Computer-implemented method for providing a training data set, computer-implemented method for training an ai system and using such a computer-implemented method, and training data set for training an ai system
Patent Information
- Authority / Receiving Office
- EP · EP
- Patent Type
- Applications
- Current Assignee / Owner
- SCHAEFFLER TECHNOLOGIES AG & CO KG
- Filing Date
- 2024-06-10
- Publication Date
- 2026-05-13
AI Technical Summary
Current methods for checking the integrity of sensors, especially in security-relevant applications like collaborative robots, are labor-intensive and require detailed knowledge of the sensor's functionality and structure, making them inefficient and difficult to implement.
A computer-implemented method generates a training data set for an AI system by creating adulterated sensor data from unadulterated data, allowing the AI to be trained without needing detailed sensor knowledge, using unadulterated data from functional sensors and adapting to specific sensors through generic or individually tailored models.
This approach reduces the effort required to train AI systems for sensor integrity checks, enabling efficient adaptation to various sensors without requiring detailed sensor knowledge, and provides a realistic image of data errors for quicker troubleshooting.
Smart Images
Figure DE2024100510_09012025_PF_FP_ABST
Abstract
Description
[0001] Computer-implemented method for providing a training data set, Computer-implemented method for training a Kl system and
[0002] Use of such a computer-implemented method, training data set for training a Kl system
[0003] The invention relates to a computer-implemented method for providing a training data set for training an artificial intelligence (AI) system. Furthermore, the invention relates to a computer-implemented method for training an AI system for assessing the integrity of sensor data and a use of such a method. Furthermore, the invention relates to a training data set for training an AI system.
[0004] The invention can be used, for example, to check the functionality of a sensor and / or the quality of the measured values it provides, i.e., the integrity of the sensor. Such checks are often required in safety-relevant applications. For example, in drive modules of collaborative robots, it may be necessary to check the integrity of the measurement data provided by torque and / or position sensors in order to prevent harm to humans working with the collaborative robots.
[0005] Such checks are usually very time-consuming and require detailed knowledge of the functionality and structure of the respective sensor.
[0006] Against this background, the task is to provide a solution that reduces the effort involved in checking the integrity of a sensor and can be used regardless of the functionality and design of the sensor.
[0007] The problem is solved by a computer-implemented method for providing a training data set for training a Kl system comprising the following steps:
[0008] - Carrying out a measurement with a sensor, in particular a torque sensor;
[0009] - Providing unadulterated sensor data that depends on the measurement performed with the sensor;
[0010] - Generating corrupted sensor data depending on the uncorrupted sensor data; and - Providing the uncorrupted sensor data and the corrupted sensor data as part of a training data set for training an AI system.
[0011] According to the invention, a training data set is provided for an artificial intelligence (AI) system which, in a trained state, can be used to check the integrity of the sensor. Through training, the AI system can be taught to the respective application, i.e. the respective sensor. Detailed knowledge of the functionality or structure of the sensor to be monitored is not required. According to the invention, the provision of the training data only requires unadulterated sensor data, which can, for example, originate from a functional sensor. Sensor data from a non-functional sensor is not required. Rather, the method according to the invention generates falsified sensor data itself as a function of the unadulterated sensor data. In this way, the effort required to train the AI system is reduced.The provision of the training data set can, for example, be carried out during the calibration of a (still) functional sensor, whereby the sensor data obtained during this process is used as unadulterated sensor data for the training data set.
[0012] Sensor data, within the meaning of the invention, refers to measurement data and / or data derived from the measurement data. Derived data can be, for example, filtered data, angle or torque signals. Furthermore, both generic AI models and parameterizations as well as models and / or parameterizations individually adapted to the respective sensor can be used for the AI system. Transfer learning methods, for example, can also be used to individually adapt a generic model to the respective sensor.
[0013] According to a preferred embodiment of the invention, it is provided that, as a further part of the training data set, a first item of classification information associated with the unadulterated sensor data and a second item of classification information associated with the falsified sensor data are provided, wherein the first and second items of classification information differ. Classification information can also be referred to as a "label," i.e., an output value of the AI system.
[0014] An advantageous embodiment of the invention provides that the first and second classification information comprises: - an indication as to whether the unadulterated or falsified sensor data to which the classification information is assigned are unadulterated or falsified.
[0015] Examples of the first classification information are "unadulterated" or "no error detected." Examples of the second classification information are "adulterated" or "error detected."
[0016] According to a preferred embodiment of the invention, it is provided that the unadulterated and falsified sensor data each comprise several sensor data channels and the first and second classification information comprises:
[0017] - an indication of how many faulty sensor data channels are present.
[0018] By providing information about the number of faulty sensor data channels, the AI system can be provided with a more realistic representation of the data. Alternatively or additionally, the severity of the error(s) can be determined and transmitted to the AI system.
[0019] An advantageous embodiment of the invention provides that the unadulterated and falsified sensor data each comprise several sensor data channels and the first and second classification information comprises:
[0020] - an indication of which sensor data channel(s) are faulty.
[0021] Troubleshooting requires precise knowledge of the fault location. Having information about which sensor data channel(s) are faulty allows for faster troubleshooting.
[0022] According to a preferred embodiment of the invention, the corrupted sensor data is generated by adding the uncorrupted sensor data to an error signal. The error signal can be a periodic or aperiodic signal and can be added to the respective uncorrupted sensor data with or without consideration of a sign.
[0023] Another object of the invention is a computer-implemented method for training a Kl system for evaluating the integrity of sensor data, comprising the following steps:
[0024] - Providing unadulterated sensor data and linking the unadulterated sensor data with a first classification information as first training data; - Providing falsified sensor data and linking the falsified sensor data with a second classification information as second training data;
[0025] - Training the KL system with the first and second training data.
[0026] In the method according to the invention for training a Kl system, the same technical advantages and effects can be achieved that have already been described in connection with the method for providing a training data set.
[0027] An advantageous embodiment of the method for training the Kl system comprises the following steps:
[0028] - Providing unadulterated sensor data and linking the unadulterated sensor data with a first classification information as first test and / or validation data;
[0029] - Providing corrupted sensor data and linking the corrupted sensor data with a second classification information as second test and / or validation data;
[0030] - Testing and / or validating the AI system after training the AI system with the first and second test and / or validation data.
[0031] A subsequent procedural step following the training of the AI system, in which testing and / or validation of the AI system is carried out, allows a review of the training carried out.
[0032] A further subject matter of the invention is a use of the computer-implemented method for training a Kl system for evaluating an integrity of sensor data and its design during the calibration of a sensor providing the unadulterated sensor data, preferably a torque sensor, particularly preferably a torque sensor in a drive module for a robot arm joint.
[0033] Another subject of the invention is a drive module, in particular for a robot arm joint, comprising
[0034] - a stress wave transmission comprising an elastic transmission element and a wave generator acting on the elastic transmission element, wherein the elastic transmission element comprises a sensor with a sensor data channel, wherein a sensor data can be measured by means of the sensor, and
[0035] - a control device with a Kl system for evaluating the integrity of sensor data, wherein a plausibility check of the sensor data can be carried out by means of the Kl system. A further subject of the invention is a training data set for training a Kl system for evaluating the integrity of sensor data, comprising:
[0036] - unadulterated sensor data and first classification information associated with the unadulterated sensor data;
[0037] - falsified sensor data and a second classification information associated with the falsified sensor data.
[0038] The advantageous embodiments and features disclosed in connection with the method for providing training data can also be applied to the method for training the AI system, the use, the training data set and the drive module.
[0039] Further details and advantages of the invention will be explained below with reference to the exemplary embodiments illustrated in the drawings. Herein:
[0040] Fig. 1 schematically shows an industrial robot having a drive module according to an embodiment of the invention;
[0041] Fig. 2 shows a schematic block diagram of a drive module according to Fig. 1;
[0042] Fig. 3 schematically shows the training of an embodiment of a Kl system according to the invention with sensor data and derived sensor data; and
[0043] Fig. 4 schematically shows the application of the already trained Kl system from Fig. 3.
[0044] Fig. 1 shows a schematic representation of an embodiment of a robot designed as an industrial robot 200 with a plurality of robot arm segments 201, each of which is rotatably connected via drive modules 100, in which the invention is implemented. Even though the industrial robot 200 shown here has three robot arm segments 201 and three drive modules 100, embodiments of the industrial robot 200 with a different number of robot arm segments 201 and drive modules 100 are conceivable, for example four, five, six, or seven. Such industrial robots 200 are often used as collaborative robots that work in close cooperation with humans and are therefore subject to particularly high safety requirements. The drive module 100 comprises a stress wave gear (not shown in Fig. 1), which has an elastic transmission element and a wave generator acting on the elastic transmission element.In addition, the elastic transmission element comprises a sensor with a strain gauge arrangement 30 having multiple sensor data channels. Furthermore, the drive module 100 has a control device with a K1 system for evaluating the integrity of sensor data, wherein a plausibility check of the sensor data can be performed using the K1 system.
[0045] Fig. 2 shows a schematic block diagram of the drive module according to Fig. 1. The sensor 30 of the drive module 100, designed as a strain gauge arrangement, comprises a plurality of, here n, strain gauges 31. Sensor data is provided to a control device 40, which includes an AI system 50, via sensor data channels 32. The AI system 50 checks the sensor data in order to draw conclusions about corrupted sensor signals 32 from the sensor 30 and then generate signals for controlling an actuator 60, in particular torque and / or angle of rotation signals. If a corrupted sensor signal is detected, a warning can be issued and / or the control device 40 can correct the sensor signal 32 identified as corrupted.
[0046] Fig. 3 schematically illustrates the training of an embodiment of an AI system 50 according to the invention, wherein the AI system 50 is provided with sensor data via sensor data channels 32 and derived sensor data 33. Furthermore, a status signal 34 is continuously provided, which may include first and / or second classification information. The classification information may include an indication as to whether the sensor data is faulty or error-free. Furthermore, the classification information may include an indication as to whether and how many faulty sensor data channels 32 are present, and which sensor data channel(s) 32 are faulty.
[0047] According to the invention, the training of the Kl system 50 is carried out using a method comprising the following steps:
[0048] - Providing unadulterated sensor data and linking the unadulterated sensor data with a first classification information as first training data;
[0049] - Providing falsified sensor data and linking the falsified sensor data with a second classification information as second training data;
[0050] - Training the AI system 50 with the first and second training data. Fig. 4 schematically shows the application of the already trained AI system from Fig. 3. This can, in particular, involve testing and / or validating the AI system 50, wherein the status signal 34 can, for example, be a warning signal.
[0051] List of reference symbols
[0052] 30 sensor / strain gauge arrangement
[0053] 31 strain gauges
[0054] 32 sensor data channels
[0055] 33 Derived sensor data
[0056] 34 Status signal
[0057] 40 Control device
[0058] 50 Kl system
[0059] 60 Actuator
[0060] 100 drive module
[0061] 200 industrial robots
[0062] 201 Robot arm
Claims
Patent claims 1. Computer-implemented method for providing a training data set for training a Kl system (50) comprising the following steps: - carrying out a measurement with a sensor (30), in particular a torque sensor; - Providing unadulterated sensor data that depends on the measurement performed with the sensor; - generating corrupted sensor data depending on the uncorrupted sensor data; and - Providing the unadulterated sensor data and the falsified sensor data as part of a training data set for training a Kl system (50).
2. Computer-implemented method according to claim 1, characterized in that as a further part of the training data set, a first classification information item associated with the unadulterated sensor data and a second classification information item associated with the falsified sensor data are provided, wherein the first and the second classification information items differ.
3. Computer-implemented method according to claim 2, characterized in that the first and second classification information comprises: - an indication as to whether the unadulterated or falsified sensor data to which the classification information is assigned are unadulterated or falsified.
4. Computer-implemented method according to one of claims 2 or 3, characterized in that the unadulterated and falsified sensor data each comprise a plurality of sensor data channels and the first and second classification information comprises: - an indication of how many faulty sensor data channels are present.
5. Computer-implemented method according to one of claims 2 to 4, characterized in that the unadulterated and falsified sensor data each comprise a plurality of sensor data channels and the first and second classification information comprises: - an indication of which sensor data channel(s) are faulty.
6. Computer-implemented method according to one of the preceding claims, characterized in that the generation of the corrupted sensor data is carried out by adding the uncorrupted sensor data with an error signal.
7. Computer-implemented method for training a Kl system (50) for evaluating the integrity of sensor data, comprising the following steps: - Providing unadulterated sensor data and linking the unadulterated sensor data with a first classification information as first training data; - Providing falsified sensor data and linking the falsified sensor data with a second classification information as second training data; - Training the Kl system (50) with the first and second training data.
8. Computer-implemented method for training a Kl system (50) according to claim 7, characterized by the following steps: - Providing unadulterated sensor data and linking the unadulterated sensor data with a first classification information as first test and / or validation data; - Providing corrupted sensor data and linking the corrupted sensor data with a second classification information as second test and / or validation data; - testing and / or validating the KL system (50) after training the KL system (50) with the first and second test and / or validation data.
9. Drive module (100), in particular for a robot arm joint, comprising - a stress wave transmission comprising an elastic transmission element and a wave generator acting on the elastic transmission element, wherein the elastic transmission element comprises a sensor (30) with a sensor data channel (32), wherein a sensor data item can be measured by means of the sensor (30), and - a control device (40) with a Kl system (50) for evaluating the integrity of sensor data, wherein a plausibility check of the sensor data can be carried out by means of the Kl system (50).
10. Training data set for training a Kl system (50) for evaluating the integrity of sensor data, comprising: - unadulterated sensor data and first classification information associated with the unadulterated sensor data; - falsified sensor data and a second classification information associated with the falsified sensor data.