Vacuum system having a diagnostic circuit and method and computer program for monitoring the soundness of such a vacuum system

Abstract

A vacuum system including at least one cryopump, a plurality of sensors associated with the cryopump, each configured to sense an operating condition of the cryopump, and a diagnostic circuit configured to receive sampled signals from the sensors, the diagnostic circuit including a diagnostic model of the cryopump, the diagnostic model being derived from historical data of a plurality of cryopumps of the same type operating over a plurality of regeneration and repair periods, and configured to relate values ​​of the sampled signals from at least some of the sensors to a probability of pump failure within a predetermined time. The diagnostic circuit is configured to apply the sampled signals to the diagnostic model and determine a probability of at least one cryopump failure within a predetermined time from an output of the model.

Description

Technical Field

[0001] The field of the present invention relates to a vacuum system for confining a cryopump and a diagnostic apparatus and method for predicting the failure time of such a pump.

Background Art

[0002] A cryopump operates based on the principle of condensing or collecting the gas to be pumped. This means that the cryopump needs to be periodically regenerated to remove the collected gas. The regeneration involves isolating the pump from the vacuum system and heating the pump while introducing a purge gas to desorb or sublime the collected gas. This step takes the pump out of operation, and thus, the end user tries to schedule the regeneration step during the planned regular preventive maintenance (PM).

[0003] During the scheduled PM, the regeneration itself (which can take up to 4 hours from start to completion) is a planned event that the end user has to deal with in the overall system preparation. However, if the regeneration or maintenance has to be performed unexpectedly and thus occurs outside the scheduled PM, this represents an unscheduled downtime for the vacuum system. This can have an adverse impact on the end user, especially on the capacity, yield, and / or productivity of the vacuum system.

[0004] Unscheduled regeneration may be triggered in response to signals indicating that the cryopump is not functioning as predicted. Typically, these signals are based on the temperatures of the first and second stages (referred to as T1 and T2 respectively). The end user sets control limits, particularly for T2, and uses statistical process control (or the equivalent) to determine whether an increase or instability trend is true and whether to take action (or not). The problem is that relying solely on T1 and T2 to detect a pump whose performance is beginning to deteriorate results in an insufficient lead time before unscheduled regeneration or, even worse, measures such as a complete replacement of the unscheduled pump and replacement with a spare part must be taken. SUMMARY OF THE INVENTION PROBLEMS TO BE SOLVED BY THE INVENTION

[0005] It is considered desirable to provide a diagnostic system that can more accurately predict future pump failures, thereby reducing the occurrence of unscheduled maintenance requirements. MEANS FOR SOLVING THE PROBLEMS

[0006] A first aspect provides a vacuum system including at least one cryopump, a plurality of sensors each configured to sense operating conditions of at least one cryopump associated therewith, and a diagnostic circuit configured to receive signals sampled from at least a portion of the plurality of sensors, the diagnostic circuit including a diagnostic model of the cryopump, the diagnostic model being derived from historical data of a plurality of cryopumps of the same type operating over a plurality of regeneration and repair periods and configured to associate values of the sampling signals from at least some of the sensors with the probability of a pump failure within a predetermined time, the diagnostic circuit being configured to apply the sampling signals to the diagnostic model and determine from the output of the model the probability of at least one cryopump failure within a predetermined time.

[0007] The inventor recognized that by associating several sensors with a cryopump, signals indicating the operation of the pump can be sampled. These signals provide a detailed indication of the current soundness of the cryopump and how it is changing. Embodiments analyze this information to predict when the pump may fail and schedule maintenance accordingly. The step of collecting additional signals from the pump can provide additional information about the current soundness of the pump, but the drawback is that the more signals there are, the more difficult it becomes for the repair technician to substantially analyze them and derive relevant information therefrom. The inventor also recognized that not only are many of the sensed signals relevant, but a model can be constructed that accurately correlates the values of various signals sampled at various times and how these signal values change over time with the probability of failure when data is collected from many machines of the same type. Such a model was able to take into account many different signals and provide an effective way to predict failures. Furthermore, when new data is collected, such a model can be updated over time using the same machine learning techniques used to develop the initial model.

[0008] In fact, such a diagnostic solution is thought to be able to provide a more advanced and accurate notification of pump performance degradation than is possible today, thereby giving the end user the opportunity to take action at the next scheduled PM rather than incurring high-cost unscheduled downtime.

[0009] This solution can generally be advantageous across currently used cryopumps, but can be particularly advantageous for next-generation cryopump systems, especially for use cases of ion implantation where the life of the cryopump is predicted to be similarly short compared to the PM interval.

[0010] It should be noted that the prediction of pump failure is the point in time when it is evaluated that the performance of the pump is considered to be lower than a given threshold value, or the operation is considered to come to a complete stop if any maintenance or sometimes regeneration is not carried out. Therefore, the prediction of failure can be used as the estimated point in time for an engineer to intervene by replacing or repairing the pump in some way.

[0011] Sensors associated with a cryopump can be configured to sample some operating conditions of the cryopump, each indicating the current soundness of the vacuum pump, and the values and / or changes in values over time to give an indication of potential pump failure. By inputting at least a portion of these signals into a diagnostic model, an accurate pump failure probability can be derived.

[0012] In some embodiments, the vacuum system includes a plurality of cryopumps, and the diagnostic circuit is configured to receive signals from each of the plurality of cryopumps and determine the probability that each cryopump will fail.

[0013] The vacuum system can enclose only one cryopump, but in many cases, there are multiple cryopumps in a vacuum system such as a system for evacuating a semiconductor processing chamber. Scheduling the replacement and maintenance of these cryopumps is important for the productivity and yield of the system. Therefore, it is advantageous to provide an accurate prediction of cryopump failure such that any maintenance and / or replacement of the pump can be carried out during the scheduled preventive maintenance period.

[0014] In some embodiments, the diagnostic circuit outputs a warning indicating that the pump must be replaced in the next scheduled preventive maintenance in response to detecting that the probability of at least one cryopump failure exceeds a predetermined threshold value over a set time period.

[0015] As described above, the prediction of pump failures can be used to replace or repair the pump during the scheduled maintenance period. In an embodiment, the system determines where the probability of a cryopump failure rises above a predetermined probability threshold over a set time period and adopts this as an indication that the pump must be replaced during the next preventive maintenance period.

[0016] In this regard, the diagnostic model is set to determine the probability of a cryopump failure within a predetermined time period, which is selected to be the period between preventive maintenance or in some cases a shorter period than that. When this predetermined time period is selected to be the period between preventive maintenance, there is a high probability that the pump will fail before the next period when a warning signal is received, so the pump must be replaced during the next preventive maintenance period. However, if the failure prediction is, for example, a failure prediction within a range of 30 days and the maintenance schedule is 45 days, which is a typical period in the case of ion implantation, and this warning occurs near the end of the operating period, for example, on the 40th day, the pump may be able to last until the next preventive maintenance schedule.

[0017] In some embodiments, the vacuum pump system includes an input for receiving signals from a remote diagnostic system and an output for outputting signals to the remote diagnostic system.

[0018] The vacuum system and the diagnostic circuit can be used as independent units, but in some embodiments, the vacuum system is used in cooperation with a remote, in some cases cloud-based system, receives signals from this system, and outputs signals to this system.

[0019] In some embodiments, the vacuum system is configured to periodically output data collected from at least some of the sensors, along with data indicating maintenance performed on at least one cryopump, to a cloud-based diagnostic system.

[0020] The remote diagnostic system can be used to maintain the diagnostic model such that the accuracy of the diagnostic model is continuously improved and is related to the pump operation. Therefore, data indicating the operation and faults of the pump can be uploaded to the system and used to improve the model.

[0021] In some embodiments, the vacuum system includes an input for receiving from a repair technician data indicating the condition of the pump replaced during the maintenance period, and the data indicating the condition of the replacement pump is configured to be output as part of the regularly output data.

[0022] To further improve the model, additional data from the repair technician can be uploaded to the cloud-based diagnostic system. In this regard, data particularly relevant to predicting pump failures is not only whether the pump was replaced in response to determining the probability that the pump must be replaced, but also the condition of the pump at that time. Without this additional information, there is a risk that the model can only be considered a failure where the pump is replaced too late, and thus the model tends to predict a maintenance period shorter than the required maintenance period. By including data from the repair technician indicating the condition of the pump at the time of replacement, this additional information can be used to determine whether it is truly necessary to replace the pump at the current time. For example, if it is determined that the pump would have lasted until the next period, the maintenance period in the database used to generate the diagnostic model can be adjusted. Further, in some cases, data indicating the reason for pump failure, such as contamination, can be output to the diagnostic model, which can be used to improve the vacuum system as a whole.

[0023] In some embodiments, the vacuum system includes an input configured to periodically receive updates to the diagnostic model.

[0024] As described above, continuously updating the diagnostic model can be advantageous in that the model is improved, and having an input for receiving such a model enables such an update to occur.

[0025] In some embodiments, the vacuum system includes an input configured to periodically receive updates to the diagnostic model from a cloud-based diagnostic system.

[0026] The model can be updated manually by a repair technician, but in some cases is updated by directly receiving a signal from a cloud-based diagnostic system. In this way, the cloud-based diagnostic system can automatically update the model when it determines that it can improve the model from the received data.

[0027] A second aspect provides a method for monitoring a vacuum system including at least one cryopump, the method comprising sampling a plurality of signals indicative of operating conditions of the at least one cryopump from sensors associated with the cryopump; inputting at least a portion of the signals into a diagnostic model derived from historical data of a plurality of cryopumps of the same type operating over a plurality of periods, at least a portion of the historical data including at least one of pump regeneration, repair, and failure, the diagnostic model associating the signals with the probability that the pump will fail; and determining a probability of at least one cryopump failure from an output of the model.

[0028] A third aspect provides a computer program including machine-readable instructions operable to control a computer to perform the method according to the second aspect of the invention when executed by the computer.

[0029] A fourth aspect of the present invention provides a method for generating a diagnostic model for a certain type of cryopump, the method comprising: inputting data collected from sensing the operating conditions of a plurality of cryopumps of this type into a machine learning algorithm from a database storing the sensed operating conditions of the plurality of cryopumps sampled over a plurality of periods including at least one of regeneration, repair, and failure, generating a plurality of probabilities that the pump will fail within a predetermined time, comparing the probability for each pump with the pump failure timing for the pump retrieved from the database, updating the parameters within the machine learning algorithm to reduce the difference between the determined probability and the pump failure timing, repeating these steps until the difference reaches either a minimum value or a predetermined value, and generating a diagnostic model from the algorithm and the parameters that gave the difference.

[0030] The diagnostic model used to diagnose the vacuum system of the embodiment can be generated using a machine learning algorithm having access to a database containing the sensed operating conditions of a plurality of pumps connected over an operating period including periods when the pumps are operating, being regenerated, and at least some of these pumps are failing. The machine learning algorithm is configured to generate a pump failure probability from the sensed and sampled operating conditions, these probabilities are then compared with the actual pump failure timing obtained from the database, and further the machine learning algorithm is updated until these probabilities and the actual pump failure timing match more closely with each other. A diagnostic model is generated from the above algorithm when an appropriate match is found, which can be where the failure prediction is considered to give a predetermined desired accuracy when compared with the actual failure, or where only a minimum difference is seen between these values.

[0031] In some embodiments, before the data is input into the machine learning algorithm, these data are filtered to remove abnormal data. In this regard, the sensed signal may have associated noise such that it is not accurate and does not represent the true operating conditions of the pump, and / or the pump may have a malfunction such that it is operated at a particularly low or high temperature, whereby the signal from this pump does not represent other pump operations.

[0032] In some embodiments, before the received signal is input into the machine learning algorithm, the signal is filtered to remove signals sampled at predetermined times before and after the regeneration of the cryopump.

[0033] The filtering of the signal can be performed to remove signals that may not be outliers but are known not to represent the normal operation of the pump. For example, signals sampled at predetermined times before and after the regeneration of the cryopump may not represent normal operation.

[0034] The fifth aspect provides a method for updating the diagnostic model generated according to the fourth aspect. The method includes receiving a further plurality of signals indicating the operating conditions of a plurality of cryopumps of the same type, receiving the timing of pump maintenance and failures for the plurality of cryopumps, adding the received data to a database, implementing the method of the fourth aspect to generate an updated diagnostic model, and outputting the updated diagnostic model.

[0035] Once the diagnostic model has been generated, it will then be used to predict cryopump failures so that the diagnostic model can diagnose cryopumps of the above type and replace these pumps without shutting down the machine. The use of this model involves sampling data indicating the operating conditions and failures of cryopumps. Therefore, in some cases, this data is collected and uploaded to the central diagnostic model generation means and then returned and used to periodically update the model. Since the model is generated from the collected data using a machine learning algorithm, the model can be updated using the same model generation method by simply including additional data in the database initially used to generate the model. In this regard, the additional data can be added to the original database or can replace some of the old data in the database.

[0036] In some embodiments, the method further includes receiving data indicating the condition of the replaced pump, which the diagnostic model indicates that the pump is likely to fail within a predetermined time, and inputting the data into the machine learning algorithm during the update of the diagnostic model.

[0037] When the repair technician inputs data indicating the condition of the pump after the failure, this information can be beneficial for improving the model. Therefore, this information can be included in the data input into the machine learning algorithm. In this regard, the machine learning algorithm can be adapted to receive this data, or when the time point at which the pump is predicted to fail is different from the time point of pump replacement, the failure time point can be corrected in the database from this replacement time point to this prediction time point.

[0038] A sixth aspect provides a computer program comprising machine-readable instructions operable to control a computer to perform the method according to the fourth or fifth aspect when executed by the computer.

[0039] A seventh aspect provides a remote diagnosis module including a computer configured to execute a computer program according to the sixth aspect.

[0040] An eighth aspect provides a system including a vacuum system according to the first aspect and a remote diagnosis module according to the seventh aspect, wherein the remote diagnosis module is configured to receive a signal output by the vacuum system, update a database of cryopump operation with the received signal, and generate an updated diagnosis module by inputting data from the updated database into a machine learning algorithm.

[0041] Still other specific and preferred aspects are set forth in the appended independent and dependent claims. The features of the dependent claims can be combined in combinations other than those specified in the claims in combination with the features of the independent claims as required.

[0042] When explaining that a device feature is operable to provide a function, it will be recognized that this device feature includes a device feature that provides that function or a device feature adapted or configured to provide that function.

[0043] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings.

Brief Description of the Drawings

[0044]

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Fig. 5

Fig. 6A

Fig. 6B

[0045] Before discussing the embodiments in more detail, an overview is first provided.

[0046] The embodiments seek to provide a system that can predict the failure of a cryopump and thereby predict its necessary maintenance or regeneration. Such pumps require regular regeneration / maintenance and are typically scheduled to be removed from a vacuum system such as a processing chamber and replaced every 45 to 60 days. If the need for maintenance can be accurately predicted well in advance, these maintenance operations can be coordinated with the scheduled maintenance period, reducing or even eliminating unscheduled maintenance periods.

[0047] The embodiments provide a diagnostic system that receives inputs from sensors associated with a cryopump, inputs these into a model that models the operation of the cryopump, and predicts future failures from the values of the received signals and / or their changes. The model is generated from an analysis of historical data collected from sensors associated with a plurality of cryopumps of the same type operating over a period that includes at least a portion of the scheduled maintenance period. The model can be periodically updated by analyzing new received data from the cryopump being diagnosed by the system using machine learning techniques.

[0048] In this way, a diagnostic system that continuously improves using the collected data from the currently operating cryopump is created.

[0049] Embodiments provide a combination of batch information and real-time information that gives a quantitative probability of "failure" in the range of a near-term specified interval (preferably longer than the PM interval) for a pump system on a particular tool or across multiple tools. This can provide very valuable guidance to the end user or customer repair technician for making decisions regarding when to replace a pump, and can avoid unplanned outages.

[0050] FIG. 1 shows a vacuum system according to an embodiment. In this vacuum system 5, there are three cryopumps 10 controlled by a supervision node 20. The supervision node 20 includes a diagnostic circuit shown as 22 containing a logic section including a diagnostic model for predicting failures of the vacuum pumps.

[0051] During operation, sensors (not shown) associated with the vacuum pumps 10 sense the operating conditions of the pumps and send signals indicating these operating conditions to the supervision node 20. The diagnostic circuit 22 within the supervision node 20 samples at least a portion of the signals and inputs the sampled signals into the diagnostic model as input data. These operating conditions can include the temperature of the first stage of the vacuum pump, the temperature of the second stage of the vacuum pump, the time required to reach the desired temperature at each of the vacuum pump stages, the speed of the vacuum pump motor, and other variables that specifically indicate the operation of the vacuum pump 10.

[0052] In some embodiments, the vacuum pump is a two-stage cryopump having a first stage array and a radiation shield both coupled to the first stage of a refrigerator and a second stage array coupled to the second stage of the refrigerator. Such a cryopump has temperature sensors for monitoring these temperatures with respect to the first stage array and the second stage array. During regeneration of the cryopump, the cryopump is heated so that the condensed gas is released. There is a heater that can include an electric heater circuit for heating these arrays during this procedure. Inputs to the diagnostic system can include the temperature readings of the temperature sensors with respect to these arrays, inputs to the heater such as the current supplied to the heater circuit, and the speed of the motor.

[0053] At least some of these signals are input into a diagnostic model that predicts the probability of failure for each of the pumps from the changes and values of these signals. In this regard, in this embodiment, the diagnostic model determines when to predict the probability of failure of one of the pumps within a predetermined period. This predetermined period can be, for example, 45 days, and this period can be the time between scheduled maintenance. Alternatively, the predetermined period can be shorter than the time between scheduled maintenance depending on the model. A warning signal is generated indicating that generally the corresponding pump must be replaced in the next scheduled maintenance if the probability does not exceed a predetermined value, in this example greater than 50%, and remains above this value for a predetermined length of time, in this case 6 days. Note that the threshold value and the set time are not predetermined and can be changed and optimized based on the usage pattern and failure pattern of the pump.

[0054] If the period for failure prediction is, for example, 30 days, the maintenance period is longer than this period, for example, 45 days, and this signal is triggered towards the end of the operating period, this warning signal can indicate that the pump must be replaced but not until the next maintenance period.

[0055] Following the replacement of the pump, in some embodiments, the repair technician will inspect the soundness of the replaced pump and input information indicating the condition of this pump and its predicted life prior to an actual failure at times into the supervision node 20. In this regard, this pump may have been replaced too early, and this information may be advantageous when optimizing the model. Information regarding the cause of the pump failure can be input if it can be derived.

[0056] Therefore, by using the diagnostic model 22 within the supervision node 20, an accurate prediction of the pump failure at any future predetermined time point can be determined, and then, instead of having to stop the entire vacuum system to urgently replace the pump, the pump can be replaced during the scheduled maintenance.

[0057] In this embodiment, there is also a connection to a remote system or a cloud-based system 30. Within the remote system 30, there is a streaming data management tool 32 that sends data to the data storage 34 or the diagnostic model 33. The diagnostic model 33 includes a log parser 33a, a data storage 33b, and a machine learning engine 35. Within the remote system 30, several data operations occur, which include, but are not limited to, the management and analysis of the streaming data regarding each pump in the system, long-term storage in a structure such as the data lake 34, and processing by a very specialized machine learning algorithm 35 for quickly outputting the failure probability. These outputs can be sent to the host 25 (e.g., the tool control system and / or the factory process data management system), repair technicians, or end-users in various forms and on various platforms including mobile devices.

[0058] The above operations can be implemented on a local server in many embodiments where they are implemented within the cloud. However, the cloud offers significant advantages regarding the function of frequently updating the machine learning model and / or software as needed, and at the same time, provides an expandable and inexpensive standard architecture (e.g., Amazon Web Services, IBM Watson, or equivalents) that is the framework for operation.

[0059] A remote system or cloud-based system is configured to store data from the operation of many pumps of the same type in a database or data lake 34. These data include, but are not limited to, temperature, motor speed, heater input, regeneration parameters, years of use, etc. The remote system or cloud-based system 30 includes a logic unit 33 for generating and updating a diagnostic model using these data. The cloud-based logic unit 33 accesses a data lake 34 containing data indicating the operating conditions of a plurality of pumps of the same type during a predetermined time including the time when these pumps are regenerated, repaired, and replaced.

[0060] The logic unit 33 can generate a diagnostic model from these data. To generate the diagnostic model, the logic unit 33 samples the data lake 34 and inputs it into a machine learning algorithm 35 that predicts the failure probability of a plurality of pumps based on the data indicating the operating conditions of the collected pumps. The logic unit 33 includes a log parser 33a for receiving data, a data store 33b, and a machine learning engine 35. This prediction is compared with the actual failure rate of the pumps stored, and the parameters of the model are changed until the probability and the failure rate match to a desirable degree. At this point, the model is considered to be accurate enough and is sent for use in the diagnostic circuit 22 within the supervision node 20.

[0061] During operation, the supervision node 20 can periodically transmit the operating conditions of the pump from the sensors associated with the pump, as well as the information input by the repair technician regarding the conditions of the pump at the time when the collected data for detecting regeneration and failure are exchanged, and these data can be added to the data lake 34. This process is controlled by the streaming data management circuit 32. Instead of and / or in addition to this, the data can be uploaded manually or by the streaming management circuit 32 in response to the repair technician noticing that the accuracy of the model has fallen below a predetermined level.

[0062] The remote logic unit 33 can regenerate the machine learning algorithm 35 using additional data based on a request to generate an updated and improved model, and periodically upload it to the supervision node 20 to replace the previously stored model. In this way, the model will be continuously improved and adapted to the updated conditions.

[0063] In some embodiments, the new collected data is simply added to the data lake 34, while in other embodiments, some of the data can be replaced. The replaced data can be selected as the oldest data, and / or it can be selected as the data that least represents the characteristics of the corresponding type of pump.

[0064] Alternatively and / or in addition, the logic unit 33 can filter the data before inputting it into the learning algorithm 35, and can remove any outliers in the received data during this process. In this regard, the operating conditions of the pump may not exhibit the characteristics of normal operation of the pump at certain points in time, such as immediately before or after regeneration. Therefore, such data can be filtered out from the data added to the machine learning algorithm 35.

[0065] FIG. 2 shows how a diagnostic model or prediction model 155 is formed from data collected from various pumps. The data input into the machine learning algorithm to construct the model includes initial pump data 100 that can include data regarding a specific pump, and the pump's serial number, part number, model, whether the pump has been removed and replaced / removed, the date / time when the pump was installed and put into operation, the date / time when the most recent pump data was recorded, the duration regarding the pump, the initial time of the pump at installation, the factory where the pump was installed, the name of the tool where the pump was installed, the manufacturer of the tool, the type of the tool and the tool station name, and whether the pump is marked as having been repaired and adjusted.

[0066] In addition to this, an initial pump program 110 indicating the operation of the pump is input into the machine learning algorithm. This initial pump program 110 can include the weighted average of the temperature of the first stage for the most recent 10 days, 30 days, and 60 days, the weighted average of the temperature of the second stage for the most recent 10 days, 30 days, and 60 days, the weighted average of the RPM for the most recent 10 days, 30 days, and 60 days, the maximum value of the temperature of the first stage, the maximum value of the temperature of the second stage, and the maximum value of the RPM.

[0067] The initial pump manufacturing inspection 120 can also be input. This manufacturing inspection 120 can include the pass / fail count value of the pump and the aggregated statistical value of the passes. The regeneration data 130 can also be input. This regeneration data 130 includes the regeneration stage of the pump, the temperature of the first stage, the temperature of the second stage, the motor status, the purge valve status, the rough valve status, the heater 1 status, the heater 2 status, the heater 1 percent on, the heater 2 percent on, the motor RPM, the duration of the pump, the time from the last high-speed regeneration, the time from the last complete regeneration, the date of the real-time record, the time required for regeneration, the pump time when regeneration started, the date / time when regeneration started, the date / time when regeneration ended, the temperature of the first stage at the end of regeneration, the temperature of the second stage at the end of regeneration, the base pressure setting for regeneration, the increase rate limit setting for regeneration, the number of increase rate cycles implemented, the time required for the pump to reach rough finishing conditions, the time required for the pump to cool, the time between the most recent complete regeneration and the time when the pump started to regenerate, the weighted average of the temperature of the first stage for the most recent 10 days, 30 days, and 60 days, the weighted average of the temperature of the second stage for the most recent 10 days, 30 days, and 60 days, the weighted average of the RPM for the most recent 10 days, 30 days, and 60 days, the maximum value of the temperature of the first stage, the maximum value of the temperature of the second stage, and the maximum value of the RPM. These regeneration data are inspected for missing data and can be made into standard regeneration values for generating aggregated regeneration data 131 in some cases.

[0068] At the start of the process for constructing the first diagnostic model using machine learning techniques, the pump data as outlined above is collected from several different information sources. This data includes the operating pump data collected from sensors associated with the pump, and further includes data on pumps that have operated for more than 300 hours. Similar to the data from the inspections performed on the pump during manufacturing and the data collected during the regeneration of the pump, a pump log indicating how the data changes over time is also collected. Next, these data are filtered to remove unnecessary attributes, outliers, and conditions where the data is considered inappropriate or not required, or conditions that generally do not indicate the characteristics of the pump's operation, and aggregated pump data 101, aggregated pump log 111, aggregated manufacturing inspection 121, and aggregated regeneration data 131 are provided from the filtered data.

[0069] Next, these data are processed to remove the data for the last two months regarding the unexchanged pumps and the data immediately before (the previous two hours) and immediately after regeneration (the next four hours).

[0070] Next, the collected data is divided into training data 140, which includes data from 300 exchanged pumps and 100 unexchanged pumps in this example, and prediction data 150, which includes data from 39 exchanged pumps and 9 unexchanged pumps. These data are used for the generation of the prediction model 155.

[0071] The training data 140 is used to define the machine learning algorithm in the generation of features and to train the data using the model. The prediction data 150 is used to apply the machine learning algorithm.

[0072] A prediction model 155 is generated from these processes.

[0073] In summary, the collected training data 140 is input into a machine learning algorithm and used as training data for generating a diagnostic model to predict pump failures from the data collected from sensors. The model is configured by comparing the values generated from the training data 140 with the actual data, and when these two are considered to be in agreement, this model is output for use in a vacuum system.

[0074] There are many different types of machine learning algorithms that can be used. In fact, the training data 140 can be input into several different models, and the one that gives the diagnostic model with the most accurate predictions is selected. In this regard, this process is an iterative process, and the accuracy of the model can be determined by comparing the predicted failure rate with the actual failure rate of the data. As described above, the predicted failure of the pump is a failure for which the pump requires some repair. Thus, the actual "failure" may be related to the actual "situation" where the operation of the pump is outside the required limits, or it may simply be the point in time when the pump is replaced so that it can function.

[0075] Figure 3 shows an example of one machine learning algorithm that has been found to generate a diagnostic model that gives particularly good results when used in combination with the data described above. This algorithm is a random forest simplified machine learning algorithm where, in this case, multiple forests with hundreds of trees are used, each tree makes a prediction, and each forest finds a probability. Next, the lowest failure probability from the multiple forests is determined, and this lowest failure probability is used to generate a warning that is selected to be triggered when the probability reaches or exceeds 0.5. If such a probability continues for six consecutive days, replacement at the next preventive maintenance time point is recommended.

[0076] Therefore, starting from instance 200, the random forest technique is used to reach trees 1 201, 2 202, ..., n 20n. From these trees, different classes 205a, 205b, 205n are derived, and the majority vote 209 is used to reach the final class 210.

[0077] Figure 4 shows a flowchart illustrating the steps in a method for generating or updating a diagnostic model of an embodiment. First, at step S10, data indicating the operating conditions of a plurality of pumps is input into a machine learning algorithm. These data can be data from a database such as the data lake 34 of FIG. 1. At step S20, the failure probability of each of the pumps is derived from the machine learning algorithm and compared at step S30 with the actual failures including the regeneration data and / or repair data for these pumps. In this regard, the pumps may not fail as predicted but may be repaired or regenerated to avoid failure, and thus these data are used. Next, the accuracy of the predicted probability is determined.

[0078] Next, in step S40, the actual algorithm of the parameter and / or model that can include the weight coefficient is changed, and the failure probability updated in step S50 is determined. Next, by comparing the generated probability with the actual failure, the accuracy of the updated probability is determined. If it is determined at D5 that this probability is more accurate than the previous measure ("Yes"), the parameter is changed in the same direction in S40 and recalculation is performed. If the accuracy of the updated measure is lower ( "No" at D5), the diagnostic model is generated in step S70 using the previous values. This flow is a highly simplified schematic of a complex procedure. As will be understood, the steps of changing the parameter can be done in several steps, each including a subset of the parameter and / or the calculations performed, and / or these variations can be done first at a coarse level and then, when the minimum value is found from the coarse level, next at a fine level. In any case, the method is iterative and a diagnostic model can be generated from these parameters when no further improvement is felt or when the desired accuracy is reached.

[0079] FIG. 5 is a flowchart showing a method for predicting pump failure according to an embodiment. The method includes sampling, at step S100, a signal indicating the operating conditions of the pump from a sensor associated with the pump. At step S110, the signal is input into the diagnostic model, and at step S120, the pump failure probability is determined from this model. At D15, it is determined whether this probability exceeds a predetermined value and actually remains above this value for a predetermined time. If so ("Yes"), a warning indicating that the pump must be repaired or replaced is output at step S130. If the probability does not exceed the threshold, no warning is output.

[0080] Figure 6 shows the values of this approach for an example of a single pump. Variations in several monitored values are shown, namely, T1 which is the temperature of the first stage of the pump, T2 which is the temperature of the second stage of the pump, the speed of the pump motor 300, the value of T1 at the end of the regeneration cycle T1end and the value of T2 at the end of the regeneration cycle T2end, and the variation in the time 310 to cool to the required temperature. From these values, a failure probability 330 is detected using the diagnostic model of the embodiment. This probability can be monitored following its trend and can generate a warning / alarm to replace the pump when it exceeds a predetermined threshold value.

[0081] In FIG. 6A, the start of the red area predicts that the pump may fail and indicates the place where it should be replaced. In this case, the pump is not replaced and FIG. 6A shows how these values change thereafter. FIG. 6B shows another embodiment where the pump is replaced when there is a possibility of pump failure.

[0082] Embodiments can monitor pump performance over time and be used in conjunction with conventional repair provisions that provide repairs to the pump. In conventional repair provisions, data is collected from the pump. These data can include, but are not limited to, temperature, motor speed, beater input, regeneration parameters, years of use, etc. These data are collected and interpreted by an embedded system. The data summary is packaged and sent by email at predetermined intervals. The data can be viewed either in an unprocessed form or visually as a set of graphs. The repair technician uses the latter to identify trends in the main control parameters such as T1, T2, and the pump speed in rpm. Next, the repair technician combines this information with the preliminary knowledge of this equipment and uses it to manage the end-user's pump group in terms of proposing pump replacement or other corrective measures. In other words, such a system provides insights into pump performance but requires the active involvement of an experienced repair technician who can recognize trends and patterns.

[0083] The solution of this proposal is composed based on the same basic data, but uses state-of-the-art and highly proprietary machine learning algorithms to calculate the possibility that a specific pump may fail at any fixed interval (for example, 30 days, 45 days, 60 days). These algorithms utilize one or more machine learning (ML) methods (including but not limited to random forest, neural network, principal component analysis, etc.). The ML algorithms are constructed and "trained" on this database where past replacement events have been clearly identified, and then verified against another subset of the database.

[0084] The main advantages of this approach include the following: · A function to predict the performance degradation of a pump that predicts performance degradation much earlier than the time when it is considered that the prediction is reflected in normal indicators such as T1, T2, and pump speed, because the prediction depends on "features" or transformation variables that represent other aspects of pump operation. · Avoid or at least reduce the (often incorrect) dependence on strict control limits for T1 and T2, which may lead to premature replacement of the pump by the end user. · Generate a platform for incorporating additional sensors or diagnostic information that can improve the accuracy and / or lead time of model predictions. · Provide guidelines regarding the proactive replacement of the pump using probability functions, and be able to eliminate (in principle) or at least reduce unplanned downtime if the end user follows the guidelines.

[0085] Although the exemplary embodiments of the present invention have been disclosed in detail with reference to the accompanying drawings herein, it is understood that the present invention is not limited to the exact embodiments, and those skilled in the art can make various modifications and alterations to these embodiments without departing from the scope of the present invention defined by the claims and their equivalents.

[0086] Reference symbol 5 Vacuum system 10 Cryopump 20 Supervisor node 22 Diagnostic circuit 25 Host 30 Remote or cloud-based system 32 Data streaming management circuit 33 Diagnostic model generation logic unit 33a Log parser 33b Data storage 34 Data lake 35 Machine learning algorithm

Claims

1. At least one cryopump, A plurality of sensors associated with the at least one cryopump, each of the plurality of sensors being configured to sense the operating status of the at least one cryopump, A diagnostic circuit configured to receive signals sampled from at least a portion of the plurality of sensors, the diagnostic circuit including a diagnostic model of the cryopump, the diagnostic model being derived from historical data of a plurality of cryopumps of the same type operating over a plurality of regeneration and repair periods, and being configured to associate the value of the signals sampled from the at least a portion of the sensors with the probability of failure of the at least one cryopump within a predetermined time, the diagnostic circuit being configured to apply the sampled signals to the diagnostic model and determine the probability of failure of the at least one cryopump within the predetermined time from the output of the diagnostic model, A vacuum system characterized by including the above.

2. The vacuum system according to claim 1, characterized in that at least a portion of the signals includes at least a portion of the first temperature of the first stage of the cryopump, the second temperature of the second stage of the cryopump, the time since the most recent regeneration, the speed of the motor of the cryopump, the input to the heater circuit, and the time for the cryopump to cool to the first temperature.

3. The vacuum system includes a plurality of cryopumps, The diagnostic circuit is configured to receive signals from each of the plurality of cryopumps and determine the probability of each of the failures of the at least one cryopump. The vacuum system according to any one of claims 1 to 2, characterized by the above.

4. The vacuum system according to any one of claims 1 to 3, characterized in that the diagnostic circuit outputs a warning indicating that the cryopump should be replaced in the next scheduled preventive maintenance in response to detecting that the probability of failure of the at least one cryopump is above a predetermined threshold over a set time.

5. The vacuum system according to any one of claims 1 to 4, characterized by comprising an input for receiving a signal from a remote diagnostic system and an output for outputting a signal to the remote diagnostic system.

6. The vacuum system according to claim 5, characterized in that it is configured to periodically output to the remote diagnostic system data collected from at least some of the sensors together with data indicating maintenance performed on the at least one cryopump.

7. Including an input for receiving from a repair technician data indicating the status of a pump replaced during the maintenance period, configured to output the data indicating the status of the replaced pump as part of the periodically output data, The vacuum system according to claim 6, characterized by this.

8. The vacuum system according to any one of claims 1 to 7, characterized by including an input configured to periodically receive an update to the diagnostic model.

9. The vacuum system according to claim 8, when dependent on any one of claims 5 to 7, characterized by including an input configured to periodically receive the update to the diagnostic model from the remote diagnostic system.

10. A method of monitoring a vacuum system including at least one cryopump, comprising: sampling a plurality of signals indicating the operating status of the at least one cryopump from sensors associated with the cryopump; inputting at least a portion of the signals into a diagnostic model of the cryopump, wherein the diagnostic model is derived from historical data of a plurality of cryopumps of the same type operating over a plurality of periods, at least a portion of which includes at least one of regeneration, repair, and failure of the cryopump, and the diagnostic model is inputting the signals in association with the probability of failure of the cryopump; determining the probability of failure of the at least one cryopump from the output of the diagnostic model; A method characterized by including.

11. Machine-readable instructions operable to control the computer to perform the method according to claim 10 when executed by the computer, A computer program characterized by including.

12. A method of generating a diagnostic model for a cryopump, comprising: Inputting data collected from sensing the operating status of a plurality of cryopumps, which stores the sensed operating status of the plurality of cryopumps sampled over a plurality of periods including at least one of regeneration, repair, and failure, into a machine learning algorithm, and generating a plurality of probabilities of failure of the cryopumps within a certain period of time; Comparing the probabilities for each cryopump with the pump failure timing for the cryopump retrieved from the database, and updating the parameters within the machine learning algorithm to reduce the difference between the determined probability and the pump failure timing; Repeating the above steps until the difference reaches either a minimum value or a predetermined value; Generating the diagnostic model from the machine learning algorithm and the parameters that gave the difference; A method characterized by including the above steps.

13. The method according to claim 12, wherein before inputting the data into the machine learning algorithm, the data is filtered to remove abnormal data.

14. The method according to claim 12 or claim 13, wherein before inputting the data into the machine learning algorithm, the data is filtered to remove signals sampled at predetermined times before and after regeneration of the cryopump.

15. A method for updating the diagnostic model generated according to any one of claims 12 to 14, comprising: Receiving a plurality of data indicating the operating status of a plurality of cryopumps of the same type; Receiving pump maintenance and failure timing for the plurality of cryopumps; Adding the data to the database, and implementing the method according to any one of claims 12 to 14 to generate an updated diagnostic model using the data from the updated database; Outputting the updated diagnostic model; A method characterized by including the above steps.

16. Further including the step of receiving data indicating the status of a cryopump replaced following the diagnostic model indicating that the cryopump is likely to fail within a predetermined period of time. The data is input to the machine learning algorithm during the update of the diagnostic model. The method according to claim 15, characterized in that.

17. Machine-readable instructions operable to control the computer to perform the method according to any one of claims 12 to 16 when executed by the computer. A computer program comprising the above.

18. A computer configured to execute the computer program according to claim 17. A remote diagnostic module comprising the above.

19. A system comprising the vacuum system according to any one of claims 5 to 7 or claim 9 and the remote diagnostic module according to claim 18, The remote diagnostic module is configured to receive a signal output by the vacuum system, update the database of cryopump operation using the received signal, and generate an updated diagnostic model by inputting data from the updated database to the machine learning algorithm. The system is characterized by the above.

Images