Method for error and / or anomaly detection in a handling or manipulation task of an industrial robot

Abstract

The present invention relates to a method for fault and / or anomaly detection. To solve the problem of providing a method for fault and / or anomaly detection which eliminates the disadvantages of the prior art and in particular allows high precision in a handling or manipulation task of an industrial robot, the method comprises obtaining first torque-determining data (D1) for a first joint (G1) and second torque-determining data (D2) for a second joint (G2) of an industrial robot (1), and determining whether a fault or anomaly exists in the handling or manipulation task of the industrial robot (1) based on the first torque-determining data (D1) and the second torque-determining data (D2).Furthermore, the invention relates to a device comprising means for carrying out the method according to the invention, a system comprising an industrial robot (1) with at least one first joint (G1) and a second joint (G2), and means for carrying out the method according to the invention for error and / or anomaly detection in a handling or manipulation task of the industrial robot (1), a computer program product comprising program instructions to execute and / or control the method according to the invention when the program is executed on a processor (201), and a computer-readable storage medium (209, 210, 211, 212, 213, 214, 215) comprising a computer program product according to the invention.

Description

[0001] The present invention relates to a method for error and / or anomaly detection. Furthermore, the invention relates to a device comprising means for executing the method according to the invention, a system comprising an industrial robot with at least one first joint and one second joint, and means for executing the method according to the invention, a computer program product comprising program instructions for executing and / or controlling the method according to the invention when the program is executed on a processor, and a computer-readable storage medium comprising a computer program product according to the invention.

[0002] Methods for error and / or anomaly detection are particularly well known in the areas of collision monitoring and condition monitoring for industrial robots. Within the scope of collision monitoring, external forces and moments for robot movements are determined and evaluated to avoid collisions with people and objects. If the external forces and moments exceed a certain limit, for example, a potential collision is detected and reactive measures are initiated. Methods for error and / or anomaly detection also play a major role in condition monitoring. These methods are intended to detect deterioration and failures of robots and their components in a timely manner.Finally, an unforeseen failure of a single robot can disrupt the operation of an entire production line, resulting in significant downtime and economic and production losses.

[0003] In the field of robotics, handling and manipulation tasks play a particularly important role. For example, in semiconductor technology, due to the increased weight of optical elements for semiconductor technology systems and increasing automation in manufacturing, robot systems are increasingly being used to handle optical elements or other components for semiconductor technology systems. Especially with robot-assisted handling tasks of objects with such high accuracy requirements, it is difficult to guarantee flawless task performance despite component wear, load influences, and other external factors. Therefore, precise monitoring of the current execution of handling tasks such as pick-and-place, joining and assembly processes, or peg-in-hole tasks is necessary.

[0004] However, prior art methods for error and / or anomaly detection generally lack precision, which is insufficient, especially in safety-critical applications or in automated operations for handling and manipulation tasks in semiconductor technology. To increase precision, for example, in condition monitoring methods, attempts have been made in the past to incorporate a variety of different signal types into error and / or anomaly detection. However, this is associated with increased computational effort, may require manual limit setting, and, in particular, cannot be easily transferred to handling and manipulation tasks for industrial robots.Furthermore, known methods are often based on clustering approaches such as the Gaussian Mixture Model or k-Means Clustering, which are not well-suited for online applications and require the number of clusters to be known in advance at best. Furthermore, these models offer poor traceability of the decisions made and are therefore ill-suited for safety-critical applications.

[0005] Against this background, the present invention is based on the object of providing a method for error and / or anomaly detection that eliminates the above-described disadvantages of the prior art and, in particular, allows for high precision in a handling or manipulation task of an industrial robot. Furthermore, an advantageous device, an advantageous system, an advantageous computer program product, and an advantageous computer-readable storage medium are to be provided.

[0006] The above-mentioned object is achieved by a method for error and / or anomaly detection in a handling or manipulation task of an industrial robot, the method comprising: - Obtaining first torque-determining data for a first joint and second torque-determining data for a second joint of the industrial robot, and - Determine whether there is an error or an abnormality in the handling or manipulation task of the industrial robot based on the first torque-determining data and the second torque-determining data.

[0007] The method is, in particular, a computer-implemented method. The method allows, in particular, error and / or anomaly detection during a handling or manipulation task of an industrial robot without additional sensors. Rather, the determination of whether an error or an anomaly exists in the handling or manipulation task of the industrial robot is based on the torque-determining data of the industrial robot.

[0008] The method is generally applicable to industrial robots with at least two joints. An industrial robot with only two joints is considered an industrial robot with two degrees of freedom (DOF), i.e., a 2-DOF industrial robot. The method is applicable to n-DOF industrial robots with n joints, where n is at least two. For example, the method can also be applied to a 6-DOF industrial robot with six joints.

[0009] The industrial robot is particularly suitable for performing a handling or manipulation task. A handling or manipulation task can, for example, involve the assembly or disassembly of a component in a system, particularly a machine. Handling or manipulation tasks can include pick-and-place, joining and assembly, or peg-in-hole tasks.

[0010] The method is particularly suitable for error and / or anomaly detection during a handling or manipulation task of an industrial robot. An anomaly can be understood as a deviation in the execution of a handling or manipulation task. Anomalies can occur in various forms, for example, they can be a point anomaly, a contextual anomaly, or a collective anomaly. An error refers in particular to a deviation that leads to at least a partial failure of the handling or manipulation task.

[0011] The method comprises obtaining first torque-determining data for a first joint and second torque-determining data for a second joint of the industrial robot. In particular, the method comprises obtaining n torque-determining data for n joints of an n-DOF industrial robot. The method can, for example, comprise obtaining first torque-determining data for a first joint, second torque-determining data for a second joint, third torque-determining data for a third joint, fourth torque-determining data for a fourth joint, fifth torque-determining data for a fifth joint, and sixth torque-determining data for a sixth joint of a 6-DOF industrial robot. In particular, torque-determining data is obtained for each joint of an industrial robot. The torque-determining data is recorded, in particular, at the individual joints and preferably stored.

[0012] The method further includes determining whether an error or an anomaly exists in the handling or manipulation task of the industrial robot. The determination is made based on the first torque-determining data and the second torque-determining data. In particular, the determination of whether an error or an anomaly exists in the handling or manipulation task of an n-DOF industrial robot is made based on the obtained n torque-determining data for the n joints of the n-DOF industrial robot.

[0013] Preferably, the method for determining whether an error or an anomaly exists in the handling or manipulation task of the industrial robot based on the first torque-determining data and the second torque-determining data comprises the application of at least one machine learning model. In particular, the machine learning model can receive the n torque-determining data of the n joints of an n-DOF industrial robot as input data and thus detect whether a handling or manipulation task of the n-DOF industrial robot is being performed correctly or incorrectly. It has been found that data-based models, particularly compared to physical models, have the advantage of simple configurability and flexible adaptability.

[0014] It has been found that, using the method according to the invention, errors and / or anomalies in handling and manipulation tasks of industrial robots can be reliably detected. In particular, it has been shown that precise error and / or anomaly detection can be provided specifically for the application area of ​​handling and manipulation tasks of industrial robots. The accuracy and / or AUROC value are in particular more than 96%, preferably more than 98%, particularly preferably more than 99%.

[0015] According to an advantageous embodiment of the method, the first torque-determining data comprises force, torque, and / or motor current data related to the first joint, and the second torque-determining data comprises force, torque, and / or motor current data related to the second joint. The torque-determining data, such as motor current, applied force, or torque, are recorded and stored, in particular, at the individual joints of the industrial robot. It has been recognized that torque-determining data of the individual joints of an industrial robot contain, in particular, information about external torques occurring at the tool center point (TCP) as well as internal changes and can be advantageously used for sensorless error and / or anomaly detection during handling or manipulation tasks of an industrial robot.In particular, it has been shown that, for example, only motor current data or only force curves at the individual joints of the industrial robot are sufficient to precisely and reliably detect errors and / or anomalies in handling or manipulation tasks of an industrial robot using the method.

[0016] According to an advantageous embodiment of the method, the method further comprises: - fusing the first torque-determining data and the second torque-determining data into a fused data set, wherein the determination of whether an error or an anomaly exists in the handling or manipulation task of the industrial robot is made based on the fused data set.

[0017] In particular, the method comprises fusing torque-determining data for all n joints of an n-DOF industrial robot into a fused dataset, wherein the determination of whether an error or an anomaly exists in the handling or manipulation task of the n-DOF industrial robot is made based on the fused dataset. By fusing the torque-determining data, the torque-determining data of the various joints of the industrial robot are combined into a fused dataset. In this way, a consistent, precise, and reliable dataset can be obtained.

[0018] The fusion of the first torque-determining data and the second torque-determining data, in particular the torque-determining data of the n joints of an n-DOF industrial robot, into a fused data set preferably comprises processing the first and / or the second torque-determining data before merging the first torque-determining data and the second torque-determining data. In particular, the method comprises processing at least a portion of the torque-determining data of the n joints of an n-DOF industrial robot before merging the torque-determining data. The processing may comprise segmenting the torque-determining data into individual data sequences. The processing of the torque-determining data may comprise interpolating the torque-determining data.The processing of the torque-determining data may comprise filtering the torque-determining data, in particular by means of a low-pass filter.

[0019] The fusion of the first torque-determining data and the second torque-determining data, in particular the torque-determining data of the n joints of an n-DOF industrial robot, into a fused data set is preferably carried out using an information fusion method. The fusion of the first torque-determining data and the second torque-determining data, in particular the torque-determining data of the n joints of an n-DOF industrial robot, into a fused data set can, for example, comprise fusion using signal-level fusion. Signal-level fusion is understood to mean merging the raw data into a fused data set. Signal-level fusion is particularly suitable for real-time applications. Disadvantages include a high data volume and slow processing speed.When using raw data, the torque-determining data for each joint are preferably first interpolated to the same number of data points and then combined into one data set using signal-level fusion.

[0020] The fusion of the first torque-determining data and the second torque-determining data, in particular the torque-determining data of the n joints of an n-DOF industrial robot, into a fused dataset can comprise fusion using signal-level fusion. In particular, the fusion can be performed using feature extraction and optionally feature selection. In feature-level fusion, the torque-determining data for the individual joints are combined into a single feature set using suitable normalization, transformation, and / or reduction methods. The extraction and selection of statistical key figures (features) from data series aims to remove redundant or irrelevant information, optimize the computing speed and memory requirements of machine learning models, and maximize their accuracy.

[0021] Preferably, the merging of the first torque-determining data and the second torque-determining data comprises: - Extracting at least one statistical key figure from the first torque-determining data and the second torque-determining

[0022] Data, wherein the merged data set comprises the at least one statistical key figure.

[0023] In particular, merging the first torque-determining data and the second torque-determining data involves merging using feature extraction. Feature extraction describes the transformation of the original input data into new statistical key figures (features) that preferably contain only the most important information. The goal is not to lose any information from the input data, but rather to present the information summarized in a smaller dimension. For this purpose, the at least one statistical key figure can be extracted from the raw data, and the feature points can be linked for n joints. In particular, extracting the at least one statistical key figure involves extracting at least one statistical key figure that describes a time series independently of movement.

[0024] Extraction is understood, in particular, to mean calculation. Extraction includes, in particular, calculating at least one statistical key figure from the first torque-determining data for the first joint and the at least one statistical key figure from the second torque-determining data for the second joint. Extraction includes, in particular, calculating at least one statistical key figure from the torque-determining data for each joint of the industrial robot. The at least one statistical key figure is extracted, in particular, from the respective torque-determining data for each joint.

[0025] Preferably, the at least one statistical key figure is or comprises a statistical key figure from the time domain, the frequency domain and / or the time-frequency domain, in particular the Katz Fractility D K, the Magnitude Fluctuation Index (MFI) of the signal envelope MFI and / or the Signal-to-Noise Ratio (SNR) SNR Feature ,

[0026] In particular, the fusion of the first torque-determining data and the second torque-determining data, in particular the n torque-determining data of an n-DOF industrial robot, comprises extracting a plurality of statistical key figures, for example at least five, in particular more than five, for example ten to 50, preferably 20 to 50, more preferably 30 to 40 statistical key figures, per joint, which are preferably collected in the time, frequency and time-frequency domain.

[0027] The at least one statistical key figure from the first torque-determining data and the at least one statistical key figure from the second torque-determining data, in particular the at least one statistical key figure from the n torque-determining data for the n joints of the industrial robot, are preferably merged into a merged data set, so that the merged data set comprises the at least one statistical key figure, in particular for each joint of the industrial robot. The merging can, in particular, comprise combining the statistical key figures into an input vector for a machine learning model.

[0028] According to a further advantageous embodiment of the method, the merging of the first torque-determining data and the second torque-determining data further comprises: - Selecting at least one selected statistical key figure from the extracted at least one statistical key figure, wherein the at least one statistical key figure comprised by the merged data set is or comprises the at least one selected statistical key figure.

[0029] In particular, merging the first torque-determining data and the second torque-determining data involves merging using feature selection. The feature selection method selects the most relevant dimensions of the torque-determining data and provides a reduced fused dataset. Feature selection is therefore a method for reducing the dimensionality of the dataset. If statistical metrics are deemed relevant, they are adopted; if statistical metrics are redundant or irrelevant to determining a prediction, they are not adopted as input data.

[0030] It has been shown that the statistical indicators of the signal power in the time domain P Zeit , the signal power in the frequency range P Frequ , the Katz Fractility D K, the Magnitude Fluctuation Index (MFI) of the signal envelope MFI and the Signal-to-Noise Ratio (SNR) SNR Feature are particularly suitable for the method for error and / or anomaly detection in a handling or manipulation task of an industrial robot.

[0031] In particular, the fusion of the first torque-determining data and the second torque-determining data, in particular the n torque-determining data for the n joints of the industrial robot, comprises extracting a plurality of statistical key figures from the torque-determining data for each joint of the industrial robot, selecting from the plurality of extracted statistical key figures, and merging the selected statistical key figures into a fused data set.

[0032] According to a further advantageous embodiment of the method, the determination of whether an error or an anomaly exists in the handling or manipulation task of the industrial robot is carried out by applying a machine learning model, in particular an unsupervised machine learning model, to the fused data set. Applying a machine learning model to the fused data set allows, in particular, a classification of the handling or manipulation task of the industrial robot based on the obtained and fused first and second torque-determining data, in particular the obtained and fused n torque-determining data of an n-DOF industrial robot. With the aid of the machine learning model, a distinction can preferably be made between the presence and absence of an error or anomaly.

[0033] Unsupervised learning refers specifically to a process in which previously unknown relationships are discovered from a data set. The machine learning model learns autonomously, meaning no human supervision or correction is necessary. It has been shown that the use of an unsupervised machine learning model is particularly advantageous when uncertain training data is available, in particular when assigning training data to a class is not possible. An unsupervised machine learning model is used in error and / or anomaly detection processes, especially when there are insufficient faulty signals available for training the models.

[0034] The unsupervised machine learning model preferably comprises a one-class classification model. One-class classification models are models that are trained only with data of a first type, for example, error-free torque-determining data, and can subsequently distinguish data of a second type, particularly faulty torque-determining data.

[0035] To train the machine learning model, in particular the one-class classification model, first torque-determining data for the first joint and second torque-determining data for the second joint, in particular n torque-determining data for n joints of an n-DOF industrial robot, are obtained. The torque-determining data are preferably based at least substantially, in particular exclusively, on error-free movements of the industrial robot. In particular, error-free reference signals of the torque-determining parameters are obtained for training the machine learning model, which optimally represent the spectrum of possible torque-determining data and the entire movement of the industrial robot.To train the machine learning model, a fused dataset based on obtained torque-determining data for each of the joints of the industrial robot can be obtained using signal-level fusion or feature-level fusion, whereby the fused dataset is used as a training dataset for training the machine learning model.

[0036] The trained machine learning model can then be tested with a test dataset based on torque-determining data for each joint during both error-free and error-inducing joint movements to assess the precision and reliability of the machine learning model. If faulty signals are available that can be used for testing, they are preferably introduced. If no faulty signals are present, manually modified signals to which errors have been introduced, for example, using generative artificial intelligence or by adding noise or serial outliers, can be used to test the machine learning model.

[0037] The trained and tested machine learning model can be saved and applied to new, undefined movements and / or torque-determining data of the industrial robot. The machine learning model is particularly suitable for determining errors and / or anomalies during handling or manipulation tasks of the industrial robot. The machine learning model can be trained flexibly and quickly for each new movement of an n-DOF industrial robot. Using the model, anomalies can be detected without prior knowledge. In particular, the present method eliminates the need for manual limit setting or the creation of a kinematic and dynamic model of the industrial robot.

[0038] Unsupervised machine learning models include, in particular, Local Outlier Factor (LOF), Elliptic Envelope, or One-Class Support Vector Machine (OC-SVM). LOF is a machine learning algorithm that uses the density of data points in the distribution as a key factor for outlier detection. LOF compares the density of any data point with the density of its neighbors. Since outliers originate from sparsely populated areas, the ratio is higher for anomalous data points. The LOF model is very robust even with varying payloads and, due to its functionality, can be used for datasets with high variance.

[0039] The Elliptic Envelope algorithm creates an imaginary elliptical region around a given data set. Values ​​within the envelope are considered normal data, and anything outside the envelope is returned as an outlier. The algorithm works best when the data has a Gaussian distribution.

[0040] One-Class Support Vector Machine (OC-SVM) is an unsupervised learning technique and is based on the basic idea of ​​minimizing the hypersphere of a single class of examples in the training data and considering all other samples outside the hypersphere as outliers or outside the training data distribution.

[0041] It is also conceivable that the unsupervised machine learning model includes a deep learning model, in particular an LSTM encoder-decoder model. Deep learning models such as LSTM encoder-decoder models are particularly advantageous for anomaly detection in time series when a large training data set is available, especially a large amount of torque-determining data based on error-free movements. The application of clustering algorithms, such as the Gaussian Mixture Model or k-Means Clustering, is also conceivable.

[0042] According to a further advantageous embodiment of the method, the method further comprises: - determining a first decision based on the first torque-determining data and a second decision based on the second torque-determining data, and - fusing the first decision and the second decision into a fused decision by means of a decision algorithm, wherein the determination of whether an error or an anomaly exists in the handling or manipulation task of the industrial robot is made based on the fused decision.

[0043] In particular, the method comprises fusing the first torque-determining data and the second torque-determining data, in particular the n torque-determining data of the n joints of an n-DOF industrial robot, by means of decision-level fusion. Within the scope of decision-level fusion, decisions based on individual data sources are processed and combined in order to achieve an optimal overall decision. In the present case, the method preferably comprises determining a first decision based on the first torque-determining data and a second decision based on the second torque-determining data, in particular determining n decisions based on n torque-determining data for n joints of an n-DOF industrial robot. A decision means, for example, a classification, in particular a determination of whether or not an error or an anomaly exists for a movement of a joint.

[0044] In a further step, the method comprises merging the first decision and the second decision, in particular the n decisions, using a decision algorithm to form a fused decision. Determining whether an error or an anomaly exists in the handling or manipulation task of the industrial robot can be done based on the fused decision. It has been shown that merging the torque-determining data using decision-level fusion is particularly advantageous for a method for error and / or anomaly detection in a handling or manipulation task of an industrial robot. In particular, this allows particularly high precision to be achieved in determining whether an error or an anomaly exists in the handling or manipulation task of the industrial robot.

[0045] Preferably, the decision algorithm includes a weighting of the first and second decisions. In particular, the decision algorithm includes a weighting of the n decisions for n joints of an n-DOF industrial robot. For example, decisions based on the torque-determining data of those joints that have a greater influence on the overall movement of the industrial robot can be given greater weight.

[0046] It is also conceivable for the decision algorithm to comprise a comparison of the number of decisions of a first type and the number of decisions of a second type, whereby the first decision and the second decision, in particular the n decisions for the n joints of the n-DOF industrial robot, are each assigned to the first or second type. A decision of the first type means, for example, the presence of an error or anomaly in an individual movement of the respective joint (bad decision), whereas a decision of the second type means, for example, that an error or anomaly does not exist in the individual movement of the respective joint (good decision). In particular, the decision algorithm is based on a majority decision. The decision of the type made for the majority of the joints can, in particular, be adopted as a fused decision.Determining whether there is an error or anomaly in the handling or manipulation task of the industrial robot can be done based on the fused decision.

[0047] According to a further advantageous embodiment of the method, the first decision is determined using a first unsupervised machine learning model, and the second decision is determined using a second unsupervised machine learning model. With the aid of respective unsupervised machine learning models, a decision can preferably be made for each joint of the industrial robot as to whether or not an error or an anomaly is present in the movement of the joint. Determining a first decision comprises, in particular, applying a first unsupervised machine learning model to the first torque-determining data for the first joint, and determining a second decision comprises, in particular, applying a second unsupervised machine learning model to the second torque-determining data for the second joint.

[0048] To train a first machine learning model, in particular a first one-class classification model, first torque-determining data for error-free movements of the first joint can be obtained. The first torque-determining data is preferably first processed, for example, by segmentation into individual data sequences, low-pass filtering, interpolation, and / or data normalization. Based on the first torque-determining data of error-free movements, a first training data set can be obtained for the first joint and used to train the first machine learning model. In this way, the first machine learning model learns error-free movements of the first joint of the industrial robot.The trained first machine learning model can then be tested with initial test data based on initial torque-determining data for error-free and faulty movements of the first joint, in particular to determine the precision and reliability of the first machine learning model. The trained and tested first machine learning model can be saved and applied to new, indeterminate movements of the first joint of the industrial robot. The steps can be performed analogously for the second machine learning model to determine the second decision, in particular for all n machine learning models to determine the n decisions for all n joints of an n-DOF industrial robot.

[0049] Preferably, the first unsupervised machine learning model and the second unsupervised machine learning model, in particular all n unsupervised machine learning models for n joints of an n-DOF industrial robot, each comprise a one-class classification model, in particular Local Outlier Factor (LOF), Elliptic Envelope or One-Class Support Vector Machine (OC-SVM).

[0050] According to a further advantageous embodiment of the method, the method further comprises: - Determine whether the industrial robot's handling or manipulation task was successful based on the presence of an error or anomaly.

[0051] In particular, the method comprises determining that the handling or manipulation task of the industrial robot was successful if it has been determined that an error or anomaly does not exist. Furthermore, the method comprises, in particular, determining that the handling or manipulation task of the industrial robot was unsuccessful if it has been determined that an error or anomaly exists. Determining whether the handling or manipulation task of the industrial robot was successful or unsuccessful allows, in particular, conclusions to be drawn about the handling or manipulation task of the industrial robot without, for example, requiring a visual inspection of the result. This is particularly advantageous for a handling or manipulation task in the field of semiconductor lithography.

[0052] The handling or manipulation task of the industrial robot preferably includes positioning and / or orienting a component in a system. In particular, the handling or manipulation task of the industrial robot includes positioning and / or orienting a component in a system for semiconductor technology. The positioning and / or orienting of the component can be performed, in particular, using at least one zero-point clamping system. Zero-point clamping systems are particularly advantageous for the secure and highly precise positioning and / or orientation of components using an industrial robot.However, the components used for imaging in semiconductor technology systems, especially optical elements, are becoming increasingly larger and heavier, for example due to increasing requirements and the associated increase in numerical aperture, which makes positioning and / or orienting the component in a semiconductor technology system more difficult. Visual inspection of the correct positioning and / or orientation of the component is hardly possible anymore, especially when accessibility is limited, for example, when using a zero-point clamping system.It was recognized that with the aid of the method for error and / or anomaly detection, these disadvantages can be overcome and, for a handling or manipulation task of an industrial robot, in particular for positioning and / or orienting a component in a system, in particular a semiconductor technology system, a reliable statement can be made as to whether the handling or manipulation task of the industrial robot was successful or not.

[0053] The aforementioned object is also achieved by a device according to the invention comprising means for carrying out the method according to the invention. It can be provided that some or all of the steps of the method are carried out. The device can be a data processing device. The device can comprise hardware and / or software components. The device can, for example, comprise at least one memory, in particular with instructions of a computer program product, and / or at least one processor, in particular designed to execute instructions from the at least one memory.Accordingly, a device comprising at least one processor and at least one memory with instructions is also to be understood as disclosed, wherein the at least one memory and the instructions are configured, together with the at least one processor, to cause the device to execute the method according to the invention. In this disclosure, a processor is to be understood as including control units, microprocessors, microcontrol units such as microcontrollers, digital signal processors (DSPs), application-specific integrated circuits (ASICs), or field-programmable gate arrays (FPGAs).

[0054] Alternatively or additionally, the device may further comprise one or more communication interfaces (e.g., one or more wired and / or wireless communication interfaces, e.g., a wireless communication interface in the form of a radio interface) and / or one or more user interfaces (e.g., a keyboard, a mouse, a screen, a touchscreen, a speaker, a microphone, etc.). It is understood that the disclosed device may also comprise other means not listed.

[0055] The above-mentioned object is also achieved by a system according to the invention comprising: - an industrial robot having at least a first and a second joint, and - Means for carrying out the method according to the invention for error and / or anomaly detection in a handling or manipulation task of the industrial robot.

[0056] The system comprises, in particular, an n-DOF industrial robot with n joints. An industrial robot with a first and a second joint is, for example, a 2-DOF industrial robot. An industrial robot with six joints is, for example, a 6-DOF industrial robot. The industrial robot is used, in particular, to perform a handling or manipulation task, for example, to position and / or orient a component in a system, in particular in a system for semiconductor technology.

[0057] The system further comprises means for executing the method according to the invention for detecting errors and / or anomalies during the handling or manipulation task of the industrial robot. For example, the device according to the invention can be part of the system according to the invention.

[0058] The aforementioned object is additionally achieved by a computer program product according to the invention comprising program instructions for carrying out and / or controlling the method according to the invention when the program is executed on a processor. A computer program product according to the present disclosure can, for example, be distributable via a network such as the Internet, a telephone or mobile radio network and / or a local area network. The computer program product can be at least partially software and / or firmware of a processor. It can equally be implemented at least partially as hardware. The computer program product can, for example, be stored on a computer-readable medium, in particular a storage medium, e.g. a magnetic, electrical, optical and / or other type of storage medium.The storage medium can, for example, be part of the processor, such as a (non-volatile or volatile) program memory of the processor or a portion thereof. The storage medium can, for example, be a tangible or physical storage medium.

[0059] The aforementioned object is also achieved by a computer-readable storage medium according to the invention comprising a computer program product according to the invention. A computer-readable storage medium can be designed, for example, as a magnetic, electrical, electromagnetic, optical and / or other type of storage medium. Such a computer-readable storage medium is preferably tangible (i.e., "touchable"), for example, it is designed as a data storage device. Such a data storage device is, for example, portable or permanently installed in a device. Examples of such a data storage device are volatile or non-volatile random access memories (RAM), such as NOR flash memories, or sequential access memories such as NAND flash memories and / or read-only memories (ROM) or read-write access memories. Computer-readable should be understood, for example, to mean that the storage medium can be read by a computer ora data processing system, for example, a processor, that can be read and / or written to. The computer-readable storage medium can also be a data communications network that enables downloading of program code, such as the Internet or a data cloud.

[0060] At least one embodiment of the invention is described below with reference to the drawing. The drawing shows: Fig. 1 schematically shows a meridional section of a semiconductor technology plant for EUV projection lithography, Fig. 2 schematically shows a 6-DOF industrial robot for positioning and / or orienting a component in a semiconductor technology system, Fig. 3a, Fig. 3b shows an embodiment of a method according to the invention for error and / or anomaly detection in a handling or manipulation task of an industrial robot, Fig. 4a, Fig. 4b shows a further embodiment of a method according to the invention for error and / or anomaly detection in a handling or manipulation task of an industrial robot, Fig. 5a, Fig. 5b shows a further embodiment of a method according to the invention for error and / or anomaly detection in a handling or manipulation task of an industrial robot, Fig. 6 a schematic block diagram of an embodiment of a device according to the invention, and Fig. 7 a schematic representation of embodiments of tangible and non-transitory computer-readable media.

[0061] In the following description of the various embodiments according to the invention, components and elements with the same function and the same mode of operation are provided with the same reference numerals, even if the components and elements in the various embodiments may have differences in their dimensions, shape or nature.

[0062] In the following, with reference to the Fig. 1 describes, by way of example, the essential components of a semiconductor technology system 100 for microlithography. The description of the basic structure of the semiconductor technology system 100 and its components is not intended to be limiting.

[0063] One embodiment of an illumination system 102 of the semiconductor technology system 100 has, in addition to a light or radiation source 103, an illumination optics 104 for illuminating an object field 111 in an object plane 112. In an alternative embodiment, the light source 103 can also be provided as a separate module from the rest of the illumination system. In this case, the illumination system does not include the light source 103.

[0064] A reticle 113 arranged in the object field 111 is exposed. The reticle 113 is held by a reticle holder. The reticle holder can be displaced, particularly in a scanning direction, via a reticle displacement drive.

[0065] In the Fig. 1 shows a Cartesian xyz coordinate system for explanation. The x-direction is perpendicular to the plane of the drawing. The y-direction is horizontal and the z-direction is vertical. The scanning direction is in the Fig.1 along the y-direction. The z-direction runs perpendicular to the object plane 112.

[0066] The semiconductor technology system 100 comprises a projection optics 10. The projection optics 117 serves to image the object field 111 into an image field 11 in an image plane 115. The image plane 115 runs parallel to the object plane 112. Alternatively, an angle other than 0° between the object plane 112 and the image plane 115 is also possible.

[0067] A structure on the reticle 113 is imaged onto a light-sensitive layer of a wafer 116 arranged in the image plane 115 in the region of the image field 114. The wafer 116 is held by a wafer holder. The wafer holder can be displaced, particularly along the y-direction, via a wafer displacement drive. The displacement of the reticle 113 via the reticle displacement drive, on the one hand, and of the wafer 116 via the wafer displacement drive, on the other hand, can be synchronized with each other.

[0068] The radiation source 103 is an EUV radiation source. The radiation source 103 emits, in particular, EUV radiation 101, which is also referred to below as useful radiation, illumination radiation, or illumination light. The useful radiation has, in particular, a wavelength in the range between 5 nm and 30 nm. The radiation source 103 can be a plasma source, for example, an LPP source (laser produced plasma) or a DPP source (gas discharged produced plasma). It can also be a synchrotron-based radiation source. The radiation source 103 can be a free-electron laser (FEL).

[0069] The illumination radiation 101 emanating from the radiation source 103 is focused by a collector 105. The collector 105 can be a collector with one or more ellipsoidal and / or hyperboloidal reflection surfaces. The at least one reflection surface of the collector 105 can be exposed to the illumination radiation 101 at grazing incidence (GI), i.e., at angles of incidence greater than 45°, or at normal incidence (NI), i.e., at angles of incidence less than 45°. The collector 105 can be structured and / or coated, on the one hand, to optimize its reflectivity for the useful radiation and, on the other hand, to suppress stray light.

[0070] After the collector 105, the illumination radiation 101 propagates through an intermediate focus in an intermediate focal plane 106. The intermediate focal plane 106 can represent a separation between a radiation source module, comprising the radiation source 103 and the collector 105, and the illumination optics 104.

[0071] The illumination optics 104 comprises a deflecting mirror 107 and, downstream of this in the beam path, a first facet mirror 108. The deflecting mirror 107 can be a flat deflecting mirror or, alternatively, a mirror with a beam-influencing effect beyond the pure deflection effect. Alternatively or additionally, the deflecting mirror 107 can be designed as a spectral filter that separates a useful light wavelength of the illumination radiation 101 from stray light of a different wavelength. If the first facet mirror 108 is arranged in a plane of the illumination optics 104 that is optically conjugated to the object plane 112 as the field plane, it is also referred to as a field facet mirror. The first facet mirror 108 comprises a plurality of individual first facets 109, which are also referred to below as field facets. Of these facets 109, Fig. 1 only some examples are shown.

[0072] The first facets 109 can be designed as macroscopic facets, in particular as rectangular facets or as facets with an arcuate or partially circular edge contour. The first facets 109 can be designed as flat facets or, alternatively, as convexly or concavely curved facets.

[0073] As is known, for example, from DE 10 2008 009 600 A1, the first facets 109 themselves can also be composed of a plurality of individual mirrors, in particular a plurality of micromirrors. The first facet mirror 108 can, in particular, be designed as a microelectromechanical system (MEMS system). For details, reference is made to DE 10 2008 009 600 A1.

[0074] Between the collector 105 and the deflecting mirror 107, the illumination radiation 101 runs horizontally, i.e. along the y-direction.

[0075] In the beam path of the illumination optics 4, a second facet mirror 110 is arranged downstream of the first facet mirror 108. If the second facet mirror 110 is arranged in a pupil plane of the illumination optics 104, it is also referred to as a pupil facet mirror. The second facet mirror 110 can also be arranged at a distance from a pupil plane of the illumination optics 104. In this case, the combination of the first facet mirror 108 and the second facet mirror 110 is also referred to as a specular reflector. Specular reflectors are known from US 2006 / 0132747 A1, EP 1 614 008 B1, and US Pat. No. 6,573,978.

[0076] The second facet mirror 110 comprises a plurality of second facets. In the case of a pupil facet mirror, the second facets 213 are also referred to as pupil facets.

[0077] The second facets can also be macroscopic facets, which can be round, rectangular, or hexagonal, for example, or alternatively facets composed of micromirrors. Reference is also made to DE 10 2008 009 600 A1 in this regard.

[0078] The second facets can have flat or alternatively convex or concave curved reflection surfaces.

[0079] The illumination optics 104 thus forms a double-faceted system. This basic principle is also known as a fly's-eye integrator.

[0080] It may be advantageous to arrange the second facet mirror 110 not exactly in a plane that is optically conjugated to a pupil plane of the projection optics 117. In particular, the pupil facet mirror 110 can be arranged tilted relative to a pupil plane of the projection optics 17, as described, for example, in DE 10 2017 220 586 A1.

[0081] With the help of the second facet mirror 110, the individual first facets 109 are imaged into the object field 111. The second facet mirror 110 is the last beam-forming mirror, or actually the last mirror for the illumination radiation 101 in the beam path before the object field 111.

[0082] In a further, not shown embodiment of the illumination optics 104, a transmission optics can be arranged in the beam path between the second facet mirror 110 and the object field 111, which transmission optics contributes in particular to the imaging of the first facets 109 into the object field 111. The transmission optics can have exactly one mirror, but alternatively also two or more mirrors, which are arranged one behind the other in the beam path of the illumination optics 104. The transmission optics can in particular comprise one or two mirrors for normal incidence (NI mirrors, normal incidence mirrors) and / or one or two mirrors for grazing incidence (GI mirrors, grazing incidence mirrors).

[0083] The illumination optics 104 has in the embodiment shown in the Fig. 1, after the collector 105 there are exactly three mirrors, namely the deflection mirror 107, the field facet mirror 108 and the pupil facet mirror 110.

[0084] In a further embodiment of the illumination optics 104, the deflection mirror 107 can also be omitted, so that the illumination optics 104 can then have exactly two mirrors after the collector 105, namely the first facet mirror 108 and the second facet mirror 110.

[0085] The imaging of the first facets 109 by means of the second facets or with the second facets and a transmission optics into the object plane 112 is usually only an approximate imaging.

[0086] The projection optics 117 comprises a plurality of mirrors Mi, which are numbered according to their arrangement in the beam path of the semiconductor technology system 100.

[0087] In the Fig.In the example shown in Figure 1, the projection optics 117 comprises six mirrors M1 to M6. Alternatives with four, eight, ten, twelve, or a different number of mirrors M1 are also possible. The projection optics 117 is a doubly obscured optics. The penultimate mirror M5 and the last mirror M6 each have a passage opening for the illumination radiation 101. The projection optics 117 has an image-side numerical aperture that is greater than 0.5 and can also be greater than 0.6, for example, 0.7 or 0.75.

[0088] Reflection surfaces of the mirrors Mi can be designed as freeform surfaces without a rotational symmetry axis. Alternatively, the reflection surfaces of the mirrors Mi can be designed as aspherical surfaces with exactly one rotational symmetry axis of the reflection surface shape. The mirrors Mi, like the mirrors of the illumination optics 104, can have highly reflective coatings for the illumination radiation 16. These coatings can be designed as multilayer coatings, in particular with alternating layers of molybdenum and silicon.

[0089] The projection optics 117 has a large object-image offset in the y-direction between a y-coordinate of a center of the object field 111 and a y-coordinate of the center of the image field 114. This object-image offset in the y-direction can be approximately as large as a z-distance between the object plane 112 and the image plane 115.

[0090] The projection optics 117 can be designed in particular to be anamorphic. In particular, it has different image scales β x , β y in the x- and y-direction. The two magnifications β x , β y of the projection optics 117 are preferably at (β x , β y ) = (+ / - 0.25, / +- 0.125). A positive magnification β means an image without image inversion. A negative sign for the magnification β means an image with image inversion.

[0091] The projection optics 117 thus leads to a reduction in the ratio 4:1 in the x-direction, i.e. in the direction perpendicular to the scanning direction.

[0092] The projection optics 117 results in a reduction of 8:1 in the y-direction, i.e. in the scanning direction.

[0093] Other magnifications are also possible. Magnifications with the same sign and absolutely identical in the x and y directions, for example, with absolute values ​​of 0.125 or 0.25, are also possible.

[0094] The number of intermediate image planes in the x- and y-directions in the beam path between the object field 111 and the image field 114 can be the same or can be different, depending on the design of the projection optics 117. Examples of projection optics with different numbers of such intermediate images in the x- and y-directions are known from US 2018 / 0074303 A1.

[0095] Each of the pupil facets is assigned to exactly one of the field facets 109 to form a respective illumination channel for illuminating the object field 111. This can, in particular, result in illumination according to the Köhler principle. The far field is divided into a plurality of object fields 111 using the field facets 109. The field facets 109 generate a plurality of images of the intermediate focus on the pupil facets assigned to them.

[0096] The field facets 109 are each imaged onto the reticle 113 by an associated pupil facet, overlapping one another, to illuminate the object field 111. The illumination of the object field 111 is, in particular, as homogeneous as possible. It preferably has a uniformity error of less than 2%. Field uniformity can be achieved by superimposing different illumination channels.

[0097] By arranging the pupil facets, the illumination of the entrance pupil of the projection optics 117 can be geometrically defined. By selecting the illumination channels, in particular the subset of the pupil facets that guide light, the intensity distribution in the entrance pupil of the projection optics 117 can be adjusted. This intensity distribution is also referred to as the illumination setting or illumination pupil fill.

[0098] A likewise preferred pupil uniformity in the area of ​​defined illuminated sections of an illumination pupil of the illumination optics 104 can be achieved by redistributing the illumination channels.

[0099] Further aspects and details of the illumination of the object field 111 and in particular of the entrance pupil of the projection optics 117 are described below.

[0100] The projection optics 117 can, in particular, have a homocentric entrance pupil. This can be accessible. It can also be inaccessible.

[0101] The entrance pupil of the projection optics 117 cannot usually be precisely illuminated with the pupil facet mirror 110. When imaging the projection optics 117, which telecentrically images the center of the pupil facet mirror 110 onto the wafer 116, the aperture rays often do not intersect at a single point. However, a surface can be found in which the pairwise determined spacing of the aperture rays is minimized. This surface represents the entrance pupil or a surface conjugate to it in spatial space. In particular, this surface exhibits a finite curvature.

[0102] It is possible that the projection optics 117 have different entrance pupil positions for the tangential and sagittal beam paths. In this case, an imaging element, in particular an optical component of the transmission optics, should be provided between the second facet mirror 110 and the reticle 113. With the help of this optical element, the different positions of the tangential entrance pupil and the sagittal entrance pupil can be taken into account.

[0103] In the Fig. In the arrangement of the components of the illumination optics 104 shown in Figure 1, the pupil facet mirror 110 is arranged in a surface conjugated to the entrance pupil of the projection optics 117. The field facet mirror 108 is arranged tilted relative to the object plane 112. The first facet mirror 108 is arranged tilted relative to an arrangement plane defined by the deflection mirror 107.

[0104] The first facet mirror 108 is arranged tilted to an arrangement plane which is defined by the second facet mirror 110.

[0105] Fig. 2 shows an industrial robot 1, wherein the industrial robot 1 is a 6-DOF industrial robot. The 6-DOF industrial robot 1 comprises six joints G1, G2, G3, G4, G5, G6 and accordingly six degrees of freedom. A motor (not shown) and a gear (not shown) are attached to each of the six joints G1, G2, G3, G4, G5, G6, by means of which the respective joint G1, G2, G3, G4, G5, G6 is moved. The motor current data at the individual joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1 are continuously recorded and stored as series of measurements. The 6-DOF industrial robot 1 also has a programmable logic controller (PLC) with an interface (not shown) for providing the motor current data of the individual joints G1, G2, G3, G4, G5, G6.

[0106] With the help of the 6-DOF industrial robot 1, handling or manipulation tasks, such as loading and unloading, can be accomplished. Fig. The 6-DOF industrial robot 1 shown in Figure 2 is used for positioning and / or orienting components in semiconductor technology systems, such as the mirrors M1 to M6 in the Fig. 1. With the help of the 6-DOF industrial robot 1, a component is picked up and transferred to the semiconductor technology system 100. This is done in the present case using a zero-point clamping system (not shown). Transferring the component to the semiconductor technology system 100 involves moving a pin (not shown) attached to the component into a zero-point clamping system. Conversely, an unloading process can involve extending a pin (not shown) attached to the component from a zero-point clamping system.

[0107] Fig. 3a shows an exemplary embodiment of a method for error and / or anomaly detection in the handling task of the Fig. 2 shown 6-DOF industrial robot 1. A flawlessly executed handling task is defined as a contactless retraction movement into the zero-point clamping system. Fig.The method shown in Figure 3a comprises, in a first step S1, obtaining torque-determining data D1, D2, D3, D4, D5, D6 for each of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1. The first torque-determining data D1 in this case are motor current data related to the first joint G1, the second torque-determining data D2 in this case are motor current data related to the second joint G2, the third torque-determining data D3 in this case are motor current data related to the third joint G3, the fourth torque-determining data D4 in this case are motor current data related to the fourth joint G4, the fifth torque-determining data D5 in this case are motor current data related to the fifth joint G5 and the sixth torque-determining data D6 in this case are Motor current data related to the sixth joint G6.

[0108] In a second step S2, the method comprises merging the first torque-determining data D1, the second torque-determining data D2, the third torque-determining data D3, the fourth torque-determining data D4, the fifth torque-determining data D5 and the sixth torque-determining data D6 into a fused data set D NeuThe fusion of the torque-determining data D1, D2, D3, D4, D5, and D6 is performed using signal-level fusion. For this purpose, the torque-determining data D1, D2, D3, D4, D5, and D6 are first processed. Since the torque-determining data D1, D2, D3, D4, D5, and D6 were recorded continuously, the processing first involves segmenting the torque-determining data D1, D2, D3, D4, D5, and D6 into individual data sequences. The torque-determining data D1, D2, D3, D4, D5, and D6 are also filtered using a low-pass filter to remove noise in the data. The processing further comprises interpolating at least part of the torque-determining data D1, D2, D3, D4, D5, D6 so that the same number of data points is available for each data sequence.

[0109] In step S3, the method comprises determining whether an error or an anomaly exists in the handling or manipulation task of the 6-DOF industrial robot 1. The determination is made based on the fused data set D Neu . For this purpose, the merged data set D Neu An unsupervised machine learning (ML) model was applied. The unsupervised machine learning (ML) model is a one-class classification (OCC) model, in this case a local outlier factor (LOF) model.

[0110] The one-class classification model OCC for application to the merged dataset D Neu in which Fig. The method shown in Figure 3a for error and / or anomaly detection in the handling task of the 6-DOF industrial robot 1 was implemented using the Fig. 3b. In a first step S1', the method comprises analogous to the method shown in Fig.3a, obtaining torque-determining data D1, D2, D3, D4, D5, D6 for each of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1, wherein in this case the torque-determining data D1, D2, D3, D4, D5, D6 are based exclusively on error-free movements of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1.

[0111] For training of the in the Fig. The one-class classification model OCC applied in the procedure shown in Figure 3a was applied analogously to step 2 of the Fig.3a, in a step S2', a fused data set is obtained by means of signal-level fusion based on obtained torque-determining data D1, D2, D3, D4, D5, D6 for each of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1, wherein in this case the torque-determining data D1, D2, D3, D4, D5, D6 are based exclusively on error-free movements of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1 and the fused data set is used as training data set D Train used to train the one-class classification model OCC.

[0112] The trained one-class classification model OCC is then tested with a test dataset D Testbased on torque-determining data D1, D2, D3, D4, D5, D6 for each of the six joints G1, G2, G3, G4, G5, G6 during error-free and faulty movements of the joints G1, G2, G3, G4, G5, G6, in particular to determine the precision and reliability of the one-class classification model OCC.

[0113] Subsequently, the one-class classification model OCC, which was trained and tested on the basis of the motor current data of the 6-DOF industrial robot, is saved and used analogously to the one in Fig.3a is applied to new, undefined movements of the 6-DOF industrial robot 1. The one-class classification model OCC detects errors and anomalies during the handling task of the 6-DOF industrial robot 1 and, based on the motor current data of the 6-DOF industrial robot 1, allows conclusions to be drawn as to whether the entry into the zero-point clamping system and thus the positioning and / or orientation of a component in the semiconductor technology system 100 was successful or not. The method can be applied analogously, for example, to classify an exit from a zero-point clamping system.

[0114] Fig. 4a shows a further exemplary embodiment of a method for error and / or anomaly detection in the handling task of the Fig. 2 shown 6-DOF industrial robot 1. The first step S1 of the method corresponds to step S1 of the Fig.3a and comprises obtaining torque-determining data D1, D2, D3, D4, D5, D6 for each of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1, wherein the torque-determining data D1, D2, D3, D4, D5, D6 comprise motor current data of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1.

[0115] In a second step S2, the torque-determining data D1, D2, D3, D4, D5, D6 are merged into a data set D Neufused, whereby the fusion of the torque-determining data D1, D2, D3, D4, D5, D6 is performed using feature-level fusion. The fusion of the data using feature-level fusion involves extracting statistical key figures from the torque-determining data D1, D2, D3, D4, D5, D6 using feature extraction. In this case, 39 statistical key figures are extracted for each joint G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1. The statistical key figures are time-domain statistical key figures, frequency-domain statistical key figures, and time-frequency-domain statistical key figures. The 39 statistical key figures can be found in Table 1. Table 1 Statistical key figures from the time domain 1 mean 11 energy 2 Standard deviation 12 Performance 3 Variance 13 Minimum difference between signal half 1 and signal half 2 4 Root Mean Squared Error (RMSE) 14 Maximum difference between signal half 1 and signal half 2 5 Crest Factor 15 Difference standard deviation signal half 1 and signal half 2 6 maximum 16 Difference skewness signal half 1 and signal half 2 7 minimum 17 Difference kurtosis signal half 1 and signal half 2 8 Skewness 18 Square weighted mean 9 Kurtosis 19 Area under the curve 10 Katz Fractal Dimension 20 Magnitude Fluctuation Index (MFI) Statistical indicators from the frequency range 21 mean frequency 28 3dB power bandwidth 22 Median frequency 29 Maximum amplitude of the Welch PSD approach 23 Average Band Power 30 Position of the maximum of the Welch PSD 24 Occupied Bandwidth (99% occupied bandwidth) 31 Signal-to-noise ratio (SNR) 25 Signal performance of the Power Spectral Density (PSD) 32 Variance PSD 26 Mean one-sided spectral density (PSD) 33 Skewness PSD 27 Kurtosis PSD Statistical key figures from the time-frequency uence range 34 Energy Continuous Wavelet Transform (CWT) 37 Kurtosis CWT 35 Entropy CWT 38 Standard deviation CWT 36 Skewness CWT 39 Variance CWT

[0116] From the 39 extracted statistical indicators, statistical indicators are selected using feature selection to create a reduced fused dataset DNeu Here, the signal power in the time domain P Zeit , the signal power of the Power Spectral \DensityP Frequ , the Katz Fractility D K , the Magnitude Fluctuation Index (MFI) of the signal envelope MFI and the Signal-to-Noise Ratio (SNR) SNR Feature based on the torque-determining data D1, D2, D3, D4, D5, D6 for the joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1. The SNR describes the ratio of the power at a specific frequency to the power of the other frequencies in the spectrum of the short-time Fourier transform (STFT). The statistical key figure

[0117] SNR Feature is the average value of all SNRs and is calculated using the following formula: SNRFeature=avg{SNR(t)}={PSignal(t)∑f|STFT(t, f)|−PSignal(t)},where PSignal(t)=max{|STFT(t, f)|}, with signal-to-noise ratio SNR and short-time Fourier transform at time t and frequency f STFT(t, f).

[0118] For the statistical key figure MFI, the envelope of the motor current signal is calculated, smoothed with a 5th order Butterworth filter and then the average of the difference between the local minima and local maxima is calculated: MFI=avg{e(max i)−e(min i)|e(max i)−e(min i)2|}, with e as a smoothed envelope (|Hilbert envelope|) with a 5th-order Butterworth filter. MFI thus considers the fluctuation in the envelope and thus the change in the overall signal. It has been recognized that the statistical metrics SNR and MFI reflect errors in the motor current data, regardless of the movement performed by an industrial robot.

[0119] The Katz fractal dimension is extracted from the time series. The fractal dimension of a signal of length N is calculated by sampling the distances between the individual data points according to the following formula: DK=log(N−1)log(DL)+log(N−1),where d=max [dist(1, i) and L=∑i=1ndist(i, i+1)], with Katz Fractal Dimension D K , length of the signal N, length of the curve L (sum of the Euclidean distances between consecutive points), and planar extension of the curve d (distance between the initial data point and the other data points i), as well as calculation of the Euclidean distance between two points u and v with dist(u,v). It was recognized that the fractal dimension allows conclusions to be drawn about how complex a signal is compared to other signals.

[0120] Further statistical indicators were the signal power in the time domain P Zeitand the signal power of the power spectral density of the signal P Frequ selected: PTime=fsN⋅∑i=1NI2i and PFrequ=fsN∗∑i=1NPyyi

[0121] With a sampling rate of fs = 125 Hz, N is the number of signal data points, and Pyy is the signal's power spectral density (PSD). A total of five statistical indicators are used per joint G1, G2, G3, G4, G5, and G6 to describe the motor current data, resulting in 30 data points per measurement when using the motor current data D1, D2, D3, D4, D5, and D6 of the six joints G1, G2, G3, G4, G5, and G6.

[0122] In a third step S3, the method comprises determining whether an error or an anomaly exists in the handling or manipulation task of the 6-DOF industrial robot 1. The determination is made based on the fused data set D Neu . For this purpose, the merged data set D NeuAn unsupervised machine learning (ML) model was applied. The unsupervised machine learning (ML) model includes a one-class classification (OCC) model.

[0123] The one-class classification model OCC for application to the merged dataset D Neu in which Fig. The method shown in Figure 4a for error and / or anomaly detection in the handling task of the 6-DOF industrial robot 1 was implemented using the Fig. 4b. In a first step S1', the Fig. 4b shown procedure analogous to that in Fig. 4a, obtaining torque-determining data D1, D2, D3, D4, D5, D6 for each of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1, wherein in this case the torque-determining data D1, D2, D3, D4, D5, D6 are based exclusively on error-free movements of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1.

[0124] For training of the in the Fig. The one-class classification model OCC applied in the procedure shown in Figure 4a was applied analogously to step 2 of the Fig. 4a, in a step S2', a fused data set is obtained by means of feature-level fusion based on obtained torque-determining data D1, D2, D3, D4, D5, D6 for each of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1, wherein in this case the torque-determining data D1, D2, D3, D4, D5, D6 are based exclusively on error-free movements of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1 and the fused data set is used as training data set D Train used to train the one-class classification model OCC.

[0125] The trained one-class classification model OCC is then tested with a test dataset D TestBased on torque-determining data D1, D2, D3, D4, D5, D6 for each of the six joints G1, G2, G3, G4, G5, G6, during error-free and error-prone movements of the joints G1, G2, G3, G4, G5, G6, the precision and reliability of the one-class classification model OCC was tested. It was shown that using the five statistical indicators SNR, MFI, D K , P Zeit and P Frequ Despite the strong dimensionality reduction of the dataset, a high information content could be maintained and even when training with only these five key figures, a high level of precision and reliability could still be achieved for the one-class classification model OCC.

[0126] Subsequently, the one-class classification model OCC, which was trained and tested on the basis of the motor current data of the 6-DOF industrial robot, is saved and used analogously to the one in Fig.4a is applied to new, undefined movements of the 6-DOF industrial robot 1. The one-class classification model OCC detects errors and anomalies during the handling task of the 6-DOF industrial robot 1 and, based on the motor current data of the 6-DOF industrial robot 1, allows conclusions to be drawn as to whether the entry into the zero-point clamping system and thus the positioning and / or orientation of a component in the semiconductor technology system 100 was successful or not. The method can be applied analogously, for example, to classify an exit from a zero-point clamping system.

[0127] Fig. 5a shows a further exemplary embodiment of a method for error and / or anomaly detection in the handling task of the Fig. 2. The first step S1 of the method corresponds to step S1 of the 6-DOF industrial robot shown in Fig.3a and comprises obtaining torque-determining data D1, D2, D3, D4, D5, D6 for each of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1, wherein the torque-determining data D1, D2, D3, D4, D5, D6 comprise motor current data of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1.

[0128] In a second step S2, the method comprises determining decisions E1, E2, E3, E4, E5, E6 based on the torque-determining data D1, D2, D3, D4, D5, D6. For this purpose, the torque-determining data D1, D2, D3, D4, D5, D6 are first processed into a data set for each joint G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1. Since the torque-determining data D1, D2, D3, D4, D5, D6 are continuously acquired, the processing of the torque-determining data D1, D2, D3, D4, D5, D6 initially comprises segmenting them into individual data sequences. The torque-determining data D1, D2, D3, D4, D5, D6 are also filtered using a low-pass filter to remove noise in the data.The processing also includes interpolating at least a portion of the torque-determining data D1, D2, D3, D4, D5, D6, so that the same number of data points is available for each data sequence, which is used in particular as an input vector for a machine learning model, and preferably for each of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1.

[0129] Decisions E1, E2, E3, E4, E5, and E6 are determined based on the torque-determining data D1, D2, D3, D4, D5, and D6 using unsupervised machine learning models ML1, ML2, ML3, ML4, ML5, and ML6. The machine learning models ML1, ML2, ML3, ML4, ML5, and ML6 each represent a one-class classification model OCC1, OCC2, OCC3, OCC4, OCC5, and OCC6, in this case each a LOF model. In this case, decisions E1, E2, E3, E4, E5, and E6 are binary decisions. Using the one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, OCC6, it is determined for each joint G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1 whether a decision of the first kind or a decision of the second kind exists.A decision of the first kind means the presence of an error or an anomaly in an individual movement of the respective joint (bad decision), a decision of the second kind means that an error or an anomaly does not exist in the individual movement of the respective joint (good decision).Specifically, in the second step S2, a first decision E1 is made based on the first torque-determining data D1 using the first one-class classification model OCC1, a second decision E2 is made based on the second torque-determining data D2 using the second one-class classification model OCC2, a third decision E3 is made based on the third torque-determining data D3 using the third one-class classification model OCC3, a fourth decision E4 is made based on the fourth torque-determining data D4 using the fourth one-class classification model OCC4, a fifth decision E5 is made based on the fifth torque-determining data D5 using the fifth one-class classification model OCC5, and a sixth decision E6 is made based on the sixth torque-determining data D6 using the sixth one-class classification model OCC6.To apply the one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, OCC6, the torque-determining data D1, D2, D3, D4, D5, D6 are first prepared, wherein the preparation comprises segmenting the torque-determining data D1, D2, D3, D4, D5, D6 into individual data sequences, filtering the torque-determining data D1, D2, D3, D4, D5, D6 and finally interpolating at least a portion of the torque-determining data D1, D2, D3, D4, D5, D6, so that in particular the same number of data points is available for each data sequence used as input vector for the machine learning models and for each of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1.

[0130] In a further step S3, the decisions E1, E2, E3, E4, E5, E6 are combined into a fused decision E using a decision algorithm. Neufused, whereby determining whether there is an error or an abnormality in the handling or manipulation task of the industrial robot is based on the fused decision E Neu The decision algorithm is based on a majority decision and involves comparing the number of decisions of the first type and the number of decisions of the second type, whereby decisions E1, E2, E3, E4, E5, and E6 are each assigned to the first or second type. The decision made for the movement by most models is called the fused decision E. NeuFor the method based on the majority decision of the LOF one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, OCC6 of the six joints G1, G2, G3, G4, G5, and G6 of the 6-DOF industrial robot, an accuracy of 99.89% and an AUROC value of 99.94% were achieved. Alternatively, the decision algorithm can also include weighting of the decisions E1, E2, E3, E4, E5, and E6. For example, decisions based on the torque-determining data of joints that have a greater influence on the overall motion of the 6-DOF industrial robot 1 can be given greater weight.

[0131] The one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, OCC6 for application to the torque-determining data D1, D2, D3, D4, D5, D6 in the Fig. The method shown in Figure 5a for error and / or anomaly detection in the handling task of the 6-DOF industrial robot 1 was implemented using the Fig.5b. In a first step S1', the Fig. 5b shown procedure analogous to that in Fig. 5a, obtaining torque-determining data D1, D2, D3, D4, D5, D6 for each of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1, wherein in this case the torque-determining data D1, D2, D3, D4, D5, D6 are based exclusively on error-free movements of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1.

[0132] In a step S2', a one-class classification model OCC1, OCC2, OCC3, OCC4, OCC5, OCC6 is trained for each joint G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1 with the torque-determining data D1, D2, D3, D4, D5, D6 of error-free movements of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1. For this purpose, the torque-determining data D1, D2, D3, D4, D5, D6 of the error-free movements are initially analogous to step S2 in the Fig. 5a and a training data set D1 Train , D2 Train , D3 Train , D4 Train , D5 Train , D6 TrainBased on segmented error-free movements, the one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, and OCC6 are used to train the respective one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, and OCC6. In this way, the one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, and OCC6 learn error-free movements of the respective joints G1, G2, G3, G4, G5, and G6 of the 6-DOF industrial robot 1.

[0133] The trained one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, OCC6 are then tested with a test dataset D1 Test , D2 Test , D3 Test , D4 Test , D5 Test , D6 Test based on the respective torque-determining data D1, D2, D3, D4, D5, D6 of error-free and faulty movements of each of the six joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1 and the overall decision E obtained using decision-level fusion Gescompared with the actual result of the movement of the industrial robot 1 to estimate the precision and reliability of the one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, OCC6.

[0134] The trained and tested one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, OCC6 are saved and processed analogously to the Fig. The method shown in Figure 5a is applied to new, indeterminate movements of the 6-DOF industrial robot to determine decisions E1, E2, E3, E4, E5, E6 based on the torque-determining data D1, D2, D3, D4, D5, D6 using the one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, OCC6. The one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, OCC6 detect errors and anomalies in the movements of individual joints G1, G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1.

[0135] In detail, the procedure according to Fig.5b, in a step S1', torque-determining data D1 for the first joint G1 is initially obtained based on at least one error-free movement of the first joint G1. The torque-determining data D1 for the first joint G1 is processed by segmentation, low-pass filtering, interpolation, and / or data normalization, and based on this torque-determining data D1, a training data set D1 is created for the first joint G1. Train created. With the training dataset D1 Train A one-class classification model OCC1 is trained to learn error-free movements of joint G1. These steps are performed analogously for all other joints G2, G3, G4, G5, and G6 of the 6-DOF industrial robot 1 to obtain one-class classification model OCC2, OCC3, OCC4, OCC5, and OCC6, respectively.

[0136] To test the one-class classification model OCC1, torque-determining data D1 for error-free and faulty movements of the joint G1 are collected, processed by means of segmentation, low-pass filtering, interpolation, and / or data normalization, and based on this torque-determining data D1, a test data set D1 is created for the first joint G1. Test This is done analogously for the other joints G2, G3, G4, G5, G6 of the 6-DOF industrial robot 1, for the test data sets D2 Test , D3 Test , D4 Test , D5 Test , D6 Test be created. With the test data set D1 Test The one-class classification model OCC1 is tested with the test data set D2 Test The one-class classification model OCC2 is tested with the test dataset D3 Test The one-class classification model OCC3 is tested with the test data set D4 TestThe one-class classification model OCC4 is tested with the test dataset D5 Test The one-class classification model OCC5 is tested with the test dataset D6 Test the one-class classification model OCC7 is tested and from the individual decisions E1, E2, E3, E4, E5, E6 of the models OCC2, OCC3, OCC4, OCC5, OCC6 for each test data set D1 Test , D2 Test , D3 Test , D4 Test , D5 Test , D6 Test a total decision E Ges and compared with the actual result of the movement of industrial robot 1 to estimate the precision and reliability of the decision-level fusion with the trained models OCC1, OCC2, OCC3, OCC4, OCC5, OCC6. The trained and tested one-class classification model OCC1 is saved and can then be used analogously to the one in Fig.The procedure shown in Figure 5a can be applied to new, indeterminate movements of the 6-DOF industrial robot 1. This is done analogously for the other one-class classification models 2 OCC2, OCC3, OCC4, OCC5, OCC6.

[0137] Subsequently, the one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, OCC6, which were trained and tested on the basis of the motor current data of the 6-DOF industrial robot, are saved and processed analogously to the Fig. The method shown in Figure 5a is applied to new, indeterminate movements of the 6-DOF industrial robot 1. Using each of the six one-class classification models OCC1, OCC2, OCC3, OCC4, OCC5, and OCC6, a decision is made as to whether the movement is faulty or error-free. Based on the decision algorithm, the decisions are combined into a fused decision E using decision-level fusion. NeuThis reflects the classification of the overall movement of the 6-DOF industrial robot 1 and provides information about whether the entry into the zero-point clamping system, and thus the positioning and / or orientation of a component in the semiconductor technology system 100, was successful or not. The method can be applied analogously, for example, to classify an exit from a zero-point clamping system. The best results are achieved with a majority decision of LOF models of all six joints, with a test accuracy of 99.89% and an AUROC value of 99.94%.

[0138] Fig.Figure 6 shows a schematic block diagram of a device 200, in particular a data processing device, configured to execute the method for error and / or anomaly detection in a handling or manipulation task of an industrial robot. The device 200 comprises a processor 201, a main memory 202, an instruction memory 203, an optional data memory 204, one or more optional communication interfaces 205, and an optional user interface 206.

[0139] The device 200 can, for example, be configured to execute the method for error and / or anomaly detection during a handling or manipulation task of an industrial robot or to comprise corresponding means (in particular one of the means 201 to 208) for executing the method. The device can also be a device that comprises at least one processor (in particular processor 201) and at least one memory (preferably memory 202 and / or memory 203) with instructions, wherein the at least one memory and the instructions are configured to cause a device, e.g., the device 200, with the at least one processor, at least to execute the method for error and / or anomaly detection during a handling or manipulation task of an industrial robot.

[0140] The processor 201 may, for example, execute instructions stored in the instruction memory 203, wherein the instruction memory 203 may, for example, represent a computer-readable medium containing instructions that, when executed by the processor 201, cause the processor 201 to perform the method for error and / or anomaly detection in a handling or manipulation task of an industrial robot.

[0141] Processor 201 may be a processor of any suitable type. Processor 201 may include, but is not limited to, one or more microprocessors, one or more processors with one or more associated digital signal processors, one or more processors without associated digital signal processors, one or more special-purpose computer chips, one or more field-programmable gate arrays (FPGAs), one or more controllers, one or more application-specific integrated circuits (ASICs), or one or more computers.

[0142] For example, processor 201 may control memories 202 through 204, communication interface(s) 205, and user interface 206. Memory 202 may be, for example, volatile memory. This may be, for example, random access memory (RAM) or dynamic random access memory (DRAM), to name a few non-limiting examples. It may be used, for example, by processor 201 when executing an operating system and / or computer program.

[0143] The instruction memory 203 and / or data memory 204 may, for example, be non-volatile memory. It may, for example, be a flash memory (or a portion thereof), a ROM, PROM, EPROM, or EEPROM memory (or a portion thereof), or a hard disk (or a portion thereof), to name just a few examples. The communication interface(s) 205 enable the device 200 to communicate with other devices. The communication interface(s) 205 may, for example, comprise a wireless interface (e.g., a cellular interface, a WLAN interface, and / or a BT / BLE interface) and / or a wired interface (e.g., an IP-based interface), for example, to communicate with devices over the Internet.

[0144] The user interface 206 is optional and may include a display for displaying information to a user and / or an input device (e.g., a keyboard, a keypad, a touchpad, a mouse, etc.) for receiving information from a user.

[0145] Some or all components of the data processing device 200 can be connected, for example, via a bus. Some or all components of the data processing device 200 can be combined, for example, into one or more modules.

[0146] Fig. Figure 7 shows a schematic representation of examples of tangible and non-transitory computer-readable media, in particular storage media according to the present disclosure, which can be used, for example, to implement the main memory 202, the instruction memory 203 and / or the data memory 204. For this purpose, Fig.7 a flash memory 209, which may be soldered and / or attached to a circuit board, for example, a solid-state drive 210 with a plurality of memory chips (e.g., flash memory chips), a magnetic hard disk 211, a Secure Digital (SD) card 212, a Universal Serial Bus (USB) memory stick 213, an optical storage medium 214 (such as a CD-ROM or DVD), and a magnetic storage medium 215.

[0147] The exemplary embodiments of the present invention described in this specification are also to be understood as disclosed in all combinations with one another. In particular, the description of a feature encompassed by an embodiment—unless explicitly stated otherwise—should not be understood in this case to imply that the feature is essential or essential to the function of the exemplary embodiment.

[0148] Terms used in the claims such as "comprise," "have," "include," "contain," and the like do not exclude further elements or steps. The phrase "at least partially" encompasses both "partially" and "fully." The phrase "and / or" is intended to indicate that both the alternative and the combination are disclosed, thus "A and / or B" means "(A) or (B) or (A and B)." A plurality of units, persons, or the like, in the context of this specification, means multiple units, persons, or the like. The use of the indefinite article does not exclude a plurality. A single device may perform the functions of multiple units or devices recited in the claims. Reference numerals indicated in the claims are not to be construed as limitations on the means and steps employed. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] DE 10 2008 009 600 A1 [0073, 0077] US 2006 / 0132747 A1

[0075] EP 1 614 008 B1

[0075] US 6,573,978

[0075] DE 10 2017 220 586 A1

[0080] US 2018 / 0074303 A1

[0094]

Claims

[1] Method for error and / or anomaly detection in a handling or manipulation task of an industrial robot (1), the method comprising: - Obtaining first torque-determining data (D1) for a first joint (G1) and second torque-determining data (D2) for a second joint (G2) of the industrial robot (1), and - Determining whether there is an error or an abnormality in the handling or manipulation task of the industrial robot (1) based on the first torque-determining data (D1) and the second torque-determining data (D2). [2] Method according to claim 1, wherein the first torque-determining data (D1) comprise force, torque and / or motor current data related to the first joint (G1) and the second torque-determining data (D1) comprise force, torque and / or motor current data related to the second joint (G2). [3] The method of claim 1 or 2, wherein the method further comprises: - merging the first torque-determining data (D1) and the second torque-determining data (D2) to form a fused data set (D Neu ,D Train ,D Test ), wherein determining whether there is an error or an anomaly in the handling or manipulation task of the industrial robot (1) is based on the fused data set (D Neu ,D Train ,D Test ) takes place. [4] The method of claim 3, wherein merging the first torque-determining data (D1) and the second torque-determining data (D2) comprises: - Extracting at least one statistical key figure from the first torque-determining data (D1) and the second torque-determining data (D2), wherein the fused data set (D Neu ,D Train ,D Test ) which includes at least one statistical indicator. [5] Method according to claim 4, wherein the at least one statistical key figure is or comprises a statistical key figure from the time domain, the frequency domain and / or the time-frequency domain. [6] The method according to claim 4 or 5, wherein the merging of the first torque-determining data (D1) and the second torque-determining data (D2) further comprises: - Selecting at least one selected statistical key figure from the extracted at least one statistical key figure, wherein the at least one statistical key figure derived from the merged data set (D Neu ,D Train ,D Test ) which is or includes at least one selected statistical key figure. [7] Method according to one of claims 3 to 6, wherein determining whether an error or an anomaly is present in the handling or manipulation task of the industrial robot (1) is carried out by applying an unsupervised machine learning (ML) model to the fused data set (D Neu ,D Train ,D Test ) takes place. [8] Method according to claim 7, wherein the unsupervised machine learning model (ML) comprises a one-class classification model (OCC), in particular Local Outlier Factor (LOF), Elliptic Envelope or One-Class Support Vector Machine (OC-SVM). [9] The method of claim 1 or 2, wherein the method further comprises: - determining a first decision (E1) based on the first torque-determining data (D1) and a second decision (E2) based on the second torque-determining data (D2), and - Merging the first decision (E1) and the second decision (E2) using a decision algorithm to form a merged decision (E Ges ), wherein determining whether there is an error or an abnormality in the handling or manipulation task of the industrial robot (1) is based on the fused decision (E Ges ) takes place. [10] The method of claim 9, wherein the decision algorithm comprises weighting the first decision (E1) and the second decision (E2) or comparing the number of decisions of a first type and the number of decisions of a second type, wherein the first decision (E1) and the second decision (E2) are each assigned to the first or second type. [11] Method according to claim 9 or 10, wherein the first decision (E1) is determined using a first unsupervised machine learning model (ML1) and the second decision (E2) is determined using a second unsupervised machine learning model (ML2). [12] The method according to claim 11, wherein the first unsupervised machine learning model (ML1) and the second unsupervised machine learning model (ML2) each comprise a one-class classification model (OCC1, OCC2), in particular Local Outlier Factor (LOF), Elliptic Envelope or One-Class Support Vector Machine (OC-SVM). [13] A method according to any one of claims 1 to 12, wherein the method further comprises: - Determining whether the handling or manipulation task of the industrial robot (1) was successful based on the presence of an error or anomaly. [14] Method according to one of claims 1 to 13, wherein the handling or manipulation task of the industrial robot (1) comprises positioning and / or orienting a component (2) in a system, in particular a system for semiconductor technology (100). [15] Apparatus (200) comprising means for carrying out the method according to one of claims 1 to 14. [16] System comprising: - an industrial robot (1) with at least a first joint (G1) and a second joint (G2), and - Means for carrying out the method according to one of claims 1 to 14 for error and / or anomaly detection in a handling or manipulation task of the industrial robot (1). [17] Computer program product comprising program instructions for executing and / or controlling the method according to one of claims 1 to 14 when the program is executed on a processor (201) [18] A computer-readable storage medium (209, 210, 211, 212, 213, 214, 215) comprising a computer program product according to claim 17.

Images