Method for monitoring a technical device
Patent Information
- Application Number
- EP2023822250
- Authority / Receiving Office
- EP · EP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2022-12-05
- Filing Date
- 2023-12-05
- Publication Date
- 2025-10-15
AI Technical Summary
Existing condition monitoring methods using artificial intelligence face challenges in recognizing undesirable operating states due to the need for extensive labeled data, especially for rare or dangerous conditions, which is difficult and risky to obtain, and are limited by the transferability of results to varying operating conditions and design changes.
The method involves training a machine learning algorithm on a technical device or an identical copy under specifically defined operating conditions that amplify the effects of the operating state class, reducing the data requirements and enabling efficient recognition of operating states without direct measurement, using defined conditions such as vibrations or torque curves to facilitate detection of issues like belt tension or imbalances.
This approach allows for efficient training and detection of operating states with reduced data effort and computing time, enabling early recognition of potential failures without exposing the device to hazardous conditions, and can be applied to various operating state classes like belt tension, imbalances, and wear, improving the reliability and safety of condition monitoring.
Smart Images
Figure 1.1
Abstract
Description
[0001] Procedure for monitoring a technical device
[0002] Description
[0003] The invention relates to a method for monitoring a technical device using artificial intelligence for operating states of a specific operating state class. The technical device can be, in particular, a machine and / or a system.
[0004] In the case of such technical equipment, monitoring of the technical equipment during its operation is carried out, in particular within the framework of so-called condition monitoring procedures, which serve to identify operating states of certain operating state classes.
[0005] Operating condition classes typically represent classes of undesirable operating conditions whose occurrence impairs or threatens to impair operation. A classic application, for example, is the early detection of wear conditions that could potentially lead to the failure of an element of the technical equipment. Depending on the scenario, the latter can then lead to damage – possibly of considerable magnitude – to the technical equipment and / or other property damage and / or even personal injury.
[0006] Methods of the type in question provide for monitoring using artificial intelligence. The use of artificial intelligence is considered to have the advantage of making it possible to detect the occurrence of a specific operating condition without precise knowledge of causal relationships—i.e., without knowing exactly how, for example, a wear condition affects a specific measured variable.
[0007] In such methods, at least one operating variable is monitored to detect the operating state and evaluated using a K1 algorithm, i.e., an artificially intelligent algorithm. In order to enable this K1 algorithm to detect operating states of the specific operating state class based on the at least one monitored operating variable, the K1 algorithm must be trained in practice. In particular, machine learning and / or deep learning methods are used for this purpose. In particular, a classification of operating states is performed.
[0008] This is where considerable difficulties arise in practice. In theory, training a KL algorithm is straightforward. The algorithm processes so-called labeled measurement data of at least one monitored operating variable. The labels specifically relate to the occurrence of the operating state classes to be recognized. During the training process, the KL algorithm can establish relationships between the monitored operating variable and the operating state classes.
[0009] The problem here, however, is that a sufficient amount of appropriately labeled data must be available. The nature of the problem—monitoring the technical equipment for the occurrence of operating states belonging to highly undesirable operating state classes—causes the main challenge. It is technically almost impossible to provide those labeled data sets whose generation is associated with the occurrence of an operating state of the corresponding operating state class.
[0010] For example, if a loose belt in a drive system is to be detected due to wear-related insufficient belt tension, the data sets would have to be recorded in a facility that specifically uses such a belt. However, this essentially involves accepting the very hazards that condition monitoring is intended to detect and avert early, ideally before they even occur.
[0011] If one takes into account that the operating states can occur under a wide variety of operating conditions, which can also affect the monitored operating variables, it becomes obvious that enormously extensive data must be available for training the Kl algorithm in order to ensure reliable detection of the operating state.
[0012] For example, it is possible that the operating condition class refers to the occurrence of insufficient belt tension on a drive belt. However, the corresponding drive can be operated under a wide variety of operating conditions, for example with regard to the torque and / or speed of the drive. A recorded operating variable, such as the energy content of the drive belt, which is used to identify the operating variable in question, is also influenced by torque and / or speed. Therefore, data with insufficient belt tension would also have to be available for a wide variety of torques and speeds; this data should therefore continue to be labeled according to the speeds and torques to enable appropriately targeted training.
[0013] In practice, the application of AI algorithms in condition monitoring is therefore subject to limitations. While it is possible to optimize AI algorithms to achieve results even in cases of insufficient data, this works within narrow limits; in particular, the influences of different operating conditions and / or design changes affect the transferability of the results. Therefore, even the slightest changes to the technical equipment or operating conditions often significantly limit the applicability of AI algorithms.
[0014] Alternatively, a larger amount of data, especially diverse data, can be collected using complex sensor technology. However, this essentially negates the advantage of AI algorithms to detect operating conditions "indirectly" based on specific operating variables. In practice, it is often easier and more cost-effective to measure, for example, the tension of a belt directly using a corresponding sensor system than to collect data using a large number of sensors and then evaluate it using artificial intelligence.
[0015] The invention is therefore based on the object of demonstrating a method for monitoring a technical device for operating states of a specific operating state class by means of artificial intelligence, in which the disadvantages described above do not occur or at least occur to a lesser extent.
[0016] The object is achieved by a method having the features of independent claim 1. The features of the dependent claims relate to advantageous embodiments.
[0017] The method for monitoring a technical device provides that the technical device is monitored using artificial intelligence for operating states of a specific operating state class. The operating state can in particular be an operating state that does not occur or at least should not occur during normal operation of the technical device. In this context, an operating state class refers in particular to a category of similar operating states. The method can also provide for the monitoring of the technical device for operating states of a plurality of specific operating state classes. For the sake of simplicity, however, only the singular "operating state class" is used below. To detect the operating states, at least one operating variable is monitored and evaluated using a Kl algorithm.This means in particular that an operating variable is measured and data concerning the operating variable and in particular its temporal development are obtained from the measured values, which are then evaluated using the Kl algorithm.
[0018] The method provides for the training of the KL algorithm through machine learning to enable the KL algorithm to recognize operating states of the specific operating state class based on the at least one monitored operating variable. The training can be performed on the same technical device and / or on a technical device that is at least substantially identical to the monitored technical device.
[0019] The technical equipment can be a machine or a system. Such machines or systems are usually designed and constructed relatively individually for a specific application; in other cases, they may be mass-produced products.
[0020] Particularly in cases where the technical device is a series product, it can be advantageous to use a technical device that is at least essentially identical in construction to the technical device to be monitored for teaching the Kl algorithm. In this way, the technical device to be monitored does not have to be available for teaching operation. A technical device that is at least essentially identical in construction is to be understood in particular as a technical device in which minor structural deviations from the technical device to be monitored do not affect the at least one operating variable to be monitored in such a way that the recognition of the operating states is prevented or at least noticeably made more difficult.Particularly in cases where the technical device is manufactured as a single specimen for a specific application scenario, it is advantageous to train the AI algorithm on the technical device to be monitored. In such cases, a comparable technical device would first have to be found, and then complex series of tests would have to be conducted to determine the extent to which the behavior of the operating variable is comparable depending on the occurrence of the operating states. Therefore, in such a scenario, it is more sensible to train the AI algorithm directly on the same technical device that is being monitored as part of the process.
[0021] Through training, the algorithm is enabled to recognize operating states of the specific operating state class based on the at least one monitored operating variable. Training is therefore carried out, in particular, with labeled data, which is labeled according to the occurrence of operating states of the specific operating state class. Such training with labeled data is also referred to as supervised learning.
[0022] The problem is solved, in particular, by specifically selecting and / or enacting defined operating conditions for training the class algorithm, which enable and / or facilitate the recognition of operating states of the specific operating state class through the evaluation of at least one operating variable. The operating conditions can, in particular, be a specific trajectory of a moving component of the technical device, which is specifically selected and / or enacted.
[0023] It has been shown that the recognition of operating states of a specific operating state class can be enabled and / or facilitated by the evaluation of at least one operating variable using a Kl algorithm by specifically selecting and / or bringing about defined operating conditions for training the Kl algorithm. Advantageously, these are operating conditions in which operating states of the specific operating state class have a greater impact on the at least one operating variable than under other operating conditions, in particular a greater impact than under the operating conditions of normal operation of the technical device. This makes it easier to recognize relationships between the at least one operating variable and the occurrence of an operating state of the specific operating state class when training the Kl algorithm. The amount of data required to train the Kl algorithm can thus be drastically reduced.In practice, this means that the generation of data, which is carried out in particular by means of a corresponding training operation of the technical device to be monitored and / or a technical device that is at least essentially identical in design, requires significantly less effort, particularly in terms of data volume and computing time. This enables efficient training of the AI algorithm.
[0024] The method can, in particular, provide for the defined operating conditions that enable and / or facilitate the detection of operating states of the specific operating state class to be specifically brought about during operation of the technical device in order to check for the presence of an operating state of the specific operating state class. This can enable and / or facilitate the detection of the operating states of the specific operating state class through the evaluation of at least one operating variable. This not only facilitates the training of the Kl algorithm, but also facilitates and / or enables the detection of the operating states during regular operation of the technical device.
[0025] The method can, for example, provide for the defined operating conditions to be repeatedly and specifically brought about during the normal operation of the technical device in order to detect operating states of the specific operating state class. The repeated bringing about can, for example, take place after certain time intervals and / or after the occurrence of certain events, such as when a technical device is started up and / or after a certain operating period. In this case, the targeted bringing about of the operating conditions is to be understood in particular as meaning that the operating conditions do not correspond to the intended normal operation of the technical device, but are brought about solely for the purpose of carrying out the method for monitoring the technical device.
[0026] The method can in particular provide for a test operation to be carried out in order to train the Kl algorithm. During the test operation, data is acquired and analyzed with regard to the effect of operating states of the specific operating state class on operating variables in order to determine the at least one operating variable and / or the defined operating conditions. In other words, the test operation serves to find out, for an operating state class, which operating variable an operating state of this operating state class affects in such a way that this operating variable is suitable for recognizing the operating state. Alternatively and / or additionally, the test operation can be used to determine operating conditions under which operating states of the specific operating state class affect the at least one operating variable in such a way that recognition of an operating state of the specific operating state class is enabled and / or facilitated.The evaluation of data with regard to the impact of operating states of a specific operating state class on operating variables is also carried out primarily using machine learning. This approach is also referred to as "feature engineering."
[0027] The method can in particular provide that the defined operating conditions are operating conditions of an electric drive of the technical device. The operating conditions of drive systems of technical devices of the type in question can usually be influenced directly and thus comparatively precisely and reproducibly. This makes drives of technical devices particularly well suited to the implementation of defined operating conditions which enable and / or facilitate recognition of the operating states in question. For example, the defined operating conditions can be a torque and / or a speed, a torque curve and / or a speed curve (i.e. a time-dependent function of the speed or torque) and / or a controller setting.
[0028] The method can provide that the at least one operating variable is an operating variable of an electric drive system of a technical device. Operating variables of technical drive systems are particularly suitable because they can be measured comparatively well. The operating variable can be measured directly or the operating variable can be derived from the measurement of other operating variables. A further advantage is that operating variables are often measured anyway for the control and / or regulation of the electric drive system. Advantageously, the at least one operating variable is therefore an operating variable that is measured anyway as part of the control and / or regulation of the drive system. In this case, no additional sensors are therefore required to carry out the method.
[0029] The method can provide that the operating condition class is the occurrence of insufficient belt tension on a drive belt. The drive belt can be a toothed belt, for example. Loose toothed belts represent a typical operating condition class, the detection of which is a desirable scenario within the framework of condition monitoring. In connection with the monitoring for operating conditions in the operating condition class of the occurrence of insufficient belt tension, the method can provide that the defined operating conditions include the superimposition of a torque curve of a drive driving the drive belt with a vibration. It has been shown that such a superimposition of a torque curve with a vibration leads to operating conditions that significantly facilitate the detection of a loose drive belt or even make it possible in the first place.
[0030] The vibration can, in particular, have a frequency that corresponds to a natural frequency of the drive belt. While it is typically avoided during operation to excite belts with natural frequencies, it has been found in connection with the present invention that excitation with natural frequencies can lead to a clear detection of insufficient belt tension.
[0031] The vibration can be excited, in particular, with a mixed-frequency signal. The mixed-frequency signal can, in particular, comprise a plurality of superimposed frequencies, which preferably contain components of the natural frequencies of the drive belt. In this way, sufficient excitation of the drive belt can be ensured by the mixed-frequency signal. It has been shown in this context that, in particular, the natural frequencies of loose drive belts change.
[0032] The vibration can have a frequency of at least 10 Hz and / or at most 1000 Hz. It has been shown in practice that, for the detection of drive belts with insufficient belt tension using the described method for monitoring a technical device, vibrations in this frequency range lead to particularly good detection of loose drive belts. This is particularly the case when the drive belt is a toothed belt. In this case, natural frequencies typically occur in this frequency range when the meshing of the teeth with the wheels leads to insufficient positive engagement due to the loose toothed belt, and the teeth are therefore deformed.
[0033] In connection with the detection of the occurrence of insufficient belt tensions on drive belts, the at least one operating variable can be in particular the frequency curve of the trajectory of the drive belt, i.e. in particular vibrations of the rotational speed and / or the speed of the drive.
[0034] The frequency response can be determined by measuring an operating variable of the drive system, such as the rotational speed and / or velocity. In particular, a frequency analysis, for example, using the Fast Fourier Transform and / or a Wavelet Transform, can be used.
[0035] In practice, it has been shown that the relationship between belt tension and the vibration behavior of the drive belt is complex. Therefore, this application of the method is particularly well suited for the use of a Kl algorithm, since a direct measurement of the condition of the drive belt via the vibration behavior, i.e., a derivation via physical relationships, without the application of a trained artificial intelligence, is not possible in practice due to the complexity. However, particularly when monitoring a drive belt according to the described method, it may be sufficient to consider only the case of sufficient belt tension for training the artificial intelligence, i.e., using the described method, the Kl algorithm can be enabled to detect insufficient belt tension in drive belts without the need for a learning operation with insufficient belt tension and the associated hazards. The operating condition class can be the occurrence of an imbalance in a drive system. Such imbalances occur in a drive system when a component of the drive system, such as a drive shaft, is unbalanced. The occurrence of unbalances also represents an operating condition class, which can have a wide variety of characteristics in practice. For this reason, the described method using a Kl algorithm is particularly advantageous for detecting unbalances.
[0036] Alternatively and / or additionally, the operating condition class can refer to the occurrence of wear on a friction wheel in a drive system. Friction wheels in drive systems can also wear out, leading to an operating condition in the "friction wheel wear" operating condition class. Even when detecting such friction wheel wear, the use of trained KL algorithms is useful for detecting friction wheel wear due to the varying effects that friction wheel wear can have on actual operating variables.
[0037] In order to enable and / or facilitate the detection of operating states of the operating state classes of the occurrence of an imbalance and the occurrence of wear on a friction wheel, the defined operating conditions can include, in particular, a constant speed of the drive system and / or the reduction of the integral component of the control of the drive system compared to the integral component used in normal operation.
[0038] A constant speed makes it possible to produce noticeable effects on operating parameters when an imbalance occurs or when wear occurs on a friction wheel.
[0039] By reducing the integral component of the drive system control compared to the integral component used in control operation, it is possible to prevent the integrating nature of the control from leading to the effects of the occurrence of the imbalance or wear of the friction wheel on at least one operating variable being reduced by the control or being influenced in a way that makes detection more difficult.
[0040] In particular, when using the method for detecting operating conditions in the operating condition class of the occurrence of an imbalance in a drive system, the at least one operating variable can be the following error of the drive system. The following error of the drive system is in particular the difference between a target position and an actual position of the drive. The target position is in particular the position specified by a control and / or regulation of the drive. The actual position of the drive is in particular measured at the drive, for example with a resolver. For control purposes, such a measurement of the actual position at a drive is often necessary anyway, i.e. the use of the following error as at least one operating variable when carrying out the method often means that no additional sensors are required to detect wear.
[0041] In particular, when using the method for detecting operating conditions in the operating condition class of the occurrence of wear on a friction wheel of a drive system, the at least one operating variable can be the torque of the drive system. The torque of the drive system is also an operating variable that is often measured as part of the control and / or regulation of a drive. The torque can be measured directly using a torque sensor or, alternatively and / or additionally, as a derived variable from an electrical operating variable of the drive system. In this case, the electrical operating variable is measured and the torque is calculated from this. The operating condition class can be the occurrence of a change in the inertia of a device driven by a drive system.
[0042] The change in inertia may be due to damage to a part of the device. For example, this could be a broken tool on a device. The breaking of a tool changes the inertia of the device. This can be used, for example, to detect damaged cutters. Alternatively and / or additionally, the change in inertia may be due to material consumption. For example, the remaining amount of unwindable paper on a roll from which paper is being unwound can be determined.
[0043] Particularly when using the method for detecting operating conditions in the operating condition class of the occurrence of a change in inertia, the defined operating conditions can include a constant acceleration of the drive system. If a drive system is operated with a constant acceleration, changes in inertia can be detected particularly well based on at least one operating variable.
[0044] Further practical embodiments and advantages of the invention are described below in conjunction with the drawing. It shows:
[0045] Fig. 1 is a simplified schematic process flow diagram of an exemplary process.
[0046] The exemplary method 10 for monitoring a technical device using artificial intelligence for operating states of a specific operating state class can provide, as in the example shown, that a test operation 12 is initially carried out. During the test operation 12, data can be acquired and analyzed with regard to the effect of operating states of the specific operating state class on operating variables. As a result, by means of the test operation 12, at least one operating variable can be determined which is affected by operating states of the specific operating state class. Alternatively and / or additionally, defined operating conditions can be determined under which operating states of the specific operating state class affect the operating variable in a way that enables and / or facilitates recognition of the operating states.
[0047] The method 10 provides for a training 14 of the Kl algorithm, wherein specifically defined operating conditions are selected and / or brought about, which enable and / or facilitate the recognition of operating states of the specific operating state class by evaluating the at least one operating variable. During the subsequent operation 16 of the technical device, according to the exemplary method 10, the at least one operating variable is monitored and evaluated by means of a Kl algorithm in order to recognize operating states of the specific operating state class. In this case, the method 10 can provide that during the operation 16 of the technical device, in order to check the presence of an operating state of the specific operating state class, the defined operating conditions, which can in particular have been determined during a test operation 12, as in the example shown, are specifically brought about.By deliberately bringing about the defined operating conditions, the recognition of the operating states of the specific operating state class is enabled and / or facilitated by the evaluation of at least one operating variable.
[0048] The features of the invention disclosed in the present description, the drawings, and the claims may be essential, both individually and in any combination, for the realization of the invention in its various embodiments. The invention is not limited to the described embodiments. It may be varied within the scope of the claims and taking into account the knowledge of the person skilled in the art.
[0049] List of reference symbols
[0050] 10 procedures
[0051] 12 Test operation
[0052] 14 Training 16 Operation of the technical equipment
Claims
Patent claims 1 . Method (10) for monitoring a technical device, in particular a machine and / or a system, by means of artificial intelligence for operating states of a specific operating state class, wherein at least one operating variable is monitored and evaluated by means of a Kl algorithm to detect the operating states, wherein a training (14) of the Kl algorithm is carried out by machine learning in order to enable the Kl algorithm to detect operating states of the specific operating state class on the basis of the at least one monitored operating variable, characterized in that for the training (14) of the Kl algorithm, specifically defined operating conditions are selected and / or brought about, which enable and / or facilitate detection of operating states of the specific operating state class by evaluating the at least one operating variable.
2. Method (10) according to claim 1, characterized in that during the operation (16) of the technical device for checking the presence of an operating state of the specific operating state class, the defined operating conditions are brought about in a targeted manner in order to enable and / or facilitate the recognition of the operating states of the specific operating state class by evaluating the at least one operating variable.
3. Method (10) according to claim 1 or 2, characterized in that a test operation (12) is carried out for learning (14) the Kl algorithm, wherein during the test operation (12) data are obtained and analyzed with regard to the effect of operating states of the specific operating state class on operating variables in order to determine the at least one operating variable and / or the defined operating conditions.
4. Method (10) according to one of the preceding claims, characterized in that the defined operating conditions are operating conditions of an electric drive of the technical device and / or of the at least one operating variable is an operating variable of an electric drive system of the technical device. Method (10) according to one of the preceding claims, characterized in that the operating condition class is the occurrence of insufficient belt tension on a drive belt. Method (10) according to claim 5, characterized in that the defined operating conditions include the superimposition of a torque curve of a drive driving the drive belt with an oscillation. Method (10) according to claim 6, characterized in that the oscillation has a frequency of at least 10 Hz and / or at most 1000 Hz.Method (10) according to one of claims 5 to 7, characterized in that the at least one operating variable is the frequency curve of the trajectory of the drive and / or oscillations of the rotational speed and / or the speed of the drive. Method (10) according to one of the preceding claims, characterized in that the operating condition class is the occurrence of an imbalance in a drive system. Method (10) according to one of the preceding claims, characterized in that the operating condition class is the occurrence of wear on a friction wheel in a drive system. Method (10) according to claim 8 or 9, characterized in that the defined operating conditions include a constant rotational speed of the drive system and / or the reduction of the integral component of the control of the drive system compared to the integral component used in normal operation, in particular the omission of the integral component.Method (10) according to one of claims 9 or 11, characterized in that the at least one operating variable is the drag error of the drive system.
13. Method (10) according to one of claims 10 or 11, characterized in that the at least one operating variable is the torque of the drive system.
14. Method (10) according to one of the preceding claims, characterized in that the operating state class is the occurrence of a change in the inertia of a device driven by a drive system.
15. Method (10) according to claim 14, characterized in that the defined operating conditions include a constant acceleration of the drive system.