Method and system for evaluating a technical plant

Abstract

A method for evaluating a technical plant is proposed. An operating data set comprising a plurality of measurement variables is recorded during an operating phase. Each measurement variable comprises a plurality of measured values, and the measured values of each measurement variable are recorded successively in time. At least one spectrum of each measurement variable of the operating data set is calculated. The calculated spectra of the operating data set are examined using a previously trained anomaly model, and a warning message is output if an anomaly in the spectra of the measurement variables is detected when examining the spectra of the operating data set. A system according to the invention for evaluating a technical plant comprises a data memory and an anomaly model. The system is configured to carry out the method according to the invention.

Description

[0001] Procedure and system for evaluating a technical plant

[0002] Description:

[0003] The invention relates to a method for evaluating a technical system, wherein an operational data set is recorded during an operating phase, comprising a plurality of measured variables, and wherein each measured variable comprises a plurality of measured values, and wherein the measured values ​​of each measured variable are recorded sequentially, and wherein at least one spectrum of each measured variable of the operational data set is calculated. The invention also relates to a system for evaluating a technical system, which is configured to carry out the method according to one of the preceding claims.

[0004] Technical systems, such as drive systems, comprise a multitude of technical components, including an electric motor, a converter or inverter to generate three-phase alternating current for the electric motor, and a gearbox to reduce the motor's speed. Examples of such technical systems include rotary tables, conveyor belts, and stacker cranes. After prolonged operation, the components of these systems can malfunction due to wear and tear.

[0005] It is known to monitor such technical systems by recording and evaluating specific measurements at set times. If the recorded measurements deviate significantly from predefined target values, a fault in the system is assumed, and a corresponding message is sent to the operator. For example, a current generated by the converter or inverter that exceeds a defined limit could indicate stiffness in the gearbox due to wear.

[0006] From EP 3 100 064 B1, a method and a device for fault detection in a machine comprising a drive unit are known. In this method, the frequency spectrum of a measured quantity characterizing the electrical power consumption of the drive unit is determined, and in an analysis step, the frequency spectrum of the measured quantity is evaluated for anomalies indicative of faults.

[0007] ISI \ EIDOPAT 22.07.2025 From DE 102006 058689 A1, a device and a method for diagnosing the condition of a machine component are known. A vibration of a machine component is detected and converted into an electrical signal, and a frequency spectrum is generated from the electrical signal.

[0008] From DE 102019 108226 A1 a motor shaft arrangement for an internal combustion engine with a motor shaft is known, which has a detection system for detecting a rotational position of the motor shaft and an angle calculation unit for determining a position signal of the motor shaft.

[0009] From EP 3 100 064 B1, a device and a method for detecting faults in a machine comprising an electric drive unit are known. The frequency spectrum of a measured quantity characterizing the electrical power consumption of the drive unit is determined, and in an analysis step, the frequency spectrum of the measured quantity is evaluated.

[0010] From DE 11 2019 007464 T5, a drive noise diagnostic system is known which includes a drive noise detection unit that detects a drive noise of an actuator or a machine to be driven, an operating state detection unit that records a drive position, a drive speed or a force generated by the drive of the actuator in a time series, a sound vibration time series spectrum acquisition unit, a feature point extraction unit and a cause determination unit.

[0011] The invention is based on the objective of further developing a method and a system for evaluating a technical plant.

[0012] The problem is solved by a method for evaluating a technical system with the features specified in claim 1. Advantageous embodiments and further developments are the subject of the dependent claims. The problem is also solved by a system for evaluating a technical system with the features specified in claim 15.

[0013] A method for evaluating a technical system is proposed. During an operational phase, an operational data set is recorded, comprising multiple measured variables. Each measured variable consists of multiple measured values, and these values ​​are recorded sequentially. At least one spectrum of each measured variable in the operational data set is calculated. The calculated spectra of the operational data set are analyzed using a previously trained anomaly model, and a warning message is issued if an anomaly is detected in the spectra of the measured variables during this analysis.

[0014] Analyzing spectra of measured quantities based on fixed, predefined criteria, such as frequency ranges or amplitude, is limited to those predefined criteria. Further criteria, which might indicate wear or defects in the technical system and are not easily defined mathematically, cannot be detected using known methods. However, examining the calculated spectra using a previously trained anomaly model allows for the detection of anomalies that are not easily defined mathematically.

[0015] According to an advantageous embodiment of the invention, the investigation of the spectra of the operational data set is carried out by a neural network configured for anomaly detection, for example an autoencoder or a variational autoencoder, a principal component analysis, a k-nearest neighbors, a one-class support vector machine, an isolation forest or ensemble methods of the preceding algorithms.

[0016] According to an advantageous embodiment of the invention, a spectrum is calculated as a frequency spectrum from the measured values ​​of each measured quantity of the operating data set.

[0017] The calculation of the frequency spectrum from the measured values ​​of the measured quantity is preferably carried out using a Fast Fourier Transform.

[0018] According to an advantageous embodiment of the invention, envelopes around the measured values ​​of each measured quantity of the operating data set are determined, and a spectrum of each measured quantity of the operating data set is calculated as an envelope spectrum from the determined envelopes.

[0019] The envelope spectrum is preferably calculated from the determined envelopes using a Fast Fourier Transform. According to an advantageous embodiment of the invention, the technical system comprises a converter or an inverter for generating an alternating voltage for an electric motor. The recorded operating data set includes rotational speed and torque as measured variables. The measured values ​​of rotational speed and torque are recorded synchronously.

[0020] The technical system in question is, for example, a drive system comprising an electric motor, a frequency converter or inverter, and a gearbox for reducing the speed of the electric motor, such as a rotary table, a conveyor belt, or a storage and retrieval machine. The aforementioned measured variables, speed and torque, are measured indirectly, for example, via the frequency converter or inverter by measuring the frequency and current of the output current to the electric motor. After prolonged operation, the components of such technical systems can malfunction due to wear and tear. Such wear can be detected early using the method according to the invention.

[0021] According to an advantageous embodiment of the invention, the operating data set is assigned to a previously defined operating point by a classifier before the spectra of the operating data set are calculated, wherein in each defined operating point the rotational speed and the torque lie within a defined range.

[0022] At such an operating point, the rotational speed, and thus the frequency of the output current, and the torque, and thus also the current intensity of the output current, are approximately constant. Operating points describe application cycles of the technical system that are regularly repeated, for example, forward travel, reverse travel, generator operation, motor operation, transport of a light load, and transport of a heavy load.

[0023] According to an advantageous embodiment of the invention, an order reduction of the recorded operating data set is performed before the spectra of the operating data set are calculated. In this process, a rotational angle is calculated from the measured rotational speed, and the torque is calculated angularly equidistant by means of interpolation. The order reduction of the recorded operating data set is performed particularly when no operating points are defined in which the rotational speed and torque lie within a defined range, i.e., when no ranges with approximately constant rotational speed and constant torque are known.

[0024] According to an advantageous embodiment of the invention, when examining the calculated spectra of the operational data set using the previously trained anomaly model, an anomaly score is determined, which is then compared with previously determined reference anomaly scores. An anomaly in the spectra of the measured quantities is detected if the determined anomaly score deviates from the previously determined reference anomaly scores.

[0025] When the anomaly model is applied to training data during a reference phase, reference anomaly scores are determined when the technical system is functioning correctly. Advantageously, by evaluating these reference anomaly scores, a model is created that classifies anomaly scores determined during the operational phase as normal or anomalous. Furthermore, the determination of anomaly scores during the operational phase is advantageously performed using cross-validation, which allows for even more specific anomaly score determination.

[0026] According to an advantageous embodiment of the invention, several reference datasets, each comprising a plurality of measured variables, are acquired during a preliminary reference phase. Each measured variable comprises a plurality of measured values, and the measured values ​​of each measured variable are acquired sequentially. At least one spectrum of each measured variable from each reference dataset is calculated, and the anomaly model is trained with the calculated spectra of the reference datasets to analyze future acquired operational datasets.

[0027] For example, a frequency spectrum is calculated from the measured values ​​of each measurand in each reference data set. For example, envelopes around the measured values ​​of each measurand in each reference data set are determined, and an envelope spectrum is calculated from the determined envelopes for each measurand in each reference data set. The calculation of the frequency spectrum from the measured values ​​of the measurand and the calculation of the envelope spectrum from the determined envelopes are preferably performed using a Fast Fourier Transform. According to an advantageous embodiment of the invention, the technical system comprises a converter or an inverter for generating an alternating voltage for an electric motor, and the recorded reference data sets each include a rotational speed and a torque as measured values. The measured values ​​of the rotational speed and torque are recorded synchronously.

[0028] According to an advantageous embodiment of the invention, a preprocessing algorithm is used to check whether the rotational speed and torque are constant in the recorded reference data sets during at least a defined time window.

[0029] The aforementioned check using the preprocessing algorithm is performed particularly when no order reduction of the recorded reference data sets is carried out before the spectra of the reference data sets are calculated. If order reduction of the recorded reference data sets is performed, a rotational angle is calculated from the measured rotational speed, and the torque is then calculated angularly equidistantly using interpolation.

[0030] According to an advantageous embodiment of the invention, if the rotational speed and torque are constant during at least one defined time window, a spectrum of the measured quantity of each reference data set is calculated solely from the measured values ​​of each measured quantity of each reference data set recorded during said time window.

[0031] According to an advantageous embodiment of the invention, reference data sets that exhibit a relatively high degree of similarity to one another are recognized as belonging to a common cluster. The reference data sets belonging to a common cluster are assigned to a common operating point, and a classifier is trained to assign future acquired operational data sets to the previously defined operating points.

[0032] At such an operating point, the rotational speed and torque are therefore approximately constant. Operating points describe application cycles of the technical system that are regularly repeated, for example, forward travel, reverse travel, generator operation, motor operation, transport of a light load, and transport of a heavy load. According to an advantageous embodiment of the invention, reference anomaly scores are determined from the anomaly model based on the calculated spectra of the reference data sets; these reference anomaly scores are comparable with anomaly scores determined in the future.

[0033] In this process, the anomaly model is applied to training data during the reference phase, thereby determining reference anomaly scores when the technical system is functioning correctly. Advantageously, by evaluating these reference anomaly scores, a model is created that classifies anomaly scores determined during the operational phase as normal or anomalous. Furthermore, the determination of the reference anomaly scores during the reference phase is advantageously performed using cross-validation, which allows for even more specific determination of the reference anomaly scores.

[0034] A system according to the invention for evaluating a technical plant comprises a data storage device and an anomaly model. The system is configured to execute the method according to the invention.

[0035] The system according to the invention allows an evaluation of the technical system based on a previously trained anomaly model, and in particular the detection of anomalies that are not easily definable mathematically.

[0036] The invention is not limited to the combination of features stated in the claims. For a person skilled in the art, further meaningful combinations of claims and / or individual claim features and / or features of the description and / or the figures will become apparent, in particular from the problem statement and / or the problem arising from a comparison with the prior art.

[0037] The invention will now be explained in more detail with reference to the illustrations. The invention is not limited to the embodiments shown in the illustrations. The illustrations only depict the subject matter of the invention schematically. They show:

[0038] Figure 1: a flowchart of the recording of a reference dataset during a reference phase,

[0039] Figure 2: a flowchart of the processing of recorded reference datasets during the reference phase and

[0040] Figure 3: a flowchart of the processing of recorded operational data records during an operational phase.

[0041] Figure 1 shows a flowchart of the acquisition of a reference dataset 10 in a technical plant during a reference phase. The reference phase is performed as a preparatory step to generate spectra of measured variables and to train an anomaly model, which is later needed for the evaluation of the technical plant.

[0042] The technical system in question is a drive system. The drive system comprises an electric motor, a converter or inverter for generating a three-phase alternating voltage for the electric motor, and a gearbox for reducing the speed of the electric motor.

[0043] During the reference phase, a reference data set 10 is recorded. This recorded reference data set 10 comprises a number of measured variables. These measured variables include, in particular, the rotational speed and torque of the electric motor. These measured variables, rotational speed and torque, are, for example, measured indirectly via the converter or via the inverter by measuring the frequency and current of the output current to the electric motor.

[0044] Each measured quantity comprises a plurality of measured values, for example, 2048 measured values. The measured values ​​of each measured quantity are recorded sequentially and synchronously. For example, the measured values ​​of each measured quantity are recorded in equidistant time intervals of 5 ms each. The recorded reference data set 10 thus comprises, for example, two different measured quantities, each with 2048 measured values. In an optional first step 101, a preprocessing algorithm is used to check whether the rotational speed and torque are constant in the recorded reference data set 10 during a defined time window. Such a time window with constant rotational speed and constant torque corresponds to a constant operating speed of the technical system, i.e., operation with a constant load and constant speed.

[0045] If such time windows with constant rotational speed and constant torque are found, only the measured values ​​of the parameters in reference data set 10 recorded during these time windows are retained. The remaining measured values, which lie outside these time windows, are deleted and not retained.

[0046] The recorded reference data set 10 is stored in a data storage device 20 in a second step 102. If measurement values ​​were deleted in the optional step 101, the deleted measurement values ​​are not stored in the data storage device 20.

[0047] During the reference phase, several more reference data sets 10 are recorded. The second step 102 is repeated, and the recorded reference data sets 10 are stored in the data memory 20. Optionally, the first step 101, in which measured values ​​are deleted, is also repeated. Another reference data set 10 is recorded when a defined trigger condition is met. Such a trigger condition is, for example, the start of an application cycle of the technical system, such as when the speed of an electric motor in the technical system exceeds a predefined minimum speed. Another such trigger condition is, for example, when the recorded measured values ​​of the measured variables in the reference data set 10 remain at least approximately constant for a period of time of, for example, 10 seconds.

[0048] Figure 2 shows a flowchart of the processing of recorded reference data sets 10 during the reference phase. The processing of the recorded reference data sets 10 according to Figure 2 takes place only once. In step 103, the previously recorded and stored reference data sets 10 are loaded from the data storage 20.

[0049] Reference data records 10, which exhibit a relatively high degree of similarity to each other, belong to a common cluster and are recognized as belonging to a common cluster. In step 104, reference data records 10, which belong to a common cluster and have been recognized as belonging to a common cluster, are assigned to a common operating point.

[0050] The clusters are automatically detected using a clustering method, such as DBSCAN. Alternatively, the clusters are detected by manually assigning an operating point. Reference data sets 10 that do not exhibit a sufficiently high degree of similarity to any of the other reference data sets 10 are discarded.

[0051] In step 105, a classifier is trained using the reference data sets 10 and the operating points assigned to the reference data sets 10 in step 104. The classifier is trained using a machine learning method to assign future recorded operational data sets 12 to the previously defined operating points. The goal is an error-free assignment of recorded operational data sets 12 to previously defined operating points.

[0052] In step 106, the classifier and the operating points assigned to the reference data sets 10 are stored in the data storage 20.

[0053] Optionally, in step 107, an order reduction of the recorded reference data sets 10 is performed. In this process, a rotational angle is calculated from the measured rotational speed, and the torque is calculated angularly equidistantly by means of interpolation.

[0054] In step 108, spectra of each measured quantity of each reference data set 10 are calculated. For example, a spectrum is calculated as a frequency spectrum from the measured values ​​of each measured quantity of each reference data set 10. Alternatively or additionally, envelopes around the measured values ​​of each measured quantity of each reference data set 10 are determined, and a spectrum is calculated as an envelope spectrum from the determined envelopes of each measured quantity of each reference data set 10.

[0055] In step 109, an anomaly model is trained using the calculated spectra of the reference data sets 10. This anomaly model is trained to analyze future recorded operational data sets 12. The anomaly model is also stored in the data storage 20. Figure 3 shows a flowchart of the processing of recorded operational data sets 12 during an operational phase. During the operational phase, the evaluation of the technical system is performed.

[0056] During the operational phase, an operational data record 12 is recorded. The recorded operational data record 12 is stored in the data storage 20 in a first step 121.

[0057] The recorded operating data set 12 comprises a number of measured variables. These measured variables include, in particular, the rotational speed and torque of the electric motor. These measured variables, rotational speed and torque, are, for example, measured indirectly via the converter or inverter by measuring the frequency and current of the output current to the electric motor.

[0058] Each measured quantity comprises a plurality of measured values, for example, 2048 measured values. The measured values ​​of each measured quantity are recorded sequentially and synchronously. For example, the measured values ​​of each measured quantity are recorded at equidistant time intervals of 5 ms each. The recorded operational data set 12 thus comprises, for example, two different measured quantities, each with 2048 measured values.

[0059] In step 122, the recorded operational data record 12 is fed to the classifier. The classifier then assigns operational data record 12 to one of the previously defined operating points. Operational data record 12 is assigned to the operating point to which it has the greatest similarity.

[0060] In step 123, an order reduction of the recorded operating data set 12 is performed. In this process, a rotational angle is calculated from the measured rotational speed, and the torque is calculated angularly equidistantly by means of interpolation.

[0061] In step 124, spectra of each measured quantity of each operational data set 12 are calculated. For example, a spectrum is calculated as a frequency spectrum from the measured values ​​of each measured quantity of each operational data set 12. Alternatively or additionally, envelopes around the measured values ​​of each measured quantity of each operational data set 12 are determined, and a spectrum is calculated as an envelope spectrum from the determined envelopes of each measured quantity of each operational data set 12. In step 125, the anomaly model is loaded from the data memory 20. In particular, the anomaly model that corresponds to the operating point to which the operational data set 12 was assigned by the classifier in step 122 is loaded.

[0062] In step 126, the spectra of the operational data set 12 calculated in step 124 are examined using the previously trained anomaly model.

[0063] In step 127, a warning message is issued if an anomaly is detected in the spectra of the measured quantities during the examination of the spectra of the operational data set 12.

[0064] Steps 121 to 127 are executed repeatedly. Specifically, these steps are repeated when a defined trigger condition is met and when a defined period of time has elapsed since the previous operational data record was recorded.

[0065] One such trigger condition is, for example, the start of an application cycle of the technical system when, for example, the speed of an electric motor of the technical system exceeds a predetermined minimum speed.

[0066] The time period is defined, for example, as six hours. Therefore, operational data records 12 are only recorded and processed four times a day.

[0067] Reference symbol list

[0068] 10 Reference data set 12 Operational data set

[0069] 20 data storage devices

[0070] 101-109 steps during the reference phase

[0071] 121-127 steps during the operational phase

Claims

Patent claims:

1. A method for evaluating a technical plant, wherein an operational data set (12) is recorded during an operational phase, comprising a plurality of measured variables, and wherein each measured variable comprises a plurality of measured values, and wherein the measured values ​​of each measured variable are recorded sequentially, and wherein at least one spectrum of each measured variable of the operational data set (12) is calculated, characterized in that the calculated spectra of the operational data set (12) are examined using a previously trained anomaly model, and that a warning message is issued if an anomaly in the spectra of the measured variables is detected during the examination of the spectra of the operational data set (12).

2. Method according to one of the preceding claims, characterized in that the investigation of the spectra of the operational data set (12) is carried out by a neural network configured for anomaly detection, for example an autoencoder or a variational autoencoder, a principal component analysis, a k-nearest neighbors, a one-class support vector machine, an isolation forest or ensemble methods of the preceding algorithms.

3. Method according to one of the preceding claims, characterized in that a spectrum is calculated as a frequency spectrum from the measured values ​​of each measured quantity of the operating data set (12).

4. Method according to one of the preceding claims, characterized in that envelopes around the measured values ​​of each measured quantity of the operating data set (12) are determined, and that a spectrum is calculated as an envelope spectrum from the determined envelopes for each measured quantity of the operating data set (12).

5. Method according to one of the preceding claims, characterized in that the technical system comprises a converter or an inverter for generating an alternating voltage for an electric motor, and that the recorded operating data set includes a speed and a torque as measured variables, and that the measured values ​​of the speed and torque are recorded synchronously.

6. Method according to claim 5, characterized in that the operating data set (12) is assigned to a previously defined operating point by a classifier before the spectra of the operating data set (12) are calculated, wherein in each defined operating point the rotational speed and the torque lie within a defined range.

7. Method according to one of claims 5 to 6, characterized in that an order reduction of the recorded operating data set (12) is carried out before the spectra of the operating data set (12) are calculated, wherein a rotation angle traveled is calculated from the measured rotational speed, and the torque is calculated angularly equidistantly by means of an interpolation. - 16 - 8. Method according to one of the preceding claims, characterized in that, when examining the calculated spectra of the operational data set (12) using the previously trained anomaly model, an anomaly score is determined which is compared with previously determined reference anomaly scores, and that an anomaly in the spectra of the measured quantities is detected if the determined anomaly score deviates from the previously determined reference anomaly scores.

9. Method according to one of the preceding claims, characterized in that several reference data sets (10) are recorded in preparation during a reference phase, each comprising a plurality of measured quantities, and that each measured quantity comprises a plurality of measured values, and that the measured values ​​of each measured quantity are recorded sequentially, and that at least one spectrum of each measured quantity of each reference data set (10) is calculated, and that the anomaly model is trained with the calculated spectra of the reference data sets (12) in order to examine future recorded operational data sets (12).

10. Method according to claim 9, characterized in that the technical system comprises a converter or an inverter for generating an alternating voltage for an electric motor, and that the recorded reference data sets (10) each comprise a rotational speed and a torque as measured variables, and that the measured values ​​of the rotational speed and torque are recorded synchronously in time.

11. Method according to claim 10, characterized in that a preprocessing algorithm is used to check whether the rotational speed and torque are constant in the recorded reference data sets (10) during at least one defined time window. - 17 - 12. Method according to claim 11, characterized in that if the rotational speed and torque are constant during at least one defined time window, a spectrum of the measured quantity of each reference data set (10) is calculated solely from the measured values ​​of each measured quantity of each reference data set (10) recorded during said time window.

13. Method according to one of claims 10 to 12, characterized in that reference data sets (10) which have a relatively high similarity to each other are recognized as belonging to a common cluster, and that the reference data sets (10) belonging to a common cluster are assigned to a common operating point, and that a classifier is trained to assign future recorded operational data sets (12) to the previously defined operating points.

14. Method according to one of claims 10 to 13, characterized in that reference anomaly scores are determined from the anomaly model from the calculated spectra of the reference data sets (12), which are comparable with anomaly scores determined in the future.

15. System for evaluating a technical plant, comprising a data storage device (20) and an anomaly model, wherein the system is configured to perform the method according to one of the preceding claims.

Images