A multi-working-condition intermittent process fault monitoring method

Abstract

The application discloses a kind of multi-working-condition intermittent process fault monitoring method, first according to process feature, stage division is carried out, then the same stage data is clustered according to operation using k-means method, due to the randomness, non-directionality of operation, the same class data also obeys Gaussian distribution, therefore, data NGS method is proposed to process data, so that it obeys Gaussian distribution, while considering the nonlinearity in the process, the corresponding class of kernel principal component analysis model is constructed, when online monitoring, first select the most suitable model from the multiple models constructed, calculate the corresponding statistics and its corresponding control line, determine whether there is fault by determining whether it is out of limit. The application has the characteristics of reducing the complexity of calculation, speeding up the fault monitoring speed and improving the accuracy of monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rescue equipment, more particularly, the present application relates to a multi-condition intermittent process fault monitoring method. BACKGROUND

[0002] The rapid development of science and technology makes modern industrial processes gradually tend to be automated, diversified and complex, and the safety of industrial production processes and the reliable product quality have been widely watched by all sectors of society. If a fault occurs in the industrial production process, not only will it waste enterprise resources and cause property losses, but also it may even cause casualties and have a negative impact on society, so fault monitoring and diagnosis technology is increasingly valued by researchers.

[0003] In industry, intermittent processes occupy a dominant position, and various processes in production change with time, how to ensure the stability, continuity and safety of production in intermittent industrial processes is the focus of current research. For some complex industrial processes, such as polymerization process, steelmaking process of refining furnace (LF), continuous stirred tank reactor system (CSTR), etc., they have nonlinear and multi-operation characteristics, and in batch processes under different operating conditions, data often have randomness due to manual operation and do not conform to Gaussian distribution, and each batch process runs under different operating conditions, due to the complexity of the operation sequence, the operation time is obviously not equal.

[0004] For intermittent processes with multiple conditions, dynamics, nonlinearity and non-Gaussian characteristics, the principal component analysis method (PCA) in the traditional monitoring method is only applicable to linear and Gaussian distribution, and due to its failure to consider local intrinsic characteristics, there is a certain false alarm. The global-local analysis method considers both global and local features, and can highly preserve the spatial feature structure in the original data, but it is not suitable when the process data is nonlinear. In recent years, the introduction of kernel method can solve the nonlinearity problem, but when the data volume is large, the computational complexity will increase. SUMMARY

[0005] The purpose of the present application is to design and develop a multi-condition intermittent process fault monitoring method, which reduces the computational complexity and improves the monitoring efficiency and accuracy through the nearest neighbor group standardized kernel principal component analysis model.

[0006] The technical scheme provided by the present application is as follows:

[0007] A multi-condition intermittent process fault monitoring method, comprising the following steps:

[0008] Step one, collecting data of I normal batches Divide them according to the operation stage, and cluster the batch data of the same operation in each stage;

[0009] wherein c s is the number of classes for each stage, and c = 1, 2,... s , is the number of batches belonging to the same class, M is the number of process variables, and N i is the reaction time of each batch in the same class;

[0010] Step two, expanding the data of the same class into two-dimensional matrix data

[0011] Step three, finding the class composed of the k points with the nearest distance to each sample in the two-dimensional matrix data, and converting the data belonging to the same class in the two-dimensional matrix data based on the nearest neighbor group standardization method:

[0012]

[0013] wherein x mn is a sample in the two-dimensional matrix data, is the class composed of the k points with the nearest distance to the sample x mn , is the average vector of , is the standard deviation vector of ;

[0014] Step four, constructing the kernel principal component analysis model of each class;

[0015] Step five, calculating the control limit of the T 2 statistic of each class and the control limit of the SPE statistic of each class;

[0016] wherein the control limit of the T 2 statistic of each class satisfies:

[0017]

[0018] wherein is the control limit of the statistic T 2 , q is the number of principal components, N is the number of samples, F q,N-q,α is the critical value of F distribution corresponding to the kernel check level α and the degrees of freedom q, N-q;

[0019] the control limit of the SPE statistic of each class satisfies:

[0020]

[0021] wherein SPE α is the control limit of the statistic SPE, θ i is the second intermediate variable, h0 is the first intermediate variable, and i = 1, 2, 3;

[0022] Step six, collecting online monitoring sample data x test ∈R 1×M , according to the operation to determine the stage of the online monitoring sample data and select the kernel principal component analysis model of the class with the shortest distance between the class center in the same stage;

[0023] Step seven, normalizing the online monitoring data;

[0024] Step eight, extracting the principal component information of x test using the offline model of the class to which it belongs, and calculating the statistics T 2 and SPE;

[0025] Step nine, comparing the statistics T 2 and SPE with the control limit of the T 2 statistic and the control limit of the SPE statistic, respectively, in the monitoring process, if the statistics T 2 or SPE continuously for three values exceed the control limit of the T 2 statistic and / or the control limit of the SPE statistic, it is determined that the batch is abnormal.

[0026] Preferably, the k-means clustering algorithm is used for clustering in step one.

[0027] Preferably, the variable unfolding method is used to unfold the data of the same class into a two-dimensional matrix data in step two.

[0028] Preferably, the mean vector satisfies:

[0029]

[0030] In the formula, E(X) is the mean vector of data X, N is the number of samples, and x n is the nth sample in data X.

[0031] Preferably, the standard deviation vector satisfies:

[0032]

[0033] In the formula, S(X) is the standard deviation vector of data X.

[0034] Preferably, the kernel function of the kernel principal component analysis model satisfies:

[0035] K ij = K(x i , x j ) = <Φ(x i ), Φ(x j )>, (i, j = 1, 2, …, M)

[0036] wherein K ij is a kernel matrix, K(x i , x j ) is a kernel function, Φ(x i ) is the i-th element of the kernel matrix, and Φ(x i ) is the j-th element of the kernel matrix. j j is a mapping value in a high-dimensional space.

[0037] Preferably, the statistic T 2 satisfies:

[0038] T 2 = [t1, t2, …, t q ]Λ -1 [t1, t2, …, t q ] T

[0039] wherein t k is the score of the k-th principal component, and k = 1, 2, …, q.

[0040] Preferably, the statistic SPE satisfies:

[0041]

[0042] wherein is a reconstructed feature vector, Φ(x) is a feature vector before reconstruction, and n is the number of non-zero eigenvalues.

[0043] Preferably, the number of principal components satisfies:

[0044]

[0045]

[0046] wherein η i is the contribution rate corresponding to each principal component, is a cumulative variance contribution rate, and λ i is the i-th eigenvalue.

[0047] Preferably, the seventh step is the same as the third step.

[0048] The beneficial effects of the present application are:

[0049] ​This invention presents a fault monitoring method for intermittent processes under multiple operating conditions. By standardizing non-Gaussian data using nearest neighbor groups, it transforms non-Gaussian data into data that follows a Gaussian distribution. Simultaneously, it constructs a kernel principal component analysis model for the corresponding clusters to address the nonlinearity of the data and reduce computational complexity. Finally, by limiting statistical quantities, it accelerates fault monitoring speed and improves monitoring accuracy. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the multi-condition intermittent process fault monitoring method described in this invention.

[0051] Figure 2 This is a schematic diagram illustrating the stage division of a batch according to the present invention.

[0052] Figure 3 This is a schematic diagram of the same-stage clustering process for multiple batches as described in this invention.

[0053] Figure 4a , 4b This is a normal distribution fitting diagram for the two working modes described in this invention, based on Z-normalization and NGS-normalization.

[0054] Figure 5a , 5b 5c represents the confidence plots of the data with outlier values ​​described in this invention after unstandardization, Z-standardization, and NGS standardization.

[0055] Figure 6a , 6b This is a normal probability plot of the key variables (4 and 13) in the embodiments of the present invention after Z-standardization and NGS-standardization.

[0056] Figure 7a , 7b The HotellingT is the first-stage normal batch first-class clustering model in the embodiments of the present invention. 2 Monitoring charts for SPE statistics.

[0057] Figure 8a , 8b The HotellingT is the second type of clustering model for the first stage normal batch in the embodiments of the present invention. 2 Monitoring charts for SPE statistics.

[0058] Figure 9a , 9b The HotellingT is the third type clustering model of the first stage normal batch in the embodiments of the present invention. 2 Monitoring charts for SPE statistics.

[0059] Figure 10a , 10bThe fault monitoring result of the normal batch under MKPCA in the embodiment of the present application.

[0060] Figure 11a 、 11b The fault monitoring result of the normal batch under NGS-KPCA in the embodiment of the present application.

[0061] Figure 12a 、 12b The fault monitoring result of the fault 1 batch under MKPCA in the embodiment of the present application.

[0062] Figure 13a 、 13b The fault monitoring result of the fault 1 batch under NGS-KPCA in the embodiment of the present application.

[0063] Figure 14a 、 14b The fault monitoring result of the fault 2 batch under MKPCA in the embodiment of the present application.

[0064] Figure 15a 、 15b The fault monitoring result of the fault 2 batch under NGS-KPCA in the embodiment of the present application. DETAILED DESCRIPTION

[0065] The present application will be further described in detail below, so that those skilled in the art can implement the present application according to the description.

[0066] Process fault monitoring is a comprehensive technology for identifying the running state of a production process, which mainly studies the reaction of the change of the running state in the diagnostic information, analyzes and processes the measured new signals in combination with the historical data of the state, and further diagnoses whether the production process is abnormal or faulty, and further predicts the future state to determine the necessary countermeasures to be taken.

[0067] Process fault monitoring includes offline systems and online systems. The offline system, also known as a fault patrol system, works in an offline analysis and periodic patrol monitoring mode, while the online system works in a mode of real-time monitoring of the working conditions and processing and analysis of the signals by a computer to control the data collection and field conditions, and further realizes fault diagnosis, such as shown in FIG. 1, which is a specific offline and online process monitoring flowchart. Figure 1

[0068] ​For production process including continuous process and batch process, continuous process is production process in a pipeline and each link is continuous, uninterrupted input and output, production product; batch process is a process that a limited amount of raw materials is processed according to a certain procedure in one or several instruments and devices to obtain a limited amount of product, and the product obtained by one batch operation is called batch. The difference between the two is shown in Table 1.

[0069] Table 1 Comparison of continuous process and batch process

[0070]

[0071] In actual industrial process, due to complex production process, various execution schemes and influence of external factors, batch process has the characteristics of multiple working conditions, nonlinearity and non-Gaussian, multiple time periods and multiple operation.

[0072] Therefore, the present application provides a batch process fault monitoring method under multiple working conditions, which specifically comprises the following steps:

[0073] Step 1, collecting historical data of I normal batches The data is divided into stages according to operation, and the batch data of the same operation in each stage is clustered, but the data under the same operation in different batches is also different. In actual implementation, each stage can be regarded as an operation;

[0074] Wherein, c = 1, 2, …, c s , c s is the number of classes of each stage, is the corresponding number of batches belonging to the class, M is the number of process variables, and N i ' is the reaction time of each batch in the class.

[0075] In specific implementation, the number of process variables is determined according to the variable data collected by different production processes. The process variables can be temperature, pressure, etc., or the number of variable data automatically generated by simulation software in the simulation process.

[0076] Due to the randomness and uncertainty of operation in batch process, the numerical value of operation has no regularity, so the model cannot be constructed by using the sample data of the whole batch process, and the data belonging to the same operation needs to be constructed into a model separately according to the different operations. First, the whole batch process needs to be segmented according to the operation. At present, most of the segmentation methods are carried out without manual operation. In the process with more manual operation, due to the randomness of the operation time and numerical value of the operator and the influence of external factors in the operation process, the difference between batches is large, and the samples between batches are seriously unequal in length, so these methods are no longer applicable, such as Figure 2As shown, generally speaking, the operation will change with the change of the process, and each step of operation can be regarded as a new stage, so the time when the operator performs the operation can be considered as a stage division point.

[0077] In a single batch process, the data of each stage is small, and due to the randomness and non-directionality of the operation, the difference between the batch processing data can be very large, so the k-means clustering algorithm is used to cluster all batch data in each stage.

[0078] In this embodiment, as shown, Figure 3 There are a total of eight batches, and according to the different operations, the eight batch data in the first stage are clustered into three classes, and the data in the same class indicates that the operations of these batches are similar. Although the data in the same class is similar in operation, the data still follows a non-Gaussian distribution, for example, the batch data whose current value is set to 32000A and 32500A in the steelmaking process. Although the set values are similar, these data do not follow a Gaussian distribution. In order to establish a statistical model, the sample needs to be transformed.

[0079] Step two, use variable expansion method to expand the data of the same class into two-dimensional matrix data

[0080] Where, N i is the number of samples per batch, and M is the number of variables in the normal batch process;

[0081]

[0082] Where, m = 1, 2, …, M; n = 1, 2, … N.

[0083] Step three, find the class mn to which each sample x k in X mn belongs, save the average vector E(G(x k mn )) and the standard deviation vector S(G(x mn mn )) of the class, and normalize the sample data based on the mean and standard deviation of the class;

[0084] Where, in the two-dimensional matrix data, the q nearest neighbors of each sample x mn can be found by calculating the Euclidean distance between the two-dimensional matrix data and the sample x k , and the q nearest neighbors are arranged in order of Euclidean distance The collected q neighboring points form a class G(x mn ), and the average vector E(G(x k )) of the class is calculated.mn )) and standard deviation vector S(G(x) k mn Furthermore, the non-Gaussian distribution is transformed into a Gaussian distribution using the nearest neighbor group normalization (NGS) method.

[0085]

[0086]

[0087] Comparison of NGS and Z-normalization methods:

[0088] Assume there are two operating modes: Mode 1 (N(10,0.1) and N(7,0.3)); Mode 2 (N(5,0.2) and N(6,0.4)), with 300 sampling points for each operation. The normal distribution probability plot is shown below. Figure 4a and 4b As shown, through Figure 4a It can be observed that the z-standardized dataset does not follow a normal distribution, while 4b shows that the data standardized by the NGS method approximately follows a normal distribution. At the same time, the Kolmogorov-Smirnov test is used to compare the values ​​in the data with the standard normal distribution. The result of NGS is 0, indicating that the data has a standard normal distribution after standardization. Conversely, the result of z-standardization is 1, confirming that the data does not strictly follow a normal distribution after standardization.

[0089] However, for process data with outliers, the standardization results using the above formula are not ideal, such as... Figure 5a , 5b As shown in 5c, the outlier lies between two modalities. By obtaining q nearest neighbors, the class to which the outlier belongs can be determined, such as... Figure 5a As shown, the class formed by q nearest neighbors comes from two different distributions; as Figure 5b As shown, by standardizing the dataset using the above formula, it can be seen that normal datasets cannot detect outliers and mistakenly treat outliers as normal sample points. Therefore, the above formula needs improvement. Typically, the class of a normal sample is the same as the class of its first nearest point. Since all points in the class of the first nearest point come from the same distribution, this class can be used to standardize the sample. Similarly, for outliers, since all samples in the class of its first nearest point come from the same distribution, this class can be used to standardize the outlier.

[0090]

[0091] In the formula, x represents mn The first nearest point, Class of the first nearest point.

[0092] As Figure 5c shown, the abnormal points are successfully distinguished and separated from the normal data.

[0093] Similarly, for process data with multiple abnormal points, in order to effectively identify these abnormal points, the first step is to determine whether all samples in the class of the first nearest point come from the same distribution. If not, it means that the first nearest point is also an abnormal point, and the second nearest point, the third nearest point, etc. need to be judged until a nearest point is found whose class contains all data from the same distribution. Therefore, formula (3.3) is modified as:

[0094]

[0095] wherein, is the kth nearest point to x mn , is the class to which the kth nearest point belongs, and this formula indicates that all data in the class of the kth nearest point come from the same distribution.

[0096] Step four, a plurality of kernel principal component analysis (KPCA) models are constructed corresponding to the plurality of classes:

[0097] The KPCA method is essentially to map the original data in the low-dimensional space to the high-dimensional space through the kernel function processing, and perform linear analysis on the data in the high-dimensional space. The kernel function is used to solve the inner product in the high-dimensional space, replacing the original nonlinear function solution, which facilitates the nonlinear data processing of unknown nonlinear functions, thereby reducing the complexity of nonlinear calculation.

[0098] Kernel principal component analysis is to map nonlinear input data to a high-dimensional space, and use a kernel function to solve its eigenvalues and eigenvectors. The specific process is as follows:

[0099] The input sample data is mapped to a high-dimensional space denoted as Φ(x k ), and The covariance of this space is:

[0100]

[0101] C F v = λv (4.2)

[0102] In the formula, C F is the covariance of the high-dimensional space, Φ(x j ) is the mapping vector of x j in the high-dimensional space, and Φ T (x j ) is Φ(x j), v is the eigenvector corresponding to the nonzero eigenvalue, and λ is the nonzero eigenvalue of C F i

[0103] Mapping all eigenvectors on the sample data, solving according to equation (4.2) will be in the sample subspace, which can be expressed as:

[0104]

[0105] In the formula, α i is the eigenvector corresponding to the ith eigenvalue.

[0106] Substituting equation (4.3) into equation (4.2), multiplying Φ(x k ) on both sides of the equation simultaneously, we get:

[0107]

[0108] Define the kernel function matrix K ij as M x M, then:

[0109] K ij = K(x i , x j ) = <Φ(x i ), Φ(x j )>, (i, j = 1, 2, …, M) (4.5)

[0110] Where K(x i , x j ) is the kernel function, which is to take the inner product of two vectors in the feature space, and to convert the problem of extracting nonlinear features into the problem of solving eigenvalues and eigenvectors, then:

[0111]

[0112] Where α = [α1, α2, …, α M ] T , the eigenvectors α1, α2, …, α M corresponding to the eigenvalues are λ1≥λ2≥…≥λ M , the eigenvectors are normalized to get:

[0113] λ i <α i , α i > = 1 (4.7)

[0114] In the feature space, t is the principal component score vector, in order to extract the principal component feature, the projection of the matrix X in the high-dimensional space mapping Φ(X) in the principal component direction, then:

[0115]

[0116] To meet the assumption condition The kernel matrix needs to be normalized, which is essentially a centering-variance normalization process. First, the kernel function matrix K ij is centered, that is:

[0117]

[0118] where L N is an MxM identity matrix, and the coefficient is 1 / M.

[0119] Then the variance normalization process is performed to obtain:

[0120]

[0121] In the formula, is the normalized kernel function matrix, is the centered kernel function matrix.

[0122] Step five, calculate the control limit of each type of T 2 statistic and the control limit of SPE statistic;

[0123] The calculation of the control limit of T 2 and SPE statistic is as follows:

[0124] The control limit of T 2 statistic is:

[0125]

[0126] In the formula, is the control limit of statistic T 2 , q is the number of pivots, N is the number of samples, F q,N-q,α is the critical value of F distribution corresponding to the kernel check level α and the degrees of freedom q, N-q;

[0127] The control limit of SPE statistic satisfies:

[0128]

[0129] In the formula, α is the significance level, usually 0.05 or 0.01, C α is the critical value of normal distribution at the significance level α, λ j represents the jth eigenvalue, θ i is the second intermediate variable, and h0 is the first intermediate variable.

[0130] Step six, collect online monitoring sample data x test ∈R 1×MAccording to the operation, the stage to which the online monitoring sample data belongs is judged, and the most suitable model in the class is selected by setting a value;

[0131] Step seven, finding the first nearest point in the model of the class to which the sample belongs and point to which the sample belongs The mean and standard deviation of the class to which the sample belongs are calculated Standardize x test ;

[0132] Step eight, extract the principal component information of x test using the offline model of the class to which the sample belongs, and calculate the statistics T 2 and SPE 2 ;

[0133] KPCA is similar to PCA in the calculation of statistics, and uses the Hotelling T 2 and SPE statistics, and the calculation of the T 2 and SPE statistics is as follows:

[0134] The formula for calculating the T 2 statistic is:

[0135] T 2 = [t1, t2, …, t q ]Λ -1 [t1, t2, …, t q ] T (8.1)

[0136] Where t k (k = 1, 2, …, q) can be obtained from formula (4.8), q is the number of principal components selected by the variance contribution rate method, Λ -1 is the inverse of the diagonal matrix composed of the first q characteristic values corresponding to the eigenvectors, and the monitoring statistic SPE of the residual space of the last N-q dimensions, also called the Q statistic, is calculated as follows:

[0137]

[0138] Where, is the reconstruction eigenvector,

[0139] Step nine, compare the statistics T 2 and SPE with the control limit of the T 2 statistic and the control limit of the SPE statistic, respectively, and if T 2 and SPE exceed the threshold value for three consecutive values in the monitoring process, it is determined that the batch is abnormal.

[0140] Embodiment

[0141] The 120t LF refining process of a certain plant is shown in Table 2. After the steel is discharged from the furnace, the ladle is lifted by the crane to the ladle car of the refining furnace treatment heating station, at the same time the pipeline for blowing argon at the bottom of the ladle is connected, the ladle car is opened to the refining furnace working area, the ladle cover is lowered at the same time the slagging agent is added, the electrode is lowered and then power heating is performed. After the slag is treated, it is basically white, the temperature is measured and sampling is performed, which is recorded as the first sample; after the measurement sample analysis results are obtained, they are uploaded to the computer system, the type and quantity of alloy needed to be added are calculated, then alloy fine tuning is performed, when the steel composition meets the standard, the temperature is measured again and sampling is performed again, which is recorded as the second sample; wire feeding is performed to cause the denaturation of inclusions, then soft blowing is performed with a heat preservative, when the steel temperature and various compositions meet certain values, the refining process is completed, the ladle cover and electrode are lifted, the ladle car is opened out of the refining furnace working area, the ladle is lifted away by the crane, and a cycle of the refining process is completed, with an average of 45 minutes per refining cycle.

[0142] During the entire refining process, bottom blowing argon stirring is very important and is also a relatively complex operation process. By using bottom blowing argon stirring technology, the temperature and composition of liquid steel are effectively uniform, and inclusions and harmful gases can be well removed. According to the requirements of the refining process, the argon blowing amount changes with the operation, but it needs to be noted that the argon blowing flow rate needs to be kept constant at each stage, and it is necessary and effective to form a change control loop for the argon flow rate by online monitoring the argon blowing flow rate. The argon blowing mode during the refining process is shown in Table 2, wherein A represents 40-120 NL / min; B represents 120-260 NL / min; and C represents 260-420 NL / min. C is used only for 3-5 minutes during the addition of reducing agents, deoxidizers and alloy fine tuning.

[0143] Table 2 LF refining process description

[0144]

[0145] The transformer of the refining furnace is divided into 13 gear voltage positions, the input voltage is 35000V, the first four gears are constant power segments, which meet the situation of rapid heating of the molten steel; the last eight gears are constant current segments, which also meet the requirement of molten steel temperature preservation, the specific operation parameters of the transformer are shown in Table 3, according to the operation requirements of the refining process, the parameter constant power segment and the voltage value range of the constant current segment of the transformer operation are designed, that is, it can meet the requirement of providing large power for rapid melting during the rapid heating of the molten steel, and it can also meet the requirement of molten steel temperature preservation in the later stage, in order to ensure that the three-phase impedance imbalance coefficient does not exceed 4%, a triangular arrangement short network system is adopted, the impedance parameter of the short network itself is obtained from the design manual: X d = 2.4 x 10 -3 Ω, r d = 0.6 x 10 -3Ω.

[0146] Table 3 LF furnace transformer operating parameters

[0147]

[0148]

[0149] Through the above research, considering that in the LF refining process, some operations (adding reactants, adjusting set values, etc.) are often affected by human factors, leading to the diversity of data. In addition, the whole refining process depends on the rhythm of the continuous casting production line, and the set values (current, argon flow) between batches are quite different, and the process data is severely subject to non-Gaussian distribution. To solve this problem, according to the changes of multiple working conditions, the whole LF refining process is divided into 5 stages: processes 1 and 2 are the first stage; process 3 is the second stage; the addition of a small amount of deoxidizer can be used as the third segmentation point, at which time the process completes the third stage; the temperature sampling for the second time is the fourth segmentation point, at which time the fourth stage is completed; the remaining operation is the fifth stage. In addition, in order to weaken the differences between the same phase data, according to the voltage, current and argon flow, all batch data are classified, then the batch data belonging to the same class are expanded by the variational method, and the NGS method is used for standardization, and then a KPCA model is established for each class.

[0150] The LF steelmaking process model simulates an actual industrial process and is applied to the prediction control of the composition and temperature of the liquid steel. The simulation model of the decentralized control strategy can simulate the power supply system, the cooling water system, the hydraulic system, the electrode, and the argon system under various working conditions.

[0151] A total of 40 furnaces were simulated to create normal reference batches, and Table 4 lists 13 variables selected for the monitoring scheme, which are closely related to the thermal system, of which the short net cooling water outlet temperature is taken as an example. When the standard value is 40 degrees, 35-38 degrees can be clustered into one class, 39-42 degrees can be clustered into one class, and 42-45 degrees can be clustered into one class.

[0152] Table 4 LF refining process variable description

[0153]

[0154]

[0155] The refining time varies from 40 minutes to 55 minutes, the sampling interval is 6s, the MKPCA and NGS-KPCA models are constructed from 40 reference batches, the new batch is monitored with a 99% confidence level, and the cumulative variance contribution rate method is used to obtain the number of principal components. The MKPCA and NGS-KPCA methods are used to test the LF steelmaking process in three states in Table 5.

[0156] Table 5. LF Refining Process Description

[0157]

[0158] The LF refining process is divided into five stages based on changes in operating conditions. Within these stages, the first and second stages are clustered into three classes each using the k-means algorithm, based on differences in voltage, current, and argon flow rate. The third and fourth stages are clustered into four classes, and the fifth stage into two classes. For each sample in each of the five stages using the NGS-KPCA method, q nearest neighbors are selected, resulting in 20, 20, 30, 30, and 20 points respectively. The number of principal components is obtained using the cumulative variance contribution rate method (>85%).

[0159] To verify the feasibility of this method and determine whether it follows a Gaussian distribution, we can compare the z-standardization method with the NGS standardization method. We choose the Lilliefers test to test the null hypothesis that the data comes from a normally distributed set. The effects of the two methods are illustrated using graphs, with the normal probability plots for the key variables (4 and 13) shown below. Figure 6a and 6b As shown, after z-standardization, the two variables fail the test, as... Figure 6a As shown, the results indicate that both variables still follow a non-Gaussian distribution; Figure 6b As shown, compared with z-standardization, the normal distribution probability map of the NGS standardization method approximately follows a Gaussian distribution. Therefore, it can be determined that after the NGS standardization method is applied, the data tends to be similar to a Gaussian distribution.

[0160] To obtain more accurate monitoring results at each stage, multiple clustering models were constructed. For online monitoring, the first step is to determine the cluster model to which a new batch of data belongs. The process of selecting the most suitable monitoring model from multiple models is as follows: Figure 7a , 7b As shown in Figures 8a, 8b, 9a, and 9b, in the first stage of a normal batch, the class model corresponding to the shortest distance to the center of the new data subset can be calculated as the most suitable monitoring model.

[0161] from Figure 7a , 7b As can be seen from 8a and 8b, because the setpoints of this batch process do not match the setpoints of the first two clustering models, this batch process was mistakenly regarded as an abnormal batch after monitoring by the first and second clustering models. In fact, neither of the first two clustering models is a suitable monitoring model, thus leading to incorrect results. Figure 9a and 9b The monitoring results of the third clustering model indicate that the batch process is normal. The correct monitoring results were obtained through the third clustering model, so it is the most suitable monitoring model. Based on this method, the third clustering model is selected as the most suitable monitoring model.

[0162] By SPE and T 2 The statistical quantity monitors each time point of the normal batch, and the monitoring results of the MKPCA and the NGS-KPCA are as shown in Figure 10a 、 10b , 11a, 11b, it can be seen from the figure that, with the change of the operating condition, the SPE statistical quantity in the MKPCA method occurs twice false alarm, and the process is within the control range of the acceptable error; for the NGS-KPCA method, the monitoring result shows that the statistical quantity is lower than the value of the control limit, and no false alarm occurs. Therefore, the monitoring value of the batch in the two monitoring methods is below the control limit, which confirms that the batch is a normal batch.

[0163] Figure 12a 、 12b , 13a, 13b are the monitoring results of the MKPCA and the NGS-KPCA on fault 1, in which the furnace cover cooling water flow linearly decreases by 10% from 30 minutes, and the decrease of the cooling water flow causes the furnace cover temperature to rise, thereby reducing the service life of the furnace cover, as shown in Figure 12a 、 12b , for the MKPCA method, the T 2 statistic quantity monitors the abnormality after 30 minutes, but the SPE statistical quantity fails within 30 minutes, and false alarm occurs; as shown in Figure 13a 、 13b , compared with the MKPCA, the NGS-KPCA method proposed in the application accurately monitors the abnormality of the T 2 statistic quantity after 30 minutes. The monitoring results of fault 2 are as shown in Figure 14a 、 14b , 15a, 15b, in fault 2, the proportional coefficient of the proportional valve is stepwise decreased by 10% from 21 minutes to 37 minutes, as shown in Figure 14a 、 14b , for the MKPCA, the monitoring chart does not accurately monitor the fault; in contrast, as shown in Figure 15a 、 15b , the statistical quantities T 2 and SPE of the NGS-KPCA both exceed the threshold value at 21 minutes and immediately decrease to the normal operating state at 37 minutes.

[0164] The experimental results show that the NGS-KPCA method described in the application has better monitoring effect than the MKPCA method in the process monitoring process, and the method is very useful for overcoming the multi-working condition, nonlinearity and non-Gaussian characteristics in the batch process, so that the PCA still has robustness in the case of multiple operating characteristics of the process data.

[0165] The application discloses a kind of multi-working-condition intermittent process fault monitoring method, by nearest neighbor group standardization, non-gaussian data is converted into the data of Gaussian distribution, while the kernel principal component analysis model of corresponding cluster is constructed to solve the nonlinear problem of data, reduce the operation complexity, finally through the limitation of statistical quantity, accelerate the speed of fault monitoring, improve the accuracy of monitoring.

[0166] Although embodiments of the application have been disclosed in connection with the above specification and drawings, it will be understood that they are not limited to the specific details of the foregoing description, since various modifications, changes and adaptations will become apparent to those skilled in the art upon reading the foregoing description.

Claims

1. A method for fault monitoring in intermittent processes under multiple operating conditions, characterized in that, Includes the following steps: Step 1: Collect data from one normal batch. The data is divided into stages according to the operations, and batches of data with the same operations in each stage are clustered. In the formula, c s Let c be the number of classes in each stage, and c = 1, 2, ..., c s , M represents the number of batches belonging to the same category, M represents the number of process variables, and N represents the number of batches belonging to the same category. i ′ represents the reaction time of each batch within the same category; Step 2: Expand data of the same type into two-dimensional matrix data. Step 3: Find the class formed by the k nearest points to each sample in the two-dimensional matrix data, and transform the data belonging to the same class in the two-dimensional matrix data based on the nearest neighbor group normalization method: In the formula, x mn For samples in a two-dimensional matrix data, To compare with sample x mn The class formed by the k nearest points. for The average vector, for The standard deviation vector; Step 4: Construct a kernel principal component analysis model for each class; Step 5: Calculate various types of T 2 Control limits for statistical measures and control limits for various SPE statistical measures; Among them, the various types of T 2 The control limits of the statistic are satisfied as follows: In the formula, For the statistic T 2 The control limits are given by: q is the number of principal components, N is the number of samples, and F... q,N-q,α This is the critical value of the F-distribution corresponding to the verification level α and degrees of freedom q and Nq; The control limits for all types of SPE statistics satisfy the following: In the formula, SPE α C is the control limit for the statistic SPE. α λ is the critical value of a normal distribution at a significance level of α. j Let θ represent the j-th eigenvalue. i h0 is the second intermediate variable, and i = 1, 2, 3; Step Six: Collect online monitoring sample data x test ∈R 1×M Based on the operation, determine the stage to which the online monitoring sample data belongs and select the kernel principal component analysis model of the class with the closest distance to the class center in the same stage; Step 7: Standardize the online monitoring sample data; Step 8: Extract x using the offline model of the class it belongs to. test The principal component information was obtained, and the statistic T was calculated. 2 and SPE; Step 9: Calculate the statistic T 2 and SPE are respectively with T 2 Compare the control limits of the statistic T with the control limits of the SPE statistic. During the monitoring process, if the statistic T... 2 Or SPE values ​​exceed T for three consecutive times 2 If the control limits of the statistical measure and / or the control limits of the SPE statistical measure are not met, the batch is considered abnormal.

2. The multi-condition intermittent process fault monitoring method as described in claim 1, characterized in that, In step one, the k-means clustering algorithm is used for clustering.

3. The multi-condition intermittent process fault monitoring method as described in claim 2, characterized in that, Step two uses the variable expansion method to expand the data of the same type into a two-dimensional matrix.

4. The multi-condition intermittent process fault monitoring method as described in claim 3, characterized in that, The average vector satisfies: In the formula, E(X) is the average vector of the data X, N is the number of samples, and x n Let X be the nth sample in the data X.

5. The multi-condition intermittent process fault monitoring method as described in claim 4, characterized in that, The standard deviation vector satisfies: In the formula, S(X) is the standard deviation vector of the data X.

6. The multi-condition intermittent process fault monitoring method as described in claim 5, characterized in that, The kernel function of the kernel principal component analysis model satisfies: K ij =K(x i ,x j )=<Φ(x i ),Φ(x j )>,(i,j=1,2,…,M) In the formula, K ij Let K(x) be the kernel function matrix. i ,x j ) is the kernel function, Φ(x) i ) represents the i-th element x in the kernel function matrix. i The mapping value in high-dimensional space, Φ(x) j () represents the j-th element x in the kernel function matrix. j Mapping value in high-dimensional space.

7. The multi-condition intermittent process fault monitoring method as described in claim 6, characterized in that, The statistic T 2 satisfy: T 2 =[t1,t2,…,t q ]Λ -1 [t1,t2,…,t q ] T In the formula, t k Let q be the score of the k-th principal component, where k = 1, 2, ..., q.

8. The multi-condition intermittent process fault monitoring method as described in claim 7, characterized in that, The statistic SPE satisfies: In the formula, To reconstruct the feature vector, Φ(x) is the feature vector before reconstruction, and n is the number of non-zero feature values.

9. The multi-condition intermittent process fault monitoring method as described in claim 8, characterized in that, The number of principal components satisfies: In the formula, η i The contribution rate corresponding to each principal element. λ represents the cumulative variance contribution rate. i Let be the i-th eigenvalue.

10. The multi-condition intermittent process fault monitoring method as described in claim 9, characterized in that, Step seven is the same as step three.

Images