[go: up one dir, main page]

CN111367253B - Chemical system multi-working-condition fault detection method based on local adaptive standardization - Google Patents

Chemical system multi-working-condition fault detection method based on local adaptive standardization Download PDF

Info

Publication number
CN111367253B
CN111367253B CN202010098141.7A CN202010098141A CN111367253B CN 111367253 B CN111367253 B CN 111367253B CN 202010098141 A CN202010098141 A CN 202010098141A CN 111367253 B CN111367253 B CN 111367253B
Authority
CN
China
Prior art keywords
data
train
local
valid
moving window
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010098141.7A
Other languages
Chinese (zh)
Other versions
CN111367253A (en
Inventor
赵劲松
吴昊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tsinghua University
Original Assignee
Tsinghua University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tsinghua University filed Critical Tsinghua University
Priority to CN202010098141.7A priority Critical patent/CN111367253B/en
Publication of CN111367253A publication Critical patent/CN111367253A/en
Application granted granted Critical
Publication of CN111367253B publication Critical patent/CN111367253B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B23/00—Testing or monitoring of control systems or parts thereof
    • G05B23/02—Electric testing or monitoring
    • G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
    • G05B23/0259—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterized by the response to fault detection
    • G05B23/0262—Confirmation of fault detection, e.g. extra checks to confirm that a failure has indeed occurred
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00—Program-control systems
    • G05B2219/20—Pc systems
    • G05B2219/24—Pc safety
    • G05B2219/24065—Real time diagnostics

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Automation & Control Theory (AREA)
  • Testing Or Calibration Of Command Recording Devices (AREA)

Abstract

本发明涉及一种基于局部自适应标准化的化工系统多工况故障检测方法，属于化工过程监控、工业数据处理和过程系统工程技术领域。本方法提出了局部自适应标准化方法，并应用了深度神经网络的变分自动编码器技术，通过计算局部移动窗口内数据的平均值，作为局部自适应标准化的平均值参数，针对不同的数据使用不同的平均值，具有自适应能力。本方法利用局部自适应标准化处理，通过检测局部移动窗口内数据是否发生偏离趋势来进行故障检测。本方法能够适用于任意工况，而且具有更高的准确率和更强的泛化能力，能够满足实时检测的需求，通过早期预警故障避免化工事故发生或者减小事故带来的危害。

Figure 202010098141

The invention relates to a multi-working-condition fault detection method for a chemical system based on local adaptive standardization, and belongs to the technical fields of chemical process monitoring, industrial data processing and process system engineering. This method proposes a local adaptive normalization method, and applies the variational auto-encoder technology of deep neural network. By calculating the average value of the data in the local moving window, it is used as the average value parameter of local adaptive normalization, which is used for different data. Different averages, with adaptive capability. The method utilizes local adaptive normalization to detect faults by detecting whether the data in the local moving window deviates from the trend. The method can be applied to any working condition, has higher accuracy and stronger generalization ability, can meet the needs of real-time detection, and can avoid chemical accidents or reduce the harm caused by accidents through early warning faults.

Figure 202010098141

Description

Chemical system multi-working-condition fault detection method based on local adaptive standardization
Technical Field
The invention relates to a chemical system multi-working-condition fault detection method based on local adaptive standardization, and belongs to the technical field of chemical process monitoring, industrial data processing and process system engineering.
Background
The safe production in the petrochemical industry relates to each link in the life cycle of chemicals, and because the production link has large chemical quantity and centralized personnel distribution, serious property loss, casualties and environmental damage are caused once an accident occurs. With the continuous progress, popularization and implementation of informatization technology, the chemical industry enters a big data era. The fault detection technology is a basic key technology in the field of chemical process safety, and aims to distinguish whether a chemical system is in a normal operation state or has a fault by collecting and analyzing real-time data of an industrial process.
With the continuous improvement of the automation degree of chemical plants, most of the chemical plants are provided with advanced process control systems and industrial large data storage platforms, so that in recent years, a data-driven chemical fault detection method becomes a research hotspot in academia and industry. Data-driven failure detection mainly includes two types of methods. The first category of methods is multivariate statistical process monitoring methods, including Principal Component Analysis (PCA) and Partial Least Squares (PLS). Because the chemical process has the characteristics of multivariable, dynamic property, nonlinearity and the like, researchers provide a dynamic method and a kernel method based on PCA and PLS in order to be more applied to the chemical process. The second category of methods is deep neural network based methods, including deep belief networks, convolutional neural networks, and Variational auto-encoders (VAEs). The VAE method can train and obtain a monitoring model for chemical engineering fault detection by using normal operation data only. Compared with a multivariate statistical process monitoring method, the deep neural network has higher accuracy, recall rate and higher generalization capability. In recent years, with the development of hardware computing capabilities such as CPUs, GPUs and the like, the computing speed of the method can meet the real-time requirement of industrial data monitoring, and the method has great advantages in practical application. However, with the influence of factors such as raw materials, markets, environment and the like, the chemical device needs to continuously adjust the operation conditions in the production link, namely, the multi-working-condition characteristic exists. The existing chemical fault detection method based on the deep neural network has the problems that normal process variable data are generally assumed to be subjected to normal distribution or unimodal distribution, and the data need to be standardized before being input into a model, so that the model can only be suitable for a single working condition. In the face of the characteristic of multiple working conditions of chemical industry, the existing deep neural network method cannot effectively deal with the problem and cannot complete the fault detection task of the chemical process.
For the multi-working-condition characteristics of chemical industry, the current research generally utilizes a local neighbor standardization combined multivariate statistical process monitoring method, the method uses local neighbor standardization to preprocess data, and then uses the standardized data to model a PCA (principal component analysis) or PLS (partial least squares) method. The original standardization method is to estimate the distribution of variables by using the average value and standard deviation of historical normal operation data, and when online data standardization is carried out, the fixed historical average value and standard deviation are used for calculation, but the differences between the average value and the standard deviation under different working conditions are huge, so that the method can only be used for a single working condition. Local neighbor normalization is performed by finding a local neighbor set of current data in historical normal operating data, and calculating with the mean and standard deviation of the local neighbor set. The local neighbor standardization can find that the current data belongs to a certain working condition, and then the data of the working condition is utilized to carry out standardization processing, so that the data of a plurality of working conditions can be mapped to approximate unimodal distribution, and further the fault detection of historical working conditions can be completed. The method has the problems that the local neighbor standardization still uses historical data to calculate the average value and the standard deviation, is highly dependent on historical working conditions and can only be applied to the historical working conditions. Once the chemical process runs under a new working condition, the fault detection task cannot be finished if neighbor data does not exist in the historical data. Up to now, no universal fault detection method capable of monitoring all working conditions of a chemical process has appeared.
Disclosure of Invention
The invention aims to provide a chemical system multi-working-condition fault detection method based on local adaptive standardization, which is used for overcoming the defects of the existing method, applies a variational automatic encoder technology of a deep neural network, processes data of a local moving window by utilizing the local adaptive standardization, inputs the window data into a variational automatic encoder model to detect whether the deviation trend exists or not, and judges whether the process data is in a normal running state or has a fault or not, so that early warning is carried out when the early data deviates, and the possibility of occurrence of chemical accidents is reduced to the maximum extent.
The invention provides a chemical system multi-working-condition fault detection method based on local adaptive standardization, which comprises the following steps of:
(1) obtaining normal operation data set D under N working conditions from historical database of chemical systemhistoryData set DhistoryThe method comprises the following steps of (1) totally m rows and n columns of data, wherein m represents a process variable of a chemical system, and n represents total operation time;
(2) setting the normal operation data set D in the step (1)historyInto training sets DtrainAnd a verification set DvalidTraining set DtrainComprising m rows ntrainColumn data, validation set DvalidComprising m rows nvalidColumn data, in which training set DtrainD of historical normal operation data sethistoryIn a ratio of
Figure BDA0002385908640000021
60％≤a≤90％；
(3) Training set D in step (2)trainAnd a verification set DvalidCarrying out local self-adaptive standardization processing to obtain a transformed training set TtrainAnd a verification set TvalidThe method comprises the following specific steps:
(3-1) Using the training set D of step (2)trainThe global mean standard deviation gmstd (D) of m process variables in the chemical system is calculated by using the following formulatrain) Including m numbers:
Figure BDA0002385908640000031
wherein i represents the working condition serial number of the chemical process, i is more than or equal to 1 and less than or equal to N, and Dtrain，iRepresentative training set DtrainNormal operating data of the ith operating mode, std (D)train，i) Representative training set DtrainThe standard deviation vector of the ith working condition comprises m numerical values, and std (D) is obtained by calculating the standard deviation of the corresponding variabletrain，i)，ntrain，iRepresentative training set DtrainThe amount of normal operation data for the ith condition,
(3-2) training set D for step (2)trainThe k-th normal operation data xk, k in (b) represents the training set DtrainRun time number of (1), 2, ntrain，xkComprises m variable values with time sequence number k, and local moving window data w with time window t is selected forward by calculating timek，wkThe total m rows and t columns of data are provided, wherein t is a time window, t is more than or equal to 10 and less than or equal to 100:
Figure BDA0002385908640000032
utilizing local moving window datawkCalculating wkAverage value of m variables in (1) to obtain mean (w)k)，mean(wk) Comprises m numbers;
(3-3) Using gmstd (D) in step (3-1)train) And mean (w) in step (3-2)k) For the local moving window data w of step (3-2)kPerforming local adaptive normalization to wkApproximately converting m variables of the data into standard normal distribution to obtain local self-adaptive standardized local moving window data
Figure BDA0002385908640000033
Figure BDA0002385908640000034
(3-4) repeating the step (3-2) and the step (3-3), and sequentially calculating D in the training settrainObtaining a local adaptive standardized training set T by each normal operation datatrain；
(3-5) authentication set D for step (2)validP-th normal operation data x in (1)pAnd p represents the verification set DvalidOperating time number of (1), 2, nvalid，xpThe local moving window data w with time window t is selected forward in time and comprises m variable values with time sequence number pp，wpThere are m rows and t columns of data, where t is the time window in step (3-2):
Figure BDA0002385908640000035
using local moving window data wpCalculating wpAverage value of m variables in (1) to obtain mean (w)p)，mean(wp) Comprises m numbers;
(3-6) Using gmstd (D) in step (3-1)train) And mean (w) in step (3-5)p) For the local moving window data w of step (3-5)pThe local self-adaptive standardization is carried out,let wpApproximately converting m variables of the data into standard normal distribution to obtain local self-adaptive standardized local moving window data
Figure BDA0002385908640000041
Figure BDA0002385908640000042
(3-7) repeating the steps (3-5) and (3-6), and sequentially calculating D in the verification setvalidObtaining a local self-adaptive standardized verification set T from each normal operation datavalid；
(4) Constructing a variational automatic encoder which comprises an encoder part and a decoder part and utilizing the training set T obtained in the step (3-4)trainTraining the variational automatic encoder to obtain the trained variational automatic encoder, and specifically comprising the following steps:
(4-1) designing and constructing an encoder by using a convolutional neural network, a cyclic neural network or a deep belief network, and carrying out local adaptive normalization on the local moving window data obtained in the step (3-3)
Figure BDA0002385908640000043
As input to the encoder, the mapping is derived
Figure BDA0002385908640000044
Feature vector σ ofkAnd mukFeature vector σkAnd mukThere are l numbers, l represents the dimension of the feature vector, l is greater than or equal to m and less than or equal to 4 m:
Figure BDA0002385908640000045
Figure BDA0002385908640000046
(4-2) Using the feature vector of step (4-1)σkAnd mukCarrying out reparameterization to obtain
Figure BDA0002385908640000047
Hidden feature vector h ofk，hkIncludes l values:
hk＝μk+σk⊙∈
where e is normally distributed from the standard
Figure BDA0002385908640000048
A random sample results,. indicates multiplication of corresponding elements of the vector;
(4-3) designing and constructing a decoder by using a convolutional neural network, a cyclic neural network or a deep belief network, and enabling the hidden feature vector h in the step (4-2)kAs input to the decoder, reconstructing to obtain the data corresponding to step (3-3)
Figure BDA0002385908640000049
Reconstructed data with the same dimensionality
Figure BDA00023859086400000410
There are m rows and t columns of data:
Figure BDA00023859086400000411
(4-4) utilizing the eigenvector σ of step (4-1) according to the following loss functionkAnd mukAnd the reconstructed data of step (4-3)
Figure BDA00023859086400000412
Calculating the locally adaptive normalized local moving window data of step (3-3)
Figure BDA00023859086400000413
Error of (2)
Figure BDA00023859086400000414
Figure BDA00023859086400000415
Figure BDA00023859086400000416
I.e. the loss function of the variational automatic encoder, the loss function
Figure BDA00023859086400000417
Including reconstruction losses
Figure BDA0002385908640000051
And KL divergence loss
Figure BDA0002385908640000052
λ is the weighting factor of KL divergence loss versus reconstruction loss, 103≤λ≤106The loss of the two parts is calculated as follows, wherein j represents the variable serial number of the chemical process, and j is more than or equal to 1 and less than or equal to m:
Figure BDA0002385908640000053
Figure BDA0002385908640000054
(4-5) repeating the step (4-1) to the step (4-4), and sequentially combining the training sets T in the step (3-4)trainEach data of
Figure BDA0002385908640000055
Inputting the variational automatic encoder to carry out error calculation, and training the variational automatic encoder through an error back propagation algorithm to obtain a trained variational automatic encoder;
(5) utilizing the trained variational automatic encoder obtained in the step (4) and the verification set T obtained in the step (3-7)validBy estimating the verification set TvalidTo obtain a variational automatic encoder for fault detection taskThe specific steps of the time monitoring threshold eta are as follows:
(5-1) local moving window data for local adaptive normalization of step (3-6)
Figure BDA0002385908640000056
Mapping the input of the variational automatic encoder trained in the step (4) to obtain
Figure BDA0002385908640000057
Feature vector σ ofpAnd mupFeature vector σpAnd mupThere are l values, respectively, where l represents the dimension of the feature vector:
Figure BDA0002385908640000058
Figure BDA0002385908640000059
(5-2) Using the feature vector σ of step (5-1)pAnd mupTo, for
Figure BDA00023859086400000510
Carrying out reparameterization to obtain
Figure BDA00023859086400000511
Hidden feature vector h ofp，hpIncludes l values:
hp＝μp+σp⊙∈
where e is normally distributed from the standard
Figure BDA00023859086400000512
A random sample results,. indicates multiplication of corresponding elements of the vector;
(5-3) combining the hidden feature vector h of the step (5-2)pAs the input of the decoder in the variational automatic encoder trained in the step (4), reconstructing to obtain the result of the step (3-6)
Figure BDA00023859086400000513
Reconstructed data with the same dimensionality
Figure BDA00023859086400000514
Figure BDA00023859086400000515
There are m rows and t columns of data:
Figure BDA00023859086400000516
(5-4) utilizing the feature vector σ of step (5-1) according to the following abnormality score calculation formulapAnd mupAnd the reconstructed data of step (5-3)
Figure BDA00023859086400000517
Calculating the local adaptive standardized local moving window number in the step (3-6)
Figure BDA00023859086400000518
Is abnormal score of
Figure BDA00023859086400000519
Figure BDA0002385908640000061
Abnormal score
Figure BDA0002385908640000062
Including reconstruction losses
Figure BDA0002385908640000063
And KL divergence loss
Figure BDA0002385908640000064
λ is a weighting coefficient of KL divergence loss with respect to reconstruction loss, the same as λ of step (4-4); two-part loss calculationWherein j represents the variable serial number of the chemical process, j is more than or equal to 1 and less than or equal to m:
Figure BDA0002385908640000065
Figure BDA0002385908640000066
(5-5) repeating the steps (5-1) to (5-4), and sequentially adding the verification sets T of the steps (3-7)validEach data of
Figure BDA0002385908640000067
Input variable automatic encoder for calculating abnormal score
Figure BDA0002385908640000068
Get the verification set TvalidIs abnormal score data set Svalid；
(5-6) abnormal score data set SvalidObtaining abnormal score data set S according to normal distributionvalidThe abnormal fraction with the normal distribution confidence coefficient alpha is used as a monitoring threshold eta of the chemical system, and alpha is more than or equal to 99% and less than or equal to 99.99%;
(6) and (3) carrying out online fault detection on the process data of the chemical system under different working conditions by using the variational automatic encoder trained in the step (4) and the monitoring threshold eta obtained in the step (5), wherein the method comprises the following steps:
(6-1) collecting process data from a real-time database of the chemical system at the current detection moment q, and selecting local moving window data w with a time window t from the time to the frontq，wqThere are m rows and t columns of data, where t is the time window in step (3-2):
Figure BDA0002385908640000069
using local moving window data wqCalculating wqAverage of m variables inValue, to obtain mean (w)q)，mean(wq) Comprises m numbers;
(6-2) Using gmstd (Dtrain) in step (3-1) and mean (w) in step (6-1)q) For the local moving window data w of step (6-1)qPerforming local adaptive normalization to wqApproximately converting m variables of the data into standard normal distribution to obtain local self-adaptive standardized local moving window data
Figure BDA00023859086400000610
Figure BDA00023859086400000611
(6-3) local moving window data for local adaptive normalization of step (6-2)
Figure BDA00023859086400000612
Mapping the input of the encoder in the variational automatic encoder trained in the step (4) to obtain
Figure BDA0002385908640000071
Feature vector σ ofqAnd muqThe two feature vectors have values of l, wherein l represents the dimension of the feature vector and has the same size as l in the step (4-1):
Figure BDA0002385908640000072
Figure BDA0002385908640000073
(6-4) Using the feature vector σ of step (6-3)qAnd muqCarrying out reparameterization to obtain
Figure BDA0002385908640000074
Hidden feature vector h ofq，hqComprisingIndividual values:
hq＝μq+σq⊙∈
where e is normally distributed from the standard
Figure BDA0002385908640000075
A random sample results,. indicates multiplication of corresponding elements of the vector;
(6-5) hiding the feature vector h of the step (6-4)qAs the input of the decoder in the variational automatic encoder trained in the step (4), reconstructing to obtain the result corresponding to the step (6-2)
Figure BDA0002385908640000076
Reconstructed data with the same dimensionality
Figure BDA0002385908640000077
There are m rows and t columns of data:
Figure BDA0002385908640000078
(6-6) utilizing the feature vector σ of step (6-3) according to the following abnormality score calculation formulaqAnd muqAnd the reconstructed data of step (6-5)
Figure BDA0002385908640000079
Calculating the locally adaptive normalized local moving window data of step (6-2)
Figure BDA00023859086400000710
Is abnormal score of
Figure BDA00023859086400000711
Figure BDA00023859086400000712
Abnormal score
Figure BDA00023859086400000713
Including reconstruction losses
Figure BDA00023859086400000714
And KL divergence loss
Figure BDA00023859086400000715
λ is a weighting coefficient of KL divergence loss with respect to reconstruction loss, the same as λ of step (4-4); the loss of the two parts is calculated as follows, wherein j represents the variable serial number of the chemical process, j is more than or equal to 1 and less than or equal to m:
Figure BDA00023859086400000716
Figure BDA00023859086400000717
(6-7) scoring the abnormality of step (6-6)
Figure BDA00023859086400000718
Comparing with the monitoring threshold eta obtained in the step (5), if so
Figure BDA00023859086400000719
The current chemical system is in a normal operation state, the step (6-1) is returned to continue monitoring the online real-time data, and if the online real-time data is not monitored, the chemical system is in a normal operation state
Figure BDA00023859086400000720
The system fault of the current chemical system is indicated, and fault warning is sent out, so that the multi-working-condition fault detection of the chemical system based on local adaptive standardization is realized.
The invention provides a chemical system multi-working-condition fault detection method based on local adaptive standardization, which has the advantages that:
the invention discloses a chemical system multi-working-condition fault detection method based on local adaptive standardization, which is different from the existing fault detection method by providing a local adaptive standardization method and applying a variational automatic encoder technology of a deep neural network. The method is different from other existing detection methods which detect the deviation degree of the current data and the normal operation data, and the fault detection is carried out by detecting whether the data in the local moving window deviates or not by utilizing local self-adaptive standardization processing. Therefore, the method can be suitable for any working condition, and not only can be applied to the historical existing working condition, but also can be applied to the historical non-occurring working condition. In addition, the invention combines and applies the variational automatic encoder to detect whether the current window data has the deviation trend, and has higher accuracy and stronger generalization capability compared with the traditional multivariate statistical method. The invention can meet the requirement of real-time detection, can be applied to the fault detection task of the chemical system under all working conditions of the chemical process, and avoids the occurrence of chemical accidents or reduces the harm brought by the accidents by early warning of the faults.
Drawings
FIG. 1 is a block diagram of the overall process of the method of the present invention.
Fig. 2 is a schematic diagram of a variational auto-encoder configuration in an embodiment of the present invention.
Fig. 3 is a schematic diagram of a fault detection result under different working conditions according to an embodiment of the present invention.
Detailed Description
The invention provides a chemical system multi-working-condition fault detection method based on local adaptive standardization, which has an overall flow diagram shown in figure 1 and comprises the following steps:
(1) obtaining normal operation data set D under N working conditions from historical database of chemical systemhistoryData set DhistoryThere are m rows and n columns of data, where m represents process variables of the chemical system, such as temperature, time, pressure, etc., and n represents total run time;
(2) setting the normal operation data set D in the step (1)historyInto training sets DtrainAnd a verification set DvalidTraining set DtrainComprising m rows ntrainColumn data, validation set DvalidComprising m rows nvalidColumn data, in which training set DtrainD of historical normal operation data sethistoryIn a ratio of
Figure BDA0002385908640000081
60％≤a≤90％；
(3) Training set D in step (2)trainAnd a verification set DvalidCarrying out local self-adaptive standardization processing to obtain a transformed training set TtrainAnd a verification set TvalidThe method comprises the following specific steps:
(3-1) Using the training set D of step (2)trainThe global mean standard deviation gmstd (D) of m process variables in the chemical system is calculated by using the following formulatrain) Including m numbers:
Figure BDA0002385908640000091
wherein i represents the working condition serial number of the chemical process, i is more than or equal to 1 and less than or equal to N, and Dtrain，iRepresentative training set DtrainNormal operating data of the ith operating mode, std (D)train，i) Representative training set DtrainThe standard deviation vector of the ith working condition comprises m numerical values, and std (D) is obtained by calculating the standard deviation of the corresponding variabletrain，i)，ntrain，iRepresentative training set DtrainThe amount of normal operation data for the ith condition,
(3-2) training set D for step (2)trainThe k-th normal operation data xk, k in (b) represents the training set DtrainRun time number of (1), 2, nttain，xkComprises m variable values with time sequence number k, and local moving window data w with time window t is selected forward by calculating timek，wkThe total m rows and t columns of data are provided, wherein t is a time window, t is more than or equal to 10 and less than or equal to 100:
Figure BDA0002385908640000092
using local moving window data wkCalculating wkAverage value of m variables in (1) to obtain mean (w)k)，mean(wk) Comprises m numbers;
(3-3) Using gmstd (D) in step (3-1)train) And mean (w) in step (3-2)k) For the local moving window data w of step (3-2)kPerforming local adaptive normalization to wkApproximately converting m variables of the data into standard normal distribution to obtain local self-adaptive standardized local moving window data
Figure BDA0002385908640000093
Figure BDA0002385908640000094
(3-4) repeating the step (3-2) and the step (3-3), and sequentially calculating D in the training settrainObtaining a local adaptive standardized training set T by each normal operation datatrain；
(3-5) authentication set D for step (2)validP-th normal operation data x in (1)pAnd p represents the verification set DvalidOperating time number of (1), 2, nvalid，xpThe local moving window data w with time window t is selected forward in time and comprises m variable values with time sequence number pp，wpThere are m rows and t columns of data, where t is the time window in step (3-2):
Figure BDA0002385908640000101
using local moving window data wpCalculating wpAverage value of m variables in (1) to obtain mean (w)p)，mean(wp) Comprises m numbers;
(3-6) Using gmstd (D) in step (3-1)train) And mean (w) in step (3-5)p) For the local moving window data w of step (3-5)pPerforming local adaptive normalization to wpApproximately converting m variables of the data into standard normal distribution to obtain local self-adaptive standardized local moving window data
Figure BDA0002385908640000102
Figure BDA0002385908640000103
(3-7) repeating the steps (3-5) and (3-6), and sequentially calculating D in the verification setvalidObtaining a local self-adaptive standardized verification set T from each normal operation datavalid；
(4) Constructing a variational automatic encoder which comprises an encoder part and a decoder part and utilizing the training set T obtained in the step (3-4)trainTraining the variational automatic encoder to obtain the trained variational automatic encoder, and specifically comprising the following steps:
(4-1) designing and constructing an encoder by using a convolutional neural network, a cyclic neural network or a deep belief network, and carrying out local adaptive normalization on the local moving window data obtained in the step (3-3)
Figure BDA0002385908640000104
As input to the encoder, the mapping is derived
Figure BDA0002385908640000105
Feature vector σ ofkAnd mukFeature vector σkAnd mukThere are l numbers, l represents the dimension of the feature vector, l is greater than or equal to m and less than or equal to 4 m:
Figure BDA0002385908640000106
Figure BDA0002385908640000107
(4-2) Using the feature vector σ of step (4-1)kAnd mukCarrying out reparameterization to obtain
Figure BDA0002385908640000108
Hidden feature vector h ofk，hkIncludes l values:
hk＝μk+σk⊙∈
where e is normally distributed from the standard
Figure BDA0002385908640000109
A random sample results,. indicates multiplication of corresponding elements of the vector;
(4-3) designing and constructing a decoder by using a convolutional neural network, a cyclic neural network or a deep belief network, and enabling the hidden feature vector h in the step (4-2)kAs input to the decoder, reconstructing to obtain the data corresponding to step (3-3)
Figure BDA00023859086400001010
Reconstructed data with the same dimensionality
Figure BDA00023859086400001011
Figure BDA00023859086400001012
There are m rows and t columns of data:
Figure BDA00023859086400001013
(4-4) utilizing the eigenvector σ of step (4-1) according to the following loss functionkAnd mukAnd the reconstructed data of step (4-3)
Figure BDA0002385908640000111
Calculating the locally adaptive normalized local moving window data of step (3-3)
Figure BDA0002385908640000112
Error of (2)
Figure BDA0002385908640000113
Figure BDA0002385908640000114
Figure BDA0002385908640000115
I.e. the loss function of the variational automatic encoder, the loss function
Figure BDA0002385908640000116
Including reconstruction losses
Figure BDA0002385908640000117
And KL divergence loss
Figure BDA0002385908640000118
λ is the weighting factor of KL divergence loss versus reconstruction loss, 103≤λ≤106The loss of the two parts is calculated as follows, wherein j represents the variable serial number of the chemical process, and j is more than or equal to 1 and less than or equal to m:
Figure BDA0002385908640000119
Figure BDA00023859086400001110
(4-5) repeating the step (4-1) to the step (4-4), and sequentially combining the training sets T in the step (3-4)trainEach data of
Figure BDA00023859086400001111
Inputting the variational automatic encoder to carry out error calculation, and training the variational automatic encoder through an error back propagation algorithm to obtain the variational automatic encoderThe trained variational automatic encoder is used for a fault detection task of the chemical system;
(5) utilizing the trained variational automatic encoder obtained in the step (4) and the verification set T obtained in the step (3-7)validBy estimating the verification set TvalidThe abnormal score confidence interval of the variable automatic encoder is used for obtaining a monitoring threshold eta when the variable automatic encoder is used for a fault detection task, and the method specifically comprises the following steps:
(5-1) local moving window data for local adaptive normalization of step (3-6)
Figure BDA00023859086400001112
Mapping the input of the variational automatic encoder trained in the step (4) to obtain
Figure BDA00023859086400001113
Feature vector σ ofpAnd mupFeature vector σpAnd mupThere are l numbers, respectively, l represents the dimension of the feature vector, and has the same size as l of step (4-1):
Figure BDA00023859086400001114
Figure BDA00023859086400001115
(5-2) Using the feature vector σ of step (5-1)pAnd mupTo, for
Figure BDA00023859086400001116
Carrying out reparameterization to obtain
Figure BDA00023859086400001117
Hidden feature vector h ofp，hpIncludes l values:
hp＝μp+σp⊙∈
where e is normally distributed from the standard
Figure BDA00023859086400001118
A random sample results,. indicates multiplication of corresponding elements of the vector;
(5-3) combining the hidden feature vector h of the step (5-2)pAs the input of the decoder in the variational automatic encoder trained in the step (4), reconstructing to obtain the result of the step (3-6)
Figure BDA00023859086400001119
Reconstructed data with the same dimensionality
Figure BDA00023859086400001120
Figure BDA00023859086400001121
There are m rows and t columns of data:
Figure BDA00023859086400001122
(5-4) utilizing the feature vector σ of step (5-1) according to the following abnormality score calculation formulapAnd mupAnd the reconstructed data of step (5-3)
Figure BDA0002385908640000121
Calculating the local adaptive standardized local moving window number in the step (3-6)
Figure BDA0002385908640000122
Is abnormal score of
Figure BDA0002385908640000123
Figure BDA0002385908640000124
Abnormal score
Figure BDA0002385908640000125
Including reconstruction losses
Figure BDA0002385908640000126
And KL divergence loss
Figure BDA0002385908640000127
λ is a weighting coefficient of KL divergence loss with respect to reconstruction loss, the same as λ of step (4-4); the loss of the two parts is calculated as follows, wherein j represents the variable serial number of the chemical process, j is more than or equal to 1 and less than or equal to m:
Figure BDA0002385908640000128
Figure BDA0002385908640000129
(5-5) repeating the steps (5-1) to (5-4), and sequentially adding the verification sets T of the steps (3-7)validEach data of
Figure BDA00023859086400001210
Input variable automatic encoder for calculating abnormal score
Figure BDA00023859086400001211
Get the verification set TvalidIs abnormal score data set Svalid；
(5-6) abnormal score data set SvalidObtaining abnormal score data set S according to normal distributionvalidThe abnormal fraction with the normal distribution confidence coefficient alpha is used as a monitoring threshold eta of the chemical system, and alpha is more than or equal to 99% and less than or equal to 99.99%;
(6) and (3) carrying out online fault detection on the process data of the chemical system under different working conditions by using the variational automatic encoder trained in the step (4) and the monitoring threshold eta obtained in the step (5), wherein the method comprises the following steps:
(6-1) collecting process data from a real-time database of the chemical system at the current detection moment q, and selecting local moving window data w with a time window t from the time to the frontq，wqThere are m rows and t columns of data, where t is the time window in step (3-2):
Figure BDA00023859086400001212
using local moving window data wqCalculating wqAverage value of m variables in (1) to obtain mean (w)q)，mean(wq) Comprises m numbers;
(6-2) Using gmstd (Dtrain) in step (3-1) and mean (w) in step (6-1)q) For the local moving window data w of step (6-1)qPerforming local adaptive normalization to wqApproximately converting m variables of the data into standard normal distribution to obtain local self-adaptive standardized local moving window data
Figure BDA00023859086400001213
Figure BDA0002385908640000131
(6-3) local moving window data for local adaptive normalization of step (6-2)
Figure BDA0002385908640000132
Mapping the input of the encoder in the variational automatic encoder trained in the step (4) to obtain
Figure BDA0002385908640000133
Feature vector σ ofqAnd vqThe two feature vectors have values of l, wherein l represents the dimension of the feature vector and has the same size as l in the step (4-1):
Figure BDA0002385908640000134
Figure BDA0002385908640000135
(6-4) Using the feature vector σ of step (6-3)qAnd muqCarrying out reparameterization to obtain
Figure BDA0002385908640000136
Hidden feature vector hq, h ofqIncludes l values:
hq＝μq+σq⊙∈
where e is normally distributed from the standard
Figure BDA0002385908640000137
A random sample results,. indicates multiplication of corresponding elements of the vector;
(6-5) hiding the feature vector h of the step (6-4)qAs the input of the decoder in the variational automatic encoder trained in the step (4), reconstructing to obtain the result corresponding to the step (6-2)
Figure BDA0002385908640000138
Reconstructed data with the same dimensionality
Figure BDA0002385908640000139
There are m rows and t columns of data:
Figure BDA00023859086400001310
(6-6) utilizing the feature vector σ of step (6-3) according to the following abnormality score calculation formulaqAnd muqAnd the reconstructed data of step (6-5)
Figure BDA00023859086400001311
Calculating the locally adaptive normalized local moving window data of step (6-2)
Figure BDA00023859086400001312
Is abnormal score of
Figure BDA00023859086400001313
Figure BDA00023859086400001314
Abnormal score
Figure BDA00023859086400001315
Including reconstruction losses
Figure BDA00023859086400001316
And KL divergence loss
Figure BDA00023859086400001317
λ is a weighting coefficient of KL divergence loss with respect to reconstruction loss, the same as λ of step (4-4); the loss of the two parts is calculated as follows, wherein j represents the variable serial number of the chemical process, j is more than or equal to 1 and less than or equal to m:
Figure BDA00023859086400001318
Figure BDA00023859086400001319
(6-7) scoring the abnormality of step (6-6)
Figure BDA0002385908640000141
Comparing with the monitoring threshold eta obtained in the step (5), if so
Figure BDA0002385908640000142
The current chemical system is in a normal operation state, the step (6-1) is returned to continue monitoring the online real-time data, and if the online real-time data is not monitored, the chemical system is in a normal operation state
Figure BDA0002385908640000143
The system fault of the current chemical system is indicated, and fault warning is sent out, so that the chemical system based on local adaptive standardization is moreAnd detecting working condition faults.
Embodiments of the method of the present invention are described below with reference to the accompanying drawings:
(1) acquiring a normal operation data set D under 2 working conditions from a historical database of a chemical systemhistoryData set DhistoryThe method comprises the following steps of (1) sharing m rows and n columns of data, wherein m is 42 to represent a process variable of a chemical system, n is 16000 to represent total operation time, and each working condition comprises 8000 operation times;
(2) setting the normal operation data set D in the step (1)historyInto training sets DtrainAnd a verification set DvalidTraining set DtrainComprising m rows ntrainColumn data, validation set DvalidComprising m rows nvalidColumn data, in which training set DtrainD of historical normal operation data sethistoryThe ratio of a to 75%, ntrain＝12000，nvalid＝4000；
(3) Training set D in step (2)trainAnd a verification set DvalidCarrying out local self-adaptive standardization processing to obtain a transformed training set TtrainAnd a verification set TvalidThe method comprises the following specific steps:
(3-1) Using the training set D of step (2)trainThe global mean standard deviation gmstd (D) of m process variables in the chemical system is calculated by using the following formulatrain) And m is 42 values:
Figure BDA0002385908640000144
wherein i represents the working condition serial number of the chemical process, i is more than or equal to 1 and less than or equal to N, and Dtrain，iRepresentative training set DtrainNormal operating data of the ith operating mode, std (D)train，i) Representative training set DtrainThe standard deviation vector of the ith working condition comprises m numerical values, and std (D) is obtained by calculating the standard deviation of the corresponding variabletrain，i)，ntrain，iRepresentative training set DtrainNumber of normal operating data of the ith working condition, ntrain，i＝6000；
(3-2) training set D for step (2)trainThe k-th normal operation data xk, k in (b) represents the training set DtrainRun time number of (1), 2, ntrain，xkComprises m variable values with time sequence number k, and local moving window data w with time window t is selected forward by calculating timek，wkThere are m rows and t columns of data, where t is the time window, m is 42, t is 30:
Figure BDA0002385908640000151
using local moving window data wkCalculating wkAverage value of m variables in (1) to obtain mean (w)k)，mean(wk) Comprises m numbers;
(3-3) Using gmstd (D) in step (3-1)train) And mean (w) in step (3-2)k) For the local moving window data w of step (3-2)kPerforming local adaptive normalization to wkApproximately converting m variables of the data into standard normal distribution to obtain local self-adaptive standardized local moving window data
Figure BDA0002385908640000152
Figure BDA0002385908640000153
(3-4) repeating the step (3-2) and the step (3-3), and sequentially calculating D in the training settrainObtaining a local adaptive standardized training set T by each normal operation datatrain；
(3-5) authentication set D for step (2)validP-th normal operation data x in (1)pAnd p represents the verification set DvalidOperating time number of (1), 2, nvalid，xpLocal movement comprising m variable values with time sequence number p, with time forward selecting time window tWindow data wp，wpThere are m rows and t columns of data, where t is the time window in step (3-2), m is 42, t is 30:
Figure BDA0002385908640000154
using local moving window data wpCalculating wpAverage value of m variables in (1) to obtain mean (w)p)，mean(wp) Comprises m numbers;
(3-6) Using gmstd (D) in step (3-1)train) And mean (w) in step (3-5)p) For the local moving window data w of step (3-5)pPerforming local adaptive normalization to wpApproximately converting m variables of the data into standard normal distribution to obtain local self-adaptive standardized local moving window data
Figure BDA0002385908640000155
Figure BDA0002385908640000156
(3-7) repeating the steps (3-5) and (3-6), and sequentially calculating D in the verification setvalidObtaining a local self-adaptive standardized verification set T from each normal operation datavalid；
(4) Constructing a variational automatic encoder which comprises an encoder part and a decoder part and utilizing the training set T obtained in the step (3-4)trainTraining the variational automatic encoder to obtain the trained variational automatic encoder, and specifically comprising the following steps:
(4-1) designing and constructing an encoder by using a bidirectional Long Short-term memory (BilSTM) and a linear layer, wherein the encoder has a structure shown in figure 2 and comprises two layers of BilSTM and the linear layer, and the local adaptive standardized local moving window data in the step (3-3) is processed
Figure BDA0002385908640000161
As input to the encoder, the mapping is derived
Figure BDA0002385908640000162
Feature vector σ ofkAnd mukFeature vector σkAnd mukThere are l values, l represents the dimension of the feature vector, l is 50:
Figure BDA0002385908640000163
Figure BDA0002385908640000164
(4-2) Using the feature vector σ of step (4-1)kAnd mukCarrying out reparameterization to obtain
Figure BDA0002385908640000165
Hidden feature vector h ofk，hkIncludes l values, l 50:
hk＝μk+σk⊙∈
where e is normally distributed from the standard
Figure BDA0002385908640000166
A random sample results,. indicates multiplication of corresponding elements of the vector;
(4-3) designing and constructing a decoder by using a bidirectional Long Short-term Memory (BilSTM) and a linear layer, wherein the decoder has a structure shown in figure 2 and comprises two layers of BilSTM and the linear layer, and the hidden feature vector h in the step (4-2) is processedkAs input to the decoder, reconstructing to obtain the data corresponding to step (3-3)
Figure BDA0002385908640000167
Reconstructed data with the same dimensionality
Figure BDA0002385908640000168
Figure BDA0002385908640000169
There are m rows and t columns of data, m is 42, t is 30:
Figure BDA00023859086400001610
(4-4) utilizing the eigenvector σ of step (4-1) according to the following loss functionkAnd mukAnd the reconstructed data of step (4-3)
Figure BDA00023859086400001611
Calculating the locally adaptive normalized local moving window data of step (3-3)
Figure BDA00023859086400001612
Error of (2)
Figure BDA00023859086400001613
Figure BDA00023859086400001614
Figure BDA00023859086400001615
I.e. the loss function of the variational automatic encoder, the loss function
Figure BDA00023859086400001616
Including reconstruction losses
Figure BDA00023859086400001617
And KL divergence loss
Figure BDA00023859086400001618
λ is the weighting factor of KL divergence loss versus reconstruction loss, λ is 105The loss of the two parts is calculated as follows, wherein j represents the variable serial number of the chemical process, j is more than or equal to 1 and less than or equal to m, and m is 42:
Figure BDA00023859086400001619
Figure BDA00023859086400001620
(4-5) repeating the step (4-1) to the step (4-4), and sequentially combining the training sets T in the step (3-4)trainEach data of
Figure BDA00023859086400001621
Inputting the variational automatic encoder to carry out error calculation, and training the variational automatic encoder through an error back propagation algorithm to obtain a trained variational automatic encoder;
(5) utilizing the trained variational automatic encoder obtained in the step (4) and the verification set T obtained in the step (3-7)validBy estimating the verification set TvalidThe abnormal score confidence interval of the variable automatic encoder is used for obtaining a monitoring threshold eta when the variable automatic encoder is used for a fault detection task, and the method specifically comprises the following steps:
(5-1) local moving window data for local adaptive normalization of step (3-6)
Figure BDA0002385908640000171
Mapping the input of the variational automatic encoder trained in the step (4) to obtain
Figure BDA0002385908640000172
Feature vector σ ofpAnd mupFeature vector σpAnd mupThere are l values, l represents the dimension of the feature vector, l is 50:
Figure BDA0002385908640000173
Figure BDA0002385908640000174
(5-2) Using the feature vector σ of step (5-1)pAnd mupTo, for
Figure BDA0002385908640000175
Carrying out reparameterization to obtain
Figure BDA0002385908640000176
Hidden feature vector h ofp，hpIncludes l values, l 50:
hp＝μp+σp⊙∈
where e is normally distributed from the standard
Figure BDA0002385908640000177
A random sample results,. indicates multiplication of corresponding elements of the vector;
(5-3) combining the hidden feature vector h of the step (5-2)pAs the input of the decoder in the variational automatic encoder trained in the step (4), reconstructing to obtain the result of the step (3-6)
Figure BDA0002385908640000178
Reconstructed data with the same dimensionality
Figure BDA0002385908640000179
Figure BDA00023859086400001710
There are m rows and t columns of data, m is 42, t is 30:
Figure BDA00023859086400001711
(5-4) utilizing the feature vector σ of step (5-1) according to the following abnormality score calculation formulapAnd mupAnd the reconstructed data of step (5-3)
Figure BDA00023859086400001712
Computing step (3-6) local adaptationNormalized local moving window data
Figure BDA00023859086400001713
Is abnormal score of
Figure BDA00023859086400001714
Figure BDA00023859086400001715
Abnormal score
Figure BDA00023859086400001716
Including reconstruction losses
Figure BDA00023859086400001717
And KL divergence loss
Figure BDA00023859086400001718
λ is a weighting coefficient of KL divergence loss with respect to reconstruction loss, and is the same as λ in step (4-4), where λ is 105(ii) a The loss of the two parts is calculated as follows, wherein j represents the variable serial number of the chemical process, j is more than or equal to 1 and less than or equal to m, and m is 42:
Figure BDA00023859086400001719
Figure BDA00023859086400001720
(5-5) repeating the steps (5-1) to (5-4), and sequentially adding the verification sets T of the steps (3-7)validEach data of
Figure BDA0002385908640000181
Input variable automatic encoder for calculating abnormal score
Figure BDA0002385908640000182
Get the verification set TvalidIs abnormal score data set Svalid；
(5-6) abnormal score data set SvalidObtaining abnormal score data set S according to normal distributionvalidThe abnormal fraction with the normal distribution confidence coefficient alpha is used as a monitoring threshold eta of the chemical system, and the alpha is 99.9 percent;
(6) and (3) carrying out online fault detection on the process data of the chemical system under different working conditions by using the variational automatic encoder trained in the step (4) and the monitoring threshold eta obtained in the step (5), wherein the method comprises the following steps:
(6-1) collecting the process data of 4 working conditions from the database of the chemical system as a test set DtestIn total, m rows ntestColumn data, m 42, ntest4 (2000+4 × 1650), wherein the 4 working conditions include 2 historical working conditions and 2 new working conditions in the step (1), and are used for testing the fault detection effect of the invention under different working conditions. Each condition includes normal operating data and 4 types of failed operating data. Wherein, each operating mode includes 2000 normal operating data, and each operating mode of the operating data that breaks down includes 4 fault types, and each fault type includes 1650 operating data, and preceding 450 operating data still belongs to normal operating data, introduces the trouble from 450 operating data, and 1200 last operating data belong to trouble operating data, and 4 fault types are as shown in the following table:
table 1 4 fault types in test data
Figure BDA0002385908640000183
For test set DtestQ-th normal operation data x in (1)qAnd q represents test set DtestOperating time sequence number q 1, 2, ntest，xqComprises m variable values with time sequence number q, and local moving window data w with time window t is selected forward by calculating timeq，wqThere are m rows and t columns of data, where t is the time window in step (3-2), m is 42, t is 30:
Figure BDA0002385908640000184
using local moving window data wqCalculating wqAverage value of m variables in (1) to obtain mean (w)q)，mean(wq) Comprises m numbers;
(6-2) Using gmstd (Dtrain) in step (3-1) and mean (w) in step (6-1)q) For the local moving window data w of step (6-1)qPerforming local adaptive normalization to wqApproximately converting m variables of the data into standard normal distribution to obtain local self-adaptive standardized local moving window data
Figure BDA0002385908640000191
Figure BDA0002385908640000192
(6-3) local moving window data for local adaptive normalization of step (6-2)
Figure BDA0002385908640000193
Mapping the input of the encoder in the variational automatic encoder trained in the step (4) to obtain
Figure BDA0002385908640000194
Feature vector σ ofqAnd muqThe two eigenvectors have values of l, l represents the dimension of the eigenvector and has the same size as l in step (4-1), and l is 50:
Figure BDA0002385908640000195
Figure BDA0002385908640000196
(6-4) advantageUsing the feature vector σ of step (6-3)qAnd muqCarrying out reparameterization to obtain
Figure BDA0002385908640000197
Hidden feature vector h ofq，hqIncludes l values, l 50:
hq＝μq+σq⊙∈
where e is normally distributed from the standard
Figure BDA0002385908640000198
A random sample results,. indicates multiplication of corresponding elements of the vector;
(6-5) hiding the feature vector h of the step (6-4)qAs the input of the decoder in the variational automatic encoder trained in the step (4), reconstructing to obtain the result corresponding to the step (6-2)
Figure BDA0002385908640000199
Reconstructed data with the same dimensionality
Figure BDA00023859086400001910
There are m rows and t columns of data, m is 42, t is 30:
Figure BDA00023859086400001911
(6-6) utilizing the feature vector σ of step (6-3) according to the following abnormality score calculation formulaqAnd muqAnd the reconstructed data of step (6-5)
Figure BDA00023859086400001912
Calculating the locally adaptive normalized local moving window data of step (6-2)
Figure BDA00023859086400001913
Is abnormal score of
Figure BDA00023859086400001914
Figure BDA00023859086400001915
Abnormal score
Figure BDA00023859086400001916
Including reconstruction losses
Figure BDA00023859086400001917
And KL divergence loss.
Figure BDA00023859086400001918
λ is a weighting coefficient of KL divergence loss with respect to reconstruction loss, and is the same as λ in step (4-4), where λ is 105(ii) a The loss of the two parts is calculated as follows, wherein j represents the variable serial number of the chemical process, j is more than or equal to 1 and less than or equal to m, and m is 42:
Figure BDA0002385908640000201
Figure BDA0002385908640000202
(6-7) scoring the abnormality of step (6-6)
Figure BDA0002385908640000203
Comparing with the monitoring threshold eta obtained in the step (5), if so
Figure BDA0002385908640000204
The current chemical system is in a normal operation state, the step (6-1) is returned to continue monitoring the online real-time data, and if the online real-time data is not monitored, the chemical system is in a normal operation state
Figure BDA0002385908640000205
The system fault of the current chemical system is indicated, and fault warning is sent out, so that the multi-working-condition fault detection of the chemical system based on local adaptive standardization is realized.
According to the above determination rule, fig. 3 shows the effect of fault detection in the present embodiment under the 4 conditions of step (6-1). Wherein, the working condition 1 and the working condition 2 represent 2 historical working conditions of the step (1), and the working condition 3 and the working condition 4 represent new working conditions which are not used in the step (4) and the step (5). Fig. 3 (a) to (e) show the monitoring effects of the fault detection model on the normal operation data and the fault operation data of the faults 1 to 4, respectively. The abscissa of each sub-graph represents the running time, the ordinate represents the anomaly score, the abscissa represents the monitoring threshold η obtained in step (5), and the vertical dotted line represents the introduction of a fault starting from the 450 th running data in step (1). If the black solid line is lower than the monitoring threshold represented by the dotted line, the normal operation of the chemical system is indicated; if the solid black line is higher than the monitoring threshold represented by the horizontal dashed line, it indicates that the chemical system is malfunctioning. As shown in fig. 3, in (a), the black solid line of the normal operation data under 4 working conditions is below the monitoring threshold (horizontal dotted line), which proves that the method can correctly determine the normal operation of the chemical system; and (b) to (e), the black solid line of the 4 working condition fault operation data is higher than the monitoring threshold (horizontal dotted line) from 450 (vertical dotted line), so that the method can be used for correctly judging that the chemical system has faults. The invention has similar fault detection results for the operation data of the working conditions 1-4, and proves that the fault detection method based on the local adaptive standardization has better detection effect under all the working conditions.

Claims (1)

1.一种基于局部自适应标准化的化工系统多工况故障检测方法，其特征在于，包括以下步骤：1. a chemical system multi-condition fault detection method based on local self-adaptive standardization, is characterized in that, comprises the following steps: (1)从化工系统的历史数据库中获取N种工况下的正常运行数据集Dhistory，数据集Dhistory共有m行n列数据，其中，m代表化工系统的过程变量，n代表总运行时间；(1) Obtain the normal operation data set D history under N working conditions from the historical database of the chemical system. The data set D history has m rows and n columns of data, where m represents the process variables of the chemical system, and n represents the total running time ; (2)将步骤(1)中的正常运行数据集Dhistory划分成训练集Dtrain和验证集Dvalid，训练集Dtrain包括m行ntrain列数据，验证集Dvalid包括m行nvalid列数据，其中训练集Dtrain占历史正常运行数据集的Dhistory比例为
Figure FDA0002385908630000011
60％≤a≤90％；(2) Divide the normal operation data set D history in step (1) into a training set D train and a validation set D valid , the training set D train includes m rows and n train columns of data, and the validation set D valid includes m rows and n valid columns Data, in which the proportion of the training set D train to the D history of the historical normal operation data set is:
Figure FDA0002385908630000011
60%≤a≤90%; (3)将步骤(2)中的训练集Dtrain和验证集Dvalid进行局部自适应标准化处理，得到转化后的训练集Ttrain和验证集Tvalid，具体步骤如下：(3) Perform local adaptive normalization on the training set D train and the verification set D valid in step (2) to obtain the transformed training set T train and verification set T valid , and the specific steps are as follows: (3-1)利用步骤(2)的训练集Dtrain中的正常运行数据，利用下式计算化工系统中m个过程变量的全局平均标准差gmstd(Dtrain)，包括m个数值：(3-1) Using the normal operation data in the training set D train of step (2), use the following formula to calculate the global average standard deviation gmstd(D train ) of m process variables in the chemical system, including m values:
Figure FDA0002385908630000012
Figure FDA0002385908630000012
 其中，i代表化工过程的工况序号，1≤i≤N，则Dtrain，i代表训练集Dtrain中第i种工况的正常运行数据，std(Dtrain，i)代表训练集Dtrain中第i种工况的标准差向量，包括m个数值，通过计算对应变量的标准差得到std(Dtrain，i)，ntrain，i代表训练集Dtrain中第i种工况的正常运行数据的数量，Among them, i represents the working condition number of the chemical process, 1≤i≤N, then D train, i represents the normal operation data of the ith working condition in the training set D train , std(D train, i ) represents the training set D train The standard deviation vector of the ith working condition in the middle, including m values, by calculating the standard deviation of the corresponding variables to get std(D train, i ), n train, i represents the normal operation of the ith working condition in the training set D train the amount of data, (3-2)对于步骤(2)的训练集Dtrain中的第k个正常运行数据xk，k代表训练集Dtrain中的运行时间序号，k＝1、2...、ntrain，xk包括时间序号为k的m个变量数值，计算时间向前选取时间窗口为t的局部移动窗口数据wk，wk共有m行t列数据，其中t为时间窗口，10≤t≤100：(3-2) For the kth normal running data x k in the training set D train in step (2), k represents the running time sequence number in the training set D train , k=1, 2..., n train , x k includes m variable values with time serial number k, and the calculation time forward selects local moving window data w k with time window t, w k has m rows and t columns of data, where t is the time window, 10≤t≤100 :
Figure FDA0002385908630000013
Figure FDA0002385908630000013
 利用局部移动窗口数据wk，计算wk中的m个变量的平均值，得到mean(wk)，mean(wk)包括m个数值；Using the local moving window data w k , calculate the mean value of m variables in w k to obtain mean(w k ), where mean(w k ) includes m values; (3-3)利用步骤(3-1)中的gmstd(Dtrain)和步骤(3-2)中的mean(wk)，对步骤(3-2)的局部移动窗口数据wk进行局部自适应标准化，使wk数据的m个变量近似转化为标准正态分布，得到局部自适应标准化的局部移动窗口数据
Figure FDA0002385908630000021
(3-3) Using gmstd(D train ) in step (3-1) and mean(w k ) in step (3-2), perform localization on the local moving window data w k of step (3-2) Adaptive standardization, so that the m variables of the wk data are approximately converted into standard normal distribution, and the local moving window data of local adaptive standardization is obtained.
Figure FDA0002385908630000021
Figure FDA0002385908630000022
Figure FDA0002385908630000022
 (3-4)重复步骤(3-2)和步骤(3-3)，依次计算训练集中Dtrain内每个正常运行数据，得到局部自适应标准化的训练集Ttrain；(3-4) repeat step (3-2) and step (3-3), calculate successively each normal operation data in training set D train , obtain the training set T train of local adaptive standardization; (3-5)对于步骤(2)的验证集Dvalid中的第p个正常运行数据xp，p代表验证集Dvalid中的运行时间序号，p＝1、2...、nvalid，xp包括时间序号为p的m个变量数值，时间向前选取时间窗口为t的局部移动窗口数据wp，wp共有m行t列数据，其中t是步骤(3-2)中的时间窗口：(3-5) For the p-th normal running data x p in the verification set D valid in step (2), p represents the running time sequence number in the verification set D valid , p=1, 2..., n valid , x p includes m variable values with time serial number p, and selects local moving window data w p with time window t forward in time, w p has m rows and t columns of data, where t is the time in step (3-2) window:
Figure FDA0002385908630000023
Figure FDA0002385908630000023
 利用局部移动窗口数据wp，计算wp中的m个变量的平均值，得到mean(wp)，mean(wp)包括m个数值；Using the local moving window data w p , calculate the mean value of m variables in w p to obtain mean(w p ), where mean(w p ) includes m values; (3-6)利用步骤(3-1)中的gmstd(Dtrain)和步骤(3-5)中的mean(wp)，对步骤(3-5)的局部移动窗口数据wp进行局部自适应标准化，使wp数据的m个变量近似转化为标准正态分布，得到局部自适应标准化的局部移动窗口数据
Figure FDA0002385908630000024
(3-6) Using gmstd(D train ) in step (3-1) and mean(w p ) in step (3-5), perform localization on the local moving window data w p of step (3-5) Adaptive standardization, so that the m variables of the wp data are approximately converted into standard normal distribution, and the local moving window data of local adaptive standardization is obtained.
Figure FDA0002385908630000024
Figure FDA0002385908630000025
Figure FDA0002385908630000025
 (3-7)重复步骤(3-5)和步骤(3-6)，依次计算验证集中Dvalid内每个正常运行数据，得到局部自适应标准化的验证集Tvalid；(3-7) repeat step (3-5) and step (3-6), calculate successively each normal operation data in the verification set D valid , obtain the verification set T valid of local adaptive standardization; (4)构建一个变分自动编码器，包括编码器和解码器两部分，并利用步骤(3-4)得到的训练集Ttrain对该变分自动编码器进行训练，得到训练完毕的变分自动编码器，具体步骤如下：(4) Construct a variational auto-encoder, including an encoder and a decoder, and use the training set T train obtained in step (3-4) to train the variational auto-encoder to obtain the trained variation Autoencoder, the specific steps are as follows: (4-1)利用卷积神经网络、循环神经网络或深度置信网络，设计并构建编码器，将步骤(3-3)的局部自适应标准化的局部移动窗口数据
Figure FDA0002385908630000026
作为编码器的输入，映射得到
Figure FDA0002385908630000027
的特征向量σk和μk，特征向量σk和μk分别有l个数值，l表示特征向量的维度，m≤l≤4m：(4-1) Using a convolutional neural network, a recurrent neural network or a deep belief network, an encoder is designed and constructed to normalize the local moving window data of the local adaptation in step (3-3)
Figure FDA0002385908630000026
As the input to the encoder, the mapping gets
Figure FDA0002385908630000027
The eigenvectors σ k and μ k of , the eigenvectors σ k and μ k have l values respectively, l represents the dimension of the eigenvector, m≤l≤4m:
Figure FDA0002385908630000031
Figure FDA0002385908630000031
Figure FDA0002385908630000032
Figure FDA0002385908630000032
 (4-2)利用步骤(4-1)的特征向量σk和μk，进行重参数化得到
Figure FDA0002385908630000033
的隐藏特征向量hk，hk包括l个数值：(4-2) Using the eigenvectors σ k and μ k of step (4-1), re-parameterization is performed to obtain
Figure FDA0002385908630000033
The hidden feature vector h k of , h k includes l values: hk＝μk+σk⊙∈h k = μ k +σ k ⊙∈ 其中∈是从标准正态分布
Figure FDA0002385908630000034
随机采样得到，⊙表示向量对应元素相乘；where ∈ is from the standard normal distribution
Figure FDA0002385908630000034
It is obtained by random sampling, and ⊙ means that the corresponding elements of the vector are multiplied; (4-3)利用卷积神经网络、循环神经网络或深度置信网络，设计并构建解码器，将步骤(4-2)的隐藏特征向量hk作为解码器的输入，重构得到与步骤(3-3)的
Figure FDA0002385908630000035
具有相同维度的重构数据
Figure FDA0002385908630000036
Figure FDA0002385908630000037
共有m行t列数据：(4-3) Use convolutional neural network, recurrent neural network or deep belief network to design and build a decoder, take the hidden feature vector h k of step (4-2) as the input of the decoder, and reconstruct to obtain the same value as step ( 3-3) of
Figure FDA0002385908630000035
Reconstructed data with the same dimensions
Figure FDA0002385908630000036
Figure FDA0002385908630000037
There are m rows and t columns of data:
Figure FDA0002385908630000038
Figure FDA0002385908630000038
 (4-4)根据如下损失函数，利用步骤(4-1)的特征向量σk和μk以及步骤(4-3)的重构数据
Figure FDA0002385908630000039
计算步骤(3-3)局部自适应标准化的局部移动窗口数据
Figure FDA00023859086300000310
的误差
Figure FDA00023859086300000311
(4-4) According to the following loss function, use the feature vectors σ k and μ k of step (4-1) and the reconstructed data of step (4-3)
Figure FDA0002385908630000039
Computation step (3-3) Local adaptive normalized local moving window data
Figure FDA00023859086300000310
error
Figure FDA00023859086300000311
Figure FDA00023859086300000312
Figure FDA00023859086300000312
Figure FDA00023859086300000313
即为变分自动编码器的损失函数，损失函数
Figure FDA00023859086300000314
包括重构损失
Figure FDA00023859086300000315
和KL散度损失
Figure FDA00023859086300000316
λ是KL散度损失相对于重构损失的权重系数，103≤λ≤106，两部分损失计算如下，其中j代表化工过程的变量序号，1≤j≤m：
Figure FDA00023859086300000313
is the loss function of the variational autoencoder, the loss function
Figure FDA00023859086300000314
including reconstruction loss
Figure FDA00023859086300000315
and KL divergence loss
Figure FDA00023859086300000316
λ is the weight coefficient of the KL divergence loss relative to the reconstruction loss, 10 3 ≤λ≤10 6 , the two parts of the loss are calculated as follows, where j represents the variable number of the chemical process, 1≤j≤m:
Figure FDA00023859086300000317
Figure FDA00023859086300000317
Figure FDA00023859086300000318
Figure FDA00023859086300000318
 (4-5)重复步骤(4-1)-步骤(4-4)，依次将步骤(3-4)的训练集Ttrain的每个数据
Figure FDA00023859086300000319
输入变分自动编码器进行误差计算，通过误差反向传播算法训练变分自动编码器得到训练完毕的变分自动编码器；(4-5) Repeat step (4-1)-step (4-4), and sequentially convert each data of the training set T train of step (3-4)
Figure FDA00023859086300000319
Input the variational autoencoder for error calculation, and train the variational autoencoder through the error back-propagation algorithm to obtain the trained variational autoencoder; (5)利用步骤(4)得到的训练完毕的变分自动编码器和步骤(3-7)得到的验证集Tvalid，通过估计验证集Tvalid的异常分数置信度区间，得到变分自动编码器用于故障检测任务时的监控阈值η，具体步骤如下：(5) Using the trained variational auto-encoder obtained in step (4) and the validation set T valid obtained in step (3-7), the variational auto-encoder is obtained by estimating the confidence interval of the abnormal score of the validation set T valid The monitoring threshold η when the controller is used for the fault detection task, the specific steps are as follows: (5-1)将步骤(3-6)的局部自适应标准化的局部移动窗口数据
Figure FDA00023859086300000320
作为步骤(4)中训练完毕的变分自动编码器的输入，映射得到
Figure FDA00023859086300000321
的特征向量σp和μp，特征向量σp和μp分别有l个数值，l表示特征向量的维度：(5-1) Local adaptive normalized local moving window data of step (3-6)
Figure FDA00023859086300000320
As the input of the trained variational autoencoder in step (4), the mapping obtains
Figure FDA00023859086300000321
The eigenvectors σ p and μ p of , the eigenvectors σ p and μ p have l values respectively, and l represents the dimension of the eigenvector:
Figure FDA0002385908630000041
Figure FDA0002385908630000041
Figure FDA0002385908630000042
Figure FDA0002385908630000042
 (5-2)利用步骤(5-1)的特征向量σp和μp，对
Figure FDA0002385908630000043
进行重参数化，得到
Figure FDA0002385908630000044
的隐藏特征向量hp，hp包括l个数值：(5-2) Using the eigenvectors σ p and μ p of step (5-1), for
Figure FDA0002385908630000043
Reparameterization is carried out to get
Figure FDA0002385908630000044
The hidden feature vector h p of , h p includes l values: hp＝μp+σp⊙∈h p = μ p +σ p ⊙∈ 其中∈是从标准正态分布
Figure FDA00023859086300000421
随机采样得到，⊙表示向量对应元素相乘；where ∈ is from the standard normal distribution
Figure FDA00023859086300000421
It is obtained by random sampling, and ⊙ means that the corresponding elements of the vector are multiplied; (5-3)将步骤(5-2)的隐藏特征向量hp作为步骤(4)训练完毕的变分自动编码器中解码器的输入，重构得到与步骤(3-6)的
Figure FDA0002385908630000045
具有相同维度的重构数据
Figure FDA0002385908630000046
Figure FDA0002385908630000047
共有m行t列数据：(5-3) Use the hidden feature vector h p of step (5-2) as the input of the decoder in the variational autoencoder trained in step (4), and reconstruct to obtain the same value as step (3-6).
Figure FDA0002385908630000045
Reconstructed data with the same dimensions
Figure FDA0002385908630000046
Figure FDA0002385908630000047
There are m rows and t columns of data:
Figure FDA0002385908630000048
Figure FDA0002385908630000048
 (5-4)根据如下异常分数计算公式，利用步骤(5-1)的特征向量σp和μp以及步骤(5-3)的重构数据
Figure FDA0002385908630000049
计算步骤(3-6)局部自适应标准化的局部移动窗口数据
Figure FDA00023859086300000410
的异常分数
Figure FDA00023859086300000420
(5-4) According to the following abnormal score calculation formula, use the feature vectors σ p and μ p of step (5-1) and the reconstructed data of step (5-3)
Figure FDA0002385908630000049
Computation step (3-6) Local adaptive normalized local moving window data
Figure FDA00023859086300000410
abnormal score of
Figure FDA00023859086300000420
Figure FDA00023859086300000412
Figure FDA00023859086300000412
 异常分数
Figure FDA00023859086300000413
包括重构损失
Figure FDA00023859086300000414
和KL散度损失
Figure FDA00023859086300000415
λ是KL散度损失相对于重构损失的权重系数，与步骤(4-4)的λ相同；两部分损失计算如下，其中j代表化工过程的变量序号，1≤j≤m：abnormal score
Figure FDA00023859086300000413
including reconstruction loss
Figure FDA00023859086300000414
and KL divergence loss
Figure FDA00023859086300000415
λ is the weight coefficient of the KL divergence loss relative to the reconstruction loss, which is the same as λ in step (4-4); the two parts of the loss are calculated as follows, where j represents the variable number of the chemical process, 1≤j≤m:
Figure FDA00023859086300000416
Figure FDA00023859086300000416
Figure FDA00023859086300000417
Figure FDA00023859086300000417
 (5-5)重复步骤(5-1)-步骤(5-4)，依次将步骤(3-7)的验证集Tvalid的每个数据
Figure FDA00023859086300000418
输入变分自动编码器，计算异常分数
Figure FDA00023859086300000419
得到验证集Tvalid的异常分数数据集Svalid；(5-5) Repeat step (5-1)-step (5-4), and sequentially convert each data of the validation set T valid in step (3-7)
Figure FDA00023859086300000418
Input variational autoencoder, compute anomaly score
Figure FDA00023859086300000419
Obtain the abnormal score dataset S valid of the validation set T valid ; (5-6)异常分数数据集Svalid服从正态分布，取异常分数数据集Svalid的正态分布置信度为α的异常分数作为该化工系统的监控阈值η，99％≤α≤99.99％；(5-6) The abnormal score data set S valid obeys the normal distribution, and the abnormal score whose normal distribution reliability is α of the abnormal score data set S valid is taken as the monitoring threshold η of the chemical system, 99%≤α≤99.99% ; (6)利用步骤(4)中训练完毕的变分自动编码器，以及步骤(5)得到的监控阈值η，在不同工况下对该化工系统的过程数据进行在线故障检测，包括以下步骤：(6) utilize the variational autoencoder trained in step (4), and the monitoring threshold η that step (5) obtains, carry out online fault detection to the process data of this chemical system under different operating conditions, including the following steps: (6-1)在当前检测时刻q，从该化工系统的实时数据库中收集过程数据，时间向前选取时间窗口为t的局部移动窗口数据wq，wq共有m行t列数据，其中t为步骤(3-2)中的时间窗口：(6-1) At the current detection time q, collect process data from the real-time database of the chemical system, select the local moving window data w q with the time window t forward in time, and w q has m rows and t columns of data, where t is the time window in step (3-2):
Figure FDA0002385908630000051
Figure FDA0002385908630000051
 利用局部移动窗口数据wq，计算wq中的m个变量的平均值，得到mean(wq)，mean(wq)包括m个数值；Using the local moving window data w q , calculate the average value of m variables in w q to obtain mean(w q ), where mean(w q ) includes m values; (6-2)利用步骤(3-1)中的gmstd(Dtrain)和步骤(6-1)中的mean(wq)，对步骤(6-1)的局部移动窗口数据wq进行局部自适应标准化，使wq数据的m个变量近似转化为标准正态分布，得到局部自适应标准化的局部移动窗口数据
Figure FDA0002385908630000052
(6-2) Using gmstd(D train ) in step (3-1) and mean(w q ) in step (6-1), perform localization on the local moving window data w q of step (6-1) Adaptive standardization, so that the m variables of the w q data are approximately converted into standard normal distribution, and the local moving window data of local adaptive standardization is obtained.
Figure FDA0002385908630000052
Figure FDA0002385908630000053
Figure FDA0002385908630000053
 (6-3)将步骤(6-2)的局部自适应标准化的局部移动窗口数据
Figure FDA0002385908630000054
作为步骤(4)训练完毕的变分自动编码器中编码器的输入，映射得到
Figure FDA0002385908630000055
的特征向量σq和μq，两个特征向量分别都有1个数值，l表示特征向量的维度，与步骤(4-1)的l具有相同大小：(6-3) The local moving window data normalized by the local adaptation of step (6-2)
Figure FDA0002385908630000054
As the input of the encoder in the variational autoencoder trained in step (4), the mapping is obtained
Figure FDA0002385908630000055
The eigenvectors σ q and μ q of , the two eigenvectors each have 1 value, l represents the dimension of the eigenvector, and has the same size as l in step (4-1):
Figure FDA0002385908630000056
Figure FDA0002385908630000056
Figure FDA0002385908630000057
Figure FDA0002385908630000057
 (6-4)利用步骤(6-3)的特征向量σq和μq，进行重参数化得到
Figure FDA0002385908630000058
的隐藏特征向量hq，hq包括l个数值：(6-4) Using the eigenvectors σ q and μ q of step (6-3), perform reparameterization to obtain
Figure FDA0002385908630000058
The hidden feature vector h q of , h q includes l values: hq＝μq+σq⊙∈h q = μ q +σ q ⊙∈ 其中∈是从标准正态分布
Figure FDA0002385908630000059
随机采样得到，⊙表示向量对应元素相乘；where ∈ is from the standard normal distribution
Figure FDA0002385908630000059
It is obtained by random sampling, and ⊙ represents the multiplication of the corresponding elements of the vector; (6-5)将步骤(6-4)的隐藏特征向量hq作为步骤(4)中训练完毕的变分自动编码器中解码器的输入，重构得到与步骤(6-2)的
Figure FDA00023859086300000510
具有相同维度的重构数据
Figure FDA00023859086300000513
共有m行t列数据：(6-5) Use the hidden feature vector h q of step (6-4) as the input of the decoder in the variational autoencoder trained in step (4), and reconstruct to obtain the same value as step (6-2).
Figure FDA00023859086300000510
Reconstructed data with the same dimensions
Figure FDA00023859086300000513
There are m rows and t columns of data:
Figure FDA00023859086300000512
Figure FDA00023859086300000512
 (6-6)根据如下异常分数计算公式，利用步骤(6-3)的特征向量σq和vq以及步骤(6-5)的重构数据
Figure FDA0002385908630000061
计算步骤(6-2)局部自适应标准化的局部移动窗口数据
Figure FDA0002385908630000062
的异常分数
Figure FDA0002385908630000063
(6-6) According to the following abnormal score calculation formula, use the feature vectors σ q and v q of step (6-3) and the reconstructed data of step (6-5)
Figure FDA0002385908630000061
Computation step (6-2) Local adaptive normalized local moving window data
Figure FDA0002385908630000062
abnormal score of
Figure FDA0002385908630000063
Figure FDA0002385908630000064
Figure FDA0002385908630000064
 异常分数
Figure FDA0002385908630000065
包括重构损失
Figure FDA0002385908630000066
和KL散度损失
Figure FDA0002385908630000067
λ是KL散度损失相对于重构损失的权重系数，与步骤(4-4)的λ相同；两部分损失计算如下，其中j代表化工过程的变量序号，1≤j≤m：abnormal score
Figure FDA0002385908630000065
including reconstruction loss
Figure FDA0002385908630000066
and KL divergence loss
Figure FDA0002385908630000067
λ is the weight coefficient of the KL divergence loss relative to the reconstruction loss, which is the same as λ in step (4-4); the two parts of the loss are calculated as follows, where j represents the variable number of the chemical process, 1≤j≤m:
Figure FDA0002385908630000068
Figure FDA0002385908630000068
Figure FDA0002385908630000069
Figure FDA0002385908630000069
 (6-7)将步骤(6-6)的异常分数
Figure FDA00023859086300000610
与步骤(5)得到的监控阈值η进行比较，若
Figure FDA00023859086300000611
则当前化工系统处于正常运行状态，返回步骤(6-1)继续监控在线实时数据，若
Figure FDA00023859086300000612
则表明当前化工系统发生了系统故障，并发出故障警告，实现基于局部自适应标准化的化工系统多工况故障检测。(6-7) Set the abnormal score of step (6-6)
Figure FDA00023859086300000610
Compare with the monitoring threshold η obtained in step (5), if
Figure FDA00023859086300000611
Then the current chemical system is in a normal operation state, return to step (6-1) to continue monitoring online real-time data, if
Figure FDA00023859086300000612
It indicates that a system fault has occurred in the current chemical system, and a fault warning is issued to realize the multi-condition fault detection of the chemical system based on local adaptive standardization.
CN202010098141.7A 2020-02-18 2020-02-18 Chemical system multi-working-condition fault detection method based on local adaptive standardization Active CN111367253B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010098141.7A CN111367253B (en) 2020-02-18 2020-02-18 Chemical system multi-working-condition fault detection method based on local adaptive standardization

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010098141.7A CN111367253B (en) 2020-02-18 2020-02-18 Chemical system multi-working-condition fault detection method based on local adaptive standardization

Publications (2)

Publication Number Publication Date
CN111367253A CN111367253A (en) 2020-07-03
CN111367253B true CN111367253B (en) 2021-03-16

Family

ID=71208052

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010098141.7A Active CN111367253B (en) 2020-02-18 2020-02-18 Chemical system multi-working-condition fault detection method based on local adaptive standardization

Country Status (1)

Country Link
CN (1) CN111367253B (en)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114692708B (en) * 2020-12-28 2025-07-22 江苏师范大学 Quality-related monitoring method for self-adaptive weighted semi-supervised probability CPLS
CN112857806B (en) * 2021-03-13 2022-05-31 宁波大学科学技术学院 Bearing fault detection method based on moving window time domain feature extraction
CN113254978B (en) * 2021-06-24 2021-09-21 国能大渡河大数据服务有限公司 Data security management system based on machine learning
CN115098330B (en) * 2022-06-27 2025-01-28 重庆大学 Device anomaly detection system and method based on cloud-edge collaboration mode
CN115688037A (en) * 2022-11-01 2023-02-03 中能融合智慧科技有限公司 New energy abnormal data processing method and system
CN115965915B (en) * 2022-11-01 2023-09-08 哈尔滨市科佳通用机电股份有限公司 Railway wagon connecting pull rod breaking fault identification method and system based on deep learning
CN119416107A (en) * 2024-10-17 2025-02-11 浙江工业大学 An anomaly detection method based on isomorphic contrastive learning of hierarchical topological graphs

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN100532892C (en) * 2005-01-25 2009-08-26 格特拉格传动机构和齿轮工厂赫尔曼·哈根迈尔有限公司&两合公司 Automatic multi-speed gearbox and method for controlling the gearbox for fault detection
CN103576594A (en) * 2013-11-11 2014-02-12 浙江工业大学 Intermittent process online monitoring method based on tensor overall-local preserving projection
CN103793601A (en) * 2014-01-20 2014-05-14 广东电网公司电力科学研究院 Turbine set online fault early warning method based on abnormality searching and combination forecasting
CN105955241A (en) * 2016-06-03 2016-09-21 北京科技大学 Quality fault locating method based on federated data driven production process
CN109299507A (en) * 2018-08-20 2019-02-01 浙江大学 A fault diagnosis and analysis method based on dynamic modeling for coal mills with different degrees under multiple working conditions
CN109726524A (en) * 2019-03-01 2019-05-07 哈尔滨理工大学 A prediction method for the remaining service life of rolling bearings based on CNN and LSTM

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH07295819A (en) * 1994-04-27 1995-11-10 Fuji Heavy Ind Ltd Fault diagnosis device
US6757665B1 (en) * 1999-09-28 2004-06-29 Rockwell Automation Technologies, Inc. Detection of pump cavitation/blockage and seal failure via current signature analysis
US9366451B2 (en) * 2010-12-24 2016-06-14 Commonwealth Scientific And Industrial Research Organisation System and method for the detection of faults in a multi-variable system utilizing both a model for normal operation and a model for faulty operation
CN105258947B (en) * 2015-11-06 2017-06-23 北京航空航天大学 A kind of Fault Diagnosis of Roller Bearings under operating mode disturbed conditions based on compressed sensing
WO2019028269A2 (en) * 2017-08-02 2019-02-07 Strong Force Iot Portfolio 2016, Llc Methods and systems for detection in an industrial internet of things data collection environment with large data sets
CN109555566B (en) * 2018-12-20 2021-04-20 西安交通大学 Steam turbine rotor fault diagnosis method based on LSTM
CN110210381B (en) * 2019-05-30 2023-08-25 盐城工学院 Domain separation self-adaptive one-dimensional convolutional neural network intelligent fault diagnosis method

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN100532892C (en) * 2005-01-25 2009-08-26 格特拉格传动机构和齿轮工厂赫尔曼·哈根迈尔有限公司&两合公司 Automatic multi-speed gearbox and method for controlling the gearbox for fault detection
CN103576594A (en) * 2013-11-11 2014-02-12 浙江工业大学 Intermittent process online monitoring method based on tensor overall-local preserving projection
CN103793601A (en) * 2014-01-20 2014-05-14 广东电网公司电力科学研究院 Turbine set online fault early warning method based on abnormality searching and combination forecasting
CN105955241A (en) * 2016-06-03 2016-09-21 北京科技大学 Quality fault locating method based on federated data driven production process
CN109299507A (en) * 2018-08-20 2019-02-01 浙江大学 A fault diagnosis and analysis method based on dynamic modeling for coal mills with different degrees under multiple working conditions
CN109726524A (en) * 2019-03-01 2019-05-07 哈尔滨理工大学 A prediction method for the remaining service life of rolling bearings based on CNN and LSTM

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
人工免疫系统在间歇化工过程故障诊断中的应用;戴一阳;《化工学报》;20090131;第60卷(第1期);第172-176页 *
基于Hopfield网络的时滞分析故障诊断策略;贺丁 等;《化工学报》;20130228;第64卷(第2期);第633-640页 *
基于深度学习的滚动轴承故障诊断方法研究;金余丰;《中国优秀硕士学位论文全文数据库 工程科技II辑》;20200215(第02期);第C029-157页 *

Also Published As

Publication number Publication date
CN111367253A (en) 2020-07-03

Similar Documents

Publication Publication Date Title
CN111367253B (en) Chemical system multi-working-condition fault detection method based on local adaptive standardization
CN109146246B (en) Fault detection method based on automatic encoder and Bayesian network
CN113298297B (en) A wind power output power prediction method based on isolated forest and WGAN network
CN115730191B (en) A coal mill fault early warning method based on attention mechanism
Li et al. False alarm reducing in PCA method for sensor fault detection in a nuclear power plant
CN110083593B (en) Power station operation parameter cleaning and repair method, repair system
CN109472097B (en) Fault diagnosis method for online monitoring equipment of power transmission line
CN111780800A (en) A sensor fault monitoring, isolation and reconstruction method and system
CN112132394A (en) A predictive state assessment method and system for a circulating water pump in a power plant
Xiu et al. Fault detection using structured joint sparse nonnegative matrix factorization
CN117688496B (en) Abnormality diagnosis method, system and equipment for satellite telemetry multidimensional time sequence data
CN112904810A (en) Process industry nonlinear process monitoring method based on effective feature selection
CN116878902A (en) Ship diesel engine fault early warning method and system based on deep learning CNN-BiLSTM
CN119067225B (en) Industrial control anomaly interpretation method and system based on difference of generated counterfactual samples
CN114995338A (en) Industrial process micro-fault detection method based on normative variable analysis and JS divergence fusion
CN119042194A (en) DCS fault prediction and diagnosis system for high-efficiency thermal power plant
CN114879612A (en) Blast furnace iron-making process monitoring method based on Local-DBKSSA
CN119310931A (en) Intelligent early warning system for DCS production process of thermal power plant based on artificial intelligence
CN119740203A (en) A regression prediction method for dynamic parameters of gas turbine based on fusion network
Liu et al. Graph attention Network-Based model for multiple fault detection and identification of sensors in nuclear power plant
CN112380763A (en) System and method for analyzing reliability of in-pile component based on data mining
Xu et al. Multiscale kernel entropy component analysis with application to complex industrial process monitoring
CN119807970A (en) Anomaly detection method for water pipeline operation status based on joint reconstruction and prediction model
CN114417704A (en) Wind turbine generator health assessment method based on improved stack type self-coding
Yang et al. A local Mahalanobis distance analysis based methodology for incipient fault diagnosis

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant