Method for predicting a remaining lifetime of a vacuum pump

Abstract

Method for operating a vacuum pump, including acquiring vibration data of the vacuum pump during normal operation, training an anomaly detection model by the acquired vibration data of the vacuum pump during normal operation, and when the anomaly detection model determines a degrading operation of the vacuum pump, predicting by a lifetime prediction model an estimated remaining lifetime of the vacuum pump on the basis of vibration data acquired during degrading operation of the vacuum pump.

Description

[0001] METHOD FOR PREDICTING A REMAINING LIFETIME OF A VACUUM PUMP

[0002] The present invention relates to a method for operating a vacuum pump, a controller for a vacuum pump implementing such method, a system comprising a controller, a remote server implementing such method and a vacuum pump system. In particular, the present invention relates to a method for providing lifetime prediction of vacuum pumps and lifetime prediction.

[0003] Background

[0004] Common vacuum pumps comprise a housing defining a pump chamber and having an inlet and an outlet. At least one rotor assembly is disposed in the pump chamber and rotated by an electromotor. The rotor assembly comprises at least one rotor element, wherein by rotating the rotor assembly, a gaseous medium is conveyed from the inlet of the vacuum pump towards the outlet. Therein, different types of vacuum pumps exist such as turbomolecular pumps or in general molecular drag pumps, or two-rotor pumps such as screw pumps, roots pumps or claw pumps having at least two rotor shafts rotating synchronously, or single shaft pumps such as rotary vane pumps or scroll pumps. All of these vacuum pumps have fast rotating rotor assemblies, which are supported by different kinds of bearings. In particular, roller bearings are cost efficient and provide a sufficient lifetime. However, since they are not contact-free, roller bearings are prone to seizure and degrade over time. In particular, individual operation states of the vacuum pump have an influence on the speed of degradation, which cannot or can be hard to predict by the manufacturer. On the other side, a failure of a bearing of the vacuum pump may have catastrophic consequences which may destroy the vacuum pump. Thus, the service intervals of the bearings are selected precautionary shorter than necessary to prevent damage but increasing the costs for services and reducing the operating times due to interruptions by maintenance and service. In addition, also cryopumps may have moving elements which are prone to degradation. Thus, also here, knowledge about the specific remaining lifetime might be beneficial.

[0005] Early failure detection in a vacuum pump is crucial in preventing system failure, increasing safety, and reducing maintenance costs. Currently, health state detection methods are based on threshold setting. If a certain parameter exceeds a predetermined threshold, maintenance is scheduled and carried out. However, these methods suffer from high false alarm rates and require complete domain knowledge in all environments. In addition, a remaining lifetime cannot be determined using existing methods.

[0006] It is an object of the present invention to provide a method and a systems for more reliable prediction of bearing failure in a vacuum pump.

[0007] The problem is solved by a method according to claim 1 , a software storage product according to claim 12, a controller according to claim 13, a system according to claim 14 and a vacuum pump system according to claim 15.

[0008] According to the present invention, a method for operating a vacuum pump comprises the steps of: acquiring vibration data of the vacuum pump during normal operation, using the acquired vibration data to train an anomaly detection model of the vacuum pump during normal operation, and when the anomaly detection model determines a degrading operation of the vacuum pump based upon the acquired vibration data, predicting, using a lifetime prediction model, preferably a machine learning-based lifetime prediction model, a remaining lifetime of the vacuum pump on the basis of vibration data acquired during degrading operation of the vacuum pump.

[0009] A vacuum pump typically comprises a housing and at least one rotor assembly disposed in the housing and rotated by an electromotor. The rotor assembly typically comprises at least one rotor element interacting with a stator of the vacuum pump and / or a rotor element of a further rotor assembly also disposed in the housing to convey a gaseous medium from an inlet to an outlet and generate a vacuum at the inlet. Therein, each rotor assembly is rotatably supported by bearings, wherein at least one bearing might be built as a roller bearing such as a ball bearing. Alternatively, the vacuum pump may be a cryo-pump having moving parts such as a displacer.

[0010] During normal operation of the vacuum pump, vibration data may be acquired by a vibration sensor. The vibration sensor may be connected to a housing of the vacuum pump. In particular, the vibration sensor may be connected to the housing of the vacuum pump in the vicinity of one of the bearings. In particular, if the vacuum pump comprises more than one roller bearing, more than one vibration sensor could be implemented for acquiring vibration data directly from each of the roller bearings. However, in another implementation, even if the vacuum pump comprises more than one roller bearing, the vibration data could be acquired by only a single vibration sensor. In particular, the vibration sensor could be built as an acceleration sensor for detecting the acceleration of the vacuum pump along one axis, more than one axis and preferably along all three axes, as vibration data.

[0011] Alternatively, the vibration data can be acquired via the motor current. In particular, current ripples may be used as vibration data. Thus, no additional sensor would be required. This is in particular described in TW201638472.

[0012] Using the acquired vibration data, an anomaly detection model is trained during normal operation of the vacuum pump. Therein, it is assumed that in the first hours of operation of a vacuum pump the vacuum pump is operating normally and exhibiting vibration associated with normal operation. If there is an anomaly detected by the anomaly detection model, i.e. a deviation from the normal operation is detected by the anomaly detection model, a degrading operation of the vacuum pump is determined. Therein, the anomaly detection model will be trained to only detect degrading operation, if there is a repeated deviation from normal operation, instead of a transient event introduced by an unusual event such as an external knock to the pump or instrument.

[0013] The anomaly detection algorithm may be a combination of machine learning or deep learning models and statistical based models.

[0014] As soon as degrading operation of the vacuum pump is determined, a lifetime prediction model estimates a remaining lifetime of the vacuum pump on the basis of vibration data acquired during degrading operation of the vacuum pump. Thus, starting with determining degrading operation of the vacuum pump, a remaining lifetime is estimated for the at least one bearing by the lifetime prediction model on the basis of the vibration data acquired during degrading operation of the vacuum pump. Thus, a more reliable prediction of the remaining lifetime of the vacuum pump and in particular the one or more roller bearings can be provided. Service or repair of the vacuum pump can be planned on the basis of the estimated remaining lifetime, providing the benefit of increased service intervals, which are adapted to and considering the actual operation conditions of the vacuum pump and need not to be set in advance so as to include a safety margin.

[0015] In particular, the present method does not rely on comparison of the acquired vibration data with a threshold. Instead, prediction of an estimated remaining lifetime is carried out, which is independent from any threshold. In particular, a threshold usually needs to be predetermined. Thus, again, a safety margin must be included in order to encompass all different operation states of the vacuum pump and further includes some safety margin. This would again lead to unnecessarily short service intervals. Hence, omitting any threshold when predicting the estimated remaining lifetime, optimal length of the service intervals can be achieved without the risk of bearing failure for the vacuum pump. No complete domain knowledge of all environments and the operation state are necessary since the anomaly detection model is trained by the vacuum pump itself and also the lifetime prediction model may not need full knowledge.

[0016] Preferably, if degrading operation is determined, a warning signal is generated. Thus, upon the determination of degrading operation by the anomaly detection model a user, operator, manufacturer or any other party may be informed about the degrading operation of the vacuum pump. The warning signal may be an optical signal, which may be shown on a display, an acoustical signal or can be otherwise indicative of the degrading operation of the vacuum pump. Further, the warning signal may be a digital signal sent to a connected controller or to a cloud via a gateway. Therein, the present invention is not limited to the location of the warning signal. Thus, a warning signal can be generated by the vacuum pump itself or by a controller connected to the vacuum pump. The warning signal can be transmitted via any data transmission line to a remote server, web interface or the like. Thus, it is possible to collect warning signals of different vacuum pumps at one location and provide an operator with a better overview over the status of its vacuum pumps.

[0017] Preferably, the warning signal contains lifetime information, i.e. information indicating the predicted remaining lifetime of the vacuum pump. Hence, the warning signal may be, for example, a text message indicating the remaining lifetime of the respective vacuum pump. Additionally, or alternatively, the warning signal may contain action information for the user, customer or operator. This action information may specify one or more expected actions and a time when the action can be expected, such as shut down of the vacuum pump after a certain time, when the remaining lifetime is below a certain threshold. Additionally, or alternatively, the action information may indicate the next actions to be performed by the user, customer or operator, such as a contact manufacturer or service provider, such as schedule a service appointment within a certain time period. Additionally, or alternatively, the action information may contain a link, such as a http-link, to a website providing further information regarding operation of the vacuum pump.

[0018] Preferably, if the estimated remaining lifetime is smaller than a predetermined threshold a warning signal is generated or the vacuum pump is shut down. Thus, if the estimated remaining lifetime is smaller than the predetermined threshold, severe risk of bearing failure exists. Thus, a warning signal is generated. Alternatively, or additionally, the vacuum pump may be shut down for safety reasons in order to prevent damage to the vacuum pump.

[0019] Preferably, the determination of the operation mode of the vacuum pump is performed repeatedly by the anomaly detection model at time intervals, for example, of between 5 seconds and 20 minutes, more preferably between 10 seconds and 10 minutes and more preferably between 30 seconds and 2 minutes. The duration of the time intervals may be fixed. Alternatively, the period and / or duration of the time intervals may change over time and may become shortened with increasing running time of the vacuum pump and / or if the vacuum pump is operating in a harsh environment. In particular, adaptive time intervals could be used to save storage space, for example: if the pump is running in standby mode (lower rotational speed) or is running with low gas load and within operational limits a longer measurement interval could be used. When running at high load or hot or near the pump operational envelope, a faster measurement rate could be used.

[0020] Preferably, the prediction of the estimated remaining lifetime may be carried out repeatedly in time intervals of between 5 seconds and 20 minutes, more preferably between 10 seconds and 10 minutes and more preferably between 30 seconds and 2 minutes. In particular, the intervals are fixed and equal over the whole time. Alternatively, the intervals may change over time and may be shortened with decreasing estimated remaining lifetime and / or if the vacuum pump is operating in a harsh application. In particular, the intervals are shorter than during normal operation. Hence, accurate prediction regarding the estimated remaining lifetime can be provided and can be used in order to schedule and plan service and repair of the vacuum pump. In particular, necessary safety margins in the service intervals can be omitted or at least reduced and the service intervals can be planned in accordance with the actual use of the vacuum pump, i.e. the actual seizure and degradation of the vacuum pump.

[0021] Preferably, before transferring the acquired vibration data to the anomaly detection model and / or the lifetime prediction model, the vibration data is transferred to the frequency domain. Different methods exist for transferring the vibration data to the frequency domain such as Fast Fourier Transform (FFT), Discrete Fourier Transform (DFT), Short Time Fourier Transform (STFT), Wavelet Transform or the like. The present invention is not limited to the specific type of transforming the vibration data to the frequency domain. However, in order to determine degrading operation of the vacuum pump and / or predicting the estimated remaining lifetime vibration data in the frequency domain is used.

[0022] Preferably, before transferring the acquired vibration data to the lifetime prediction model, monotonic features or monotones are determined from the acquired vibration data and in particular from the transferred vibration data in the frequency domain. The monotonic features are transferred to the lifetime prediction model. Monotones or monotonic features are a monotonic input information for the lifetime prediction model. Therein, the monotones of the vibration data show a consistent relationship with the target variable, i.e. the estimated remaining lifetime. Estimated remaining lifetime may be a monotonic decreasing function over time. Since the vibration data is transformed into a monotonic function over time (increasing or decreasing), improved correlation of the monotones of the vibration data as input information to the estimated remaining lifetime is possible by the lifetime prediction model. Hence, when determining monotonic features to be used in the lifetime prediction model, prediction accuracy of the estimated remaining lifetime is improved. Therein, the monotones may be determined by a threshold, wherein the monotones are monotonously increased or monotonously decreased each time a vibration data point or the difference between two subsequent vibration data points is above the threshold. Preferably, operating data is acquired during normal operation and / or degrading operation, wherein the operating data is transferred to the anomaly detection model and / or the lifetime prediction model. Thus, if operating data is acquired during normal operation, the operating data is used in the anomaly detection model in order to detect the degrading operation. Similar, if operating data is acquired during degrading operation, the operating data is used by the lifetime prediction model to predict the estimated remaining lifetime. Thus, the operating data provides further insights into the operation state of the vacuum pump. Consequently, through using operating data for determining degrading operation or predicting the estimated remaining lifetime, accuracy is increased.

[0023] Preferably, operating data is repeatedly acquired synchronously to determine the operation mode by the anomaly detection model and / or the prediction of the estimated remaining lifetime. Thus, actual operation data is used in the anomaly detection model and / or the lifetime prediction model.

[0024] Preferably, the acquired operating data comprises one or more of the motor power, the rotation speed of the vacuum pump, the operation time or pump running hours, environmental temperature at the location of the vacuum pump, temperature of the vacuum pump during operation, orientation of the vacuum pump, motor current, motor temperature, rotor temperature, bearing temperature, motor voltage, pump cycle, humidity, and pressure, or other information of connected sensors, pressure gauges and accessories. All these parameters may have an influence on the lifetime of the vacuum pump bearings, which are consequently considered as operating data during detection of the degrading operation or predicting of the estimated remaining lifetime.

[0025] Preferably, the anomaly detection model is a machine learning model or a deep learning model. In particular, the anomaly detection model may be built as cluster network.

[0026] Preferably, the lifetime prediction model is a trained deep learning regression network. In particular, the deep learning regression network may be one or more of a long short-term memory network (LSTM), a support vector machine network (SVM), a bidirectional LSTM, a peephole LSTM, a depth gated LSTM, CNN, GRU, IndRnn. In particular, the lifetime prediction model may be combination of different deep learning regression networks. The results of the deep learning regression networks may then be combined in an ensemble network to provide coherent output.

[0027] Preferably, the lifetime prediction model is a pre-trained model. Thus, the lifetime prediction model may be pre-trained with historic data in order to predict the estimated remaining lifetime. Subsequently, the lifetime prediction model may be retrained on the basis of new available data. In particular, the lifetime prediction model is trained for each specific pump type once and can then be transferred to all vacuum pumps of the same type, i.e. same family of pumps or comparable pumps, wherein the pump type relates to the pumping mechanism of the vacuum pump. Instead, the lifetime prediction model may be trained for each individual vacuum pump mark and model once and can then be transferred to all vacuum pumps of the same model. Therein, the vacuum pump model refers to vacuum pumps having comparable specifications, dimensions, and pump performance or the like. Thus, the trained lifetime prediction model can be transferred to the vacuum pumps of the same type without the necessity to retrain the model.

[0028] Preferably, training of the lifetime prediction model may be performed with historical data, which may include normal operation, degrading operation and also including a bearing failure. Therein, the historical data can be acquired from measurement or can be acquired by synthetization as artificial historical data.

[0029] In another aspect of the present invention a software storage product is provided, storing instructions which, when executed by a processor, perform the steps of the method as described before.

[0030] In another aspect of the present invention a controller for a vacuum pump is provided, wherein the controller can be connected to a vacuum pump in order to control operation of the vacuum pump. The controller comprises at least one processor and a memory storage, wherein the memory storage stores instructions which, when executed by the at least one processor, perform the steps of the method described before. Thus, determining degrading operation as well as predicting estimated remaining lifetime is performed by the controller itself. Hence, the controller integrates the control of the vacuum pump as well as the lifetime prediction of the vacuum pump. In another aspect of the present invention a system is provided comprising a controller and a remote server connected to the controller. Therein, connection between the controller and the remote server can be facilitated by any means of data connection, including wireless and wired data connection. The controller can be connected to a vacuum pump. By the controller vibrational data and preferably also operating data are acquired and transmitted to the remote sever. The remote server comprises at least one processor and a memory storage, wherein the memory storage stores instructions which, when executed by the processor, perform the steps of the method as described before. In the system of the present invention the controller controlling the vacuum pump acquires the necessary data, wherein the remote server implements the anomaly detection model as well as the lifetime prediction model in order to determine degrading operation as well as estimated remaining lifetime of the vacuum pump. Alternatively, the controller may implement the anomaly detection model in order to detect the degrading operation, wherein the remote server implements the lifetime prediction model in order to predict the estimated remaining lifetime. Thus, as long as the vacuum pump is in normal operation, the controller itself detects the normal operation and any deviation from the normal operation due to degrading operation. As soon as degrading operation is determined by the controller itself, vibration data and preferably operating data is transferred to the remote server, wherein predicting the estimated remaining lifetime is performed by the remote server. Thus, the computation intensive prediction of the estimated remaining lifetime is transferred to a remote server and computational power of the controller itself can be kept at a minimum.

[0031] In another aspect of the present invention a vacuum pump system is provided comprising a vacuum pump and a controller as described before. Alternatively, the vacuum pump system comprises a vacuum pump and a system including a controller and a remote server as described before.

[0032] Figure Description

[0033] In the following the present invention is described in more detail with reference to the accompanying figures.

[0034] The figures show: Figure 1 a flow diagram of a method according to the present invention,

[0035] Figure 2 a flow diagram of another embodiment of a method according to the present invention,

[0036] Figure 3 a flow diagram of another embodiment of a method according to the present invention,

[0037] Figure 4 a flow diagram of another embodiment of a method according to the present invention,

[0038] Figure 5 a system according to the present invention,

[0039] Figure 6 a system according to another embodiment of the present invention and

[0040] Figure 7 a system according to another embodiment of the present invention

[0041] Detailed Description

[0042] Figure 1 shows a method for operating a vacuum pump and in particular for predicting a remaining lifetime of the vacuum pump. The method includes:

[0043] - in step S01 , acquiring of vibration data of the vacuum pump during normal operation,

[0044] - in step S02, training of an anomaly detection model by the acquired vibration data of the vacuum pump during normal operation, and

[0045] - in step S03, when the anomaly detection model determines a degrading operation of the vacuum pump based upon the acquired vibration and pump data, predicting, by a lifetime prediction model, an estimated remaining lifetime of the vacuum pump on the basis of vibration data acquired during degrading operation of the vacuum pump.

[0046] Thus, the method is structured as a two stage process. In a first stage, during normal operation, the anomaly detection model is determining the operation mode / status of the vacuum pump. As soon as degrading operation is determined by the anomaly detection model, in a second stage, an estimated remaining lifetime is predicted. In step S02 the anomaly detection model is trained by the acquired vibration data of the vacuum pump during normal operation. Thus, the anomaly detection model is able to detect a deviation of the operation mode of the vacuum pump. Therein, the anomaly detection model will be trained to only detect degrading operation if there is a repeated deviation from normal operation, instead of a transient event. Therein, repeated deviation may mean that there are a number of data points within a preset time interval which are outside a threshold range or the data points are outside a threshold range for a preset amount of time. Therein the threshold may be determined from the anomaly detection model training such that if the derived value from model using vibration and operational parameter is greater than the training threshold, there exists an anomaly. Hence, upon detection of an anomaly on the basis of the acquired vibration data, the anomaly detection model is able to determine a degrading operation of the vacuum pump. Thus, in the degrading operation the bearing of the vacuum pump may start to degrade, limiting the lifetime of the vacuum pump.

[0047] Upon detection of a degrading operation by the anomaly detection model, the estimated remaining lifetime is predicted by the lifetime prediction model. Therein, the estimated remaining lifetime is predicted on the basis of acquired vibration data during degrading operation of the vacuum pump. Thus, as soon as the vacuum pump deviates from normal operation determined by the anomaly detection model, the estimated remaining lifetime is predicted by the lifetime prediction model. Therein, preferably the detection of the operation mode of the vacuum pump is performed repeatedly by the anomaly detection model in time intervals for example between 5 seconds and 20 minutes, more preferably between 10 seconds and 10 minutes, and more preferably between 30 seconds and 2 minutes. In particular, the intervals are fixed and equal over the whole time. Alternatively, the intervals may change over time and may be shortened with increasing running hours of the vacuum pump and / or if the vacuum pump is operating in a harsh application. Similar, the prediction of the estimated remaining lifetime may be carried out repeatedly in time intervals of between 5sec. and 20 min., more preferably between 10sec. and 10 min. and more preferably between 30sec. and 2min. In particular, the intervals are fixed and equal over the whole time. Alternatively, the intervals may change over time and may be shortened with decreasing estimated remaining lifetime and / or if the vacuum pump is operating in a harsh application. In particular, the intervals are shorter than during normal operation. Hence, accurate prediction regarding the estimated remaining lifetime can be provided and can be used in order to schedule and plan service and repair of the vacuum pump. In particular, necessary safety margins in the service intervals can be omitted or at least reduced and the service intervals can be planned in accordance with the actual use of the vacuum pump, i.e. the actual seizure and degradation of the vacuum pump.

[0048] As exemplified in the embodiment of Figure 1 , the method according to the present invention may generate a warning signal 100 if the vacuum pump enters degrading operation and / or a warning signal may be generated if the estimated remaining lifetime is smaller than a predetermined threshold. Further warning signals may be generated repeatedly, preferably at set intervals, or further warning signals may be generated if the remaining lifetime is smaller than a further threshold lower than the predetermined threshold. Thus, repeated reminders are provided to the customer to service the vacuum pump. Therein, there might be either only one warning signal either if the vacuum pump enters degrading operation or if the estimated remaining lifetime is smaller than a predetermined threshold. Alternatively, the subsequent warning signals are generated for each of entering the degrading operation by the vacuum pump and if the estimated remaining lifetime is smaller than the predetermined threshold. The warning signals might be identical or different. The warning signal may be an optical signal which may be shown on a display, an acoustical signal or can be otherwise indicative of the degrading operation of the vacuum pump. The present invention is not limited to the location of the warning signal. Thus, the warning signal can be generated by the vacuum pump itself, a controller connected to the vacuum pump or the warning signal can be transmitted via any data transmission line to a remote server, web interface or the like. In particular, the warning signal may be a digital signal. Thus, it is in particular possible to collect warning signals of different vacuum pumps at one location and provide an operator a better overview over the status of its vacuum pumps.

[0049] Preferably, the anomaly detection model is a machine learning model (like Autoencoders, SVM, KNN Isolation Forest, Random Forest. Deep learning methods like LSTM based models, CNN-LSTM models) or a deep learning model in particular the anomaly detection model may be built as cluster network. Further, the machine leaning model may be based on statistical methods like Z-score, Mahalanobis distance, Moving Average, Moving Median method, Median absolute deviation. Preferably, the lifetime prediction model is a trained deep learning regression network. In particular, the deep learning regression network may be one or more of a long short-term memory network (LSTM), a support vector machine network (SVM), a bidirectional LSTM, a peephole LSTM, a depth gated LSTM, CNN, GRU, IndRnn. In particular, the lifetime prediction model may be combination of different deep learning regression networks. The results of the deep learning regression networks may then be combined in an ensemble network to provide coherent output.

[0050] Preferably, the lifetime prediction model is a pre-trained model. Thus, the lifetime prediction model may be pre-trained with generic data in order to predict the estimated remaining lifetime. Subsequently, the lifetime prediction model may be trained during use in normal operation or on the basis of customer data, i.e. historic data. In particular, the lifetime prediction model is trained for each specific pump type once and can then be transferred to all vacuum pumps of the same type, wherein the pump type relates to the pumping mechanism of the vacuum pump. Instead, the lifetime prediction model may be trained for each individual vacuum pump model once and can then be transferred to all vacuum pumps of the same model. Therein, the vacuum pump model refers to vacuum pumps having the same specifications, dimensions, and pump performance or the like. Thus, the trained lifetime prediction model can be transferred to the vacuum pumps of the same type without the necessity to retrain the model.

[0051] Preferably, training of the lifetime prediction model may be performed with historical data, which may include normal operation, degrading operation and also including a bearing failure. The historical data can be acquired from measurement or can be acquired by synthe- tization as artificial historical data.

[0052] Figure 5 illustrates a vacuum pump system 10 comprising a vacuum pump 12 having an inlet 14 and an outlet 16, wherein the vacuum pump 12 is configured to pump a gaseous medium from the inlet 14 towards the outlet 16 in order to generate a vacuum in a vessel (not shown) connected to the inlet 14. Therein, the vacuum pump 12 may be a turbomolec- ular pump, a Gaede pump, a Siegbahn pump, a Holweck pump or in general a molecular drag pump. Alternatively, the vacuum pump may be a screw pump, a roots pump, a claw pump or in general a two-rotor pump. Alternatively, the vacuum pump may be a single axis pump such as a rotary vane pump or scroll pump. In fact, the present invention is not limited to any specific type of vacuum pump, having at least one rotor assembly being rotated for example by an electromotor. Therein, the rotor assembly may be supported by bearings, wherein at least one bearing is built as roller bearing such as a ball bearing, wherein preferably the vacuum pump 12 comprises more than one roller bearing. However, even if no rotor is present, such as in a cryopump, moving parts may be present which deteriorate with time.

[0053] A controller 18 is connected to the vacuum pump 12 in order to control operation of the vacuum pump 12, i.e. controlling the rotation speed of the vacuum pump 12 and the like. A vibration sensor 19 is attached to the vacuum pump 12. As exemplified schematically in Figure 5, the vibration sensor 19 is attached to the housing of the vacuum pump 12 but can be integrated into the housing and might be placed in the vicinity of or directly adjacent to the at least one roller bearing of the vacuum pump 12. Alternatively, the vibration data can be acquired by the motor current. In particular, current ripples may be used as vibration data. The vibration data acquired by the vibration sensor 19 is transmitted to the controller 18.

[0054] The controller 18 implements an anomaly detection module 20 and a lifetime prediction module 22. The anomaly detection module 20 may comprise, store or otherwise provide the anomaly detection model. Similar, the lifetime prediction module 22 may comprise, store or otherwise provide the lifetime prediction model. The anomaly detection model 20 and the lifetime prediction module 22 may be integrated into a computational structure, may be stored on the same storage device and may be executed by the same processor. Alternatively, different storage devices and / or different processors can be provided for each of the anomaly detection module 20 and the lifetime prediction module 22. Thus, the anomaly detection module 20 and / or the lifetime prediction module 22 may be implemented as software or as dedicated hardware such as ASICs or FPGAs.

[0055] By the anomaly detection module 20, anomaly detection is provided in order to detect degrading operation of the vacuum pump, i.e. degrading of the at least one roller bearing of the vacuum pump 12. If the anomaly detection module 20 determines degrading operation of the vacuum pump 12, the lifetime prediction module 22 predicts an estimated remaining lifetime of the vacuum pump.

[0056] Referring now to Figure 2, where the same method steps previously described are indicated by the same reference signs. In the embodiment of Figure 2, in step S11 the acquired vibration data is transferred to the frequency domain and then by connection 101 provided to the anomaly detection model. Similar, the vibration data transferred to the frequency domain is provided by connection 102 to the anomaly detection model. Although Figure 2 shows that the vibration data transferred to the frequency domain is fed to the anomaly detection model as well as the lifetime prediction model by connections 101 and 102, it is not necessarily both and thus the vibration data transferred to the frequency domain of step S11 could be either fed only to the anomaly detection model or the lifetime prediction model or both.

[0057] Referring now to Figure 3, in this embodiment in step S12 monotonic features are determined from the acquired vibrational data and transmitted via connection 102 to the lifetime prediction module 22. Therein, the monotonic features are in particular determined from the vibration data in the frequency domain, i.e. the vibration data, which has been transferred in step S11 to the frequency domain. Monotones or monotonic features are a monotonic input information for the lifetime prediction model. Therein, the monotones of the vibration data show a consistent relationship with the target variable, i.e. the estimated remaining lifetime. Estimated remaining lifetime may be a monotonic decreasing function over time. Since the vibration data is transferred into a monotonic function over time (increasing or decreasing), improved correlation of the monotones of the vibration data as input information to the estimated remaining lifetime is possible by the lifetime prediction model. Thus, by considering monotones of the acquired vibrational data, improved accuracy of predicting the estimated remaining lifetime can be achieved. Therein, the monotones may be determined by a threshold, wherein the monotones are increased or decreased, each time a vibration data point or the difference between two subsequent vibration data points is above the threshold. Therein, decreasing or increasing might be by one, or might be in dependence on the difference between the two subsequent vibration data points or might be dependent on the distance of the vibration data point to the threshold or the like. In the following, an illustrative example is provided how to determine the monotones:

[0058] In the example the threshold is set to 3. If the measurement of the vibration is above the threshold of 3, the monotone is increased in the present example by 1 . Of course, the example is only for illustrative purposes and might not represent real measurement data. Also the monotones may be generated differently. In particular, the present invention cannot be limited to the example shown above. However, by the monotones, the randomly distributed vibration data is mapped to a monotonic increasing or decreasing function or series of values. Since the output of the lifetime predicting model is also a monotonic (decreasing) function, accuracy of the determining the relationship between the two monotonic series of values is improved compared to determining a relationship between the randomly distributed vibration data points and the estimated remaining lifetime.

[0059] In the embodiment of Figure 4 operating data is acquired in step S04 and via connection 103 or 104 provided either to the anomaly detection model and / or the lifetime prediction model. Therein, the operating data comprises one or more of a motor power, rotation speed, operating time, environmental temperature of the vacuum pump during operation, orientation of the vacuum pump, motor current, motor temperature, rotor temperature, bearing temperature, motor voltage, pump cycle, humidity, and pressure, or other information of connected sensors, pressure gauges and accessories. All these parameters may have an influence on the lifetime and operation of the vacuum pump and thus are considered either during anomaly detection of the anomaly detection model in step S02 and / or prediction of the estimated remaining lifetime by the lifetime prediction model in step S03. Therein, this operating data can be either detected by individual sensors or can be provided by the controller 18 itself as a control parameter.

[0060] Figure 6 shows another embodiment of a vacuum pump system 10. In the embodiment of Figure 6 the controller 18 is connected to a remote server 24 via a connection line 26. Therein, the connection line 26 can be any wireless or wired data connection such as an internet gateway or the like. In the embodiment of Figure 6, via the connection 26, vibration data and preferably operating data collected by the controller 18, is transferred to the remote server 24. The remote server 24 comprises the anomaly detection module 20 as well as the lifetime prediction module 22. Hence, anomaly detection, i.e. detecting degrading operation, is performed by the remote server 24. Similar, prediction of the estimated remaining lifetime is performed by the remote server 24. Correspondingly the remote server 24 implements the anomaly detection model as well as the lifetime prediction model. Subsequently, as indicated by connection line 28 information regarding degrading operation and / or information regarding the estimated remaining lifetime may be feed back to the controller 18.

[0061] In the embodiment of Figure 7 the controller 18 implements the anomaly detection module 20, wherein the remote server 24 implements the lifetime prediction module 22. Hence, detection of the degrading operation is performed by the controller itself, wherein the more computational intensive estimated remaining lifetime prediction is performed by a remote server 24.

[0062] Reference List

[0063] 10 vacuum pump system

[0064] 12 vacuum pump

[0065] 14 inlet

[0066] 16 outlet

[0067] 18 controller

[0068] 19 vibration sensor

[0069] 20 anomaly detection module

[0070] 22 lifetime prediction module

[0071] 24 remote server

[0072] 26 connection line

[0073] 28 connection line

[0074] 100 warning signal

[0075] 101 connection

[0076] 102 connection

[0077] 103 connection

[0078] 104 connection

Claims

CLAIMS1. Method for operating a vacuum pump, comprising the steps of: acquiring vibration data of the vacuum pump during normal operation, training an anomaly detection model by the acquired vibration data of the vacuum pump during normal operation, and when the anomaly detection model determines a degrading operation of the vacuum pump based upon the acquired vibration data, predicting, using a machine learning-based lifetime prediction model, a remaining lifetime of the vacuum pump on the basis of vibration data acquired during degrading operation of the vacuum pump.

2. Method according to claim 1 , wherein if degrading operation is determined, a warning signal is generated.

3. Method according to claim 1 or 2, wherein if the estimated remaining lifetime is smaller than a predetermined threshold, a warning signal is generated and / or the vacuum pump is shut down.

4. Method according to any of claims 1 to 3, wherein before transferring the acquired vibration data to the anomaly detection model and / or the lifetime prediction model, the vibration data is transferred to the frequency domain.

5. Method according to any of claims 1 to 4, wherein before transferring the acquired vibration data to the lifetime prediction model, monotones are determined from the acquired vibration data and the monotones are transferred to the lifetime prediction model.

6. Method according to any of claims 1 to 4, wherein operating data is acquired during normal operation and / or degrading operation, wherein the operating data is transferred to the anomaly detection model and / or the lifetime prediction model.

7. Method according to claim 6, wherein the operating data comprises one or more of motor power, rotation speed, operating time, environmental temperature, temperature of the vacuum pump during operation, orientation of the vacuum pump, motor current, motor temperature, rotor temperature, bearing temperature, motor voltage, pump cycle, humidity and pressure.

8. Method according to any of claims 1 to 7, wherein the anomaly detection model is a machine learning network or a deep learning network.

9. Method according to any of claims 1 to 8, wherein the anomaly detection model can be a pretrained model or can be trained in real-time.

10. Method according to any of claims 1 to 9, wherein the lifetime prediction model is a trained machine learning network or a deep learning network or ensemble of more than one deep learning network.

11. Method according to any of claims 1 to 9, wherein the lifetime prediction model is a pretrained model.

12. Software storage product, storing instructions which, when executed by a processor, perform the steps of the method according to any of claims 1 to 11.

13. Controller for a vacuum pump, wherein the controller can be connected to a vacuum pump in order to control operation of the vacuum pump, wherein the controller comprises at least one processor and a memory storage, wherein the memory storage stores instructions which, when executed by the at least one processor, perform the steps of the method according to any of claims 1 to 11.

14. System comprising a controller and a remote server connected to the controller, wherein the controller can be connected to a vacuum pump, wherein vibrational data and preferably operating data are acquired by the controller and transmitted to the remote server, wherein the remote server comprises at least one processor and a memory storage, wherein the memory storage stores instructions which, whenexecuted by the processor, perform the steps of the method according to any of claims 1 to 11.

15. Vacuum pump system comprising a vacuum pump and a controller according to claim 13 or comprising a vacuum pump and a system according to claim 14.

Images