Detecting disturbances in robot arrangements

Abstract

Method for machine learning of a disturbance detection, for detecting disturbances in a robot arrangement which has at least one robot (100), comprising the step: - Providing (S10) undisturbed time series (1, 1') of data from a first robot arrangement comprising at least one robot (100); and the multiple alternating steps: - Generating (S20) time series (3) of data with artificial disturbances based on at least a part (1) of the provided time series assessed as undisturbed using a generator (10) of a GAN; and - Training (S30) a discriminator (20) of the GAN based on the generated time series (3) with artificial perturbations, the artificial perturbations and at least a part (1') of the provided time series assessed as undisturbed; wherein - the discriminator (20) is trained to detect time points (t1) and / or magnitudes (A(t1)) of disturbances; and - the generator (10) is trained to generate time series (3) with artificial disturbances, the extent of which the discriminator (20) underestimates.

Description

[0001] The present invention relates to a method for detecting faults in a robot arrangement having at least one robot, with the aid of fault detection, in one embodiment of at least one artificial neural network and / or other machine learning algorithm, methods for machine learning or training, in particular by means of distance- and / or clustering-based machine learning and / or with or without the use of at least one artificial neural network, for example random forests, Arma / Arima models, gradient boosting or the like, the fault detection, in particular an artificial neural network, as well as a system and computer program product for carrying out at least one of the methods.

[0002] The fault detection system used or machine-learned thus comprises, in one embodiment, one or more artificial neural networks. Additionally or alternatively, other machine learning algorithms can also be used, preferably distance- and / or clustering-based machine learning; in particular, random forests, ARMA / ARIMA models, gradient boosting, or the like can be used. Distance-based clustering methods are particularly suitable and can advantageously be combined with at least one neural network. Accordingly, in one embodiment, the fault detection system used or machine-learned comprises one or more, in particular these, machine learning algorithms, or uses them, or is based on them.

[0003] DE 10 2018 203 234 A1 discloses a method for the predictive assessment of at least one robot.

[0004] DE 10 2018 009 309 A1 discloses a numerical control device that can quantify a degree of an anomaly in response to detecting the occurrence of the anomaly and offer a countermeasure according to the magnitude of the anomaly.

[0005] DE 10 2017 003 165 A1 discloses a machine learning device and a machine learning method for learning the failure prediction of a main shaft or a motor driving the main shaft, as well as a failure prediction device and a failure prediction system comprising the machine learning device.

[0006] DE 10 2016 008 867 A1 discloses a fault checking system capable of suitably checking a fault between a machine tool and a robot in real time, even in the case where real-time characteristics of data communication are not ensured in a system including a machine tool and a robot.

[0007] DE 10 2016 008 987 A1 discloses a failure prediction system including a machine learning device that learns conditions associated with a failure of an industrial machine.

[0008] DE 11 2016 001 831 B4 discloses a method and a device for controlling / regulating an actuator-driven robot joint with error detection and corresponding reactions.

[0009] The object of the present invention is to improve the operation of a robot arrangement.

[0010] This object is achieved in particular by a method having the features of claims 1, 5, and 9, respectively. Claims 10 and 11 protect a system or computer program product for implementing a method described herein. The subclaims relate to advantageous developments.

[0011] According to a first aspect of the present invention, a method for machine learning of fault detection, in an embodiment for training an artificial neural network for detecting faults in a robot arrangement comprising one or more robots, comprises the step: - Providing time series of data assessed as undisturbed from a first robot arrangement comprising at least one robot.

[0012] In one embodiment, this provision comprises capturing, in particular measuring, and / or filtering data from the first robot arrangement. Additionally or alternatively, the provision may also comprise, in particular be, retrieving stored, in particular measured and / or filtered, data from the first robot arrangement.

[0013] The time series are, in particular, assessed as undisturbed in one version manually or by at least one person. Additionally or alternatively, they can also be assessed as undisturbed in a (partially or fully) automated manner.

[0014] According to one embodiment of the present invention, the method comprises the multiple alternating steps: - generating time series of data with artificial disturbances based on at least a part of the provided time series assessed as undisturbed by means of a generator of a GAN; and - Training a GAN discriminator based on • at least part of the time series provided, assessed as undisturbed, • the generated time series with artificial disturbances and • artificial disturbances;whereby - the discriminator is monitored or supervised, in particular on the basis of the artificial disturbances, and is trained to detect times and / or extents, in particular amplitudes, of disturbances; and - the generator is trained to generate time series with artificial disturbances whose extent is underestimated by the discriminator, in particular which the discriminator does not detect or estimates to be less than the artificial disturbances in the generated time series.

[0015] In one embodiment, the (trained) discriminator itself is used or provided (for this purpose), in one embodiment subsequently, in or for, in one embodiment as, disturbance detection.

[0016] Additionally or alternatively, in one embodiment subsequently and / or in the same way as the discriminator, at least one other artificial neural network and / or another machine learning algorithm can be trained or machine-learned on the basis of at least a portion of the provided time series assessed as undisturbed, the generated time series with artificial disturbances and / or the artificial disturbances and, in one embodiment subsequently, used or provided in or for, in one embodiment as, disturbance detection.

[0017] Accordingly, in one embodiment, the machine-learned disturbance detection comprises the discriminator of the GAN, the other artificial neural network, and / or the other machine learning algorithm.

[0018] The machine learning or training, in particular of the discriminator, based on at least a portion of the provided time series assessed as undisturbed and on the basis of the generated time series with artificial disturbances comprises, in one embodiment, feeding or inputting this portion of the provided time series assessed as undisturbed and these generated time series with artificial disturbances, in particular to or into the discriminator. The training, in particular of the discriminator, based on the artificial disturbances comprises, in one embodiment, corresponding labeling with the artificial disturbances.

[0019] In one version, • provided time series assessed as undisturbed, on the basis of which time series of data with artificial disturbances are generated by the generator; and • provided time series assessed as undisturbed, on the basis of which the disturbance detection or discriminator is trained; at least partially match. In other words, provided time series assessed as undisturbed can be used both to generate time series of data with artificial disturbances using the generator and for training, in particular the discriminator.

[0020] In one embodiment, this can advantageously increase the amount of data available for training.

[0021] Additionally or alternatively, in one version • a part of the provided time series assessed as undisturbed is used only to generate time series of data with artificial disturbances using the generator, i.e. not (also) to train the disturbance detection and / or the discriminator; and / or • a part of the provided time series assessed as undisturbed is used only for training, in particular of the discriminator, ie not (also) for generating time series of data with artificial disturbances using the generator.

[0022] This can improve training in one version.

[0023] The GAN (“Generative Adversarial Network”) in one embodiment comprises a generator comprising at least one artificial neural network and a discriminator comprising at least one artificial neural network, which in one embodiment perform a zero-sum game.

[0024] By training the generator and discriminator, both become successively better and the discriminator learns to detect (increasingly better) the time points and / or extents of disturbances based on the artificially generated time series or artificial disturbances.

[0025] Robot assemblies often exhibit few faults—which are assessed as such. This makes detection using machine-learned fault detection systems, in particular artificial neural networks, or training or machine learning of fault detection systems, in particular artificial neural networks, problematic. By training the fault detection system, in particular the discriminator, based on artificially generated time series with (artificial) faults, fault detection systems, in particular artificial neural networks, can be advantageously trained or used to detect faults in robot assemblies.

[0026] By training the disturbance detection or the discriminator on the basis of the artificial disturbances in one embodiment of their (known or predetermined by the generator) times and / or extents, disturbance detection, in particular artificial neural networks, can advantageously detect disturbances in robot arrangements, in particular detect their times and / or extents.

[0027] This allows for a particularly advantageous response to the detection in one embodiment, in particular for better diagnosis and / or prediction of potential errors or error causes, particularly in monitoring and / or predictive maintenance. Accordingly, by training, in particular the (generator and discriminator of the) GAN, in one embodiment, a particularly advantageous fault detection or a corresponding artificial neural network can be provided. This allows faults in robot assemblies to be particularly advantageously detected, thus improving the operation of the robot assemblies, in particular increasing safety and / or reducing the error rate and / or downtime.

[0028] In one embodiment, the artificial disturbances (from the generator) are generated based on noise and / or a predefined scaling, in one embodiment, a one- or multi-dimensional scaling parameter. Additionally or alternatively, in one embodiment, the artificial disturbances (from the generator) are added to provided time series that are considered undisturbed.

[0029] As a result, the generator in one embodiment can, in particular in combination, particularly advantageously generate time series with artificial disturbances, and in one embodiment can vary these particularly advantageously, in particular in order to deceive the discriminator.

[0030] In one embodiment, the discriminator, the other artificial neural network and / or the other machine learning algorithm of the disturbance detection is / are trained, in particular additionally or also, on the basis of stochastically generated time series which, in one embodiment, are independent of time series assessed as undisturbed, in particular, in addition to the provided time series assessed as undisturbed and the generated time series with artificial disturbances, stochastically generated time series which, in one embodiment, are independent of time series assessed as undisturbed, are also used to train the disturbance detection or the discriminator.

[0031] This can improve the robustness of a design.

[0032] In one embodiment, time series generated (by the generator) are stored with artificial disturbances and the discriminator, the other artificial neural network and / or the other machine learning algorithm of the disturbance detection are repeatedly trained with these stored time series.

[0033] In one embodiment, the process of generating the disturbances, in particular the generation of time series of data with artificial disturbances based on at least a portion of the provided time series assessed as undisturbed by means of the generator, and the machine learning of the disturbance detection or anomaly detection, in particular the training of the discriminator, other artificial neural network, or other machine learning algorithm, in particular based on the generated time series with artificial disturbances, the artificial disturbances, and at least a portion of the provided, in particular also stored, time series assessed as undisturbed, are decoupled. In one embodiment, this can improve the training, the disturbance detection, the trained artificial neural network, and / or the detection of disturbances.

[0034] According to a second aspect of the present invention, a method for machine learning of fault detection, in an embodiment for training one or more artificial neural networks for detecting faults in a robot arrangement comprising at least one robot, comprises the step: - Providing data of a first robot arrangement comprising at least one robot.

[0035] In one embodiment, the machine learning comprises one or more distance- and / or clustering-based machine learning methods and / or the use or non-use of at least one artificial neural network, for example random forests methods, Arma / Arima models-based methods, gradient boosting methods or the like, in particular in addition to or alternatively to the at least one artificial neural network.

[0036] As already explained with regard to the first aspect, in one embodiment, this provision may include, in particular, capturing, in particular measuring, and / or filtering data from the first robot arrangement. Additionally or alternatively, this provision may also include, in particular, retrieving stored, in particular measured and / or filtered, data from the first robot arrangement.

[0037] According to an embodiment of the second aspect, the provided data comprises data records of data that are assigned to one another in time, in one embodiment, wherein each of these data records comprises a date from a plurality of the different data channels.

[0038] In one embodiment, each of these data channels has at least, in particular only, one time series of data or values, in particular of a data channel-specific parameter of the first robot arrangement.

[0039] In one embodiment, a first data channel (a time series) comprises first data, in particular kinematic data, in particular position, speed or acceleration data, force or torque data, electrical data, in particular current or voltage data, error messages from a robot controller or following error data of an axis or a drive of the robot.

[0040] In a further development, one or more further data channels each have (a time series) of this first data, in particular temporally associated, further data, in one embodiment kinematic data, in particular position, speed or acceleration data, force or torque data, electrical data, in particular current or voltage data, error messages from a robot controller or following error data of another axis or another drive of the robot.

[0041] Additionally or alternatively, in a further development, one or more further data channels each have (a time series) of this first data, in particular temporally associated, further data, in one embodiment kinematic data, in particular position, speed or acceleration data, force or torque data, electrical data, in particular current or voltage data, error messages from a robot controller or following error data of another axis or another drive of the robot.

[0042] Thus, each of the data sets in an execution has at least • a first datum, for example a position of an axis or a drive of the robot, at or for a specific point in time; and • a data item from another data channel, for example a torque of this axis or drive of the robot, at or for this specific point in time; and / or • a data item from another data channel, for example a position of another axis or another drive of the robot, at or for this specific point in time.

[0043] In other words, the provided data of the first robot arrangement in one embodiment comprise a plurality of time series of data for different parameters of the robot arrangement, each of the data channels in one embodiment comprising one of these time series and each of the data sets comprising the data of the different parameters of the robot arrangement assigned to the same time.

[0044] According to an embodiment of the second aspect, the method comprises the step: - machine learning, in one embodiment by means of distance- and / or clustering-based machine learning and / or with or without use of at least one artificial neural network, for example random forests, Arma / Arima models, gradient boosting or the like, a disturbance detection, in one embodiment training one or more artificial neural networks on the basis of this data, wherein a date of a data set is not finally or after the machine learning classified as a date of a disturbance if the data assigned to it is or is not classified as data of a disturbance in at least a predetermined number and / or not at least in a predetermined group of the data channels.

[0045] As already mentioned, in one embodiment, in addition to or instead of (training the) at least one artificial neural network, the disturbance detection can be distance-based or machine-learned.

[0046] This second aspect is based on the realization that in robots, different data channels are redundant in that disturbances occur simultaneously in (data from) several of the data channels.

[0047] According to an embodiment of the second aspect, this finding is utilized by requiring a certain minimum redundancy for classification as a fault, in that the data is or will be classified as fault data in at least a predetermined number of the data channels, in one embodiment regardless of which data channel(s) are involved. Thus, for example, it can be provided that a datum is only classified as a fault datum if the data associated with this datum is or will also be classified as a fault in at least one or at least two other data channels.

[0048] According to another embodiment of the second aspect, this finding is utilized by requiring a certain minimum redundancy for classification as a fault, insofar as this redundancy is or will be classified as fault data in at least one predefined group of (specific) data channels. Thus, for example, it can be provided that a date is only classified as a fault date if the data associated with this date is or will also be classified as a fault in at least one or two specific or predefined data channels.

[0049] The second aspect allows a particularly advantageous response to the detection in one embodiment, in particular to better diagnose and / or predict possible errors (causes), in particular in monitoring and / or predictive maintenance. Accordingly, machine learning based on a required minimum redundancy for classification as data of a fault in one embodiment can provide particularly advantageous fault detection, thereby detecting faults in robot arrangements in one embodiment and thus improving the operation of the robot arrangements, in particular increasing safety and / or reducing error rates and / or downtime. In one embodiment, fault detection can also be machine-learned in a particularly advantageous manner without external evaluation of the data or without (external) labeling of the data, or autonomously.

[0050] In one embodiment of the second aspect, in a first phase of machine learning, a channel interference detection is machine-learned for a plurality of data channels based on the data of the respective data channel, in one embodiment using distance- and / or clustering-based machine learning and / or with or without the use of at least one artificial neural network, for example, random forests, ARMA / ARIMA models, gradient boosting, or the like. In one embodiment, at least one artificial neural network is trained. The machine-learned interference detection according to the second aspect comprises, in one embodiment, one or more of these (individual or channel-specific) channel interference detections.

[0051] In one embodiment, this is done at least partially in parallel.

[0052] This can advantageously reduce the learning or training time in one implementation.

[0053] Additionally or alternatively, in one embodiment, in the first phase, an extreme anomaly score is automatically determined for several of the data channels, each based on the data of the respective data channel, for the date that is not yet classified as a fault date. Subsequently, a threshold value for classification as a fault date is modified, in particular increased or reduced, relative to this extreme anomaly score in such a way that at least one date that is not (yet) classified as a fault date based on the extreme anomaly score is now classified as a fault date. In this way, more false negatives are initially generated as a precautionary measure.

[0054] In a subsequent second phase of machine learning, a date in a data set that was initially classified as a date of a fault is / are no longer classified as a date of a fault or is / are again classified as an undisturbed date if the other data in its data set assigned to it were not classified as data of a fault in at least one or the specified number or group of data channels in the first phase, in particular the artificial neural networks are (further) trained accordingly.

[0055] Preferably, in the first phase, a low minimum value or high maximum value is used for classifying a date (still) as a date of a fault, in particular a minimum value that is lower than a minimum value used in the second phase, which must be undershot in order to be classified as a date of a fault, or a maximum value that is higher than a maximum value used in the second phase, which must be exceeded in order to be classified as a date of a fault.

[0056] As a result, more false negatives and fewer false positives are generated in the first phase, or (initially) data that is not actually data of a fault is more likely to be classified as data of a fault (false negatives) than, conversely, data that is actually data of a fault is not classified as data of a fault (false positives).

[0057] In one embodiment, this can advantageously increase the safety when detecting faults in robot arrangements.

[0058] In an embodiment of the second aspect, the number or group of data channels is specified based on a user input.

[0059] This allows the user's (expert) knowledge of the (required) redundancy to be advantageously utilized, thus improving machine learning. For example, based on their knowledge that a fault in one axis usually results in both kinematic and force / torque data exhibiting or indicating a fault, or that a certain type of fault is always reflected in the data of several, especially specific, axes, a user can specify corresponding groups or a corresponding number of data channels.

[0060] Additionally or alternatively, in one embodiment of the second aspect, the number or group of data channels is specified partially or fully automatically, and in one embodiment, it is varied. In one embodiment, a particularly favorable number or group of data channels can thus be determined, and in one embodiment, it can be machine-learned.

[0061] According to an embodiment of the first or second aspect of the present invention, a method for detecting faults in a robot arrangement comprising at least one robot comprises the steps of: - Providing a fault detection, in an embodiment of one or more artificial neural networks, using a method described here, in particular a fault detection machine-learned using a method described here; and - Detecting faults in the first and / or another robot arrangement comprising at least one robot, using this fault detection, in particular this or these provided artificial neural networks, in particular detecting faults in one or more robots of this (first or other) robot arrangement, in particular in time series of their data.

[0062] In one embodiment, this provision comprises providing, in particular using, the machine-learned disturbance detection or the (trained) artificial neural network(s), in one embodiment the (trained) discriminator of the GAN, other artificial neural network or other machine learning algorithm.

[0063] Likewise, the provision may also include, in particular, parameterizing a fault detection or one or more artificial neural networks based on the machine-learned fault detection or (trained) artificial neural network(s), in an embodiment of the (trained) discriminator of the GAN, other artificial neural network or other machine learning algorithm, and / or copying the machine-learned fault detection or (trained) artificial neural network(s), in an embodiment of the (trained) discriminator of the GAN, other artificial neural network or other machine learning algorithm.

[0064] In particular, in one embodiment, the discriminator can be trained on the basis of time series of data from the first robot arrangement that are assessed as undisturbed, and subsequently the discriminator, a copy of the (trained) discriminator, or an artificial neural network parameterized on the basis of the (trained) discriminator, can detect disturbances in this first robot arrangement, in particular while it is being further trained.

[0065] Likewise, in one embodiment, a fault detection can be or be machine-learned based on data of the first robot arrangement and subsequently this fault detection, a copy of the (machine-learned) fault detection or a fault detection parameterized based on the (machine-learned) fault detection, in particular while it is further trained, can detect faults in this first robot arrangement.

[0066] Through this specific training, faults in the first robot arrangement can be diagnosed and / or predicted particularly advantageously, in particular quickly, reliably and / or precisely.

[0067] Additionally or alternatively, in one embodiment, the discriminator can be trained on the basis of time series of data from the first robot arrangement that are assessed as undisturbed, and subsequently the discriminator, a copy of the (trained) discriminator, or an artificial neural network parameterized on the basis of the (trained) discriminator, can detect disturbances in another robot arrangement, in particular while it is being further trained.

[0068] Likewise, in one embodiment, a fault detection can be or be machine-learned based on data of the first robot arrangement and subsequently this fault detection, a copy of the (machine-learned) fault detection or a fault detection parameterized based on the (machine-learned) fault detection, in particular while it is further trained, can detect faults in another robot arrangement.

[0069] By means of this transmission, malfunctions of robot arrangements can be diagnosed and / or predicted in a particularly advantageous manner, in particular on the basis of small amounts of data from these robot arrangements and / or with less commissioning effort, in particular time.

[0070] In one embodiment of the first or second aspect, the data comprises kinematic data, in one embodiment position data, speed data and / or acceleration data, of the first robot arrangement, in one embodiment of one or more robots of the robot arrangement, in particular axis position, speed and / or acceleration data and / or Cartesian pose data of at least one robot-fixed reference, in particular of an end effector or TCP of the robot(s).

[0071] Additionally or alternatively, in one embodiment of the first or second aspect, the data comprise force data and / or torque data of the first robot arrangement, in one embodiment of one or more robots of the robot arrangement, in particular axis and / or drive forces and / or torques and / or external forces and / or torques exerted on the first robot arrangement, in one embodiment of one or more robots.

[0072] Additionally or alternatively, in an embodiment of the first or second aspect, the data comprise electrical data, in particular current and / or voltage data of the first robot arrangement, in an embodiment of one or more robots of the robot arrangement, in particular of drives of the first robot arrangement, in an embodiment of one or more robots.

[0073] Additionally or alternatively, in an embodiment of the first or second aspect, the data comprise error messages from a robot controller and / or following error data or deviations between target and actual values ​​of the first robot arrangement, in an embodiment of one or more robots of the robot arrangement, in particular from controllers of the first robot arrangement, in an embodiment of one or more robots.

[0074] Additionally or alternatively, in an embodiment of the first or second aspect, the data comprise temperature data of the first robot arrangement, in an embodiment of one or more robots, in a further development of robot arms, of the robot arrangement.

[0075] The or one or more robots of the first and / or other robot arrangement have, in one embodiment, a robot arm, in one embodiment multi-axis, and can in particular be such.

[0076] On the basis of such data, in an embodiment of the first or second aspect, malfunctions of a robot arrangement can be diagnosed and / or predicted particularly advantageously, in particular quickly, reliably and / or precisely.

[0077] In an embodiment of the first or second aspect, the data comprises data from two or more data channels.

[0078] This allows a larger amount of data to be used in one implementation, thereby improving training, the trained artificial neural network and / or the detection of disturbances.

[0079] According to one embodiment of the present invention, a system, in particular hardware and / or software, in particular program technology, is set up to carry out a method described here according to an embodiment of the first and / or second aspect.

[0080] According to one embodiment of the present invention, the system comprises: - means for providing time series of data, assessed as undisturbed, of a first robot arrangement comprising at least one robot; - Means for multiple alternating - generating time series of data with artificial disturbances based on at least a part of the provided time series assessed as undisturbed by means of a generator of a GAN; and - training a discriminator of the GAN on the basis of the generated time series with artificial disturbances, the artificial disturbances and at least a part of the provided time series assessed as undisturbed; where - the discriminator is trained to detect times and / or magnitudes of disturbances; and - the generator is trained to generate time series with artificial disturbances whose extent is underestimated by the discriminator, in particular which the discriminator does not detect or estimates to be less than the artificial disturbances in the generated time series.

[0081] In one embodiment, the system or its means comprises: - means for generating the artificial disturbances based on noise and / or a predetermined scaling and / or for adding the artificial disturbances to provided time series assessed as undisturbed; and / or - means for storing generated time series with artificial disturbances and repeatedly training the discriminator, another artificial neural network and / or another machine learning algorithm for disturbance detection with these stored time series; and / or - Means for training the discriminator, another artificial neural network and / or another machine learning algorithm for disturbance detection based on stochastically generated time series.

[0082] According to one embodiment of the present invention, the system comprises: - means for providing data of a first robot arrangement comprising at least one robot, wherein said data comprises data sets associated with one another, in particular chronologically, wherein each of said data sets comprises a respective datum from a plurality of different data channels; and - Means for machine learning, in one embodiment by means of distance- and / or clustering-based machine learning and / or with or without use of at least one artificial neural network, for example random forests, Arma / Arima models, gradient boosting or the like, a disturbance detection, in particular training of at least one artificial neural network, on the basis of the data, wherein a date of a data set is ultimately not classified as a date of a disturbance if the data assigned to it is not or will not be classified as data of a disturbance in at least a predetermined number and / or group of the data channels.

[0083] In one embodiment, the system or its means comprises: - Means for, in particular at least partially parallelized, machine learning, in particular by means of distance- and / or clustering-based machine learning and / or with or without the use of at least one artificial neural network, for example Random Forests, Arma / Arima Models, Gradient Boosting or the like, in each case a channel interference detection based on the data of the respective data channel in a first phase of machine learning for several of the data channels, in particular for training at least one artificial neural network in the first phase; and - means for classifying a date of a data set initially classified as a date of a fault as no longer a date of a fault if the data assigned to it have not been classified as data of a fault in at least a predetermined number and / or group of the data channels in the first phase, in a subsequent second phase of machine learning, in particular for the corresponding further training of the artificial neural networks.

[0084] In one embodiment, the system or its means comprises: Means for specifying the number or group of data channels on the basis of a user input and / or at least partially automated, in particular for at least partially automated variation.

[0085] According to one embodiment of the present invention, the system comprises: - means for providing fault detection, in one embodiment of at least one artificial neural network, using a method or means described herein; and - means for detecting faults in the first and / or another robot arrangement comprising at least one robot, with the aid of said fault detection, in particular said at least one artificial neural network.

[0086] A means within the meaning of the present invention can be designed in hardware and / or software, in particular a processing unit, in particular a microprocessor unit (CPU), graphics card (GPU) or the like, which is preferably connected to a memory and / or bus system in terms of data or signals, and / or can have one or more programs or program modules. The processing unit can be designed to execute instructions implemented as a program stored in a memory system, to detect input signals from a data bus and / or to output signals to a data bus. A memory system can have one or more, in particular different, storage media, in particular optical, magnetic, solid-state and / or other non-volatile media. The program can be designed in such a way that it embodies the methods described here oris capable of executing such methods, so that the processing unit can execute the steps of such methods and thus, in particular, detect faults in robot arrangements having at least one robot with the aid of fault detection, in particular at least one artificial neural network, or can machine-learn this fault detection or train this artificial neural network. In one embodiment, a computer program product can have, in particular be, a storage medium, in particular a non-volatile one, for storing a program or with a program stored thereon, wherein executing this program causes a system or a controller, in particular a computer, to execute a method described here or one or more of its steps.

[0087] In one embodiment, one or more, in particular all, steps of the method are carried out fully or partially automatically, in particular by the system or its means. In one embodiment, the system comprises the first and / or other robot arrangement.

[0088] Further advantages and features emerge from the subclaims and the exemplary embodiments. The following shows, partly schematically: Fig. 1: a system according to an embodiment of the present invention; Fig. 2: a generator of the system or GANs of the Fig. 1; Fig. 3: a time series of data; Fig. 4: a method according to an embodiment of the present invention; Fig. 5: a system according to a further embodiment of the present invention; Fig. 6: a method according to a further embodiment of the present invention; and Fig. 7: a classification of data in two data channels in a first and second phase of a machine learning method.

[0089] Fig. 1 shows a system according to an embodiment of the present invention.

[0090] The system comprises a robot (arm) 100 and a robot controller 110, which provides undisturbed evaluated time series 1 of data of the robot 100 ( Fig. 4: Step S10).

[0091] A generator 10 of a GAN generates time series 3 of data with artificial disturbances ( Fig. 4: Step S20).

[0092] For this purpose, values ​​of the scaling parameter α are generated stochastically and input together with the noise 2 and the time series 1 into an artificial neural network with several layers 11.

[0093] This has a softmax layer 12 and a sign layer 13, each containing the size and length of the time series, with the sign layer 13 outputting +1 or -1 at each position.

[0094] In a final layer 14, the (outputs of) the softmax layer 12 and the sign layer 13 are multiplied component-wise or per time point, and the result is then multiplied (component-wise) by the scaling parameter α.

[0095] The artificial disturbance thus generated is added to the respective time series provided, which is considered to be undisturbed.

[0096] The Softmax layer 12 ensures that the disturbance is never completely equal to 0, and the Sign layer 13 ensures that in each component or each time point disturbance can occur in both positive and negative directions.

[0097] The time series 3 with artificial disturbances thus generated by generator 10, together with the artificial disturbances themselves generated by generator 10, in particular the times and associated extents of the disturbances, are made available to a discriminator 20 of the GAN, which also receives other time series 1' of data from robot 100 that are evaluated as undisturbed and attempts to (correctly) detect the times and extents of the disturbances. The time series generated by generator 10 are labeled with the underlying times and associated extents of the artificial disturbances ( Fig. 4: Step S30).

[0098] In a manner known per se, generator 10 and discriminator 20 of the GAN are now trained in that generator 10 attempts to generate time series with artificial disturbances in such a way that discriminator 20 underestimates their extent, in particular those that discriminator 20 does not detect or estimates to be less than the artificial disturbances generated by generator 10, and conversely, discriminator 20 attempts to (correctly) detect the times and extents of the disturbances.

[0099] With the discriminator 20 trained in this way or a correspondingly parameterized artificial neural network, disturbances of the robot 100 or of another robot can then be detected in a step S40.

[0100] The discriminator 20 can determine the extent and timing of the disturbances using the same artificial neural network or can have two separate artificial neural networks for this purpose.

[0101] Fig. Figure 3 shows an example of a time series of data d of the robot 100, for example a torque lag error or the like, over time t, which is evaluated as undisturbed (thin in Fig. 3) and a time series generated by the generator 10 with artificial disturbances, which are Fig. 3 are highlighted in bold and have, for example, the magnitude A(t1) at time t1.

[0102] Fig. 5 shows a system according to another embodiment of the present invention.

[0103] The system comprises a robot (arm) 100 and a robot controller 110, which provides data of the robot 100 in data channels, of which Fig. 5 two data channels 1000, 2000 are indicated as examples ( Fig. 6: Step S100).

[0104] In each case, a date from one data channel 1000 and a date from the other data channel 2000 are temporally assigned to each other. This is indicated by different shapes for a date 1.1 and a date 2.1 of one data set, a date 1.2 and a date 2.2 of another data set, and a date 1.3 and a date 2.3 of another data set (data set (1.1, 2.1): square; data set (1.2, 2.2): triangle; data set (1.3, 2.3): cross).

[0105] For example, the data 1.1, 1.2 and 1.3 can be position, speed or acceleration data of an axis of the robot 100 at or for specific (measurement) time(s) and the data 2.1, 2.2 and 2.3 can be force or torque data of the robot 100, in particular of this axis, at or for these (measurement) time(s), or position, speed or acceleration data of another axis of the robot 100 at or for these (measurement) time(s), wherein further corresponding data at or for other (measurement) time(s) in Fig. 5 are indicated by circles.

[0106] In a first phase ( Fig. 6: Step S200), a channel interference detection is machine-learned for each data channel, in particular at least partially parallelized, based on the data of the respective data channel, in the exemplary embodiment an artificial neural network is trained. Examples of this are shown in Fig. 5 two artificial neural networks 30, 40 are indicated.

[0107] In the first phase, the classification is learned in such a way that data that is not a fault data (false positives) is more likely to be classified as fault data than data that is a fault data (false negatives).

[0108] This is in the sequence of figures Fig. 7(a) → Fig. 7(b) is schematically indicated, which shows how the classification area (hatched in Fig. 7) for data that are not classified as data of a disturbance, is first reduced.

[0109] In this way, (at least) (most) of the data that are actually data of a fault (in Fig. 5, Fig. 7, for example, the data 1.3, 2.3 indicated by a filled cross are classified as data of a disturbance, but also (relatively many) data that are not data of a disturbance (in Fig. 5, Fig. 7 shows, as an example, the data 1.1, 2.2 of the two data channels 1000, 2000 - false negatives, indicated by an empty square or triangle.

[0110] In a second phase ( Fig. 6: Step S300), for the data classified as a fault date in the first phase, it is checked whether, in a predetermined number and / or group of the other data channels, the data assigned to the respective date classified as a fault date were also classified as a fault data in the first phase.

[0111] This is in Fig. Figure 7(b) illustrates that the temporally related data 1.3 and 2.3 are both classified as data from a fault. Accordingly, they continue to be classified as data from a fault in the second phase.

[0112] However, for the two dates 1.1 and 2.1, which are temporally associated with each other, and for the two dates 1.2 and 2.2, which are temporally associated with each other, only one of the two dates is classified as a date of a fault, the other is not.

[0113] Accordingly, these dates 1.1, 2.2 are no longer classified as dates of a disruption in the second phase.

[0114] This is in the Fig. 7(c) is schematically indicated, which shows how the classification area is modified accordingly.

[0115] With the fault detection learned in this way, in the exemplary embodiment the two artificial neural networks 30, 40 trained in this way, faults of the robot 100 or of another robot can then be detected in a step S400.

[0116] Although exemplary embodiments have been explained in the preceding description, it should be noted that a large number of modifications are possible.

[0117] Artificial neural networks were used above. Additionally or alternatively, other machine learning methods or algorithms can be used, including, in one implementation, random forests, ARMA / ARIMA models, and / or gradient boosting. Distance- and / or clustering-based machine learning is particularly suitable.

[0118] Furthermore, it should be noted that the exemplary embodiments are merely examples and are not intended to limit the scope of protection, applications, or structure in any way. Rather, the preceding description provides the skilled person with a guide for implementing at least one exemplary embodiment, whereby various modifications, particularly with regard to the function and arrangement of the described components, may be made without departing from the scope of protection as defined by the claims. List of reference symbols 1, 1' time series evaluated as undisturbed 2 noise 3 time series with artificial disturbances 10 Generator 11 shift 12 SoftMax layer 13 Sign layer 14 final layer 20 Discriminator 30, 40 artificial neural network 100 robots (arms) 110 Robot control 1000, 2000 data channels (1.1, 2.1) Data set (1.2, 2.2) Data set (1.3, 2.3) Data set A extent d data t time t1 time point α scaling parameter

Claims

[1] Method for machine learning of fault detection, for detecting faults in a robot arrangement comprising at least one robot (100), comprising the step: - Providing (S10) time series (1, 1') of data of a first robot arrangement, which has at least one robot (100), which are assessed as undisturbed; and the multiply alternating steps: - generating (S20) time series (3) of data with artificial disturbances based on at least a part (1) of the provided time series assessed as undisturbed by means of a generator (10) of a GAN; and - training (S30) a discriminator (20) of the GAN on the basis of the generated time series (3) with artificial disturbances, the artificial disturbances and at least a part (1') of the provided time series evaluated as undisturbed; wherein - the discriminator (20) is trained to detect times (t1) and / or extents (A(t1)) of disturbances; and - the generator (10) is trained to generate time series (3) with artificial disturbances, the extent of which the discriminator (20) underestimates. [2] Method according to claim 1, characterized by that the artificial disturbances are generated on the basis of a noise (2) and / or a predetermined scaling (α) and / or are added to provided time series (1) which are assessed as undisturbed. [3] Method according to one of the preceding claims, characterized by that generated time series are stored with artificial disturbances and the discriminator, another artificial neural network and / or another machine learning algorithm for disturbance detection is repeatedly trained with these stored time series. [4] Method according to one of the preceding claims, characterized bythat the discriminator, another artificial neural network and / or another machine learning algorithm for disturbance detection is trained on the basis of stochastically generated time series. [5] Method for machine learning of fault detection, for detecting faults in a robot arrangement comprising at least one robot (100), comprising the steps: - Providing (S100) data of a first robot arrangement having at least one robot (100), wherein said data comprises data sets ((1.1, 2.1), (1.2, 2.2), (1.3, 2.3)) of mutually associated data (1.1, 2.1, 1.2, 2.2, 1.3, 2.3), wherein each of said data sets comprises a respective datum from a plurality of different data channels (1000, 2000); and - machine learning (S200) of a fault detection based on the data, whereby a date of a data set is finally not classified as a date of a fault if the data assigned to it are not or will not be classified as data of a fault in at least a predetermined number and / or group of the data channels. [6] Method according to the preceding claim, characterized by , that - in a first phase (S200) of machine learning, a channel fault detection is machine-learned for several of the data channels on the basis of the data of the respective data channel; and - in a subsequent second phase (S300) of machine learning, a date of a data set initially classified as a date of a fault is no longer classified as a date of a fault if the data assigned to it were not classified as data of a fault in at least a predetermined number and / or group of data channels in the first phase. [7] Method according to one of the two preceding claims, characterized by that the number or group of data channels is specified on the basis of user input and / or at least partially automatically. [8] Method according to one of the preceding claims, characterized by that the data - kinematic data; - force and / or torque data; - electrical data; - Error messages from a robot controller and / or tracking error data; and / or - Temperature data of the first robot arrangement and / or - Data from at least two data channels and / or - contain data from at least two robots in the robot arrangement. [9] Method for detecting faults in a robot arrangement comprising at least one robot (100), comprising the steps of: - Providing (S10-S30; S100-S300) a fault detection using a method according to one of the preceding claims; and - detecting (S40; S400) faults in the first and / or another robot arrangement comprising at least one robot (100) using said fault detection. [10] System arranged to carry out a method according to any one of the preceding claims. [11] A computer program product comprising a program code stored on a computer-readable medium for carrying out a method according to any one of claims 1 to 9.

Images