NEG vacuum pump predictive maintenance method using a deep neural network

Abstract

A method to predict the time before reactivation is needed for a reactivable NEG pump connected to a vacuum system comprising at least one pressure sensor operatively connected to a computer, said method comprising the cyclic repetition of many steps through a trained Deep Neural Network model.

Description

[0001] NEG VACUUM PUMP PREDICTIVE MAINTENANCE METHOD USING A

[0002] DEEP NEURAL NETWORK

[0003] Nowadays, the advantages of using Non-Evaporable Getter (NEG) pumps in High Vacuum (HV) and Ultra-High Vacuum (UHV) pumping systems are well-known in the scientific community; in fact, NEG pumps are compact, less demanding in energy consumption and do not cause vibrations.

[0004] Considering what has been said above, the demands on reliable NEG pumps in vacuum systems are continuously increasing so, even more than before, there is a need to optimize their performance and maintenance.

[0005] In EP 3751147 there is shown a computer-implemented method for determining the operability of a machinery; in particular, it describes how to check whether or not compressor damages are compromising the overall functionality of the system through a machine-learning algorithm.

[0006] A diagnostic analysis method for the failure protection of a generic vacuum pump is described in EP 1839151. The method estimates the best-fitted model parameters from the minimum and maximum values evaluating the state variables of the system.

[0007] In US 20030132514 an electronic device sealed under vacuum containing a getter is presented; in particular it describes how the NEG reactivation is based on the comparison between the pressure value read by a sensor and a pressure reference value.

[0008] The above-cited prior art documents do not address the problem of estimating instantaneously the residual lifetime of the main active component of the system, i.e. the getter pump, providing the user or customer with an immediate information for a forwardlooking use and a scheduled maintenance, such as NEG pump reactivation.

[0009] In “An automated sensor fusion approach for the RUL prediction of electromagnetic pumps” by Akpudo Ugochukwu Ejike et al. published in the journal IEEE Access vol. 9, on 2 March 2021 the predictive maintenance of an electromagnetic pump is linked to its fluidic contamination status.

[0010] The prediction on the pump Residual Useful Life (RUL) (i.e. the remaining time before reactivation) of a NEG pump can be accomplished by a digital controller according to the present invention connected to the vacuum system which comprises at least one NEG pump, said controller comprising at least one pressure sensor and being integrated into a computer for user interface and capable to run a Deep Neural Network (DNN) algorithm.

[0011] Once the controller has successfully concluded the data elaboration, the user will be able to read on a display as outputs the estimated time to reactivation and possibly other parameters related to the status of consumption of the getter cartridge, such as pumping speed and residual capacity. Adding the possibility of networking the digital controller, information on the use of the NEG pumps could be shared with the pumps suppliers, in order to further improve the programming and relative codes behind the controller’s logic and deepen knowledge on the application systems of the NEG pumps, and more generally, on the vacuum products. A single digital controller may be used to control, collect and elaborate data from multiple NEG pumps connected thereto. Even though not limited to a specific number of NEG pumps, the minimal configuration is 1 : 1 (one controller for one NEG pump). Moreover, the controller is configured also to detect any anomalies related to some parameters of the vacuum pump, such as the pumping speed and the saturation rate, which is an indicator of the NEG consumption speed between an activation and the next one. In particular, the digital controller according to the present invention can detect if the bulk of the NEG has reached complete saturation (the total capacity has been exploited according to the number of reactivations and to the estimated amount of gas absorbed) and the presence of possible spikes in pressure readings from the sensors that might cause a misleading estimation on the pump consumption status. Moreover, the controller can detect if the activation of the NEG has not been successful, for example when the threshold power is not delivered for the entire duration; this will be described in detail below.

[0012] In a preferred solution according to the present invention, the controller is configured to perform, through a DNN algorithm, a computer-implemented method as described in the following paragraph.

[0013] The present invention provides a predictive method for the maintenance of a NEG pump in a vacuum system which includes at least one pressure sensor. In a preferred embodiment according to the present invention the sensor is a Bayard-Alpert gauge, but other pressure sensors could be easily used such as extractor gauge, cold-cathode and, when an ion pump is also installed in the vacuum system, its current.

[0014] The invention is described in the following with reference to a non-limiting example, with the aid of the attached figures wherein:

[0015] Fig.l shows an example of RUL estimation according to the invention with the parameters described in detail further on (the Time axis must be intended in arbitrary units).

[0016] Fig, 2 is an example of estimated RUL compared with expected (true) RUL instant by instant (the Time axis must be intended in arbitrary units).

[0017] Fig, 3 shows the Root Mean Square Error (RMSE) of an Operating Parameter (OP) prediction reconstructed using different network models, as detailed further on.

[0018] Fig, 4 is a boxplot representing the standard deviation of the k coefficient defined in equation [4] below with the parameters described in detail further on.

[0019] Figs.5-7 show three boxplots representing the error computed as in equation [6] below on RUL prediction with the parameters described in detail further on.

[0020] Fig, 8 is a histogram representing the distribution of AOP values, which will be discussed in detail further on, obtained in an experimental test.

[0021] Before starting the monitoring phase through the neural network, the system checks the NEG pump activation; this function comprises the steps of: a) setting pump model and working parameters such as: technical specification of the pump model, working values of voltage / current, thermocouple reading (when present), the ramping time to reach the activation temperature true, the maintenance time on which the NEG is kept at the activation temperature thoid,' b) starting the activation of the pump manually or programming it at a certain time tclock > trise thoid , c) confirming the activation success and sending a success signal S to the system while updating the activation counter; d) waiting for NEG readiness before starting data collection, readiness determined by: dl) lapse of a cooling time, ^cooling Of d2) temperature equal to or below a threshold temperature, T threshold-

[0022] According to the present invention, the prediction method comprises:

[0023] - Training phase - Monitoring phase

[0024] During the training phase, a training dataset comprising a set of pressure evolution curves in time is used. Each curve is converted into an operating parameter (OP) curve, which is a decreasing straight line ranging from 1 to 0, computed as:

[0025] Where t z is the time it takes for the NEG pump to drop below 10% of its full pumping efficiency, and RUL is the difference between tfan and the current time instant (i.e. the time needed to reach t from the current instant). From the OP values it is possible to have an overview of the instantaneous status, in terms of pumping efficiency (i.e., for example, pumping speed and sorption capacity), of the NEG pump in a 1-0 range. The set of the pressure evolution curves, along with the set of associated OP curves, are given as input and output respectively to a Deep Neural Network model for the training phase.

[0026] In particular, such model might consist of a Long Short-Term Memory (LSTM) Neural Network, comprising one or more hidden layers L made of a number N of LSTM cells, trained for a sufficiently large number X of epochs. LSTM models can take into account a significant input and store it in the long-term state, learn to preserve it for as long as it is needed and learn to extract it whenever it is needed. This structure provides a reliable way to capture long-time dependency in the pressure evolution curve. Once trained, the model will provide as output an instantaneous estimation OP of the OP value up to the reactivation of the NEG pump.

[0027] This method is complementary to the one described in the unpublished Italian patent application 102023000012843, in the applicant’s name, and is more apt in making accurate predictions in case of vacuum systems experiencing sudden variations in the operating parameters and conditions.

[0028] The monitoring phase comprises some operative steps repeated cyclically as follows:

[0029] 1. Storing the instant pressure readout from at least one pressure sensor.

[0030] 2. Estimating an instant OP value from said instant pressure sensor(s) readout using the Deep Neural Network model. 3. Checking whether the difference between two consecutive OP values falls within a given range and, if not, reporting a warning,

[0031] 4. Creating an OP curve decreasing from 1 to 0, where 1 represents the start of pump activation and 0 the pump failure time, from said estimated instant OP value.

[0032] 5. In order to estimate the RUL, i.e. the time needed to get to OP=0 from the current status: a. a straight line is fitted passing through the first value of the OP curve and the current value of the OP estimated by the model - linearization method 0, or b. a straight line is fitted passing through a first value corresponding to operating parameter equal to 1 (at the first time instant) and the current value of the OP estimated by the model (Figure 1) - linearization method 1, or c. a straight line is fitted on the whole collection of the past values of OP, from the start of the monitoring phase, up to the current one - linearization method 2, or d. a straight line is fitted on the collection of only the M most recent values of OP up to the current one - linearization method 3, or e. a straight line is used with an angular coefficient which is obtained by a weighted average of the M angular coefficients obtained by the M lines of the M most recent consecutive time instants (where the M lines are constructed through linearization method 1) - linearization method 4.

[0033] The estimation described in step 2 above is obtained through a DNN model trained using pressure curves evolving in time, where the NEG performances and the experimental conditions are known, said model consisting in numerical weights which are multiplied by the instant pressure sensor(s) readout to obtain an operating parameter OP representing the instant NEG performance.

[0034] As described in step 3 above, at any time during the monitoring phase, the controller according to the present invention is able to detect any anomalies related to the pumping efficiency by observing the trend of instant OP values. In a preferred embodiment the check is performed through the following steps: a) compute where OPtand OPt-are respectively the OP values at instant t and at instant t-1 b) check if the LOP computed falls within a range delimited by a lower boundary LB and an upper boundary UB both described in detail below. c) if the AO / 3] does not fall inside the range cited above, the system reports a warning, otherwise the saturation rate of the vacuum pump is considered to be ’’constant” with respect to the relatively recent dynamics of the vacuum system, represented by the collection of the L most recent LOP^ . In the event of a warning, if the A P] is lower than LB the saturation rate is considered “faster” (than expected), if it is greater than UB the saturation rate is considered “slower” (than expected). d) After the first warning the controller repeats the computation:

[0035] AOP2= OPt+1- OPt[2b] which is calculated at the following instant with respect to [2a], If also A P2does not fall inside the above range the system reports an anomaly, otherwise the first warning disappears.

[0036] For the calculation of the LB and UB boundaries, a descriptive statistics method about the interquartile range was exploited, often used for the identification of outliers when data is not necessarily Gaussian-distributed. The steps are: a) compute the interquartile range IQR as: IQR = q3-ql where ql is the first quartile of A P distribution and q3 is the third quartile of A P distribution. b) compute LB and UB as: LB = ql- KLQR and UB = q3+ KLQR where K may vary, here taken as K =3,5.

[0037] Any person skilled in the art might prefer other anomaly detecting methods, e.g. statistical methods like z-score, machine learning methods like DBSCAN, isolation forests, Local Outlier Factor (LOF), One-Class Support vector Machine (OCSVM) or deep learning methods like autoencoders and neural networks.

[0038] The controller according to the present invention is capable of setting the number of warnings after which the system generates an anomaly. Moreover, the user can decide how the anomaly should be overcome: with his action or with a further check by the controller.

[0039] The estimated RUL is given by the difference between the time instant when the evolution of the estimated OP given by the straight line (as described in step 5 above) reaches the value OP=0 and the current time instant. In Figure 2, there is shown the RUL estimation during the monitoring phase, using as an example linearization method 1.

[0040] Each complete pressure evolution curve is stored in the database for future retraining of the network, with the aim of obtaining better accuracy of the model. The retraining phase is done periodically with an extended training dataset together with an optimization of N parameter.

[0041] During the monitoring phase, the linearization (as described in step 5 above) of the estimated instant operating parameter allows to estimate the RUL of the NEG pump, while the stand-alone DNN wouldn’t be able to perform long term predictions.

[0042] In order to validate the model, a subset of the entire dataset, called validation set,

[0043] (different from the set used for training the Deep Neural Network) is used for evaluating the accuracy of the model. Such model is trained for a number X of epochs and optimized by varying the parameter N of the network (so training the same model with different N values). The most accurate model is chosen according to some metrics for evaluating the prediction error of the model: in particular, the RMSE is considered and computed as: where Ntrainis the amount of data of the training dataset, OP are the true values (computed with [1]) and OPt are the predicted values of the operating parameters computed by the LSTM Deep Neural Network.

[0044] The validation process includes two main sequential steps:

[0045] 1) Choosing the optimal value of N cells.

[0046] 2) Once fixed N, choosing the optimal linearization method.

[0047] As regards step 1, the neural network has been trained for 10000 epochs varying the number of cells, considering N=10, 20, 30, 40, 50, 100 and 150. In order to avoid overfitting on data, the training error is compared with the RMSE evaluated during the validation phase of the network, computed as: where Nvaiidationis the number of data of the validation dataset, OPj are the true values (computed with [1]) and OPj are the predicted values of the operating parameters computed by the LSTM Deep Neural Network. The optimal value of parameter N is chosen according to the minimum error on OP, computed as in [4],

[0048] The error computed for data used to train the network (as in [3]) should be low as well as the error computed for data used for validation (as in [4]), meaning that the model is accurate on data used for both training and validation.

[0049] Once the best value of N has been chosen, the second validation step comprises the evaluation of the best linearization method. In this case, the criterion is based on the analysis of the standard deviation of k, i.e. the angular coefficient of the line drawn for making a prediction of the RUL based on OP (as explained in step 4 of the monitoring phase), computed as: where Ntis the number of data collected for each test, ktare the instant coefficients of the linearization line and k is the average of all kt.

[0050] Such parameter allows to select the linearization method which guarantees a lower fluctuation of RUL estimation during the monitoring phase. Moreover, the error on RUL predictions computed as: is used for evaluating and comparing the accuracy of linearization methods, by varying the time instant at which the monitoring phase starts (an example of comparison is shown in figure 5). A secondary result of such evaluations, valid for all linearization methods, regards the evolution of the error on predictions: it becomes lower and lower as the considered time instant from which RUL estimations start increases.

[0051] With reference to the attached figures, the following detailed parameters are provided:

[0052] Fig.l shows an example of OP estimation made by the LSTM neural network compared with true (expected) OP (thick line). The dashed line is built through the linearization method 1 : it passes through the points OP=1 and, as an example, OP=0.4. It is relative to an experiment with gaseous composition of 5% CO, 80% H2, 15% O2 and the LSTM Network has been trained for X=10000 epochs, with N=100 cells and L=1 layer.

[0053] Fig, 2 is an example of estimated RUL compared with expected (true) RUL instant by instant. The RUL estimation starts after 5 time steps, it is relative to an experiment with gaseous composition of 10% CO, 85% H2, 5% O2 and it is done using linearization method 1. The LSTM Network has been trained for X=10000 epochs, with N=100 cells and L=1 layer.

[0054] Fig, 3 shows the RMSE of the OP prediction reconstructed using different LSTM network models, each one trained with a different numbers N of cells, in particular, N=30, N=50, N=100, N=150. The number of epochs of training is 10000 and L=1 layer.

[0055] Fig, 4 is a boxplot representing the standard deviation of k computed as in [5] for each curve of the validation set starting from 5 time steps. Results for linearization methods 1 and 2 are shown.

[0056] Figs.5-7 show three boxplots representing the error computed as in [6] on RUL prediction for all curves of the validation set, considering different time instants from which the error is computed (5, 300 and 500 time steps). Linearization methods 1 and 2 are compared: for predictions starting at earlier time steps, linearization 1 is more accurate; for RUL estimations made at later time steps, methods 1 and 2 become comparable in accuracy.

[0057] Fig, 8 shows the results of an experiment related to an absorption test of a gas mixture of 5% CO, 90% H2, 5% O2, performed with a SAES NEG vacuum pump CTZ200. As it can be seen from the histogram, there are extremely negative values of AOP (as computed in [2a]): this is because during the monitoring an unexpected return of air into the system (leak) occurred, with a consequent increase in pressure and faster consumption of the NEG vacuum pump. The AOP values distribution derives from the 200 most recent pressure readouts.

[0058] Any person skilled in the art could easily understand that, in order to achieve a predictive model for a vacuum pump, it is possible to use other supervised machine learning algorithms such as: linear regression, logistic regression, support vector machine, k-Nearest Neighbors regression. These algorithms are all reliable for the purpose, even if the optimal choice for dealing with sequential data could be an LSTM DNN which adapts more rapidly when sudden variations of pressure in the system occur, therefore it is more suitable for dynamic systems (such as vacuum benches, microscopes, systems with degassing components).

[0059] Other models with similar aim of an LSTM DNN could be used, such as Recurrent Neural Network (RNN), even though they are subject to the vanishing and exploding gradient problem, Gated Recurrent Unit (GRU) and Transformers.

[0060] According to another aspect of the present invention, a digital controller configured to perform the computer-implemented method according to the present invention is described.

[0061] The method according to the present invention can be implemented in an electronic device, which is able to receive relevant vacuum system information such as the pressure values measured by sensors. Said information can be transmitted via analog or digital signals, via serial interfaces or data networks, including wired and wireless data transmission networks, LAN, Wi-Fi, Bluetooth and others. This electronic device can include memories, Arithmetic Logic Unit (ALU), microprocessors, programmable logic elements, FPGAs, personal computers; in some embodiments the electronic device can also be part of a more complex system including the power supply for the pumps and / or their controllers.

[0062] The electronic device can be equipped with a local or remote user interface, which makes monitoring and managing the vacuum system easier for the user.

[0063] During monitoring, experimental data can be collected and they can be used to feed the training dataset of the method of the present invention or a system with a different application.

[0064] This approach, when applied to multiple customers’ vacuum systems, allows to collect information from multiple systems about the vacuum pump status and performance, and after the processing of the method of the present invention, provide back valuable information to the customer creating a two-way direct line.

Claims

CLAIMS1. A method to predict the time before reactivation is needed for a reactivable NEG pump connected to a vacuum system comprising at least one pressure sensor operatively connected to an electronic control unit, capable of storing and processing data, and a feedback control system, said method comprising the cyclic repetition of the following steps: a) storing in said electronic control unit an instant pressure readout from said at least one pressure sensor, b) estimating and storing in the electronic control unit an instant operating parameter OP value in terms of pumping efficiency obtained from said instant pressure sensor(s) readout through a trained artificial intelligence algorithm using a Deep Neural Network model trained using pressure curves evolving in time, with known NEG performances and known experimental conditions, c) checking whether the difference between two consecutive OP values falls within a given range and, if not, report a warning, d) creating an OP curve decreasing from 1 to 0, where 1 represents the start of pump activation and 0 the pump failure time, from said instant operating parameter OP value estimated in step b), e) linearizing said OP curve through any of a plurality of linearization methods stored in the electronic control unit, f) estimating the pump residual useful life RUL as the difference between the current time instant and the time instant when the instant operating parameter OP value given by the linearized OP curve reaches the value OP = 0.

2. The method of claim 1, wherein the step e) is performed by fitting a straight line passing through the first value of the OP curve or the value corresponding to OP = 1 and the current value of the OP estimated through said Deep Neural Network model.

3. The method of claim 1, wherein the step e) is performed by fitting a straight line on all the computed values of OP, from the start of the monitoring phase to the current one.

4. The method of claim 1, wherein the step e) is performed by fitting a straight line on the series of only the M most recent values of OP up to the current one.

5. The method according to claim 2, wherein the straight line has an angular coefficient that is obtained by a weighted average of the M angular coefficients obtained by the M lines of the M most recent consecutive time instants.

6. The method according to any of the preceding claims, wherein if a warning is reported in step c) then the check is repeated with the following OP value and if the difference again does not fall inside the range the system reports an anomaly, otherwise the first warning disappears.

7. The method according to any of the preceding claims, wherein the Deep NeuralNetwork model is a Long Short-Term Memory Neural Network.

8. A digital controller, comprising at least one pressure sensor, configured to perform the method of any of the preceding claims.

Images