a first technical condition, the third training results and the fourth training results relate to a second technical condition. The third model 6 comprises a first module and a second module, wherein the first module is formed from the first training results and the second training results, and the second module is formed from the third training results and the fourth training results. Instances of the first module and the second module are formed in the first model instance 19 and in the second model instance 20.
[0041] The first federalization step 16 is a process of federated machine learning in which the first model 4, the second model 5, the fourth model 7 and the fifth model 8 are each trained with local data sets without input data for the first model training step 14 being exchanged. Rather, the first training results, the second training results, the third training results and the fourth training results are exchanged in the form of updated model parameters between the first computing unit 1 and the second computing unit 2 at regular intervals. It is a decentralized approach to federated learning in which the first computing unit 1 and the second computing unit 2 coordinate themselves.
[0042] The first model 4, the second model 5, the first module of the third model 6, the second module of the third model 6, the fourth model 7, and the fifth model 8 are designed as autoencoders, which are artificial neural networks. The autoencoders are used to learn compressed representations for image data relating to the vehicle bodies (encoding). Parameters that characterize anomalous or faulty properties of the vehicle bodies (e.g., scratches or cracks) are extracted. The autoencoders each comprise three layers: an input layer, an encoding layer, and an output layer.
[0043] The input layer comprises first neurons that describe an image dataset. The encoding layer creates a compressed representation of the image dataset. The output layer comprises second neurons that represent a reconstruction of the first neurons and have the meaning of the first neurons.
[0044] Using autoencoders, error-free as well as faulty or damaged properties of the vehicle bodies are learned from image data and finally, based on training results, deviations from error-free properties of the vehicle bodies are detected.
[0045] By means of the first model 4 and by means of the first module of the third model 6, the first technical property is then detected on the basis of a first detection image data set (detection step 21), wherein a first detection result is formed by means of the first model 4 and a first control detection result is formed by means of the first module of the third model 6.
[0046] The first detection result and the first control detection result are compared. A first detection error is detected if the first detection result and the first control detection result differ from one another (first control step 22). The first detection error is detected if, for example, the first detection result indicates a defect-free body surface, but the first control detection result indicates damage.
[0047] In the detection step 21, the second technical quality is further detected by means of the fourth model 7 and by means of the second module of the third model 6 on the basis of a second detection image data set, wherein a second detection result is formed by means of the fourth model 7 and a second control detection result is formed by means of the second module of the third model 6. The second detection result and the second control detection result are compared with one another. A second detection error is detected in the first control step 22 if the second detection result and the second control detection result differ from one another.
[0048] In the detection step 21, the first technical quality is further detected by means of the second model 5 and the first module of the third model 6 based on a third detection image data set, wherein a third detection result is formed by means of the second model 5 and a third control detection result is formed by means of the first module of the third model 6. The third detection result and the third control detection result are compared with one another. A third detection error is detected in the first control step 22 if the third detection result and the third control detection result differ from one another.
[0049] In the detection step 21, the second technical quality is further detected by means of the fifth model 8 and by means of the second module of the third model 6 on the basis of a fourth detection image data set, wherein a fourth detection result is formed by means of the fifth model 8 and a fourth control detection result is formed by means of the second module of the third model 6. The fourth detection result and the fourth control detection result are compared with one another. A fourth detection error is detected in the first control step 22 if the fourth detection result and the fourth control detection result differ from one another.
[0050] The first detection result, the second detection result, the third detection result, the fourth detection result, the first control detection result, the second control detection result, the third control detection result, and the fourth control detection result are generated for visual quality control for the image-based detection of anomalies, defects, and / or damage to vehicle bodies, i.e., technical devices, in a vehicle manufacturing facility. This allows for the automated detection of body damage.
[0051] The first detection result, the second detection result, the third detection result, the fourth detection result, the first control detection result, the second control detection result, the third control detection result, the fourth control detection result and, if applicable, the first detection error, the second detection error, the third detection error and / or the fourth detection error are output via a first output unit 24 connected to the first computing unit 1, as shown by way of example in Fig. 2 and Fig. 3, and a second output unit 25 connected to the second computing unit 2, as shown by way of example in Fig. 2 and Fig. 3.
[0052] By means of detection step 21, the first technical condition and the second technical condition are detected, which are damages to vehicle bodies. By means of detection step 21, the first technical condition cannot be distinguished from the second technical condition; it can only be determined that damage is present (since the first model 4 and the second model 5 are trained only on the basis of the first images, which show scratches, and the fourth model 7 and the fifth model 8 are trained only on the basis of the second images, which show cracks). A distinction between the first technical condition and the second technical condition (between scratches and cracks) is only made possible by classification step 18.
[0053] According to the invention, however, it is also conceivable that the detection step 21 and the classification step 18 are combined, for example, into a single method step, for which purpose the first model 4, the second model 5, the third model 6, the fourth model 7 and the fifth model 8 can comprise, for example, classification modules which can be designed, for example, as support vector machines or as Bayesian classifiers, etc.
[0054] According to the invention, it is further conceivable that the detection step 21 and the classification step 18 are based, for example, on a technical first state and a technical second state, wherein, for example, by means of the detection step 21 a state is detected in which a friction brake exceeds a maximum permissible temperature and by means of the classification step 18, for example, on the basis of a duration of the maximum permissible temperature being exceeded by the friction brake, a criticality of the exceedance is classified, wherein the technical first state can, for example, be a non-critical maximum temperature exceedance and the technical second state can be a critical maximum temperature exceedance, etc . According to the invention, it is also conceivable that the third model 6 is in a via a fourth signal line 12, as shown by way of example in Fig.3, is formed with the first computing unit 1 and, via a fifth signal line 13, as is shown by way of example in Fig. 3, with the second computing unit 2, the third computing unit 3 of the data processing system, as is shown by way of example in Fig. 3, being connected in a signal-transmitting manner, wherein the first training results and the third training results from the first computing unit 1 and the second training results and the fourth training results from the second computing unit 2 are transferred to the third computing unit 3. In an alternative, centralized approach to federated learning, the third computing unit 3 can be used to coordinate learning steps in the first computing unit 1 and the second computing unit 2.The first control detection result, the second control detection result, the third control detection result and the fourth control detection result of the third model 6 can be transferred from the third computing unit 3 to the first computing unit 1 and to the second computing unit 2.
[0055] The first technical quality and the second technical quality are mechanical quality properties of vehicle bodies, wherein the vehicle bodies are inspected automatically on the basis of detection image data formed on the vehicle bodies by means of digital cameras, comprising the first detection image data set, the second detection image data set, the third detection image data set and the fourth detection image data set, in order to form the first detection result, the second detection result, the third detection result, the fourth detection result, the first control detection result, the second control detection result, the third control detection result and the fourth control detection result.
[0056] These detection image data relating to the first technical quality and the second technical quality are encoded in signals which represent digital detection images and which are evaluated to form the first detection result, the second detection result, the third detection result, the fourth detection result, the first control detection result, the second control detection result, the third control detection result and the fourth control detection result.
[0057] According to the invention, it is also conceivable that the first technical state or the first technical quality is a state or quality of a signal or a signal pattern that represents an image and that is evaluated to generate the first detection result and the first control detection result. This allows, for example, anomalies in a medical X-ray image to be detected and subsequently classified, etc.
[0058] According to the invention, it is also possible for the technical first state and the technical second state or the technical first quality and the technical second quality to relate, for example, to electrical and / or electronic quality properties of technical devices. An electrical quality property can relate, for example, to the winding quality of a coil of a relay; an electronic quality property can relate, for example, to the assembly of a circuit board with electronic components that conforms to the plan or does not conform to the plan, etc.
[0059] In the detection step 21, the first detection result is assigned as a first detection label to the first detection image data set relating to the first detection result, the second detection result is assigned as a second detection label to the second detection image data set relating to the second detection result, the third detection result is assigned as a third detection label to the third detection image data set relating to the third detection result, and the fourth detection result is assigned as a fourth detection label to the fourth detection image data set relating to the fourth detection result.
[0060] If a detection error is detected in a detection result in the first control step 22 (e.g. the first detection error relating to the first detection result), the detection label assigned to the corresponding detection result is declared invalid (e.g. the first detection label in the presence of the first detection error).
[0061] The classification step 18, by means of which the technical first condition and the technical second condition or the technical first state and the technical second state are classified, is carried out by means of a first classification model 26 implemented in the first computing unit 1, as shown by way of example in Fig. 2 and Fig. 3, and a second classification model 27 implemented in the second computing unit 2, as shown by way of example in Fig. 2 and Fig. 3.
[0062] The first classification model 26 and the second classification model 27 are designed as support vector machines, with which a plurality of objects are divided into classes via hyperplanes.
[0063] The first classification model 26 and the second classification model 27 are trained to distinguish the technical first condition from the technical second condition or the technical first state from the technical second state on the basis of detection results of the first model 4, the second model 5, the fourth model 7 and the fifth model 8 in a second model training step 15.
[0064] To distinguish between scratches and cracks in vehicle bodies, the first classification model 26 is trained on the basis of the first detection result and the second detection result as well as further detection results formed by means of the first model 4 and the fourth model 7, and the second classification model 27 is trained on the basis of the third detection result and the fourth detection result as well as further detection results formed by means of the second model 5 and the fifth model 8.
[0065] A first training process of the second model training step 15 on the first classification model 26 is monitored by the third model 6 on the basis of the first control detection result, the second control detection result, and further control detection results relating to the further detection results formed by the first model 4 and the fourth model 7 in order to detect anomalous or erroneous first training inputs into the first classification model 26. This detection result relates to the first detection result, the second detection result, and the further detection results formed by the first model 4 and the fourth model 7. Detection results which have invalid detection labels are thus not taken into account in the first training process.
[0066] A second training process of the second model training step 15 on the second classification model 27 is monitored by the third model 6 on the basis of the third control detection result, the fourth control detection result, and further control detection results relating to the further detection results formed by the second model 5 and the fifth model 8 in order to detect anomalous or erroneous second training inputs into the second classification model 27 relating to the third detection result, the fourth detection result, and the further detection results formed by the second model 5 and the fifth model 8. Detection results which have invalid detection labels are thus not taken into account in the second training process.
[0067] According to the invention, it is also possible for detection results which have invalid detection labels to be stored in an error data memory of the first computing unit 1, the second computing unit 2 and / or the third computing unit 3.
[0068] On the basis of classification training results from the second model training step 15 on the first classification model 26 and on the second classification model 27, a federal third classification model 28 is formed (second federalization step 17), of which a first classification model instance 29 is formed in the first computing unit 1 and a second classification model instance 30 is formed in the second computing unit 2.
[0069] The third classification model 28 is shown as an example in Fig. 2 and Fig. 3, the first classification model instance 29 and the second classification model instance 30 are shown as an example in Fig. 2.
[0070] The second federalization step 17 is a process of decentralized federal machine learning, in which the first classification model 26 and the second classification model 27 are each trained with local classification data sets without exchanging input data for the second model training step 15. Rather, first classification training results relating to the first classification model 26 and second classification training results relating to the second classification model 27 are exchanged in the form of updated classification model parameters between the first computing unit 1 and the second computing unit 2 at regular intervals.
[0071] Once the first classification model 26 and the second classification model 27 have been trained and federated, classifications are performed in classification step 18 using the first classification model 26, the second classification model 27, and the third classification model 28. For example, in classification step 18, the first image data set would be evaluated using the first classification model 26 to the effect that it indicates a scratch on a vehicle body (a first technical condition) and not a tear in the vehicle body (second technical condition). Thus, the first classification model 26 and the second classification model 27 can be used to differentiate between the first technical condition and the second technical condition, or between the first technical condition and the second technical condition.
[0072] By means of the third classification model 28, control classification results are formed, via which classification results formed by the first classification model 26 and the second classification model 27 are controlled and / or improved (second control step 23).
[0073] According to the invention, it is also conceivable that the third classification model 28 is formed in the third computing unit 3. In this case, the first classification training results and the second classification training results are transferred from the first computing unit 1 and the second computing unit 2 to the third computing unit 3. The control classification results are transferred from the third computing unit 3 to the first computing unit 1 and the second computing unit 2.
[0074] The third computing unit 3 can thus be used, in an alternative, centralized approach to federal learning, to coordinate learning steps in the first computing unit 1 and the second computing unit 2.
[0075] In Fig. 2, an exemplary first embodiment of a data processing system according to the invention with a first computing unit 1 and a second computing unit 2 is shown schematically.
[0076] The first computing unit 1 is connected to the second computing unit 2 by means of a first signal line 9.
[0077] The data processing system comprises a first artificial intelligence model 4 , a second artificial intelligence model 5 , a federal third artificial intelligence model 6 , a fourth artificial intelligence model 7 , a fifth artificial intelligence model 8 , as well as a first classification model 26 , a second classification model 27 and a federal third classification model 28 .
[0078] The first model 4 and the fourth model 7 , a first model instance 19 of the third model 6 , the first classification model 26 and a first classification model instance 29 of the third classification model 28 are implemented in the first computing unit 1 .
[0079] The second model 5 and the fifth model 8 , a second model instance 20 of the third model 6 , the second classification model 27 and a second classification model instance 30 of the third classification model 28 are implemented in the second computing unit 2 .
[0080] To feed data from external data sources into the first computing unit 1 and into the second computing unit 2, a second signal line 10 is connected to the first computing unit 1 and a third signal line 11 is connected to the second computing unit 2.
[0081] For outputting information and / or warnings, a first output unit 24 designed as a first display is connected to the first computing unit 1 and a second output unit 25 designed as a second display is connected to the second computing unit 2.
[0082] The first computing unit 1, the second computing unit 2, the first model 4, the second model 5, the third model 6, the fourth model 7, the fifth model 8, the first model instance 19, the second model instance 20, the first classification model 26, the second classification model 27, the third classification model 28, the first classification model instance 29, the second classification model instance 30, the first signal line 9, the second signal line 10, the third signal line 11, the first output unit 24 and the second output unit 25 are means of the data processing system for carrying out a method according to the invention, as shown by way of example in Fig. 1.
[0083] According to the invention, however, it is also conceivable that the means of the data processing system for executing an alternative variant of a method according to the invention comprise, for example, only the first computing unit 1, the second computing unit 2, the first model 4, the second model 5 and the third model 6, etc. Fig. 3 shows an exemplary second embodiment of a data processing system according to the invention in a schematic representation.
[0084] This exemplary second embodiment is similar to that exemplary first embodiment of a data processing system according to the invention, as disclosed in Fig. 2.
[0085] Therefore, in Fig. 3 some of the same reference symbols are used as in Fig. 2.
[0086] In contrast to Fig. 2, the exemplary second embodiment of a data processing system according to the invention according to Fig. 3 comprises a first computing unit 1, a second computing unit 2 and a third computing unit 3.
[0087] The data processing system comprises a first artificial intelligence model 4 , a second artificial intelligence model 5 , a federal third artificial intelligence model 6 , a fourth artificial intelligence model 7 , a fifth artificial intelligence model 8 , as well as a first classification model 26 , a second classification model 27 and a federal third classification model 28 .
[0088] The first model 4 , the fourth model 7 and the first classification model 26 are implemented in the first computing unit 1 .
[0089] The second model 5 , the fifth model 8 and the second classification model 27 are implemented in the second computing unit 2 .
[0090] The third model 6 and the third classification model 28 are implemented in the third computing unit 3 .
[0091] The first processing unit 1 is connected to the second processing unit 2 via a first signal line 9. To feed data from external data sources into the first processing unit 1 and the second processing unit 2, a second signal line 10 is connected to the first processing unit 1 and a third signal line 11 is connected to the second processing unit 2.
[0092] The third computing unit 3 is connected to the first computing unit 1 via a fourth signal line 12 and to the second computing unit 2 via a fifth signal line 13.
[0093] The first computing unit 1, the second computing unit 2, the third computing unit 3, the first model 4, the second model 5, the third model 6, the fourth model 7, the fifth model 8, the first classification model 26, the second classification model 27, the third classification model 28, the first signal line 9, the second signal line 10, the third signal line 11, the fourth signal line 12 and the fifth signal line 13 are means of the data processing system for carrying out a method according to the invention, as shown by way of example in Fig. 1.
[0094] List of names
[0095] 1 First computing unit
[0096] 2 Second computing unit
[0097] 3 Third computing unit
[0098] 4 First model
[0099] 5 Second model
[0100] 6 Third model
[0101] 7 Fourth model
[0102] 8 Fifth model
[0103] 9 First signal line
[0104] 10 Second signal line
[0105] 11 Third signal line
[0106] 12 Fourth signal line
[0107] 13 Fifth signal line
[0108] 14 First model training step
[0109] 15 Second model training step
[0110] 16 First step towards federalization
[0111] 17 Second federalization step
[0112] 18 Classification step
[0113] 19 First model instance
[0114] 20 Second model instance
[0115] 21 Detection step
[0116] 22 First control step
[0117] 23 Second control step
[0118] 24 First output unit
[0119] 25 Second output unit
[0120] 26 First classification model
[0121] 27 Second classification model
[0122] 28 Third classification model
[0123] 29 First classification model instance
[0124] 30 Second classification model instance
Claims
Patent claims 1. A computer-implemented method for detecting technical conditions or technical properties, wherein at least one first artificial intelligence model (4) implemented in a first computing unit (1) of a data processing system and a second artificial intelligence model (5) implemented in a second computing unit (2) of the data processing system that is connected to the first computing unit (1) in a signal-transmitting manner are trained on the basis of training data representing technical conditions or technical properties, wherein in the data processing system, a federal third artificial intelligence model (6) is formed at least from first training results relating to the first artificial intelligence model (4) and from second training results relating to the second artificial intelligence model (5), characterized in thatthat at least one technical first state or at least one technical first property is detected at least by means of the first artificial intelligence model (4) and by means of the third artificial intelligence model (6), wherein at least the technical first state or at least the technical first property relates to a mechanical, electrical and / or electronic quality property of at least one technical device, wherein for visual quality control for image-based detection of anomalies, errors and / or damages of the at least one technical device, at least one first detection result is formed by means of the first artificial intelligence model (4) and at least one first control detection result is formed by means of the third artificial intelligence model (6), and wherein a first detection error is then detected,if at least the first detection result and the first control detection result differ from each other., 2. Computer-implemented method according to claim 1, characterized in that the at least one technical device is inspected automatically to form at least the first detection result and the first control detection result.
3. Computer-implemented method according to claim 1 or 2, characterized in that data relating to at least the technical first state or at least the technical first quality are encoded in at least one signal which represents at least one image and which is evaluated to form at least the first detection result and the first control detection result.
4. Computer-implemented method according to claim 1 or 2, characterized in that at least the technical first state or at least the technical first quality is at least one state or at least one quality of at least one signal which represents at least one image and which is evaluated to form at least the first detection result and the first control detection result.
5. Computer-implemented method according to one of claims 1 to 4, characterized in that the first detection result is assigned as a label to a data set which concerns the first detection result.
6. Computer-implemented method according to one of claims 1 to 5, characterized in that at least the first artificial intelligence model (4) is an autoencoder. 7 . Computer-implemented method according to one of claims 1 to 6, characterized in that by means of at least one first classification model (26) implemented in the first computing unit (1) at least the technical first condition or at least the first technical condition is classified.
8. Computer-implemented method according to claim 7, characterized in that in the data processing system, on the basis of at least the first classification model (26) and a second classification model (27) implemented in the second computing unit (2), a federal third classification model (28) is formed, of which a first classification model instance (29) is formed in the first computing unit (1) and a second classification model instance (30) is formed in the second computing unit (2).
9. Computer-implemented method according to claim 7 or 8, characterized in that a federal third classification model (28) is formed in a third computing unit (3) of the data processing system, which is connected to the first computing unit (1) and the second computing unit (2) in a signal-transmitting manner, on the basis of at least the first classification model (26) and a second classification model (27) implemented in the second computing unit (2).
10. Computer-implemented method according to one of claims 7 to 9, characterized in that a fourth artificial intelligence model (7) implemented in the first computing unit (1) and a fifth artificial intelligence model (8) implemented in the second computing unit (2) are trained on the basis of training data representing technical conditions or technical properties, wherein the federal third artificial intelligence model (6) is derived from the first training results and the second training results as well as from third training results relating to the fourth artificial intelligence model (7) and fourth training results relating to the fifth artificial intelligence model (8) is formed, wherein at least one technical second state or at least one technical second condition is detected at least by means of the fourth artificial intelligence model (7) and by means of the third artificial intelligence model (6), wherein at least one second detection result is formed by means of the fourth artificial intelligence model (7) and at least one second control detection result is formed by means of the third artificial intelligence model (6), and wherein a second detection error is detected when at least the second detection result and the second control detection result differ from one another.
11. Computer-implemented method according to claim 10, characterized in that at least the first classification model (26) is trained to distinguish at least the technical first state from the technical second state or at least the technical first condition from the technical second condition on the basis of at least the first detection result and the second detection result, wherein a first training process of the first classification model (26) for detecting anomalous or erroneous first training inputs into the first classification model (26) relating to at least the first detection result and the second detection result is monitored by means of the third artificial intelligence model (6) on the basis of at least the first control detection result and the second control detection result.
12. Computer-implemented method according to claim 8 or 9 and claim 10 or 11, characterized in that the second classification model (27) is based on at least one third detection result formed by means of the second model (5) of artificial intelligence relating to at least the technical first state or at least the technical first quality and a detection result formed by means of the fifth model (8) artificial intelligence is trained on the basis of a fourth detection result relating to at least the technical second state or at least the technical second condition, wherein a second training process of the second classification model (27) for detecting anomalous or erroneous second training inputs into the second classification model (27) relating to at least the third detection result and the fourth detection result is monitored by means of the third artificial intelligence model (6) on the basis of at least one third control detection result formed by means of the third artificial intelligence model (6) for checking the third detection result and one fourth control detection result formed by means of the third artificial intelligence model (6) for checking the fourth detection result.
13. Computer-implemented method according to one of claims 1 to 12, characterized in that a first model instance (19) of the third artificial intelligence model (6) is formed in the first computing unit (1) and a second model instance (20) of the third artificial intelligence model (6) is formed in the second computing unit (2), wherein the second training results are transferred from the second computing unit (2) to the first computing unit (1).
14. Computer-implemented method according to one of claims 1 to 13, characterized in that the third model (6) of artificial intelligence is formed in a third computing unit (3) of the data processing system which is connected to the first computing unit (1) and the second computing unit (2) in a signal-transmitting manner, wherein the first training results from the first computing unit (1) and the second training results from the second computing unit (2) are transmitted to the third computing unit (3) and at least the first control detection result from the third computing unit (3) into the first computing unit (1).
15. A data processing system comprising means for carrying out the method according to one of claims 1 to 14.