A computer implementation method for providing a training dataset, a computer implementation method for training an AI system and using such a computer implementation method, and a training dataset for training an AI system.

Abstract

The present invention relates to a computer implementation method for providing a training dataset for training an AI system, the computer implementation method comprising the following steps: - A step of performing a measurement using a sensor, particularly a torque sensor, - A step of providing unaltered sensor data measured based on measurements performed by the sensor, - A step of generating modified sensor data based on unchanged sensor data, - The step of providing unchanged sensor data and changed sensor data as part of a training dataset for training an AI system.

Description

Technical Field

[0001] The present invention relates to a computer-implemented method for providing a training dataset for training an artificial intelligence (AI) system. Further, the present invention relates to a computer-implemented method for training an AI system for evaluating the consistency of sensor data and the use of such a method. The present invention also relates to a training dataset for training an AI system.

Background Art

[0002] The present invention can be used, for example, to check the function of a sensor and / or the quality of the measurement values it provides, i.e., the consistency of the sensor. Such checks are often provided for safety-related applications. For example, in the drive module of a collaborative robot, it may be necessary to check the consistency of the measurement data provided by torque and / or position sensors in order to eliminate the risk of harming the people operating the collaborative robot.

[0003] Such checks usually require a great deal of effort and detailed knowledge about the function and configuration of each sensor.

Summary of the Invention

Problems to be Solved by the Invention

[0004] In view of such a background, the object is to reduce the labor of checking the consistency of sensors and to provide a solution that can be used regardless of the function and configuration of the sensors.

Means for Solving the Problems

[0005] The object is achieved by a computer-implemented method for providing a training dataset for training an AI system, the computer-implemented method comprising the following steps: - performing measurements using a sensor, particularly a torque sensor, - A step of providing unchanged sensor data based on measurements performed by the sensor, - A step of generating modified sensor data based on unchanged sensor data, - A step of providing unchanged sensor data and changed sensor data as part of a training dataset for training an AI system, Includes.

[0006] According to the present invention, a training dataset for an artificial intelligence (AI) system is provided, which can be used in a trained state to check the integrity of sensors. Training makes it possible to adapt the AI ​​system to each application, i.e., each sensor. Detailed knowledge of the function or configuration of the monitored sensors is not required. According to the present invention, in order to provide training data, only unmodified sensor data, which can be obtained from, for example, a functioning sensor, is required. Sensor data from non-functioning sensors is not required. Rather, in the method according to the present invention, modified sensor data is generated based on the unmodified sensor data itself. This reduces the effort required to train the AI ​​system. The training dataset can be supplied, for example, during the calibration of a (static) functioning sensor, and the determined sensor data is used as unmodified sensor data for the training dataset.

[0007] Sensor data in the sense of this invention should be understood as measurement data and / or data derived from measurement data. Derived data can be, for example, filtered data, angle, or torque signals. Furthermore, in addition to using general AI models and parameterizations in the AI ​​system, models and / or parameterizations adapted individually to each sensor can be used. For example, transfer learning methods can also be used in this context to adapt general models individually to each sensor.

[0008] According to a preferred embodiment of the present invention, as a further part of the training dataset, a first item of classification information associated with unmodified sensor data and a second item of classification information associated with modified sensor data are provided, wherein the first and second items of classification information are different. The items of classification information may also be referred to as "labels," i.e., output values ​​of the AI ​​system.

[0009] According to an advantageous embodiment of the present invention, the first and second items of the classification information are: -Includes an indication of whether the sensor data associated with the classification information item has remained unchanged or has been changed. Examples of characteristics for the first item in the classification information are "not changed" or "no defects detected." Examples of characteristics for the second item in the classification information are "changed" or "defect detected."

[0010] According to a preferred embodiment of the present invention, the unmodified sensor data and the modified sensor data each include a plurality of sensor data channels, and the first item and the second item of the classification information are: - Includes displaying how many faulty sensor data channels exist. Having information about the number of faulty sensor data channels can provide the AI ​​system with a more realistic representation of the data. Alternatively, or additionally, the severity of the fault can be determined and transmitted to the AI ​​system.

[0011] According to an advantageous embodiment of the present invention, the unmodified sensor data and the modified sensor data each include a plurality of sensor data channels, and the first item and the second item of the classification information are, -Includes an indication of which of the one or more sensor data channels is faulty. Troubleshooting requires precise knowledge of the fault's location. Having information about which of one or more sensor data channels is faulty means the fault can be corrected more quickly.

[0012] According to a preferred embodiment of the present invention, modified sensor data is generated by adding a fault signal to the unchanged sensor data. The fault signal can be periodic or aperiodic and can be added to each of the unchanged sensor data without considering its sign.

[0013] A further subject of the present invention is a computer implementation method for training an AI system for evaluating the integrity of sensor data, the computer implementation method comprising the following steps: - A step of providing unaltered sensor data and linking the unaltered sensor data as the first training data to the first item of the classification information, - A step of providing modified sensor data and linking the modified sensor data to a second item of classification information as second training data, - A step of training the AI ​​system using the first training data and the second training data, Includes. In the method for training an AI system according to the present invention, the same technical advantages and effects as those already described with respect to the method for providing a training dataset can be achieved.

[0014] A favorable embodiment of a method for training an AI system is as follows: - A step of providing unaltered sensor data and linking the unaltered sensor data to the first item of classification information as first test data and / or verification data, - A step of providing modified sensor data and linking the modified sensor data to a second item of classification information as second test data and / or verification data, - After training the AI system using the first test data, the second test data, and / or the verification data, testing and / or verifying the AI system; including. The method steps following the training of the AI system, during which the AI system is tested and / or verified, can be used to perform a check of the training performed.

[0015] A further subject of the invention is the use of a computer-implemented method for training an AI system for evaluating the consistency of sensor data during calibration of a sensor providing non-altered sensor data, preferably a torque sensor, particularly preferably a torque sensor of a drive module for a robot arm joint, and an embodiment thereof.

[0016] A further subject of the invention is a drive module, particularly for a robot arm joint, the drive module comprising: - A harmonic drive mechanism having an elastic transmission element and a wave generator acting on the elastic transmission element, the elastic transmission element having a sensor with a sensor data channel, the sensor data being measurable by the sensor; - A control device having an AI system for evaluating the consistency of sensor data, the validity check of the sensor data being executable by the AI system; comprising.

[0017] A further subject of the invention is a training data set for training an AI system for evaluating the consistency of sensor data, the training data set comprising: - Non-altered sensor data and a first item of classification information associated with the non-altered sensor data; - Altered sensor data and a second item of classification information associated with the altered sensor data; including.

[0018] In the method, use, training dataset, and drive module for training an AI system, the advantageous embodiments and functions disclosed in relation to the method for providing training data can also be applied.

[0019] Further details and advantages of the present invention will be described below with reference to the exemplary embodiments shown in the drawings.

Brief Description of the Drawings

[0020] [Figure 1] Schematically shows an industrial robot having a drive module according to an exemplary embodiment of the present invention. [Figure 2] Schematically shows a block diagram of the drive module according to FIG. 1. [Figure 3] Schematically shows the training of an exemplary embodiment of an AI system according to the present invention using sensor data and derived sensor data. [Figure 4] Schematically shows the application of the already trained AI system from FIG. 3.

Modes for Carrying Out the Invention

[0021] FIG. 1 schematically shows an exemplary embodiment of a robot configured as an industrial robot 200 having a plurality of robot arm segments 201 each rotatably connected via a drive module 100, in which the present invention is implemented. The industrial robot 200 shown here has three robot arm segments 201 and three drive modules 100, but embodiments of the industrial robot 200 may have a different number of robot arm segments 201 and drive modules 100, for example, 4, 5, 6, or 7. Such an industrial robot 200 is often used as a collaborative robot that works in close cooperation with humans and is thus particularly exposed to high safety requirements.

[0022] The drive module 100 includes a strain wave gear mechanism, not shown in Figure 1, which comprises an elastic transmission element and a wave generator acting on the elastic transmission element. In addition, the elastic transmission element includes a sensor having a strain gauge array 30 having multiple sensor data channels. Furthermore, the drive module 100 has a control device with an AI system for evaluating the integrity of the sensor data, and the validity check of the sensor data can be performed by the AI ​​system.

[0023] Figure 2 schematically shows a block diagram of the drive module according to Figure 1. The sensor 30 of the drive module 100, which is configured as a strain gauge array, comprises a plurality (in this case, n) of strain gauges 31. Sensor data is provided to a control device 40, which includes an AI system 50, via a sensor data channel 32. The AI ​​system 50 checks the sensor data to draw conclusions about the modified sensor signal 32 from the sensor 30, and then generates signals for controlling the actuator 60, in particular torque and / or rotation angle signals. If a modified sensor signal is detected, a warning can be issued, and / or the control device 40 can correct the sensor signal 32 that is recognized as modified.

[0024] Figure 3 schematically illustrates the training of an exemplary embodiment of the AI ​​system 50 according to the present invention, in which sensor data is provided to the AI ​​system 50 via sensor data channel 32 and derived sensor data 33. Furthermore, a status signal 34 is continuously provided, which may include a first and / or second item of classification information. The item of classification information may include an indication of whether the sensor data is faulty or not. In addition, the item of classification information may include an indication of whether there are faulty sensor data channels 32, how many faulty sensor data channels 32 there are, and which of the one or more sensor data channels 32 is faulty.

[0025] According to the present invention, the AI ​​system 50 is trained using the method, which consists of the following steps: - A step of providing unaltered sensor data and linking the unaltered sensor data as the first training data to the first item of the classification information, - A step of providing modified sensor data and linking the modified sensor data to a second item of classification information as second training data, - The step of training an AI system 50 using first training data and second training data.

[0026] Figure 4 schematically illustrates an application of the AI ​​system already trained from Figure 3. This may include, in particular, testing and / or verifying the AI ​​system 50, where the status signal 34 may be, for example, a warning signal. [Explanation of Symbols]

[0027] 30 Sensor / Strain Gauge Array 31 Strain Gauges 32 sensor data channels 33 Derived Sensor Data 34 Status signals 40 Control device 50 AI Systems 60 Actuators 100 drive module 200 Industrial Robots 201 Robot Arm

Claims

1. A computer implementation method for providing a training dataset for training an AI system (50), comprising the following steps: - A step of performing a measurement using a sensor (30), particularly a torque sensor, - A step of providing unchanged sensor data based on the measurement performed by the sensor, - A step of generating modified sensor data based on the unchanged sensor data, - The steps of providing the unchanged sensor data and the changed sensor data as part of a training dataset for training the AI ​​system (50), Computer implementation methods, including those mentioned above.

2. The computer implementation method according to claim 1, wherein a first item of classification information associated with the unchanged sensor data and a second item of classification information associated with the changed sensor data are provided as further part of the training dataset, characterized in that the first item and the second item of classification information are different.

3. The first and second items of the classification information are, - Display whether the unmodified or modified sensor data associated with the aforementioned item in the classification information has been modified or not. The computer implementation method according to claim 2, characterized by including the following:

4. The unmodified sensor data and the modified sensor data each include a plurality of sensor data channels, and the first item and the second item of the classification information are, - Displaying the number of faulty sensor data channels. A computer implementation method according to claim 2 or 3, characterized by including the following:

5. The unmodified sensor data and the modified sensor data each include a plurality of sensor data channels, and the first item and the second item of the classification information are, - Includes an indication of which of one or more sensor data channels is faulty. A computer implementation method according to any one of claims 2 to 4, characterized in that

6. The computer implementation method according to any one of claims 1 to 5, characterized in that the modified sensor data is generated by adding an error signal to the unmodified sensor data.

7. A computer implementation method for training an AI system (50) for evaluating the integrity of sensor data, comprising the following steps: - A step of providing unaltered sensor data and linking the unaltered sensor data to a first item of classification information as first training data, - A step of providing modified sensor data and linking the modified sensor data to a second item of classification information as second training data, - A step of training the AI ​​system (50) using the first training data and the second training data, Computer implementation methods, including those mentioned above.

8. The following steps: - A step of providing unaltered sensor data and linking the unaltered sensor data to a first item of classification information as first test data and / or verification data, - A step of providing modified sensor data and linking the modified sensor data to a second item of classification information as second test data and / or verification data, - A step of training the AI ​​system (50) using the first test data, the second test data, and / or verification data, and then testing and / or verifying the AI ​​system (50), A computer implementation method for training the AI ​​system (50) according to claim 7, characterized by the above.

9. In particular, a drive module (100) for a robot arm joint, - A strain wave gear mechanism comprising an elastic transmission element and a wave generator acting on the elastic transmission element, wherein the elastic transmission element has a sensor (30) having a sensor data channel (32), and the sensor data can be measured by the sensor (30), - A control device (40) having an AI system (50) for evaluating the consistency of sensor data, wherein the validity check of the sensor data can be performed by the AI ​​system (50), A drive module (100) equipped with the following.

10. A training dataset for training an AI system (50) for evaluating the integrity of sensor data, - Unmodified sensor data, and the first item of classification information associated with the unmodified sensor data, - The modified sensor data, and the second item of classification information associated with the modified sensor data, A training dataset that includes [the specified data].

Images