Method for generating and training an anomaly detection system for detecting abnormal vibrations in a robot arm based on sensor data and a robot arm

Abstract

The invention relates to a method for generating and training an anomaly detection system for detecting abnormal vibrations in a robot arm based on sensor data, wherein the robot arm has several robot arm joints, in particular six, each with at least one fiber optic sensor, wherein the method comprises at least the following method steps: - In a first method step, sensor data from the operation of the robot arm are obtained using the fiber optic sensors and at least one training data set and one validation data set are determined from the sensor data for each of the several robot arm joints, - In a second method step following the first method step, at least one classifier for detecting abnormal vibrations is determined for each of the plurality of robot arm joints, wherein for the determination, a plurality of classifier candidates are provided for each classifier and are trained on the basis of the training data set and checked using the validation data set, wherein the plurality of classifier candidates provided for the same robot arm joint are compared with one another on the basis of at least one performance metric and the classifier candidate with the highest performance metric is selected for each robot joint.

Description

The invention relates to a method for generating and training an abnormality detection system for detecting abnormal vibrations in a robot arm on the basis of sensor data. The invention further relates to a robot arm having a plurality of robot arm joints and a computing unit which is configured to carry out such a method.BACKGROUND OF THE INVENTIONIn the following, the prior art relevant to the invention will first be outlined with reference to the following listed sources.[1] K. R. d. S. Santos, E. Villani, W. R. de Oliveira, A. Dttman, Comparison of visual servoing technologies for robotized aerospace structural assembly and inspection, Robotics and Computer-Integrated Manufacturing, 73 (2022) 102237.[2] A. Jokinb, M. Petrolovinb, Z. Miljkovid, Semantic segmentation based stereo visual servoing of nonholonomic mobile robot in intelligent manufacturing environment, Expert Systems with Applications, 190 (2022) 116203.[3] Jiang, T., Cui, H., Cheng, X., A calibration strategy for vision-guided robot assembly system of large cabin, Measurement, 163 (2020) 107991.[4] X. Pu, H. Guo, Q. Tang, J. Chen, L. Feng, G. Liu, X. Wang, Y. Xi, C. Hu, Z. L. Wang, Rotation sensing and gesture control of a robot joint via triboelectrolytic quantization sensor, Nano Energy, 54 (2018) 453-460.[5] Kim, Y., Park, J., Na, K., Yuan, H., Youin, B.D., C.-s. Kang, Phase-based time domain averaging (PTDA) for fault detection of a gearbox in an industrial robot using vibration signals, Mechanical Systems and Signal Processing, 138 (2020) 106544.[6] S. Wang, X. Wei, Y. Zhao, Z. Jiang, Y. Shen, A MEMS resonant accelerometer for low-frequency vibration detection, Sensors and Actuators A: Physical, 283 (2018) 151-158.[7] C. a. Zhou, K. Guo, J. Sun, An integrated wireless vibration sensing tool holder for milling tool condition monitoring with singularity analysis, Measurement, 174 (2021) 109038.[8] B. Cloostermans, D. Pronk, B. Bruckenburg, T. Gernaert, Spiral wound rockets with fiber Bragg grating sensors, Mechanical Systems and Signal Processing, 181 (2022) 109475.[9] Z. Liang, D. Liu, X. Wang, J. Zhang, H. Wu, X. Qing, Y. Wang, FBG-based strain monitoring and temperature compensation for composite tank, Aerospace Science and Technology, 127 (2022) 107724.

[10] C. Li, L. Yang, C. Luo, H. Liu, X. Wan, Frost Having Strain Monitoring for Training Structure in Extreme Cold and High-Old Area with FBG Strain Sensors, Measurement, 196 (2022) 110918.

[11] L. Chen, J. Cao, K. Wu, Z. Zhang, Application of Generalized Frequency Response Functions and Improved Convolutional Neural Network to Fault Diagnosis of Heavy-duty Industrial Robot, Robotics and Computer-Integrated Manufacturing, 73 (2022) 102228.

[12] J. Long, Y. Qin, Z. Yang, Y. Huang, C. Li, Distinctive feature learning using a multiscale convolutional capsule network from attitude data for fault diagnosis of industrial robots, Mechanical Systems and Signal Processing, 182 (2023) 109569.

[13] Z. Pu, D. Cabera, Y. Bai, C. Li, Generative adventitial one-shot diagnosis of transmission results for industrial robots, Robotics and Computer-Integrated Manufacturing, 83 (2023) 102577.

[14] Chen, T., Liu, X., Xia, B., Wang, W., Lai, Y., Unupuputive Abnormality Detection of Industrial Robots Using Sliding-Window Convolutional Variational Autoencoder, IEEE Access, 8 (2020) 47072-47081.

[15] Z. Zhong, Y. Zhao, A. Yang, H. Zhang, D. Qiao, Z. Zhang, Industrial Robot Vibration Abnormality Detection Based on Sliding Window One-Dimensional Convolution Autoencoder, Shock and Vibration, 2022 (2022) 1179192.

[16] H. Yun, H. Kim, Y. H. Jeong, M. B. G. Jun, Autoencoder-based abnormality detection of industrial robot arm using support atmosphere based internal sound sensor, Journal of Intelligent Manufacturing, 34 (2023) 1427-1444.Industrial robots are increasingly being used as a more efficient and safer alternative to manual operations in various industrial applications, e.g. in automatic manufacture and in aerospace, cf. [1-4]. Robot joints which serve to support adjacent links during different movements are important movable components of industrial robots. Since most industrial robots are generally used in production lines to increase working efficiency, failure of a robot joint due to adverse working conditions, overloads and unexpected events could lead to expensive unscheduled downtime and large economic losses, cf. [5]. Therefore, it is necessary to develop state monitoring systems (CM) for robot joints.In state monitoring systems, the vibration signal is one of the typical signals that may reflect the fault characteristics of robot joints. Various vibration sensors, including microelectromechanical systems [6] and piezoelectric accelerometers [7], are already used for monitoring the state of industrial robots. These sensors have the disadvantage, for example, of their size, the complexity of the wiring and the susceptibility to electromagnetic interference. In contrast, fiber optic sensors such as the Fiber Bragg Grating (FBG) sensor are significantly more suitable for CM applications in robotic joints. This is because FBG sensors are easy, compact, flexible, embeddable, multiplexable, highly sensitive, corrosion resistant, and easy to connect to a sensor network. In particular, the FBG sensors have the advantage that they are insensitive to electromagnetic influences [8-10].The focus of research is primarily on the fault diagnosis of industrial robots [11-13]. Due to considerable technological advances, the quality of the machine components installed in industrial robots has significantly improved, which has led to a considerable reduction in robot failures. As a result, robot failures are more and less likely to occur, making the detection and characterization of fault data for industrial robots challenging. At the same time, anomalies in industrial robots can be detected at an early stage, so that longer downtime of the industrial robot can be avoided.Chen et al.

[14] provide an unsupervised anomaly detection method using a combination of a convolutional neural network (CNN) and a variation autoencoder (VAE). This method relies on learning the normal pattern from only normal time-series data. Zhong et al.

[15] are concerned with abnormality detection in industrial robots by developing a 1D convolutional autoenconder (CAE), integrating CNN and autoencoder (AE), and incorporating a sliding window algorithm for data expansion. Yun et al.

[16] propose an anomaly detection system that uses time-frequency characteristics from the Short Time Fourier Transform Spectrogram (STFT) of sounds picked up with stereoscopics as input to an AE model. The aforementioned deep learning architectures have been developed manually by experts for the detection of anomalies, wherein the trial and error process is time-consuming and the developed deep learning architecture may not be optimally adapted to the respective application. Therefore, it is necessary to investigate how the best deep learning architectures for detecting anomalies at robot joints can be found automatically.Disclosure of the InventionAs explained above, the models for detecting anomalies in industrial robots are based either on models tailored manually by experts to detect anomalies or on a model which is used at the same time for a plurality of applications. In particular in industrial robots with a plurality of robot joints, these two approaches do not represent a satisfactory solution. The model developed by experts is approximated to a sufficiently good model in a time-consuming manner by trial and error methods. However, in a plurality of robot joints, each of which requires its own model for detecting anomalies, this represents a high time and cost requirement. Alternatively, a developed model may be used for the plurality of robot joints, the model not being tailored to the respective different robot joints and having accordingly lower abnormality detection accuracy.Against this background, the object is to provide an abnormality detection system for a robot arm having a plurality of robot joints in a time-efficient and cost-effective manner, which abnormality detection system also has high accuracy in detecting abnormalities.To achieve the object, a method for generating and training an abnormality detection system for detecting abnormal vibrations in a robot arm on the basis of sensor data is provided, wherein the robot arm has a plurality of robot arm joints, in particular six, each having at least one fiber-optic sensor, wherein the method comprises at least the following method steps:in a first method step, sensor data from the operation of the robot arm are obtained by means of the fiber-optic sensors and at least one training data set and one validation data set are determined from the sensor data for each of the plurality of robot arm joints,In a second method step following the first method step, at least one classifier for detecting abnormal vibrations is determined for each of the plurality of robot arm joints, wherein a plurality of classifier candidates are provided for each classifier for the determination and are trained on the basis of the training data set and checked by means of the validation data set, wherein the plurality of classifier candidates provided for the same robot arm joint are compared with one another on the basis of at least one performance metric and that classifier candidate which has the highest performance metric is selected for each robot joint.The method according to the invention enables the generation and training of an anomaly detection system which can detect abnormal vibrations, i.e. anomalies, during the operation of a robot arm. According to the invention, the robot arm has a plurality of robot arm joints, each of which has at least one fiber-optic sensor. In a first method step, the fiber-optic sensors provide sensor data from the operation of the robot arm. The sensor data are preferably robot arm joint specific, wherein the sensor data measured by a fiber optic sensor in a specific robot arm joint preferably only serve for the generation and training of an anomaly detection system for the respective specific robot arm joint. In the first method step, according to the invention, at least one training data set and one validation data set are determined from the sensor data of the respective fiber-optic sensor of a robot arm joint. In a second method step following the first method step, at least one classifier for detecting abnormal vibrations, i.e. anomalies, is determined for each of the plurality of robot arm joints. To ascertain a classifier, a plurality of candidate classifiers are provided for each classifier. The classifier candidates are trained using the corresponding training data record and then checked using the validation data record. To select the respective most suitable candidate classifier, the candidate classifiers are compared with one another on the basis of at least one performance metric, such that the respective most suitable candidate classifier can be selected for each robot joint. The most suitable classifier candidate is the candidate with the respective highest performance metric. The method according to the invention can thus provide a suitable classifier for each of the plurality of robot arm joints, wherein complex trial and error methods are advantageously avoided when manually generating such classifiers. Furthermore, the accuracy of the detection of anomalies can be increased since a classifier is determined and trained for each robot arm joint, which classifier has a good performance of anomaly detection for the fiber-optic sensor of the respective robot arm joint.The classifier candidates provided in the second method step are preferably generated randomly or by means of heuristic methods or meta-heuristic methods. In this case, the number of classifier candidates provided can be a compromise between the required outlay and the positive yield therefrom. A larger number of provided classifier candidates may speed up the ascertainment of a suitable classifier candidate as a classifier.In one advantageous embodiment, a performance metric can be a fitness of the respective classifier candidate. Quantities which can describe the performance metric or fitness of a candidate classifier are, inter alia, the cross entropy error (cross entropy loss), the mean squared error (mean squared error) or the accuracy (accuracy). Advantageously, the classifier candidates are comparable to one another by the performance metrics, wherein the performance metrics contain a statement as to how accurate the predictions of the respective classifier candidates or how large the error measure of the respective classifier candidate is.According to an advantageous embodiment of the invention, it is provided that the second method step is run through repeatedly until an abort criterion is reached. An abort criterion can be a minimum value of the performance metric, or a maximum value of an error measure. Thus, a candidate classifier can be automatically provided as a classifier if it meets certain predefined requirements for the performance metric, i.e. the prediction accuracy, or for the error measure. If a candidate classifier meets the termination criterion, it is preferably defined as a classifier and can then be used for detecting abnormal vibrations for the respective robot arm joint.According to an advantageous embodiment of the invention, it is provided that, if the termination criterion is not reached in the second method step, that classifier candidate which has the highest performance metric is selected and a plurality of new classifier candidates are provided on the basis of the classifier candidate and the second method step is run through with these new classifier candidates. Preferably, a plurality of new classifier candidates are provided in each case during the repeated pass through the second method step. Particularly preferably, the plurality of new classifier candidates are generated randomly or by means of heuristic methods or metaheuristic methods. The plurality of new classifier candidates is preferably based on the selected classifier candidate of the previous pass of the second method step, which has the highest performance metric. Advantageously, the respective best classifier candidate of a robot arm joint is refined in this way and, in addition, an automatic approach to a classifier candidate which meets the termination criterion.According to an advantageous embodiment of the invention, it is provided that the fiber-optic sensors each comprise an optical waveguide which has a fiber Bragg grating. Preferably, the plurality of fiber-optic sensors are connected to one another and to a receiver unit by one or more optical waveguides. Light can be coupled into the optical waveguide, wherein the fiber Bragg grating reflects light. This reflected light can be detected, in particular by means of the receiver unit. Fiber-optic sensors advantageously require a low level of cabling complexity, it also being possible for the individual fiber-optic sensors to be connected by means of an optical waveguide. In particular, in the case of a plurality of robot arm joints, the positive effect of the lower cabling outlay is intensified.Preferably, the plurality of fiber optic sensors substantially simultaneously reflect a signal to the receiver unit. The signal may be dependent on a sampling frequency. In the first method step, the signal of a fiber-optic sensor can first be scaled by means of a normalization method. Subsequently, in the first method step, a time-frequency spectrogram can preferably be obtained on the basis of the signal by means of a time-frequency analysis. Advantageously, the signals can thus be brought into a form adapted for the candidate classifiers.In an advantageous embodiment of the invention, a robot arm or a robot arm joint has a reference sensor, wherein the reference sensor has a reference optical waveguide which has a reference fiber Bragg grating, wherein the reference sensor is arranged on the robot arm or robot arm joint in a vibration-decoupled manner. By means of the reference sensor with a reference optical waveguide and reference fiber Bragg grating, a temperature influence during the measurement can be compensated. Such temperature compensation is advisable since a fiber Bragg grating, in particular the Bragg wavelength, is dependent not only on vibrations but also on the temperature. In this context, it is advantageous if the reference sensor is coupled to the object by means of a thermally conductive material, for example a thermally conductive silicone.According to an advantageous embodiment of the invention, it is provided that the plurality of fiber-optic sensors have different center wavelengths. The different center wavelengths preferably result in the fiber-optic sensors reflecting different wavelengths of the radiated light. Thus, a signal of a fiber-optic sensor can be unambiguously assigned to a robot arm joint in which the respective sensor is arranged. Advantageously, when an abnormality occurs, the robot arm joint in which the abnormality occurs can be identified. This allows shorter downtime, since an operator of the robot arm can quickly control the respective robot arm joint without having previously controlled the other robot arm joints.According to an advantageous embodiment of the invention, it is provided that the classifier candidates each have an artificial neural network for feature extraction from the respective sensor data of the respective robot arm joint, in particular a convolutional neural network or convolutional neural network, wherein different classifier candidates preferably comprise different artificial neural networks, wherein the classifier candidates extract first features from the training data set and second features from the validation data set. Artificial neural networks, particularly a convolutional neural network (CNN), can extract and learn features from a dataset without manually extracting data or requiring human input. In the second method step, the artificial neural network of a classifier candidate can extract first features from the training data set and second features from the validation data set. The training data record preferably does not have any abnormal vibrations, wherein particularly preferably the validation data record comprises normal vibrations and abnormal vibrations. Thus, training can be carried out on the basis of the training data record and the progress can be checked by means of the validation data record. The first features and second features can also be understood as deep features, wherein these represent patterns in the data sets. The deep features are preferably automatically extracted from the training data set and / or validation data set. The classifier candidates preferably only extract features from the associated training data sets and validation data sets of the sensor data of the respective robot arm joint. For example, a candidate classifier for a classifier of a first robot arm joint does not extract features from the sensor data of a third robot arm joint. It can thus be ensured that the determined classifier is optimally adapted to the detection of abnormal vibrations of the corresponding robot arm joint.According to an advantageous embodiment of the invention, it is provided that the classifier candidates additionally carry out a K-nearest neighbor classification method, wherein the features extracted by the respective artificial neural network are supplied to the K-nearest neighbor classification method. The features extracted by the artificial neural network from the corresponding sensor data, or training data sets and validation data sets, are preferably supplied to a K-nearest neighbor classification method. The K nearest neighbor classification method may perform classification (e.g., normal or abnormal vibration) of the respective data depending on the extracted characteristics.The artificial neural networks of the plurality of candidate classifiers for a classifier for a robot arm joint preferably differ, wherein particularly preferably the K-nearest neighbor classification methods of the different candidate classifiers are the same. Furthermore, the K nearest neighbor classification method may be identical for the multiple new classifier candidates compared to the classifier candidates of the previous second method step. An improvement of the K-nearest neighbor classification method is not necessary, wherein preferably only the artificial neural network of a candidate classifier is improved.The K-nearest neighbor classification method can preferably be trained on the basis of the first features. Subsequently or alternatively, the K-nearest neighbor classification method can determine and evaluate the performance metric of the associated artificial neural network of the corresponding candidate classifier on the basis of the validation dataset or the second features. Thus, the respective artificial neural networks of the classifier candidates are advantageously comparable.According to an advantageous embodiment of the invention, it is provided that a classifier, in particular the K-nearest neighbor classification method, detects an abnormal vibration on the basis of the first and second features. If a classifier for one robot arm joint each was determined in the second method step according to the invention, an abnormal vibration can be detected or recognized by means of the classifier. For this purpose, the selected artificial neural network preferably has a performance metric which is higher than a predefined termination criterion. In particular, an anomaly can be established by means of the K-nearest neighbor classification method of a classifier to which the first features and second features are supplied by means of the artificial neural network of the respective classifier. For this purpose, the artificial neural network can extract the first features and second features from the sensor data of the respective fiber-optic sensor of the respective robot arm joint.According to an advantageous embodiment of the invention, it is provided that the robot arm is stopped when a classifier of the anomaly detection system detects an anomaly of a robot arm joint. Advantageously, more severe damage to the robot arm joint or the robot arm is avoided in this way. At the same time, the probability of longer standstill times of the robot arm can thus be reduced. The operator of the robot arm can then be able to ascertain the respective robot arm joint by means of the method according to the invention in that the anomaly has occurred and acts accordingly.A further subject matter of the invention is a robot arm having a plurality of robot arm joints, in particular six, and a computing unit, wherein a robot arm joint has in each case at least one fiber-optic sensor which comprises an optical waveguide which has a fiber Bragg grating, wherein the computing unit is configured to carry out a method according to one of the configurations described above.The robot arm is preferably a six-axis robot arm, which particularly preferably has six robot arm joints. Accordingly, the robot arm has at least six fiber-optic sensors, wherein one fiber-optic sensor is arranged in each robot arm joint.In the robot arm according to the invention, the same technical effects and advantages can be achieved as have already been described in connection with the method according to the invention.Further details and advantages of the invention will be explained below with reference to the exemplary embodiment shown in the drawings. Shown herein: FIG. 1 shows a first exemplary embodiment of a method according to the invention in a schematic flow diagram; and FIG. 2 shows a second exemplary embodiment of the method according to the invention in a schematic flow diagram.FIG. 1 shows a first exemplary embodiment of the method 1 according to the invention for generating and training an anomaly detection system for detecting abnormal vibrations in a robot arm on the basis of a schematic flow diagram. According to the invention, the detection of the abnormal vibrations takes place on the basis of sensor data which are measured at least by means of a fiber-optic sensor within a robot arm joint. The respective robot arm can have a plurality of robot arm segments which are connected to one another via a plurality of robot arm joints. According to the invention, these robot arm joints each have at least one fiber-optic sensor, wherein the fiber-optic sensor comprises in particular an optical waveguide which has a fiber Bragg grating. Advantageously, the plurality of fiber-optic sensors can be connected to one another via an optical waveguide, so that the cable outlay, in particular in the case of rotating components, is kept low. The one optical waveguide can be connected to a receiver unit, wherein this receiver unit receives all signals of the plurality of fiber-optic sensors. In order to be able to distinguish the signals of the respective sensors, the sensors have different center wavelengths. The center wavelength specifies which wavelength of the radiated light is reflected, wherein different center wavelengths advantageously result in the fact that the fiber-optic sensors or the robot arm joints can be differentiated or can be assigned exactly. The Bragg wavelength can be dependent on the center wavelength, as a result of which the Bragg wavelengths of the respective fiber-optic sensors can also differ. The fiber optic sensors with fiber Bragg gratings are light, compact, flexible, embeddable, multiplexable, highly sensitive, corrosion resistant and easy to connect to a sensor network and consequently are particularly suitable for use in a robot arm joint. It is conceivable that a robot arm with several robot arm joints is designed as a 6-axis robot arm and has at least six robot arm joints.FIG. 1 shows a first method step 2, a second method step 3 and a first query 4 a. The first method step 2 comprises the operation of the robot arm 21, a real-time vibration monitoring device 22, a sensor data processing device 23 and a generation of a training and validation data set 24. For each of the plurality of robotic arm joints, the sensor data is processed 23. Preferably, the fiber optic sensors provide an output signal that is fed back to a receiver unit. The output signal can be dependent on a sampling frequency defined by the receiver unit. The signal processing 23 can comprise two method steps. First, a normalization process can be performed in the signal processing 23, wherein the output signal of the respective sensor is scaled. Additionally or alternatively, a time-frequency analysis can be carried out, in particular according to the normalization method, wherein a time-frequency spectrogram for the respective output signal is obtained from the output signal by means of the time-frequency analysis. Furthermore, in the next step, at least one training data set and one validation data set is generated 24 from the sensor data, wherein at least one training data set and one validation data set is present for a robot arm joint.In the second method step 3, the abnormality detection or the detection of abnormal vibrations takes place. For this purpose, according to the invention, at least one classifier for detecting abnormal vibrations is determined for each robot arm joint. The ascertainment of a classifier is made possible via a plurality of candidate classifiers. The classifier candidates are preferably generated randomly or by means of heuristic methods or meta-heuristic methods. The classifier candidates are trained on the basis of the training data sets generated in the first method step 2 and checked by means of the validation data sets. The plurality of candidate classifiers are compared with one another using at least one performance metric, and that candidate classifier having the highest performance metric is selected for each robot arm joint. For example, the performance metric may be a cross entropy error, mean square error, or accuracy. Within the scope of this invention, the statement of highest performance metrics is understood such that, depending on the selection of the performance metric, the best result is included among the classifier candidates. For example, in the case of the cross entropy error or the mean square error, the classifier candidate which has the lowest error is selected. Here, the least error would correspond to the highest performance metric.If a sufficient classifier has been determined in the second method step 3, it is determined in a first query 4 a whether an abnormality or abnormal vibration is present. In the presence of an abnormal vibration j, a warning 41 is issued and the robot arm is stopped by means of an emergency stop 42. This allows the operator of the robot arm to check 43 the robot arm or the faulty robot arm joint with the abnormal vibration. Advantageously, the fault can be localized as quickly as possible and long service lives can be avoided. If no abnormal vibration is detected in the first query 4 a, the path n is traversed and the method 1 starts from the beginning.FIG. 2 shows a second exemplary embodiment of method 1 according to the invention in a schematic flow diagram. First, the first method step 2 according to the invention is carried out. Subsequently, in the second method step 3, firstly the classifier candidates 31 are provided. Furthermore, the second method step 3 comprises the evaluation of a performance metric of the candidate classifiers 32, wherein this also comprises the determination of precisely this performance metric for each candidate classifier.FIG. 2 shows the evaluation of the performance metrics of the candidate classifiers 32 (and 32') in detail next to the vertical flow chart section. For evaluating and ascertaining the performance metrics of the candidate classifiers 32 (or alternatively the new candidate classifiers 32'), at least one training data set 24 aand one validation data set 24 bfrom the respective sensor data of the corresponding robot arm joint are first required. According to the invention, these are already determined in the first method step 2 for each robot arm joint from the sensor data. The classifier candidates extract features, in particular deep features or so-called deep features, from the respective data sets 34. The artificial neural networks are particularly preferably convolutional neural networks. The classifier candidates preferably differ, wherein the artificial neural networks thus preferably also differ. The classifier candidates or the artificial neural networks extract features from the training data set 34 aand the validation data set 34 b. The extracted features from the training dataset 24 amay be considered first features. The extracted features from the validation dataset 24 bmay be understood as second features.The extracted features from the training data set 34 aare fed to a K-nearest neighbor classification method, which can additionally be part of the respective classifier candidate. The K nearest neighbor classification method can be the same for different candidate classifiers. The first features are preferably supplied to the K-nearest neighbor classification method, as a result of which the K-nearest neighbor classification method is trained 35. The classifier candidates, in particular the K-nearest neighbor classification method, can classify 36 the vibrations of the validation dataset or the extracted features of the validation dataset and divide them into normal vibrations or abnormal vibrations in a following step. The accuracy of the classification from the classification step 36 can be determined in a further following step 37. Alternatively, the performance metric may include a cross entropy error or mean square errors instead of accuracy. This performance metric of the respective candidate classifier is then output 38 in the next step. The candidate classifiers are compared using this performance metric and the candidate classifier with the highest performance metric is output 33.In the second exemplary embodiment of the method according to the invention shown, a predefined termination criterion has not yet been reached after first passing through the second method step 3. This termination criterion 4 bmay be a predefined level of the performance metric, which may represent a sufficiently good result by means of the classifier. To achieve the termination criterion 4 b, it may require one or more passes of the second method step 3 or of the new second method step 3'.If the termination criterion 4 bhas not been reached n, the second method step is run through 3' again. For this purpose, new classifier candidates are first provided 31', which are preferably based on the selected classifier candidate of the previous pass of the second method step 3. These new candidate classifiers are again evaluated 32' on the basis of their performance metrics, as already explained above. Subsequently, a new candidate classifier is selected 33' which has the highest performance metric.If the termination criterion 4 bis reached or the second query 4 bis answered in the affirmative, the respective candidate classifier can be considered as classifier 5 for the respective robot arm joint. This classifier 5 meets the requirements for the performance metric and can be automatically processed, so that time is saved in particular. Furthermore, the determination of a classifier 5 for each robot arm joint enables a higher accuracy in the classification of the vibrations. The classification of the vibrations preferably takes place by means of the K-nearest neighbor classification method of the respective classifier 5.List of reference characters1 Method 2 First method step 3 Second method step 3' New second method step 4a First query 4b Second query 5 Classifier 21 Operation of the robot arm 22 Real-time vibration monitoring 23 Sensor data processing 24 Generating a training and validation dataset 24a Training dataset 24b Validation dataset 31 Providing classifier candidates 31' Providing new classifier candidates 32 Evaluating a performance metric of the classifier candidates 32' Evaluating the performance metric of the new classifier candidates 33 Selecting the classifier candidate having the highest performance metric 33' Selecting the new classifier candidate having the highest performance metric 34 Extracting features extracted from the respective datasets 34a Extracting features extracted from the training datasets 34b Training features extracted from the validation datasets 35 k-nearest neighbor classification method 36 classifying the vibrations of the validation datasets 37 determining the accuracy of the classifications 38 outputting a performance metric of the respective candidate classifier 41 warning 42 emergency stop 43 checking the robot arm joint with the abnormal vibration j Yes n No

Claims

Method (1) for generating and training an abnormality detection system for detecting abnormal vibrations in a robot arm on the basis of sensor data, wherein the robot arm has a plurality of robot arm joints, in particular six, each having at least one fiber-optic sensor, wherein the method comprises at least the following method steps: - in a first method step (2), sensor data are obtained from the operation of the robot arm (21) by means of the fiber-optic sensors and at least one training data set (24a) and one validation data set (24b) are determined from the sensor data for each of the plurality of robot arm joints, - in a second method step (3) following the first method step (2), at least one classifier (5) for detecting abnormal vibrations is determined for each of the plurality of robot arm joints, wherein a plurality of candidate classifiers are provided (31) for each classifier (5) and trained on the basis of the training dataset (24a) and checked by means of the validation dataset (24b), wherein the plurality of candidate classifiers provided for the same robot arm joint are compared with one another on the basis of at least one performance metric and that candidate classifier which has the highest performance metric is selected (33) for each robot joint.Method according to Claim 1, characterized in that the second method step (3) is carried out repeatedly until an abort criterion is reached.Method according to Claim 2, characterized in that, if the termination criterion is not reached in the second method step (3), that classifier candidate which has the highest performance metric is selected (33) and a plurality of new classifier candidates are provided (31') on the basis of the classifier candidate, and the second method step (3') is run through with these new classifier candidates.Method according to one of the preceding claims, characterized in that the fibre-optic sensors comprise an optical waveguide which has a fibre Bragg grating, preferably wherein the plurality of fibre-optic sensors are connected to one another and to a receiver unit by one or more optical waveguides.The method of claim 4, characterized in that the plurality of fiber optic sensors have different center wavelengths.Method according to one of the preceding claims, characterized in that the classifier candidates each have an artificial neural network for feature extraction from the respective sensor data of the respective robot arm joint, in particular a convolutional neural network or convolutional neural network, wherein different classifier candidates preferably comprise different artificial neural networks, wherein the classifier candidates extract first features (34a) from the training data set (24a) and second features (34b) from the validation data set (24b).Method according to Claim 6, characterized in that the classifier candidates additionally carry out a K-nearest neighbor classification method, the features extracted by the respective artificial neural network being fed to the K-nearest neighbor classification method.Method according to Claim 6 or 7, characterized in that a classifier (5), in particular the K-nearest neighbor classification method, identifies an abnormal vibration on the basis of the first (34a) and second features (34b).Method according to one of the preceding claims, characterized in that the robot arm is stopped when a classifier of the abnormality detection system detects an abnormality at a robot arm joint.Robot arm having a plurality of robot arm joints, in particular six, and a computing unit, wherein a robot arm joint has in each case at least one fiber-optic sensor which comprises an optical waveguide which has a fiber Bragg grating, wherein the computing unit is configured to carry out a method (1) according to one of the preceding claims.

Images