An early fault detection method for air separation units based on multi-state Gaussian process autoregression with multi-scale adaptive updating

Abstract

The present invention discloses an early fault detection method for air separation units based on multi-scale adaptively updated multi-state Gaussian process autoregression. According to the multi-scale time-varying characteristics existing in the air separation process, batch adaptive and sample adaptive update methods for fast and slow scales are adopted to perform multi-scale adaptive updates on the key history matrix. In the fault diagnosis stage, a multi-state Gaussian process autoregression fault detection model suitable for the fault diagnosis framework is designed in combination with the leave-one-out reconstruction strategy to solve the linear limitations of the traditional multi-state estimation-based fault diagnosis method, and finally the air separation unit fault detection results are obtained to assist in guiding the relevant diagnostic work on site. This method adopts a multi-scale adaptive update method to solve the time-varying problem of the air separation production process. Compared with the existing methods, the present invention has better accuracy and can achieve early warning of faults, and can provide fault diagnosis and troubleshooting guidance to operating personnel in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial fault diagnosis, and in particular to an air separation unit early fault detection method based on multi-scale adaptive updating multivariate state Gaussian process autoregression. Background Art

[0002] Air separation units are an indispensable component of the metallurgical, chemical, and petrochemical industries, and are closely intertwined with modern industry, particularly high-tech industries. Within the gas supply and demand systems of air separation pipelines, process parameter failures are a common fault. These often refer to equipment that slowly degrades or operates in complex and harsh environments, resulting in performance degradation over time. For example, in an air separation unit's oxygen-nitrogen distillation system, prolonged operation can lead to an increase in the nitrogen content of the argon fraction extracted from the upper column stripping section. This nitrogen accumulates in the crude argon condenser, reducing the effective heat exchange area and heat exchange temperature difference, lowering the heat load and resulting in a decrease in argon extraction, a "nitrogen concentration" failure. Due to the lag in human judgment, remedial measures for these types of failures often take place only after serious consequences have occurred, resulting in significant unnecessary economic losses. Therefore, early and accurate fault diagnosis is crucial. This means identifying potential fault signs before they occur, providing operators with ample time to troubleshoot and address the equipment's precursors, thus preventing further deterioration and resulting in more serious consequences and losses.

[0003] However, in actual industrial production, the system characteristics of air separation units (ASUs) are not static. Due to factors such as variable loads, widely varying operating conditions, brand switching, and equipment aging, ASU system characteristics fluctuate significantly. Taking the steel industry as an example, these system characteristics can be categorized into two main scales: fast-frequency and slow-frequency. Slow-frequency variations occur when long-term operation, equipment degradation, and aging cause system characteristics to change slowly. Fast-frequency transient variations occur when brand switching and downstream user demand shift, forcing the ASU to continuously perform load switching operations, resulting in system instability at a static operating point. Due to the instantaneous nature of these switching operations, ASU equipment characteristics can change rapidly within a short period of time. Furthermore, the complex air separation process, involving a large number of equipment, exhibits strong nonlinear characteristics. These factors result in poor performance when applying traditional fault diagnosis methods to ASU fault detection, making it difficult to achieve fast, accurate, and effective early fault detection. Therefore, comprehensively considering the time-varying system characteristics and strong nonlinearity of ASUs to achieve accurate early fault detection has become a pressing need in the ASU industry. Summary of the Invention

[0004] The purpose of the present invention is to address the deficiencies of the prior art and provide an air separation unit early fault detection method based on multi-scale adaptively updated multivariate state Gaussian process autoregression.

[0005] The object of the present invention is achieved through the following technical solutions:

[0006] A method for detecting early faults in an air separation unit based on multi-state Gaussian process autoregression with multi-scale adaptive updating comprises the following steps:

[0007] S1. Obtain historical normal sample data of the air separation unit at different times, and form a historical matrix D representing the normal working status of the air separation unit system after data preprocessing;

[0008] S2. Use the multi-objective discrete Archimedean optimization algorithm to perform batch adaptive update on the history matrix, transform the update problem of the history matrix into a sample selection problem, and obtain the updated optimized history matrix D opt ;

[0009] S3, using the similarity measurement method of real-time learning to optimize the history matrix D obtained in S2 opt Perform sample adaptive update to form the local history matrix D that is ultimately used for fault diagnosis local ;

[0010] S4. In the fault detection stage, a leave-one-sample reconstruction strategy is adopted to design a multivariate state Gaussian process autoregressive early fault detection model suitable for air separation units, and online early fault diagnosis of air separation units is performed.

[0011] Preferably, the specific steps of S1 are as follows:

[0012] Historical data describing the normal operating status of the air separation unit system is read from a real-time database at the industrial site. The read normal historical data is preprocessed and stored in a historical matrix D. The data preprocessing includes filling in missing data and normalizing the data.

[0013]

[0014] Where: z i (i=1,2,3,…,n) is t i Multivariate sample data at time z j,i (i=1,2,3,…,n;j=1,2,3,…,M) is t i Multivariate sample data z at time i The value of the variable j in .

[0015] Preferably, the specific steps of S2 are as follows:

[0016] S21: Design three important indicators to measure the quality of the data matrix: η(x), λ(x), and ρ(x). η(x) calculates the sum of the diagonal elements of the variance matrix of the historical matrix, representing the diversity of the samples in the historical matrix; λ(x) calculates the L2 difference, representing the uniformity of the sample distribution; ρ(x) calculates the root mean square of the Pearson coefficient, representing the degree of redundancy of the samples. The calculation methods of the three indicators are as follows:

[0017]

[0018] Where: C is the constant used for regularization, n represents the number of samples in the history matrix, x is the sample data, p is the sample dimension, x k,i (k=1,2,3,…,n;i=1,2,3,…,p) is the value of the i-th variable of the k-th sample, x l,i (l=1,2,3,…,n;i=1,2,3,…,p) is the value of the i-th variable of the l-th sample, and denote the mean values ​​of the i-th and j-th variables respectively;

[0019] S22: Based on the metrics of the historical matrix, a multi-objective discrete Archimedean optimization algorithm is used to construct the optimization proposition for batch adaptive updating of the historical matrix according to the following formula:

[0020] min(f1(x),f2(x),f3(x))

[0021] stx=D(p)

[0022] Where: D(p) represents the sample selected from the historical matrix. And f1(x) = -η(x), f2(x) = λ(x), f3(x) = ρ(x).

[0023] S23: Whether the samples in the historical matrix are selected as optimization variables, by optimizing the above-mentioned minimization optimization proposition, we can obtain the historical matrix sample distribution that makes the three indicators η(x), λ(x) and ρ(x) optimal at the same time, forming the optimized historical matrix D opt .

[0024] Furthermore, the specific steps of S3 are as follows:

[0025] S31: According to the current sample to be detected z obs , using the instant learning strategy to measure the current sample z to be detected obs The history matrix D after batch adaptive update opt The similarity of all samples in is used to realize the adaptive update of historical matrix samples and construct the following measurement method:

[0026]

[0027] Among them, α is the weight parameter; D M_i and D cos_i Represent the samples to be tested z obs and the i-th historical sample z in the historical matrix i The Mahalanobis distance and cosine distance between them; Σ is the mutual covariance matrix; According to the above formula, D cos_i If it is negative, the corresponding sample will not be selected;

[0028] S32: To S i The values ​​are sorted in descending order and the top K largest S are selected i The most relevant historical sample corresponding to the value, constructing the final local history matrix D local .

[0029] Furthermore, the specific steps of S4 are as follows:

[0030] S41: Using a leave-one-sample reconstruction strategy, a sample set for establishing a multivariate state Gaussian process autoregressive fault detection model is reconstructed from the local history matrix;

[0031] S42: For the m-th output channel of the multivariate Gaussian process autoregressive fault detection model, each sample point in the m-th row of the local history matrix is ​​selected in turn as the training output of the multivariate Gaussian process autoregressive fault detection model. At the same time, the remaining sample points selected in each round are constructed as the training input of the multivariate Gaussian process autoregressive fault detection model. The final training sample set for the m-th channel is as follows:

[0032]

[0033] Among them, X m is the training input of the mth channel, y m is the training output of the mth channel, (x m,i y m,i ) is the i-th training sample pair of the m-th channel.

[0034] S43: Selection from the history matrix and the current sample to be tested z obs The most relevant n-1 sample points (m=1,2,3,…,M) constitutes the estimated input of the multivariate state Gaussian process autoregressive fault detection model

[0035] S44: The average modeling error during the training process is The following formula represents the error ε of the i-th training sample i and the average modeling error All obey Gaussian distribution:

[0036]

[0037] Among them, mean(ε i ) represents the error ε of the i-th training sample i The mean of cov(ε i ) represents the error ε of the i-th training sample i The variance of represents the estimated mean of the training sample, y is the label value of the training sample, is the estimated variance of the training samples, These are three matrices designed for mathematical derivation;

[0038] S45: Sample z to be tested obs The estimated value of The Gaussian distribution of Estimated mean and the estimated covariance Σ * The calculation is as follows:

[0039]

[0040] in, represents the diagonal noise matrix, There are elements m,m'=1,2,...,M, With matrix blocks Symmetric block matrix With matrix blocks

[0041] S46: Based on KL divergence, a multivariate probabilistic fault diagnosis index is proposed to calculate the difference between the average modeling error distribution and the estimated error distribution to achieve fault diagnosis. If the KL divergence index corresponding to the sample to be tested is higher than the control line KL limit , then an alarm is issued. The KL divergence indicator is calculated as follows:

[0042]

[0043] Among them, Σ model andΣ est Represents the average modeling error and the estimation error ε(X * ) variance, μ model and μ est Represent the mean vectors of the two errors respectively.

[0044] Furthermore, in the multivariate state Gaussian process autoregressive air separation unit early fault detection method in S4, during the fault online detection phase, for each sample to be detected z obs, both require real-time updating of the multivariate state Gaussian process autoregressive fault detection model for fault diagnosis.

[0045] Compared with the prior art, the present invention has the following innovative advantages and significant effects:

[0046] (1) Aiming at the time-varying problems of fast-frequency and slow-frequency system characteristics in the production process of air separation units, the present invention adopts a multi-scale adaptive update multivariate state Gaussian process autoregressive early fault detection method for air separation units, which adopts a multi-objective discrete Archimedean optimization algorithm and a real-time learning strategy to perform multi-scale batch adaptation and sample adaptive update on the historical matrix used to describe the normal working state of the air separation unit, thereby solving the influence of the time-varying problem of system characteristics on the accuracy of early fault detection of air separation units.

[0047] (2) The early fault detection method of the multi-scale adaptively updated multivariate state Gaussian process autoregressive method for air separation units of the present invention can take into account the strong nonlinearity of the air separation unit process and process noise interference, has high diagnostic accuracy, can be implemented online, and well meets the needs of the production site.

[0048] (3) In order to solve the problem of real-time detection required for on-site implementation, the early fault detection method for air separation units based on multi-scale adaptive updating of multivariate state Gaussian process autoregression of the present invention does not require a large number of fault samples for offline training, and can be directly deployed to realize online real-time diagnosis, which can better meet the complex requirements of fault diagnosis of large air separation units. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The objects, features, and advantages of the present invention can be further understood by describing preferred embodiments of the present invention in conjunction with the accompanying drawings. The present invention will be described in more detail below with reference to the accompanying drawings. However, the present invention may be implemented in many different forms and should not be considered limited to the embodiments set forth in the specification. Rather, such embodiments are provided to illustrate the implementation and completeness of the present invention and to describe the specific implementation process of the present invention to those skilled in the art.

[0050] Figure 1 This is a schematic diagram of the production process of the air separation unit in a steel plant.

[0051] In the figure, 1-air compressor; 2-air filter; 3-air cooling tower; 4-water cooling tower; 5-molecular sieve adsorber; 6-electric heater; 7-booster; 8-main heat exchanger; 9-turbine expander; 10-liquid oxygen pump; 11-upper tower; 12-lower tower; 13-main cooling; 14-subcooling tower; 15-crude argon I tower; 16-circulating crude argon pump; 17-refined argon condenser; 18-refined argon tower.

[0052] Figure 2 It is a schematic diagram of the implementation process of the present invention.

[0053] Figure 3 Flowchart of the multi-objective discrete Archimedean optimization algorithm.

[0054] Figure 4 Schematic diagram of the leave-one-sample strategy.

[0055] Figure 5 Schematic diagram of the multivariate state Gaussian process autoregressive fault diagnosis model. DETAILED DESCRIPTION

[0056] In order to better understand the technical solutions and specific implementations of the present invention, the implementations of the present invention are further described below with reference to the accompanying drawings.

[0057] Taking the steel industry as an example, due to variable operating conditions, changes in downstream demand, brand switching, and performance degradation of equipment over a long period of time, the system characteristics of the air separation unit have fast-frequency and slow-frequency time-varying problems. In addition, due to the complexity of the air separation process, the production process of the air separation unit has strong nonlinear characteristics. In response to the above situation, the present invention constructs an early fault detection method based on multi-scale adaptive updating of multivariate state Gaussian process autoregression. Figure 1 As can be seen from the schematic diagram of the air separation unit shown, the air separation unit consists of an air pre-cooling system, an argon distillation system, an oxygen and nitrogen distillation system, a water circulation system, and a molecular sieve system. The main equipment involved are: air compressors, boosters, turbo expanders, oxygen compressors, and nitrogen compressors; the main devices involved in the oxygen storage system are: oxygen pipe networks and spherical tanks, liquid oxygen storage tanks, liquefiers, and vaporizers; the main unit operations included are: air compression, purification and separation of gaseous and liquid oxygen, nitrogen, and argon required. Due to the complex actual configuration of the equipment, which is spread across various production areas of the factory, and the nonlinearity and large time lag of the production process, accurate and effective mechanism modeling is particularly difficult. In view of this, the present invention solves the related fault detection problems from the perspective of data-driven methods.

[0058] The overall framework of the air separation unit early fault detection method based on multi-scale adaptive update of multivariate state Gaussian process autoregression proposed in this invention is as follows: Figure 2 The specific implementation steps are as follows:

[0059] Step 1: Data preparation and data preprocessing

[0060] Normal operating data for an air separation unit over a period of time is extracted from a real-time relational database at an industrial site and stored in a historical matrix D. This matrix contains historical normal operating data for the air separation unit at different times. The extracted data may have quality issues, so preprocessing is required. Missing points in the data are detected and recorded, and filled with the average of the points before and after the missing point. The data is then normalized to its mean / variance.

[0061] Step 2: Batch adaptive update of history matrix

[0062] The quality of the history matrix directly impacts fault diagnosis performance. However, the production process characteristics of an air separation unit can change over time, for example, as the equipment operates over extended periods of time. Therefore, using a static history matrix for fault diagnosis will reduce the accuracy of fault diagnosis due to these slow, long-term system characteristic changes. Therefore, it is necessary to adaptively update the history matrix D at regular intervals. Specifically:

[0063] Step 2.1: Design three important indicators to measure the quality of the data matrix: η(x), λ(x), and ρ(x). η(x) calculates the sum of the diagonal elements of the variance matrix of the historical matrix, representing the diversity of the samples in the historical matrix; λ(x) calculates the L2 difference, representing the uniformity of the sample distribution; ρ(x) calculates the root mean square of the Pearson coefficient, representing the degree of redundancy of the samples. The calculation methods of the three indicators are as follows:

[0064]

[0065] Where: C is the constant used for regularization, n represents the number of samples in the history matrix, x is the sample data, p is the sample dimension, x k,i (k=1,2,3,…,n;i=1,2,3,…,p) is the value of the i-th variable of the k-th sample, x l,i (l=1,2,3,…,n;i=1,2,3,…,p) is the value of the i-th variable of the l-th sample, and denote the mean values ​​of the i-th and j-th variables respectively;

[0066] Step 2.2: Based on the above indicators, the multi-objective discrete Archimedean optimization algorithm is used to construct the optimization proposition for batch adaptive updating of the history matrix according to the following formula:

[0067] min(f1(x),f2(x),f3(x))

[0068] stx=D(p)

[0069] Where: D(p) represents the sample selected from the historical matrix. And f1(x) = -η(x), f2(x) = λ(x), f3(x) = ρ(x).

[0070] Step 2.3: If Figure 3 The optimization process shown is as follows:

[0071] Step 2.3.1: Individual initialization, initial position is p i , density den, velocity vol, acceleration acc, the number of individuals is m, and p ij =[p i1 ,p i2 ,p i3 ,…,p in ],i=1,2,…,m,p ij ∈{0,1},p ij =0 means the jth sample is selected, p ij =1 means that the jth sample is not selected, and n is the total number of samples to be selected.

[0072] Step 2.3.2: Update individual attributes. Update individual position, density, volume and acceleration in sequence according to the following formula:

[0073]

[0074]

[0075]

[0076]

[0077]

[0078]

[0079]

[0080]

[0081] P=2×rand-C4

[0082] T=C3×TF

[0083] in, and are the i-th individual and the best individual in the k-th generation respectively; C1, C2, C3, and C4 are artificially set hyperparameters; rand is a random number between 0 and 1; TF is the transfer operator; u and l are regularization variables; t is the current optimization generation; t max is the maximum optimization algebra; d is the density operator; denmr ,vol mr ,acc mr They represent the density, velocity and acceleration of the random individual mr respectively; den best ,vol best ,acc best represent the density, velocity and acceleration of the optimal individual respectively; represent the density, velocity, and acceleration of the i-th individual in the k-th generation respectively;

[0084] Step 2.3.3: Update the optimal individual and calculate the individual dominance relationship. If the current individual p i Dominate the most optimal individual p best That is, p best <p i , then update the optimal individual.

[0085] Step 2.3.4: Algorithm stopping criteria: when the maximum number of iterations or the minimum error standard is not reached, go to step 2.3.2 and repeat the procedure; otherwise, go to step 2.3.5.

[0086] Step 2.3.5: Individual transformation, transform the optimal individual obtained by optimization into the final sample screening situation, and obtain the optimized history matrix D opt .

[0087] Step 3: Adaptive update of historical matrix samples:

[0088] In addition to slow-frequency time-varying characteristics, the production process of air separation units also presents fast-frequency time-varying system characteristics, such as transient load changes and product brand switching. To address this, a sample-adaptive historical matrix update method based on real-time learning is employed. The historical samples that best match the current system state are selected as the data set for the fault detection model. The specific steps are as follows:

[0089] Step 3.1: According to the current sample z to be detected obs , using the instant learning strategy to measure the current sample z to be detected obs The history matrix D after batch adaptive update opt The similarity of all samples in is used to achieve sample adaptive update and construct the following measurement method:

[0090]

[0091] Among them, α is the weight parameter; D M_i and D cos_i Represent the samples to be tested z obs and the i-th historical sample z in the historical matrix i The Mahalanobis distance and cosine distance between them; Σ is the mutual covariance matrix; According to the above formula, D cos_iIf it is negative, the corresponding historical sample will not be selected;

[0092] Step 3.2: S i Sort the values ​​in descending order and select the first K S i The most relevant historical sample corresponding to the value, constructing the final local history matrix D local Among them, K is a preset value, which can generally be 10-50.

[0093] Step 4: Construct a multivariate state Gaussian process autoregressive fault detection model in the fault detection stage:

[0094] Step 4.1: Use Figure 4 The leave-one-sample reconstruction strategy shown in the figure reconstructs the sample set used to build the multivariate Gaussian process autoregression fault detection model from the local history matrix. That is, for the m-th output channel of the multivariate Gaussian process autoregression fault detection model, each sample point in the m-th row of the local history matrix is ​​selected in turn as the training output of the multivariate Gaussian process autoregression fault detection model. At the same time, the remaining sample points in each round of selection are constructed as the training input of the multivariate Gaussian process autoregression fault detection model. The final training sample set for the m-th channel is as follows:

[0095]

[0096] Among them, X m is the training input of the mth channel, y m is the training output of the mth channel, (x m,i y m,i ) is the i-th training sample pair of the m-th channel.

[0097] Step 4.2: Select the sample z from the history matrix obs The most relevant n-1 sample points Estimated inputs for the multivariate state Gaussian process autoregressive fault detection model

[0098] Step 4.3: The average modeling error during the training process is The following formula represents the error ε of the i-th training sample i and the average modeling error All obey Gaussian distribution:

[0099]

[0100] Among them, mean(ε i ) represents the error ε of the i-th training sample i The mean of cov(ε i ) represents the error ε of the i-th training sample iThe variance of represents the estimated mean of the training sample, y is the label value of the training sample, is the estimated variance of the training samples, These are three matrices designed for mathematical derivation;

[0101] Step 4.4: Sample z to be tested obs The estimated value of The Gaussian distribution of Estimated mean and the estimated covariance Σ * The calculation is as follows:

[0102]

[0103] in, represents the diagonal noise matrix, There are elements m,m'=1,2,...,M, With matrix blocks Symmetric block matrix With matrix blocks

[0104] Step 5: Propose a multivariate probabilistic fault detection indicator based on KL divergence for fault diagnosis:

[0105] Step 5.1: Calculate the difference between the average modeling error distribution and the estimated error distribution to achieve fault diagnosis. The KL divergence index corresponding to the sample to be tested is higher than the control line KL limit , then an alarm is issued. The KL divergence indicator is calculated as follows:

[0106]

[0107] Among them, Σ model andΣ est Represents the average modeling error and the estimation error ε(X * ) variance, μ model and μ est Represent the mean vectors of the two errors respectively.

[0108] The prediction methods of S1 to S5 above are applied to a specific example to demonstrate their technical effects.

[0109] In this example, we use the air separation unit (ASU) of a steel plant with a high level of informatization in the steel industry as an example. Six types of ASU faults are shown, as detailed in Table 1. These faults include five types of equipment-level faults and one type of process-level faults. The ASU data sampling period for this steel plant is 30 seconds. The KPCA method (Lee, JM, Yoo, CK, Choi, SW, Vanrolleghem, PA, Lee, IB, 2004. Nonlinear process monitoring using kernel principal component analysis. Chem. Eng. Sci., 59, 223-234) and the JITL-PPCA method (Wang, ZY, Liang, J., 2018. A JITL-Based ProbabilisticPrincipal Component) were implemented respectively. Analysis for Online Monitoring of NonlinearProcesses. Journal Of Chemical Engineering of Japan., 51(10), 874-889), MMO-MSET method (Lv, You., Fang, F., Yang, TT, Carlos, ER, 2020. An early fault detection method for induced draft fans based on MSET with informative memory matrixselection.ISATrans.,102,325-334), EL-MSET method (Liu, CL, Wang, ZQ, 2020. Faultearly warning of wind turbine gearbox based on MSET and ensemblelearning. Acta Energiae Solaris Sinica.41(11):228-233) and the method of the present invention, selecting FDR and FAR as two evaluation indicators, and comparing the prediction results of the five methods. The fault diagnosis accuracy results are shown in Table 2. The fault early warning effect is shown in Table 3.

[0110]

[0111]

[0112] Among them, SN() represents the number of samples that meet the conditions in the brackets, SN(z=0) represents the number of normal samples, and SN(z=1) represents the number of faulty samples. In addition, SN(KL>KL limit |z=1) represents the successfully identified fault sample, SN(KL>KL limit |z=0) The number of normal samples that are misclassified as faulty samples.

[0113] Table 1 Detailed description of six types of faults in air separation units

[0114]

[0115] Table 2 Comparison of fault diagnosis performance

[0116]

[0117]

[0118] Table 3 Results of early warning of air separation unit failure

[0119]

[0120] Thus, the method adopted by the present invention can overcome the problems of multi-time variation, strong nonlinearity, and process noise interference in the air separation production process through the adaptive updating of the multivariate state Gaussian process autoregressive model and the multi-scale history matrix, thereby realizing high-precision early fault diagnosis of the air separation unit. In addition, the general sampling period of the air separation unit is 30 seconds, and the real-time diagnosis time of this method does not exceed 5 seconds, which fully meets the application of on-site fault diagnosis of the air separation unit. The principle of this method is simple and clear, easy to implement on a computer, and has good flexibility. It does not require offline model training and can be directly deployed in the air separation unit, which well meets the requirements of the air separation unit for real-time fault detection, such as rapidity and safety.

[0121] The embodiment described above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Persons skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of protection of the present invention.

Claims

1. A method for early fault detection of air separation units based on multi-scale adaptive updating multivariate state Gaussian process autoregression, characterized in that: The steps include: S1. Obtain historical normal sample data at different times in the air separation unit, and form a historical matrix D representing the normal working status of the air separation unit after data preprocessing; S2. Use the multi-objective discrete Archimedean optimization algorithm to perform batch adaptive update on the history matrix, transform the update problem of the history matrix into a sample selection problem, and obtain the updated optimized history matrix D opt ; The S2 comprises the following steps: S21: Design three indicators η(x), λ(x) and ρ(x) to measure the quality of the data matrix. Among them, η(x) calculates the sum of the diagonal elements of the variance matrix of the historical matrix, which represents the diversity of samples in the historical matrix; λ(x) calculates the L2 difference, which represents the uniformity of the sample distribution; ρ(x) calculates the root mean square of the Pearson coefficient, which represents the redundancy of the sample. The calculation methods of the three indicators are as follows: Where: C is the constant used for regularization, n represents the number of samples in the history matrix, x is the sample data, p is the sample dimension, x k,i is the value of the i-th variable of the k-th sample, k = 1, 2, 3, ..., n; i = 1, 2, 3, ..., p, x l,i is the value of the i-th variable of the l-th sample, l=1,2,3,…,n; i=1,2,3,…,p, and denote the mean values ​​of the i-th and j-th variables respectively; S22: Based on the three metrics of the historical matrix, the multi-objective discrete Archimedean optimization algorithm is used to construct the optimization proposition for batch adaptive updating of the historical matrix according to the following formula: min(f1(x),f2(x),f3(x)) stx=D(p) Where: D(p) represents the sample selected from the historical matrix, and f1(x) = -η(x), f2(x) = λ(x), f3(x) = ρ(x); S23: Whether the samples in the historical matrix are selected as optimization variables, by optimizing the above-mentioned minimization optimization proposition, the historical matrix sample distribution that makes the three indicators η(x), λ(x) and ρ(x) optimal at the same time is obtained, forming the optimized historical matrix D opt ; S3, then uses the similarity measurement method of instant learning to optimize the history matrix D obtained in S2 opt Perform sample adaptive update to form the local history matrix D that is ultimately used for fault diagnosis local ; The specific approach of S3 is as follows: S31: According to the current sample to be detected z obs , using the instant learning strategy to measure the current sample z to be detected obs The history matrix D after batch adaptive update opt The similarity of all samples in is used to adaptively update the history matrix of samples and construct the following measurement method: Among them, α is the weight parameter; D M_i and D cos_i Represent the samples to be tested z obs and the i-th historical sample z in the historical matrix i The Mahalanobis distance and cosine distance between them; Σ is the mutual covariance matrix; According to the above formula, D cos_i If it is negative, the corresponding historical sample will not be selected; S32: To S i The values ​​are sorted in descending order and the top K largest S are selected i The most relevant historical sample corresponding to the value, constructing the final local history matrix D local , where K is the preset value; S4. In the fault detection stage, a leave-one-sample reconstruction strategy is adopted to design an early fault detection model of multivariate state Gaussian process autoregression suitable for air separation units, and online early fault diagnosis of air separation units is performed.

2. The method for early fault detection of air separation units based on multi-scale adaptive updating multivariate state Gaussian process autoregression according to claim 1 is characterized in that: The specific steps of S1 are as follows: Read historical data describing the normal operating status of the air separation unit system from a real-time database at the industrial site, preprocess the read normal historical data, and store them in a historical matrix D. The data preprocessing includes filling in missing data and normalizing the data. Where: z i t i Multivariate sample data at time z j,i t i Multivariate sample data z at time i The value of the variable m is i=1,2,3,…,n; m=1,2,3,…,M.

3. The method for early fault detection of air separation unit based on multi-scale adaptive updating multivariate state Gaussian process autoregression according to claim 1 is characterized in that: The specific steps of S4 are as follows: S41: Using the leave-one-sample reconstruction strategy, we can get the local history matrix D local Reconstructing the sample set used to establish the multivariate state Gaussian process autoregressive fault detection model; S42: For the m-th output channel of the multivariate Gaussian process autoregressive fault detection model, each sample point in the m-th row of the local history matrix is ​​selected in turn as the training output of the multivariate Gaussian process autoregressive fault detection model. At the same time, the remaining sample points after each round of selection are constructed as the training input of the multivariate Gaussian process autoregressive fault detection model. The final training sample set of the m-th channel is as follows: Among them, X m is the training input of the mth channel, y m is the training output of the mth channel, (x m,i y m,i ) is the i-th training sample pair of the m-th channel; S43: Select the sample z from the history matrix that is currently to be tested obs The most relevant n-1 sample points And finally constitute the estimated input of the multivariate state Gaussian process autoregressive fault detection model S44: The average modeling error during the training process is The following formula represents the error ε of the i-th training sample i and the average modeling error All obey Gaussian distribution: Among them, mean(ε i ) represents the error ε of the i-th training sample i The mean of cov(ε i ) represents the error ε of the i-th training sample i The variance of represents the estimated mean of the training sample, y is the label value of the training sample, is the estimated variance of the training samples, S45: Sample z to be tested obs The estimated value of The Gaussian distribution of Estimated mean and the estimated covariance Σ * The calculation is as follows: in, represents the diagonal noise matrix, There are elements With matrix blocks Symmetric block matrix With matrix blocks S46: Calculate the difference between the average modeling error distribution and the estimated error distribution to achieve fault diagnosis; when the KL divergence index corresponding to the sample to be tested is higher than the control line KL limit , then an alarm is issued; the KL divergence indicator is calculated as follows: Among them, Σ model andΣ est Represents the average modeling error and the estimation error ε(X * ) variance, μ model and μ est Represent the mean vectors of the two errors respectively.

4. The method for early fault detection of air separation units based on multi-scale adaptive updating multivariate state Gaussian process autoregression according to claim 3 is characterized in that: In the fault online detection phase, for each sample z to be detected obs , both require real-time updating of the multivariate state Gaussian process autoregressive fault detection model for fault diagnosis.

Images