Vacuum system fault prediction method and system

Abstract

The invention provides a vacuum system fault prediction method and system. The method comprises the following steps: S1, dividing a vacuum system into a plurality of subsystems; s2, collecting and preprocessing data of each subsystem, further training a fault prediction model corresponding to each subsystem, and obtaining output data of the fault prediction model; and S3, fusing output data of all the fault prediction models to obtain a fault prediction result of the vacuum system. The method can be used for a large-scale complex vacuum system, system faults are predicted in advance and maintained, and then long-term stable operation of the vacuum system is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of vacuum system fault prediction, and in particular relates to a vacuum system fault prediction method and system. BACKGROUND

[0002] With the rapid progress of today's technology, vacuum systems play a crucial role in the fields of semiconductor manufacturing, biological engineering, aerospace, etc. The composition of the vacuum system has become extremely complex, usually consisting of vacuum pumps, valves, pipelines, sensors and other complex component modules, and its running stability and reliability directly affect the overall system operation efficiency and product quality.

[0003] As the vacuum system becomes more and more complex, the risk of component failure due to aging, wear, environmental factor changes, etc. of a single component of the vacuum system is rising, leading to various faults of the vacuum system and difficulty in locating single component faults; a large amount of normal operation data and fault data need to be collected for a complex vacuum system, which cannot be quickly applied in a new system and continuously run.

[0004] Therefore, it is of great value to develop an efficient and accurate fault prediction technology for a vacuum system, which can quickly apply in a new vacuum system and accurately identify and locate faults in the early or initial stage, so as to improve the stability and reliability of the vacuum system operation.

[0005] At present, common vacuum system fault prediction technologies include signal processing-based prediction, expert system-based prediction, and model-based prediction methods. Although these methods have their own advantages, when dealing with large and complex vacuum systems, there are often problems such as long time to collect system fault data, weak self-adaptability, and difficulty in locating specific fault sources.

[0006] Patent document CN118861587A discloses a vacuum dry pump fault prediction method based on distribution feature subsets, which includes: constructing a vacuum dry pump running state vector set DPVS; constructing a data distribution vector set DPFI based on feature subsets; constructing a vacuum dry pump distribution feature subset DPFS; calculating the distance between the vacuum dry pump distribution feature subsets according to DPFS to obtain a to-be-predicted vacuum dry pump distribution feature subset distance list fs_dist; constructing a dry pump fault prediction training set FDDPD based on the distribution feature subset according to DPFI and fs_dist; and using a Logistic classifier to establish a vacuum dry pump fault prediction model BFPM. This scheme cannot solve the technical problems of long time to collect system fault data and weak self-adaptability.

[0007] This problem needs to be solved urgently. SUMMARY

[0008] In view of the defects in the prior art, the purpose of the present application is to provide a vacuum system fault prediction method and system.

[0009] According to the present application, a vacuum system fault prediction method is provided, comprising:

[0010] Step S1: dividing the vacuum system into multiple subsystems;

[0011] Step S2: collecting and preprocessing the data of each subsystem, then training the corresponding fault prediction model of each subsystem, and obtaining the output data of the fault prediction model;

[0012] Step S3: fusing the output data of all the fault prediction models to obtain the fault prediction result of the vacuum system.

[0013] Preferably, in the step S1, the vacuum system is divided into m subsystems, and the mathematical expression is:

[0014] S=[s1,s2,…,s i ,…,s m ]

[0015] Wherein, S represents the vacuum system, s i represents the i-th subsystem; i≤m, m is a constant;

[0016] The subsystems include dry pumps, molecular pumps, cryogenic pumps, ion pumps, vacuum valves, vacuum gauges and thermometers.

[0017] In the step S2, the data of the subsystems include environmental temperature T, operating current I, vibration frequency f, cooling water temperature T w , gas flow rate V q , gas pressure P q and gas temperature T q .

[0018] Preferably, in the step S2, the data of each subsystem is preprocessed to obtain the time series data of the subsystem, and then the corresponding fault prediction model of each subsystem is trained through the time series data;

[0019] The time series data of the subsystem within the time window t1 to t n is represented as:

[0020] [(X i1 , Y i1 ), (X i2 , Y i2 ), …, (X ij , Y ij ), …, (X in , Y in )]

[0021] wherein n time windows are arranged in order from near to far in time, and the time length is the same; n is a constant, (X ij , Y ij ) represents the data acquisition point of the i-th subsystem at t j ;

[0022] X ij and Y ij The mathematical expressions from top to bottom are as follows:

[0023] X ij = [s type , T, I, f, T w , V q , P q , T q , …]

[0024] Y ij = [y, y type ]

[0025] wherein X ij represents the abscissa of the data acquisition point of the i-th subsystem at t j time as input data, Y ij represents the ordinate of the data acquisition point of the i-th subsystem at t j time as output data; s type represents the fault type of the input data; y type represents the fault type of the output data; y represents whether it is a fault; T represents the environmental temperature, I represents the running current, f represents the vibration frequency, T w represents the cooling water temperature, V q represents the gas flow rate, P q represents the gas pressure, and T q represents other temperatures.

[0026] Preferably, in the step S2, a training data set is constructed, and then a fault prediction model of each subsystem is trained to obtain a fine-tuning model of the current vacuum system; the fine-tuning model of the current vacuum system is used to predict faults to obtain the fault prediction result of the vacuum system;

[0027] The samples of the training data set are organized as follows:

[0028]

[0029] wherein X represents the input matrix of the fault prediction model; Y represents the output matrix of the fault prediction model; and then (X, Y) data samples are constructed for model training, wherein n is the column number of the input matrix of the fault prediction model; k is the row number of the input matrix of the fault prediction model.

[0030] The mathematical expression of the failure prediction model is:

[0031] Y = F(X; θ base )

[0032] Wherein, F represents a supervised training model, any one of Transformer, BP neural network or Bi-LSTM, i.e. bidirectional long short-term memory network model; θbase represents parameters after training by historical data;

[0033] The fine-tuning model of the current vacuum system, the mathematical expression is:

[0034] Y = F(X; θ base + Δθ)

[0035] Wherein, Δθ represents the change amount of parameters obtained by data training.

[0036] Preferably, in the step S3, the failure prediction results of each of the subsystems are aggregated by using a weighted average or a majority voting strategy to obtain the failure prediction result of the vacuum system.

[0037] According to the vacuum system failure prediction system provided by the application, the vacuum system is divided into a plurality of subsystems, the data of each subsystem is collected and preprocessed, and the corresponding failure prediction model of each subsystem is trained to obtain the output data of the failure prediction model.

[0038] Module M1: dividing the vacuum system into a plurality of subsystems;

[0039] Module M2: collecting and preprocessing the data of each subsystem, and training the corresponding failure prediction model of each subsystem to obtain the output data of the failure prediction model;

[0040] Module M3: fusing the output data of all the failure prediction models to obtain the failure prediction result of the vacuum system.

[0041] Preferably, in the module M1, the vacuum system is divided into m subsystems, and the mathematical expression is:

[0042] S = [s1, s2, …, s i ,…, s m ]

[0043] Wherein, S represents the vacuum system, s i represents the i-th subsystem; i≤m, m is a constant;

[0044] The subsystems include dry pumps, molecular pumps, cryogenic pumps, ion pumps, vacuum valves, vacuum gauges and thermometers.

[0045] Preferably, in the module M2, data of each subsystem is preprocessed to obtain time series data of the subsystem, and then the time series data is used to train a corresponding failure prediction model of each subsystem.

[0046] The time series data of the subsystem within a time window t1 to t n is expressed as:

[0047] [(X i1 , Y i1 ), (X i2 , Y i2 ), …, (X ij , Y ij ), …, (X in , Y in )]

[0048] wherein n time windows are arranged in a sequence from near to far in time, and have the same time length; n is a constant, (X ij , Y ij ) represents a data collection point of the i-th subsystem at t j ;

[0049] The mathematical expressions of X ij and Y ij from top to bottom are as follows:

[0050] X ij = [s type , T, I, f, T w , V q , P q , T q , …]

[0051] Y ij = [y, y type ]

[0052] wherein X ij represents an abscissa of a data collection point of the i-th subsystem at t j as input data, Y ij represents an ordinate of the data collection point of the i-th subsystem at t j as output data; s type represents a fault type of the input data; y type represents a fault type of the output data; y represents whether there is a fault; T represents an environmental temperature, I represents a running current, f represents a vibration frequency, T w represents a cooling water temperature, V q represents a gas flow rate, P q represents a gas pressure, and T q represents another temperature.

[0053] Preferably, in the module M2, a training data set is constructed, and then a fault prediction model of each subsystem is trained to obtain a fine-tuning model of the current vacuum system; and the fine-tuning model of the current vacuum system is used to predict faults to obtain the fault prediction result of the vacuum system.

[0054] The samples of the training data set are organized as follows:

[0055]

[0056] Wherein, X represents an input matrix of the fault prediction model; Y represents an output matrix of the fault prediction model; and then, a (X, Y) data sample is constructed for model training, wherein n is the column number of the input matrix of the fault prediction model; and k is the row number of the input matrix of the fault prediction model.

[0057] The mathematical expression of the fault prediction model is as follows:

[0058] Y=F(X; θ base )

[0059] Wherein, F represents a supervised training model, which is any one of a Transformer, a BP neural network or a Bi-LSTM (Bi-directional Long Short-Term Memory network) model; and θ base represents parameters obtained after historical data training.

[0060] The fine-tuning model of the current vacuum system has the following mathematical expression:

[0061] Y=F(X; θ base +Δθ)

[0062] Wherein, Δθ represents the change amount of the parameters obtained after data training.

[0063] Preferably, in the module M3, the fault prediction results of each subsystem are aggregated by using a weighted average or a majority voting strategy to obtain the fault prediction result of the vacuum system.

[0064] Compared with the prior art, the present application has the following beneficial effects:

[0065] 1. The present application divides a complex vacuum system into multiple subsystems by using a divide-and-conquer strategy, quickly locates a single subsystem fault, and maintains the single subsystem before the fault occurs, thereby avoiding secondary hazards caused by the subsystem fault and the subsequent vacuum subsystem chain reaction.

[0066] 2、The system provided by the present application can realize fault prediction, maintenance before failure, avoid downtime and cause multiple continuous subsystem failures after real failure, improve the stability and reliability of the vacuum system operation, have better economy, and solve the problem of high maintenance and repair cost after original failure.

[0067] 3、The present application adopts a model pre-training fault prediction scheme, uses historical collected subsystem fault data to train the model, only needs to collect a small amount of current subsystem fault operation data to incrementally train the model, quickly replicates to other complex vacuum systems, improves the model prediction efficiency, and is applicable to vacuum systems of various complexities, and has good fault prediction expansibility and universality.

[0068] 4、The present application can be used for large and complex vacuum systems, and can predict system failure in advance and perform maintenance, thereby ensuring long-term stable operation of the vacuum system, in other words, the present application can predict system failure in advance and perform maintenance, thereby ensuring stable operation of the complex vacuum system. The method can be used for large and complex vacuum systems, can quickly locate the specific fault subsystem, has the advantages of not needing to collect a large amount of system fault data, strong self-adaptability, etc. BRIEF DESCRIPTION OF DRAWINGS

[0069] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments with reference to the attached drawings:

[0070] Figure 1 A prediction method flowchart is provided for the present application;

[0071] Figure 2 A specific flowchart of the fault prediction method is provided for the present application;

[0072] Figure 3 A system division schematic diagram is provided for the present application;

[0073] Figure 4 A subsystem fault prediction pre-training model flowchart based on the Transformer model is provided for the present application. DETAILED DESCRIPTION

[0074] The present application will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present application. These all belong to the protection scope of the present application.

[0075] The present application divides the vacuum system into multiple subsystems, predicts the failure probability and type of each subsystem at the target time in the future, predicts the aggregated failure probability and type of the vacuum system based on the failure probability and type of each subsystem, and uses the historical data of the subsystems of other vacuum systems to pre-train the basic model and uses a small amount of current vacuum system operation data to train the fine-tuning model, so that the current vacuum system can be quickly adapted to the current vacuum system operation with a small amount of data,

[0076] The present application provides a vacuum system failure prediction method, which comprises: i predicting the failure probability and type of the vacuum system S and each subsystem s i at the target time in the future.

[0077] Based on the data of each subsystem s i and the pre-processing, the subsystem s i failure prediction model training is performed to obtain the predicted failure probability and type of the subsystem s i output by the model;

[0078] For the vacuum system S, based on the failure probability and type of each subsystem s i , the predicted aggregated failure probability and type of the vacuum system S at the target time in the future are determined, and the predicted aggregated failure probability and type of the vacuum system S are described.

[0079] Based on the failure probability and type of each subsystem s i at the target time in the future and the predicted aggregated failure probability and type of the vacuum system S, the predicted failure probability and type of the vacuum system S and the subsystem s i are determined.

[0080] Specifically, according to the composition characteristics of the vacuum system S, the vacuum system S is divided into m subsystems s i , denoted as S=[s1,s2,…,s i ,…,s m ], wherein the subsystem s i represents a dry pump, a molecular pump, a cryogenic pump, an ion pump, a vacuum valve, a vacuum gauge and a thermometer. type represents the type of the subsystem s i , each subsystem s i is relatively independent and can be combined into a complete system S.

[0081] Specifically, the subsystem s iA small amount of operation input data and output data, the input data at least including ambient temperature T, operating current I, vibration frequency f, cooling water temperature T w , gas flow rate V q , gas pressure P q , gas temperature T q , the input data coming from sensors, or from the subsystem s i Operation data; the output data includes: whether there is a fault y, fault type y type ;

[0082] Based on the input data and output data, the subsystem s i is constructed at the data collection point t j , denoted as (X ij , Y ij ), wherein: X ij =[s type , T, I, f, T w , V q , P q , T q ,…], Y ij =[y, y type ]; for different s type , the y type has different fault types.

[0083] Specifically, the data preprocessing includes data cleaning and data set construction, the data cleaning at least includes missing value filling and outlier processing, based on the data collection point s i The time series data [(X i1 , Y i1 ), (X i2 , Y i2 ),…, (X ij , Y ij ),…, (X in , Y in )] in the time window t1 to t n ;

[0084] The data set construction constructs the subsystem s i Supervised model training data set as follows:

[0085] The samples of the training data set are organized as:

[0086]

[0087] Wherein, X represents the input matrix of the fault prediction model; Y represents the output matrix of the fault prediction model; and then the (X, Y) data sample is constructed for model training, wherein n is the column number of the input matrix of the fault prediction model; k is the row number of the input matrix of the fault prediction model; X i(n+k) represents the element of the nth column and the kth row of the training data set; Y i(n+k+1) represents the element of the kth row of the output matrix of the fault prediction model.

[0088] Specifically, the time window t1 to t n n is set according to the accuracy requirement of the vacuum system S, the n time windows are arranged in the order from near to far in time, and the time length is the same.

[0089] Specifically, the model is pre-trained using the historical collected subsystem fault data of other vacuum systems:

[0090] Y = F (X; θ init )

[0091] Wherein, X represents the model input, that is, the input matrix of the fault prediction model, Y represents the output matrix of the fault prediction model; θ init represents a randomly initialized parameter, and a pre-trained fault prediction model F (X; θ base ) is obtained through historical fault data training, wherein θ base represents the parameter after historical data training, so that the model has the expected prediction accuracy.

[0092] Specifically, the model F (X; θ base ) is a supervised training model, which is one of Transformer, BP neural network or Bi-LSTM, that is, a bidirectional long short-term memory network model.

[0093] The model F (X; θ base ) is based on a small amount of current subsystem acquisition data to continue training based on the parameter θ base , to obtain a fine-tuning model F (X; θ base + Δθ) of the current vacuum system.

[0094] Wherein, Δθ represents the change amount of the parameter obtained through data training, and the model F (X; θ base + Δθ) is used for fault prediction of the subsystem, and the subsystem fault prediction probability and type are given.

[0095] In some embodiments, the subsystems s i predicted fault probability and type are aggregated and predicted in a weighted average, majority voting strategy manner to determine the predicted fault probability and type of the vacuum system S.

[0096] Before the specific embodiments of the present specification are described, the application scenarios of the present specification are introduced as follows:

[0097] With the rapid development of vacuum technology in semiconductor manufacturing, biological engineering and aerospace scenes, the complexity of the vacuum system is also increasing, especially in some long-term operation scenes, it is necessary to predict the failure of the modular component in advance and handle it, so as to avoid the failure of the operation scene due to the failure of the timely and effective management. The present application is a fault prediction research of complex vacuum system, which has important social engineering application significance.

[0098] As shown in Figure 1 and Figure 2 , a vacuum system fault prediction method is disclosed, which decomposes the complex vacuum system into multiple subsystems, independently and adaptively predicts the failure probability and type of each subsystem, and then fuses the subsystem failure prediction failure probability and type to realize efficient and accurate fault prediction of the complex vacuum system.

[0099] As shown in Figure 3 , according to the structure characteristics of the highly complex vacuum system, the vacuum system S is divided into 30 subsystems s i , represented as S=[s1,s2,…,s i ,…,s 29 , s 30 ], wherein: s1, s 10 ~s 15 and s 28 represent vacuum gauges, s2 represents a thermometer, s3, s 16 and s 27 represent gas valves, s4~s9 and s 23 ~s 26 represent vacuum valves, s 17 and s 18 represent molecular pumps, s 19 and s 20 represent cryogenic pumps, s 21 represents an ion pump, s 22 and s 29 represent dry pumps, and s 30 represents a vacuum container. Each subsystem s i is relatively independent and can be combined into a complete system S.

[0100] The historical subsystem s i and the current subsystem s i collect the running state data of the subsystem s i in the stable running, early failure, mid-failure and post-failure stages. The subsystem s iFor example, a dry pump, collect a small amount of running input data and output data of the dry pump. The input data at least includes the ambient temperature T, the dry pump current I, the dry pump vibration frequency f, the cooling water temperature T w , the gas flow rate V q in the dry pump, the gas pressure P q , the gas temperature T q , etc. The data can come from sensors such as temperature sensors, pressure sensors, vibration sensors, and also from actual running data of the dry pump; the output data includes whether it is faulty y, the fault type y type . Based on the collected input and output data, a dry pump subsystem s i is constructed j at time t ij data collection point is represented as (X ij , Y ij ), wherein: X type =[s w ,T,I,f,T q ,V q ,P q ,T ij ,…], Y type =[y,y type ]. For s type is a dry pump, the common fault types y i include bearing failure, gear failure, pipe blockage, seal loosening and pump cavity gas leakage, etc.

[0101] The subsystem s i data preprocessing includes but is not limited to missing value filling, outlier processing and time series feature addition, etc., to ensure the data quality and availability of the model. The missing value filling can adopt average value filling, mode value filling, etc.; the outlier processing can set the outliers as null, filling method and interpolation method, etc.; the time series feature addition can add the statistical average value, maximum value, minimum value and kurtosis of the sliding window, etc. Based on the data collection point, the dry pump subsystem s n is obtained i1 its time window t1 to t i1 acquisition time series data [(X i2 , Y i2 ), (X ij , Y ij ), …, (X in , Y in )], wherein n in the sliding time window t1 to t n can be set according to the accuracy requirement of the complex vacuum system. The dry pump subsystem s i constructed based on the collected time series data has a supervised training model data set under the sliding time window as follows:

[0102]

[0103] As Figure 4 shown, the history collection other vacuum system dry pump subsystem fault data pre-training model Y=F(X; theta init ), wherein X represents model input, theta init Randomly initialized parameters, F can be selected as BP neural network, Bi-LSTM, i.e. bidirectional long short term memory network and Transformer and other supervised training models, can obtain pre-training fault prediction model F(X; theta base ), wherein theta base Indicates that the model parameters are trained after the history data, so that the model has a certain prediction accuracy, and a small amount of fault data can be quickly applied to the current vacuum system.

[0104] Subsystem s i Fault prediction, based on a small amount of current subsystem s i Collect fault operation data, continue training on the model F(X; theta base ), obtain current subsystem s i Fine-tuning model, i.e. F(X; theta base +Delta theta);

[0105] Wherein, Delta theta represents the parameter change amount obtained by training the collected current dry pump subsystem s i Data, based on the prediction model F(X; theta base +Delta theta) to predict the fault of the dry pump subsystem, and give the predicted fault probability and type of the dry pump subsystem.

[0106] Vacuum system S fault prediction, multiple subsystems s i Predict the fault probability and type, and aggregate the prediction results of the subsystems s i Through weighted average, majority voting and other strategy methods, realize the prediction of the fault probability and type of the complex vacuum system S, and ensure the accuracy and integrity of the fault prediction.

[0107] The above only describes the preferred embodiments of the present application and does not limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0108] The present application also provides a vacuum system fault prediction system, which can be realized by executing the flow steps of the vacuum system fault prediction method, i.e. the vacuum system fault prediction method can be understood as the preferred embodiment of the vacuum system fault prediction system by those skilled in the art.

[0109] According to the vacuum system fault prediction system provided by the present application,

[0110] Module M1: dividing the vacuum system into multiple subsystems;

[0111] Module M2: collecting and preprocessing the data of each subsystem, training the corresponding fault prediction model of each subsystem, and obtaining the output data of the fault prediction model;

[0112] Module M3: fusing the output data of all the fault prediction models to obtain the fault prediction result of the vacuum system.

[0113] Specifically, in the module M1, the vacuum system is divided into m subsystems, and the mathematical expression is:

[0114] S = [s1, s2, …, sm] i m

[0115] wherein, S represents the vacuum system, and s i represents the i-th subsystem; i≤m, m is a constant;

[0116] The subsystems include: dry pumps, molecular pumps, cryogenic pumps, ion pumps, vacuum valves, vacuum gauges, and thermometers.

[0117] Specifically, in the module M2, the data of each subsystem is preprocessed to obtain the time series data of the subsystem, and then the corresponding fault prediction model of each subsystem is trained through the time series data;

[0118] The time series data of the subsystem within the time window t1 to t n is represented as:

[0119] [(X i1 , Y i1 ), (X i2 , Y i2 ), …, (X ij , Y ij ), …, (X in , Y in )]

[0120] wherein, n time windows are arranged in the order of time from near to far, and the time length is the same; n is a constant, and (X ij , Y ij ) represents the data collection point of the i-th subsystem at time t j ;

[0121] The mathematical expressions of X ij and Y ij are as follows from top to bottom:

[0122] X ij ​​= [s type T, I, f, T w V q P q T q , …]

[0123] Y ij = [y, y type ]

[0124] wherein, X ij represents the abscissa of the data acquisition point of the i-th subsystem at t j time as input data, Y ij represents the ordinate of the data acquisition point of the i-th subsystem at t j time as output data; s type represents the fault type of the input data; y type represents the fault type of the output data; y represents whether it is a fault; T represents the ambient temperature, I represents the running current, f represents the vibration frequency, T w represents the cooling water temperature, V q represents the gas flow rate, P q represents the gas pressure, T q represents other temperatures.

[0125] Specifically, in the module M2, a training data set is constructed, and then a fault prediction model corresponding to each subsystem is trained to obtain a fine-tuning model of the current vacuum system; the fine-tuning model of the current vacuum system is used to predict faults to obtain a fault prediction result of the vacuum system;

[0126] The samples of the training data set are organized as follows:

[0127]

[0128] wherein, X represents the input matrix of the fault prediction model; Y represents the output matrix of the fault prediction model; then (X, Y) data samples are constructed for model training; wherein, n is the column number of the input matrix of the fault prediction model; k is the row number of the input matrix of the fault prediction model;

[0129] The mathematical expression of the fault prediction model is:

[0130] Y = F(X; θ base )

[0131] wherein, F represents a supervised training model; θ base represents the parameters after historical data training;

[0132] The fine-tuning model of the current vacuum system has the mathematical expression:

[0133] Y = F(X; θbase + Δθ

[0134] wherein, Δθ represents the variation of the parameter obtained through data training.

[0135] Specifically, in the module M3, the failure prediction results of each of the subsystems are aggregated by using a weighted average or a majority voting strategy to obtain the failure prediction result of the vacuum system.

[0136] Those skilled in the art know that, in addition to implementing the system provided by the present application and each device, module, unit thereof in the form of pure computer readable program code, the system provided by the present application and each device, module, unit thereof can also be implemented in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded microcontrollers, etc. to achieve the same function by logically programming the method steps. Therefore, the system provided by the present application and each device, module, unit thereof can be considered as a hardware component, and the devices, modules, units included therein for implementing various functions can also be considered as structures within the hardware component; the devices, modules, units for implementing various functions can also be considered as both software modules implementing methods and structures within hardware components.

[0137] The specific embodiments of the present application are described above. It needs to be understood that the present application is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essential content of the present application. The embodiments of the present application and the features in the embodiments can be arbitrarily combined with each other without conflict.

Claims

1. A vacuum system failure prediction method, characterized by, The method comprises the steps of: Step S1: dividing the vacuum system into multiple subsystems; Step S2: collecting and preprocessing the data of each subsystem, training the corresponding fault prediction model of each subsystem, and obtaining the output data of the fault prediction model; Step S3: fusing the output data of all the fault prediction models to obtain the fault prediction result of the vacuum system.

2. The vacuum system fault prediction method of claim 1, wherein, In the step S1, the vacuum system is divided into m subsystems, and the mathematical expression is: S = [s1, s2,..., s i ,…,s m ] where S denotes the vacuum system, s i denotes the i-th subsystem; i≤m, m being a constant; The subsystems include dry pumps, molecular pumps, cryogenic pumps, ion pumps, vacuum valves, vacuum gauges, and thermometers. In the step S2, the data of the subsystem, including: ambient temperature T, operating current I, vibration frequency f, cooling water temperature T w , gas flow rate V q , gas pressure P q and gas temperature T q .

3. The vacuum system fault prediction method of claim 1, wherein, In the step S2, the data of each subsystem is preprocessed to obtain the time series data of the subsystem, and then the time series data is used to train the corresponding fault prediction model of each subsystem. The subsystem is in timing data within a time window t1 to t n denoted as: [(X i1 , Y i1 ), (X i2 , Y i2 ),..., (X ij , Y ij ),..., (X in , Y in )] Wherein, n time windows are arranged in order from near to far in time, and the time length is the same; n is a constant, (X ij , Y ij ) represents the data acquisition point of the i th subsystem at t j time X ij With Y ij The mathematical expressions of Y from top to bottom are: X ij = [s type , T, I, f, T w , V q , P q , T q ,... ] Y ij = [y, y type ] wherein X ij represents the abscissa of the data acquisition point at time t j for the i-th subsystem as input data, Y ij represents the ordinate of the data acquisition point at time t j for the i-th subsystem as output data; s type represents the fault type of the input data; y type represents the fault type of the output data; y represents whether there is a fault or not; T represents the ambient temperature, I represents the operating current, f represents the vibration frequency, T w represents the cooling water temperature, V q represents the gas flow rate, P q represents the gas pressure, T q represents other temperatures.

4. The vacuum system fault prediction method of claim 3, wherein, In the step S2, a training data set is constructed, and then the corresponding fault prediction model of each subsystem is trained to obtain a fine-tuning model of the current vacuum system; the fine-tuning model of the current vacuum system is used to predict faults to obtain the fault prediction result of the vacuum system. The samples of the training data set are organized as follows: Wherein, X represents the input matrix of the fault prediction model; Y represents the output matrix of the fault prediction model; then (X, Y) data samples are constructed for model training; wherein n is the column number of the input matrix of the fault prediction model; k is the row number of the input matrix of the fault prediction model. The mathematical expression of the fault prediction model is: Y = F(X; θ base ) wherein F denotes a supervised training model; θ base denotes the parameters after training on historical data; The mathematical expression of the fine-tuning model of the current vacuum system is: Y = F(X; θ base + Δθ) Wherein, Δθ represents the change of the parameter obtained by data training.

5. The vacuum system fault prediction method of claim 4, wherein, In the step S3, the weighted average or majority voting strategy is used to aggregate the fault prediction results of each subsystem to obtain the fault prediction result of the vacuum system.

6. A vacuum system failure prediction system characterized by, The method comprises the steps of: Module M1: dividing the vacuum system into multiple subsystems; Module M2: collecting and preprocessing the data of each subsystem, training the corresponding fault prediction model of each subsystem, and obtaining the output data of the fault prediction model; Module M3: fusing the output data of all the fault prediction models to obtain the fault prediction result of the vacuum system.

7. The vacuum system fault prediction system of claim 6, wherein, In the module M1, the vacuum system is divided into m subsystems, and the mathematical expression is: S = [s1, s2,..., s i ,…,s m ] where S denotes the vacuum system, s i denotes the i-th subsystem; i≤m, m being a constant; The subsystems include dry pumps, molecular pumps, cryogenic pumps, ion pumps, vacuum valves, vacuum gauges, and thermometers.

8. The vacuum system fault prediction system of claim 6, wherein, In the module M2, the data of each subsystem is preprocessed to obtain the time series data of the subsystem, and then the time series data is used to train the corresponding fault prediction model of each subsystem. The subsystem is in timing data within a time window t1 to t n is represented as: [(X i1 , Y i1 ), (X i2 , Y i2 ),..., (X ij , Y ij ),..., (X in , Y in )] Wherein, n time windows are arranged in order from near to far in time, and the time length is the same; n is a constant, (X ij , Y ij ) represents the data acquisition point of the i th subsystem at t j time. X ij With Y ij The mathematical expressions of Y from top to bottom are: X ij = [s type , T, I, f, T w , V q , P q , T q ,...] Y ij = [y, y type ] wherein X ij represents the abscissa of the data acquisition point at time t j for the i-th subsystem as input data, Y ij represents the ordinate of the data acquisition point at time t j for the i-th subsystem as output data; s type represents the fault type of the input data; y type represents the fault type of the output data; y represents whether there is a fault or not; T represents the ambient temperature, I represents the operating current, f represents the vibration frequency, T w represents the cooling water temperature, V q represents the gas flow rate, P q represents the gas pressure, T q represents other temperatures.

9. The vacuum system fault prediction system of claim 8, wherein, In the module M2, a training data set is constructed, and then the corresponding fault prediction model of each subsystem is trained to obtain a fine-tuning model of the current vacuum system; the fine-tuning model of the current vacuum system is used to predict faults to obtain the fault prediction result of the vacuum system. The samples of the training data set are organized as follows: Wherein, X represents the input matrix of the fault prediction model; Y represents the output matrix of the fault prediction model; then (X, Y) data samples are constructed for model training; wherein n is the column number of the input matrix of the fault prediction model; k is the row number of the input matrix of the fault prediction model. A mathematical expression of the failure prediction model is: Y = F(X; θ base ) wherein F denotes a supervised training model; θ base denotes the parameters after training on historical data; A mathematical expression of the fine-tuning model of the current vacuum system is: Y = F(X; θ base + Δθ) Wherein, Δθ represents a change amount of the parameter obtained through data training.

10. The vacuum system fault prediction system of claim 9, wherein, In the module M3, a weighted average or a majority voting strategy is adopted to aggregate the failure prediction results of each of the subsystems to obtain the failure prediction result of the vacuum system.

Images