Neg vacuum pump predictive maintenance method using deep neural networks

Abstract

A method for predicting the time before a reactivatable NEG pump, which is connected to a vacuum system comprising at least one pressure sensor operatively connected to a computer, needs to be reactivated, the method comprising a number of steps repeated in a loop by a trained deep neural network model.

Description

[0001] Today, 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, have low energy consumption requirements, and do not cause vibration.

[0002] Given the above, the demand for reliable NEG pumps in vacuum systems is constantly increasing, thus requiring even greater optimization of NEG pump performance and maintenance than ever before.

[0003] In EP 3751147, a computer-implemented method for determining the operability of machinery is shown; in particular, the method describes how to examine whether compressor damage jeopardizes the overall functionality of the system using machine learning algorithms.

[0004] EP 1839151 describes a diagnostic analysis method for failure protection of general-purpose vacuum pumps. This method estimates the best-fit model parameters based on the minimum and maximum values ​​of the state variables of the evaluation system.

[0005] US 20030132514 proposes an electronic device containing a getter that is sealed under vacuum; in particular, it describes how to reactivate the NEG based on a comparison between a pressure value read by a sensor and a pressure reference value.

[0006] The prior art literature cited above does not address the problem of instantaneously estimating the remaining lifespan of the system’s main moving parts (i.e., the getter pump) to provide users or customers with immediate information for forward-looking use and planned maintenance (e.g., NEG pump reactivation).

[0007] In “An automated sensor fusion approach for the RUL prediction of electromagnetic pumps”, published by Akpudo Ugochukwu Ejike et al. in IEEE Access Volume 9 on March 2, 2021, predictive maintenance of electromagnetic pumps is correlated with their fluid contamination status.

[0008] Predicting the remaining pump life (RUL) (i.e., the time remaining before reactivation) of a NEG pump can be accomplished by a digital controller connected to a vacuum system according to the invention, the vacuum system including at least one NEG pump, the controller including at least one pressure sensor and integrated into a computer for a user interface and capable of running deep neural network (DNN) algorithms.

[0009] Once the controller successfully completes data processing, the user can read on the display the estimated time for reactivation and other parameters, such as pumping speed and remaining capacity, that may be related to the consumption status of the getter cartridge. Increasing the networking capability of the digital controller allows information about the use of the NEG pumps to be shared with pump suppliers, enabling further improvements to the programming and related code behind the controller's logic and deepening understanding of the application system of the NEG pumps, and more generally, a deeper understanding of vacuum products. A single digital controller can be used to control, collect, and process data from multiple NEG pumps connected to that single digital controller. While not limited to a specific number of NEG pumps, the minimum configuration is 1:1 (one controller for one NEG pump). Furthermore, the controller is configured to detect any anomalies related to certain parameters of the vacuum pump, such as pumping speed and saturation rate, which are indicators of the NEG consumption rate between one activation and the next. In particular, the digital controller according to the invention can detect whether the body of the NEG has reached full saturation (the total capacity has been fully utilized based on the number of reactivations and the estimated amount of gas absorbed), and the presence of potential spikes in pressure readings from sensors that could lead to misleading estimates of the pump's consumption status. In addition, the controller can detect whether NEG activation has not been successful, for example, if the threshold power has not been delivered for the entire duration; this will be described in detail below.

[0010] In a preferred embodiment of the invention, the controller is configured to execute the computer-implemented method described in the following paragraphs using a DNN algorithm.

[0011] This invention provides a predictive method for the maintenance of NEG pumps in a vacuum system, which includes at least one pressure sensor. In a preferred embodiment of the invention, the sensor is a Bayard-Alpert vacuum gauge, but other pressure sensors can be readily used, such as a separate extractor gauge, a cold cathode sensor, and a sensor for the current of an ion pump when such an ion pump is also installed in the vacuum system.

[0012] The present invention is described below with reference to non-limiting examples and by means of the accompanying drawings, in which:

[0013] Figure 1 An example of RUL estimation according to the invention is shown with the parameters described in detail below (the time axis should be intended to be in arbitrary units).

[0014] Figure 2 This is an example of a time-by-time comparison between the estimated RUL and the expected (true) RUL (the time axis should be intended to use arbitrary units).

[0015] Figure 3 The root mean square error (RMSE) predicted by the operating parameters (OP) reconstructed using different network models is shown below in detail.

[0016] Figure 4 This is a box plot representing the standard deviation of the k coefficient as defined in the following equation [4] under the parameter conditions described in detail below.

[0017] Figures 5 to 7 Three box plots are shown representing the error of the RUL prediction calculated in Equation [6] below, under the parameter conditions described in detail below.

[0018] Figure 8 This indicates that the results were obtained in experimental testing, which will be discussed in detail later. Histogram of the distribution of values.

[0019] Before commencing the monitoring phase via neural network, the system checks for NEG pump activation; this function includes the following steps:

[0020] a) Set the pump model and operating parameters, such as: pump model specifications, operating voltage / current values, thermocouple readings (if present), and ramp time t to reach activation temperature. 上升 The duration t of NEG at the activation temperature 保持 ;

[0021] b) Manually start pump activation or program pump activation to occur at a specific time t. 时钟 ≥t 上升 +t 保持 Down;

[0022] c) Confirm successful activation and send a success signal S to the system, while updating the activation counter;

[0023] d) Wait for NEG to be ready before starting data collection. Readiness is determined by the following:

[0024] d1) After cooling time t 冷却 ,or

[0025] d2) Temperature equal to or below the threshold temperature T 阈值 .

[0026] According to the present invention, the prediction method includes:

[0027] - Training phase

[0028] - Monitoring phase

[0029] During the training phase, a training dataset consisting of a set of stress evolution curves varying over time is used. Each curve is converted into an operating parameter (OP) curve, which is a decreasing straight line in the range of 1 to 0, and is calculated as follows:

[0030] [1]

[0031] in, It is the time it takes for the NEG pump to drop below 10% of its full pumping efficiency, and RUL is... The difference between the current time and the present time (i.e., the time from the current time to the present time) (Time required). Based on the OP value, the instantaneous state of the NEG pump in terms of pumping efficiency (i.e., pumping speed and adsorption capacity) can be summarized in the range of 1 to 0. The set of pressure evolution curves and the associated set of OP curves are given as the input and output of the deep neural network model used in the training phase, respectively.

[0032] Specifically, such a model can include a Long Short-Term Memory (LSTM) neural network comprising one or more hidden layers L consisting of N LSTM units, trained for a sufficiently large number of X epochs. The LSTM model can consider important inputs and store them in long-term states, learn to retain these important inputs for the required duration, and learn to retrieve them when needed. This structure provides a reliable method for capturing long-term correlations in stress evolution curves. Once trained, the model will provide an instantaneous estimate of the OP value. As output, until the NEG pump is reactivated.

[0033] This method is complementary to the method described in the applicant’s unpublished Italian patent application 102023000012843, and is more conducive to making accurate predictions when the vacuum system undergoes sudden changes in operating parameters and conditions.

[0034] The monitoring phase includes the following recurring operational steps:

[0035] 1. Store instantaneous pressure readings from at least one pressure sensor.

[0036] 2. Using a deep neural network model to estimate the instantaneous pressure from the instantaneous pressure sensor readings. value.

[0037] 3. Check two consecutive The system checks whether the difference between the values ​​falls within a given range, and if the difference does not fall within the range, it reports a warning.

[0038] 4. Based on the estimated instantaneous... Values ​​are used to create values ​​that decrease from 1 to 0. The curve represents the starting point of pump activation, where 1 indicates the pump failure time and 0 indicates the pump failure time.

[0039] 5. In order to estimate RUL, i.e., from the current state to... Time required to reach 0:

[0040] a. Fit across The first value of the curve and the value estimated by the model The current value of the straight line—linearization method 0, or

[0041] b. Fit through the first value corresponding to the operating parameter being equal to 1 (at the first time step) and the value estimated by the model. The straight line of the current value ( Figure 1 — Linearization method 1, or

[0042] c. Targeting the entire history from the beginning of the monitoring phase to the present. Fitting a straight line to a set of values—Linearization Method 2, or

[0043] d. For only M recent [data points] up to the present... Fitting a straight line to a set of values—Linearization Method 3, or

[0044] e. Using a straight line with angle coefficients obtained by weighted averaging of M angle coefficients obtained from M most recent consecutive moments (where the M lines are constructed by linearization method 1) – linearization method 4.

[0045] The estimation results described in step 2 above are obtained by using a DNN model trained with a pressure curve that evolves over time, where the NEG performance and experimental conditions are known. The model includes numerical weights, which are multiplied by instantaneous pressure sensor readings to obtain operating parameters representing instantaneous NEG performance. .

[0046] As described in step 3 above, at any time during the monitoring phase, the controller according to the invention is able to observe instantaneous... The trend of the value is used to detect any anomalies related to pumping efficiency. In a preferred embodiment, the check is performed by the following steps:

[0047] a) Calculation [2a]

[0048] in, and They are at time t and at time t-1, respectively. value;

[0049] b) Check the calculations Whether it falls within the range defined by the lower boundary LB and the upper boundary UB, which are described in detail below.

[0050] c) If If the value does not fall within the range described above, the system will issue a warning; otherwise, the vacuum pump's saturation rate will be relative to the vacuum system's L nearest values. The set representing the relatively recent dynamics is considered "constant". In the case of a warning, if... Below LB, the saturation rate is considered to be "faster" (than expected). If the value is greater than UB, the saturation rate is considered to be "slower" than expected.

[0051] d) After the first warning, the controller repeats the following calculation process:

[0052] [2b]

[0053] It is calculated relative to [2a] at the next time step. If If the value does not fall within the above range, the system will report an anomaly; otherwise, the first warning will disappear.

[0054] To calculate the LB and UB boundaries, a descriptive statistical method regarding the interquartile range is used. This descriptive statistical method is often used to identify outliers when the data is not necessarily Gaussian distributed. The steps are as follows:

[0055] a) The interquartile range (IQR) is calculated as: IQR = q3 - q1

[0056] Where q1 is The first quartile of the distribution and q3 is The third quartile of the distribution.

[0057] b) Calculate LB and UB as follows: and

[0058] K can vary; here we take K=3 or 5.

[0059] Anyone skilled in the art may prefer other anomaly detection methods, such as statistical methods like z-scores, machine learning methods like DBSCAN, isolated forests, local outlier factors (LOF), one-class support vector machines (OCSVM), or deep learning methods like autoencoders and neural networks.

[0060] The controller according to the invention can set the number of warnings before the system generates an anomaly. Furthermore, the user can decide how to overcome the anomaly: through user actions or through additional checks performed via the controller.

[0061] The estimate is given by the straight line (as described in step 5 above). The evolution reached a value The difference between the time when =0 and the current time gives the estimated RUL. Figure 2 In the example, linearization method 1 is used to illustrate the RUL estimation during the monitoring phase.

[0062] Each complete stress evolution curve is stored in a database for future network retraining, with the aim of achieving better model accuracy. The retraining phase is performed periodically using an expanded training dataset and optimization of the N parameters.

[0063] During the monitoring phase, the linearization of the estimated instantaneous operating parameters (as described in step 5 above) makes it possible to estimate the RUL of the NEG pump, whereas a standalone DNN cannot perform long-term predictions.

[0064] To validate the model, a subset of the entire dataset, called the validation set (different from the set used to train the deep neural network), is used to evaluate the model's accuracy. This model is trained for a number of X epochs and optimized by varying the network's parameters N (thus training the same model with different values ​​of N). The most accurate model is selected based on several metrics used to evaluate its prediction error: specifically, RMSE is considered and calculated as follows:

[0065] [3]

[0066] in, It refers to the amount of data in the training dataset. It is the true value (calculated using [1]), and These are the predicted values ​​of the operating parameters calculated using an LSTM deep neural network.

[0067] The verification process includes two main sequential steps:

[0068] 1) Select the optimal value of N units.

[0069] 2) Once N is fixed, the optimal linearization method is selected.

[0070] Regarding step 1, the neural network is trained for 10,000 epochs with varying numbers of units, considering N=10, 20, 30, 40, 50, 100, and 150. To avoid overfitting, the training error is compared to the RMSE evaluated during the network's validation phase, which is calculated as follows:

[0071] [4]

[0072] in, It verifies the number of data points in the dataset. It is the true value (calculated using [1]), and These are the predicted values ​​of the operating parameters calculated by the LSTM deep neural network. The optimal value of parameter N is selected based on the minimum error with respect to OP, as calculated in [4].

[0073] The errors computed on the data used to train the network (as in [3]) and the errors computed on the data used for validation (as in [4]) should be low, which means that the model is accurate on both the data used for training and validation.

[0074] Once the optimal value of N is selected, the second verification step involves evaluating the optimal linearization method. In this case, the criterion is based on k—that is, (as explained in step 4 of the monitoring phase) the method used for linearization. The standard deviation of the angle coefficients of the lines drawn for predicting RUL is analyzed and calculated as follows:

[0075] [5]

[0076] in, This refers to the amount of data collected in each test. These are the instantaneous coefficients of the linearized line, and It is all The average value.

[0077] These parameters allow for the selection of a linearization method that ensures low volatility in the RUL estimate during the monitoring phase. Furthermore, the error in the RUL prediction is calculated as follows:

[0078] [6]

[0079] It is used to evaluate and compare the accuracy of linearization methods by changing the start time of the monitoring phase (examples of the comparison are in...). Figure 5 (As shown in the figure). A secondary result of such an evaluation (which is valid for all linearization methods) involves the evolution of the error in the prediction: the error becomes increasingly lower as the considered RUL estimate increases from its initial time.

[0080] Referring to the attached diagram, the following detailed parameters are provided:

[0081] Figure 1 This demonstrates the process performed using an LSTM neural network. An example of estimating the actual (expected) OP (thick line). The dashed line is constructed using linearization method 1: it passes through the point OP=1, and as an example, =0.4.

[0082] It is related to an experiment in which the gas components are 5% CO, 80% H2, and 15% O2, and the LSTM network has been trained for X=10,000 epochs, with N=100 units and L=1 layer.

[0083] Figure 2 This is an example of a time-by-time comparison between the estimated RUL and the expected (true) RUL. The RUL estimation begins after 5 time steps, is related to an experiment in which the gas composition is 10% CO, 85% H2, and 5% O2, and is performed using linearization method 1. The LSTM network has been trained for X=10000 epochs, with N=100 units and L=1 layer.

[0084] Figure 3 The RMSE of the OP predictions reconstructed using different LSTM network models is shown, each LSTM network model being trained with a different number of units N, specifically N=30, N=50, N=100, and N=150. The number of training epochs is 10000 and L=1 layer.

[0085] Figure 4 This is a box plot representing the standard deviation of k for each curve on the validation set, starting from 5 time steps as calculated in [5]. Results for linearization methods 1 and 2 are shown.

[0086] Figures 5 to 7 Three box plots are shown representing the errors in RUL predictions for all curves on the validation set, as calculated in [6], taking into account different times from which the errors were calculated (5, 300, and 500 time steps). Linearization method 1 and linearization method 2 are compared: linearization 1 is more accurate for predictions starting at earlier time steps; for RUL estimates made at later time steps, method 1 and method 2 become comparable in accuracy.

[0087] Figure 8 Experimental results related to the absorption test of a gas mixture of 5% CO, 90% H2, and 5% O2 performed using a SAES NEG vacuum pump CTZ200 are shown. The histogram shows the presence of... Extremely negative values ​​(as calculated in [2a]): This is because air accidentally returned to the system during monitoring (leakage), resulting in increased pressure and faster consumption of the NEG vacuum pump. The value distribution was obtained from 200 recent pressure readings.

[0088] Anyone skilled in the art will readily understand that other supervised machine learning algorithms, such as linear regression, logistic regression, support vector machines, and k-nearest neighbor regression, can be used to implement predictive models for vacuum pumps. These algorithms are all reliable for the purpose, although the best choice for processing sequential data may be LSTM DNN, which adapts faster to sudden pressure changes in the system and is therefore more suitable for dynamic systems (e.g., vacuum stages, microscopes, systems with degassing components).

[0089] Other models with similar purposes to LSTM DNNs, such as recurrent neural networks (RNNs) (even though they are susceptible to vanishing and exploding gradient problems), gated recurrent units (GRUs), and transformers, can be used.

[0090] According to another aspect of the invention, a digital controller is described, which is configured to perform a computer-implemented method according to the invention.

[0091] The method according to the invention can be implemented in an electronic device capable of receiving relevant vacuum system information, such as pressure values ​​measured by sensors. This information can be transmitted via analog or digital signals, via a serial interface or data network (including wired and wireless data transmission networks, LAN, Wi-Fi, Bluetooth, etc.). The electronic device may include a memory, an arithmetic logic unit (ALU), a microprocessor, a programmable logic element, an FPGA, or a personal computer; in some embodiments, the electronic device may also be part of a more complex system including a power supply for the pump and / or its controller.

[0092] Electronic devices can be equipped with local or remote user interfaces, which makes it easier for users to monitor and manage vacuum systems.

[0093] During monitoring, experimental data can be collected and used to feed the training dataset of the method of the present invention or to systems with different applications.

[0094] When applied to vacuum systems for multiple customers, this method allows information about the status and performance of vacuum pumps to be collected from multiple systems, and after processing by the method of the present invention, valuable information is provided back to the customers, thereby creating a bidirectional direct line.

Claims

1. A method for predicting the time required before a reactivatable NEG pump needs to be reactivated, the reactivatable NEG pump being connected to a vacuum system including at least one pressure sensor and a feedback control system, the at least one pressure sensor being operatively connected to an electronic control unit capable of storing and processing data, the method comprising cyclically repeating the following steps: a) The electronic control unit stores instantaneous pressure readings from the at least one pressure sensor. b) Using a trained artificial intelligence algorithm and a deep neural network model, instantaneous operating parameters regarding pumping efficiency are estimated based on the instantaneous pressure sensor readings. Value, and the instantaneous operation parameters The value is stored in the electronic control unit, wherein, The deep neural network model was trained using a stress curve that evolves over time, under known NEG performance and known experimental conditions. c) Check two consecutive Check if the difference between the values ​​falls within a given range, and if the difference does not fall within the range, report a warning. d) Based on the instantaneous operating parameters estimated in step b) Values ​​are used to create values ​​that decrease from 1 to 0. The curve represents the start of pump activation, where 1 indicates the pump failure time. e) By any of the various linearization methods stored in the electronic control unit, the electronic control unit is subjected to... Linearize the curve. f) Estimate the remaining pump life RUL at the current moment and by linearization. Instantaneous operating parameters given by the curve Value reached value The difference between the moments when the time value equals 0.

2. The method according to claim 1, wherein, Step e) By fitting through the The first value of the curve or The value corresponding to =1 and the value estimated by the deep neural network model. The current value is used to execute the line.

3. The method according to claim 1, wherein, Step e) By applying all calculated data from the start of the monitoring phase to the present... The value is fitted to a straight line to perform the operation.

4. The method according to claim 1, wherein, Step e) By targeting only the M most recent ones up to the present The sequence of values ​​is fitted with a straight line to perform the operation.

5. The method according to claim 2, wherein, The straight line has angle coefficients obtained by weighted averaging of M angle coefficients obtained from M lines at M most recent consecutive moments.

6. The method according to any one of the preceding claims, wherein, If a warning was reported in step c), then use the subsequent... The check is repeated, and if the difference again does not fall within the range, the system reports an anomaly; otherwise, the first warning disappears.

7. The method according to any one of the preceding claims, wherein, The deep neural network model is a long short-term memory neural network.

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

Images