Tin smelting process abnormity monitoring method based on multi-mode dynamic manifold learning

Abstract

The invention discloses a tin smelting process abnormity monitoring method based on multi-mode dynamic manifold learning, and belongs to the technical field of fault monitoring in the industrial process. The method comprises the steps of mode division, manifold structure extraction, dynamic feature extraction and anomaly monitoring. The method comprises the following steps: firstly, carrying out modal division on industrial process data through a time sequence slow feature analysis TSFA algorithm, and identifying and distinguishing feature information of different modals; then, a manifold learning algorithm is adopted to extract a manifold structure of each modal so as to reserve original geometric and local features of modal data; then, dynamic latent variables in the multi-modal data are extracted by using a dynamic latent variable DLV technology; and finally, in combination with the dynamic characteristics of each mode, constructing a multi-mode dynamic manifold learning MM-DML model, fully considering the manifold structure and the dynamic characteristics of each mode, and improving the accuracy of industrial process anomaly monitoring by fusing multi-mode information.

Description

Technical Field

[0001] The present invention belongs to the field of fault monitoring in industrial processes, and particularly relates to an abnormal monitoring method for the tin smelting process based on multi-modal dynamic manifold learning. Background Art

[0002] Tin is one of the four strategic resources in China and is also an important raw material for realizing industrial transformation. As a key link in refined tin processing, the tin smelting process is characterized by large fluctuations in core temperature, high pressure, and large random variations in the oxygen content of the top-blown furnace lance. These factors result in non-Gaussian distributions of process variables. With the gradual decrease in the grade of tin ore, the smelting process becomes increasingly complex, and the relevant monitoring variables continue to increase, showing high-dimensional characteristics. In addition, compared with other metal smelting processes, due to its low melting point and high boiling point, tin poses more stringent requirements on the smelting process.

[0003] The entire tin smelting process flow needs to frequently switch working conditions, and the human operation factor is relatively large to cope with complex production requirements. Therefore, it has multi-modal characteristics, that is, it involves multiple operating states. Traditional process monitoring methods are usually based on the assumption that the production process state is unchanged and stable. This makes it difficult for traditional single-process monitoring methods to timely and effectively detect abnormalities in the tin smelting process, which may then cause serious production accidents. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an abnormal monitoring method for the tin smelting process based on multi-modal dynamic manifold learning. Using the data in the normal operating state of the tin smelting process as the training set, through the division of different sub-processes of the data, the sub-manifold structure, and the extraction of dynamic latent variables, and finally using Bayesian inference to fuse each process to obtain a comprehensive monitoring model, thereby realizing the real-time monitoring of the tin smelting process and having a good effect in actual production.

[0005] The technical solution of the present invention is: an abnormal monitoring method for the tin smelting process based on multi-modal dynamic manifold learning, and the specific steps are as follows:

[0006] S1: Modal division: Adopt the Time Slow Feature Analysis (TSFA) algorithm to preprocess and divide the industrial process data into modes.

[0007] The process data includes control variables (such as furnace temperature, gas flow rate), process variables (such as physical properties of materials, gas composition in the furnace), and composition variables (such as molar amount of the generated product).

[0008] S2: Manifold Structure Extraction: Apply the manifold learning algorithm to the data of each modality to extract its manifold structure, so as to preserve the local geometric features of the data and provide a basis for dynamic feature extraction.

[0009] S3: Dynamic Feature Extraction: Based on the extracted manifold structure, use the Dynamic Latent Variable (DLV) technology to extract dynamic latent variables from multi-modal data. These latent variables can capture the dynamic change trends in the process and provide support for the anomaly monitoring of valve stickiness, cooling water switch, and heat transfer oil temperature fluctuations.

[0010] S4: Anomaly Monitoring: Combine the historical normal data of the industrial process to construct a multi-modal dynamic manifold learning MM-DML model.

[0011] The MM-DML model monitors the anomalies of valve stickiness, cooling water switch, and heat transfer oil temperature fluctuations in the tin smelting process.

[0012] The above S1 is specifically as follows:

[0013] S1.1: Preprocess the data of the entire industrial process, that is, standardize the data values and normalize the data distribution. Specifically:

[0014] S1.1.1: Assume the industrial process dataset Calculate its mean value and standard deviation X std ;

[0015] S1.1.2: Standardize the data

[0016] S1.2: Use Time Series Slow Feature Analysis (TSFA) to divide different modalities. Specifically:

[0017] S1.2.1: Calculate the difference vector matrix Specifically:

[0018]

[0019] Among them, corresponds to the first-order difference. The main goal of TSFA is to judge whether a new modality is entered by constructing the corresponding relationship between the general original dataset X and the difference dataset

[0020] S1.2.2: First, make a preliminary division according to the time series of the original dataset X and the difference dataset Specifically:

[0021] (1) Set the lowest threshold samples to eliminate the influence of disturbances on the training model. Specifically:

[0022] I = {i | variable i <limit}

[0023] Remove the samples corresponding to index I from the original dataset X and the differential dataset ;

[0024] (2) Set the minimum partitioning variable length min = (m - 1) × 3 × 5, and calculate the initial number of intervals d:

[0025]

[0026] where m is the number of variables and N is the number of samples;

[0027] (3) According to the d initial data slices defined in (2), construct the sample sets for each interval, specifically:

[0028] Find the maximum value X max = max(X) and the minimum value X min = min(X) in the original dataset X; Solve the interval The i-th data slice is defined as k i = {j | X j ∈ [X min + (i - 1)δ, X min + iδ]}.

[0029] S1.2.3: If the sample size of the obtained data slice is less than the minimum partitioning variable length min , then merge it with the adjacent data slice, and finally obtain slices.

[0030] S1.2.4: According to the slices obtained in step S1.2.3, use Slow Feature Analysis (SFA) to calculate the modeling and control limits for each data slice, and merge the slices with detailed control limits to obtain slices, specifically:

[0031] (1) Calculate the covariance matrix of each slice dataset Then decompose it to obtain the corresponding eigenvectors and eigenvalues;

[0032] (2) Select the number of eigenvectors num corresponding to the top 85% of the eigenvalues, that is:

[0033]

[0034] (3) Calculate the original dataset X and the difference dataset the projected Z sum Specifically:

[0035]

[0036] (4) Calculate the covariance matrix of the difference dataset projection According to the obtained load matrix calculate the slow features and and

[0037] (5) Determine the first 90% of the difference dataset and the diagonal elements of the original dataset X as the boundary for determining slow and fast features;

[0038] (6) Calculate the slow feature statistics for each interval and determine whether they belong to the same interval based on the magnitudes of the interval statistics. For the overall difference dataset and the original dataset X, use SFA to obtain the projection matrix of the slow features and the projection matrix m of the fast features m ;

[0039] (7) According to the slices obtained in step S1.2.3, calculate the statistical control limits for each interval to determine whether the slices need to be merged, and merge the slices with similar control limits to obtain

[0040] Specifically, the above S2 is as follows:

[0041] S2.1: Use the manifold learning algorithm to extract the manifold structure of each modality respectively. Specifically:

[0042] S2.1.1: For the slices obtained by using the temporal slow feature (TSFA) analysis in S1, use the manifold learning algorithm - Neighborhood Preserving Embedding (NPE) to construct the objective function L = min(XA - WXA) T (XA - WXA) to obtain the manifold structure of each original slice and obtain the mapping matrix A. Specifically as follows:

[0043] (1) Use the knn nearest neighbor algorithm to find the nearest neighbor matrix W for the data of the first slice i ;

[0044] (2) Calculate the similarity of each slice using the KL divergence, and correct the nearest neighbor matrix W of other slices for the similarity to avoid redundant calculations;

[0045] (3) Calculate the mapping matrix A of each slice i ;

[0046] S2.1.2: Use the manifold structure model of each trained slice to reduce the dimension of the original data and obtain the reduced data Y that retains the original geometry and local features of the modal data i = X i A i ;

[0047] The above S3 is specifically as follows:

[0048] S3.1: Construct an autoregressive model (AR) to extract the dynamic latent variables in the reduced data Y with a manifold structure i ;

[0049] S3.2: Decompose the data of each slice after extracting the dynamic latent variables using traditional principal component analysis (Principal Component Analysis, PCA) and calculate the statistical control limit Ctrl of each slice i ;

[0050] The above S4 is specifically as follows:

[0051] S4.1: Use Bayesian fusion for the control limits of each slice to construct a multi-mode dynamic manifold learning (Multi-Mode Dynamic Manifold Learning, MM-DML) model;

[0052] S4.2: For the newly collected monitoring data, locate the operating state of the system according to TSFA, calculate the statistics corresponding to the state slice model to which it belongs, and determine whether an abnormal working condition occurs by comparing the size of the statistics with the control limit.

[0053] The beneficial effects of the present invention are:

[0054] (1) By introducing time series, characterize the correlation in the tin smelting process, and automatically divide the entire tin smelting process into multiple stages through the TSFA algorithm.

[0055] (2) By weighting the nearest neighbor weights of each stage using the KL divergence and constructing an NPE model for each stage, fully retain the manifold structure of each stage.

[0056] (3) By introducing an autoregressive model, the dynamic latent variables in the data after retaining the manifold structure are extracted, and Bayesian inference is used to fuse each stage to obtain a comprehensive fault monitoring model.

[0057] (4) By calculating the statistics of the data that extracts the dynamic characteristics and retains the static manifold structure, the accurate monitoring of abnormal conditions in the tin smelting process is realized. Description of the Drawings

[0058] Figure 1 This is a flowchart of the abnormal monitoring method for the tin smelting process based on multi-modal dynamic manifold learning in the present invention. Detailed Embodiments

[0059] The present invention will be further described below in conjunction with the drawings and specific embodiments.

[0060] In order to be able to monitor the abnormal conditions in the tin smelting process in a timely and efficient manner and ensure production safety, an abnormal monitoring method for the tin smelting process based on multi-modal dynamic manifold learning is proposed. First, based on TSFA, the entire tin smelting process is divided into multiple stages to form multiple sub-processes; secondly, KNN is used to construct the neighbor matrix of the first sub-process, and the Kullback-Leibler divergence is used to calculate the similarity between each sub-process, and the neighbor matrix of the subsequent sub-processes is weighted according to the similarity; then, the NPE method is used to extract the manifold structure of each sub-process; then, an autoregressive model is used to extract the dynamic latent variables of each sub-process to obtain static variables; finally, the Bayesian inference method is used to fuse the dynamic latent variables of each sub-process to construct a comprehensive model, thereby realizing the abnormal monitoring of the tin smelting process.

[0061] An embodiment of the present invention provides an abnormal monitoring method for the tin smelting process based on multi-modal dynamic manifold learning, and its process is as Figure 1 shown, and the specific steps are as follows:

[0062] S2: Manifold structure extraction: Apply the manifold learning algorithm to the data of each modality to extract its manifold structure to retain the local geometric features of the data and provide a basis for dynamic feature extraction;

[0063] S3: Dynamic feature extraction: Based on the extracted manifold structure, use the dynamic latent variable (DLV) technology to extract dynamic latent variables from multi-modal data, and these latent variables can capture the dynamic change trends in the process and provide support for abnormal monitoring;

[0064] S4: Abnormal monitoring: Combine the historical normal data in the tin smelting process to construct a multi-modal dynamic manifold learning MM-DML model to realize the monitoring of abnormalities such as valve sticking, cooling water switch, and thermal oil temperature fluctuation in the tin smelting process

[0065] Select the production data of a tin smelting process for verification;

[0066] Specifically, the above S1 is as follows:

[0067] S1.1: Preprocess the data of the entire industrial process, that is, standardize the data values and normalize the data distribution. Specifically:

[0068] S1.1.1: Assume the industrial process data set Calculate its mean value and standard deviation X std ;

[0069] S1.1.2: Standardize the data

[0070] S1.2: Use time series slow feature analysis (TSFA) to divide different modes. Specifically:

[0071] S1.2.1: Calculate the difference vector matrix Specifically:

[0072]

[0073] Among them, corresponds to the first-order difference. The main goal of TSFA is to judge whether a new mode is entered by constructing the corresponding relationship between the general original data set X and the difference data set .

[0074] S1.2.2: First, make a preliminary division according to the time series of the original data set X and the difference data set . Specifically:

[0075] (1) Set the lowest threshold sample to eliminate the influence of disturbances on the training model. Specifically:

[0076] I = {i|variable i < limit}

[0077] Remove the samples corresponding to the index I from the original data set X and the difference data set ;

[0078] (2) Set the minimum division variable length min = (m - 1) × 3 × 5, and calculate the initial number of intervals d:

[0079]

[0080] Among them, m is the number of variables and N is the number of samples;

[0081] (3) Construct the sample sets for each interval according to the d initial data slices defined in (2), specifically as follows:

[0082] Find the maximum value X max = max(X) and the minimum value X min = min(X) in the original data set X; Solve the interval The i-th data slice is defined as k i = {j|X j ∈ [X min + (i - 1)δ, X min + iδ]}.

[0083] S1.2.3: If the sample size of the obtained data slice is less than the minimum division variable length min , then merge it with the adjacent data slice, and finally obtain slices.

[0084] S1.2.4: According to the slices obtained in step S1.2.3, use Slow Feature Analysis (SFA) to calculate the modeling and the statistical control limits of the model for each data slice, and merge the slices with detailed statistical control limits of the model to obtain slices, specifically as follows:

[0085] (1) Calculate the covariance matrix of each slice data set Then decompose it to obtain the corresponding eigenvectors and eigenvalues;

[0086] (2) Select the number of eigenvectors num corresponding to the top 85% of the eigenvalues, that is:

[0087]

[0088] (3) Calculate the projections of the original data set X and the difference data set Z and Specifically as follows:

[0089]

[0090] (4) Calculate the covariance matrix of the projection of the difference data set According to the obtained load matrix calculate the slow feature and and

[0091] (5) Determine the top 90% of the difference data set The diagonal elements of the original dataset X are used as the boundary for determining slow features and fast features;

[0092] (6) Calculate the slow feature statistics for each interval, and determine whether they belong to the same interval based on the magnitudes of the interval statistics. For the overall differential dataset and the original dataset X, use SFA to obtain the projection matrix for slow features and the projection matrix for fast features Then calculate the confidence limits Ctrs m and Ctrf m corresponding to the statistics of slow features and fast features;

[0093] (7) According to the slices obtained in step S1.2.3, calculate the statistical control limits for each interval to determine whether the slices need to be merged, and merge the slices with similar control limits to obtain

[0094] Specifically, the above S2 is as follows:

[0095] S2.1: Use the manifold learning algorithm to extract the manifold structure of each modality respectively, specifically:

[0096] S2.1.1: For the slices obtained by using the time-series slow feature (TSFA) analysis in S1, use the manifold learning algorithm - Neighborhood Preserving Embedding (NPE) to construct the objective function L = min(XA - WXA) T (XA - WXA) to obtain the manifold structure of each original slice, and obtain the mapping matrix A, specifically as follows:

[0097] (1) Use the knn nearest neighbor algorithm to find the nearest neighbor matrix W for the data of the first slice i ;

[0098] (2) Use the KL divergence to calculate the similarity of each slice, and correct the nearest neighbor matrix W of other slices for the similarity to avoid redundant calculations;

[0099] (3) Calculate the mapping matrix A for each slice i ;

[0100] S2.1.2: Use the manifold structure model of each trained slice to reduce the dimension of the original data and obtain the dimension-reduced data Y that retains the original geometry and local features of the modal data i = X i A i ;

[0101] Specifically, S3 is as follows:

[0102] S3.1: Construct an autoregressive model (AR) to extract the dimensionality-reduced data Y with a manifold structure i and the dynamic latent variables in

[0103] S3.2: Decompose the data of each slice after extracting the dynamic latent variables using traditional principal component analysis (PCA) and calculate the statistical control limit Ctrl of each slice i ;

[0104] Specifically, S4 is as follows:

[0105] S4.1: Use Bayesian fusion for the control limits of each slice to construct a multi-modal dynamic manifold learning (MM-DML) model;

[0106] S4.2: For newly collected monitoring data, locate the operating state of the system according to TSFA, calculate the statistic corresponding to the state slice model it belongs to, and determine whether an abnormal condition occurs by comparing the statistic with the control limit.

Claims

1. A method for monitoring abnormality in tin smelting process based on multimodal dynamic manifold learning, characterized in that: include: S1. Modal division: The time series slow feature analysis (TSFA) algorithm is used to preprocess and modal divide the process data in the tin smelting process, wherein the process data includes control variables, process variables, and component variables; S2, manifold structure extraction: Apply the manifold learning algorithm to the data of each modality to extract its manifold structure to retain the local geometric features of the data and provide a basis for dynamic feature extraction; S3. Dynamic feature extraction: Based on the extracted manifold structure, the dynamic latent variable (DLV) technology is used to extract dynamic latent variables from multimodal data. The dynamic latent variables can capture the dynamic change trend in the process and provide support for abnormal monitoring of valve sticking, cooling water switching and thermal oil temperature fluctuations. S4. Abnormal monitoring: Combined with the historical normal data of the tin smelting process, a multimodal dynamic manifold learning (MM-DML) model is constructed to monitor the abnormalities of valve sticking, cooling water switching, and thermal oil temperature fluctuations during the tin smelting process.

2. The abnormality monitoring method according to claim 1, characterized in that: The S1 is specifically: S1.

1. Preprocess the industrial process data in the tin smelting process, including data value standardization and data distribution normalization, specifically: S1.1.

1. Assumptions of industrial process data set Calculate its mean and standard deviation X std ; S1.1.

2. Standardize the data S1.

2. Use the time series slow feature analysis (TSFA) to divide the preprocessed data into different modes, specifically: S1.2.

1. Calculate the difference vector matrix That is, the difference data set, specifically: in, Corresponding to the first-order difference, the goal of TSFA is mainly to construct a general original data set X and a difference data set The corresponding relationship between the two modes is then used to determine whether to enter a new mode; S1.2.

2. First, based on the original data set X and the differential data set The temporal sequence of the data is preliminarily divided to obtain the number of data slice samples; S1.2.

3. If the number of data slice samples obtained is less than the minimum partition variable length min , then merge it with the adjacent data slices and finally get slices; S1.2.4, obtained according to the steps in S1.2.3 Slices are calculated using slow feature SFA for each data slice and the statistical control limits of the model. The detailed slices of the model statistical control limits are merged to obtain slices.

3. The abnormality monitoring method according to claim 2, characterized in that: The S1.2.4 is specifically: (1) Calculate the covariance matrix of each slice data set Then decompose it Get the corresponding Eigenvectors and Eigenvalue; (2) Select the number of eigenvectors num corresponding to the first 85% of the eigenvalues, that is: (3) Calculate the original data set X and the difference data set After projection, Z and Specifically: (4) Calculate the difference data set projection The covariance matrix of According to the load matrix obtained Computational slow features and (5) Determine the first 90% of the differential data set The diagonal elements of the original dataset X are used as the dividing line between slow features and fast features; (6) Calculate the slow feature statistics of each interval, and determine whether they belong to the same interval based on the size of the statistics of each interval. Use SFA with the original dataset X to get the projection matrix of the slow features and the projection matrix of the fast feature Then calculate the confidence limits Ctrs corresponding to the statistics of slow features and fast features m and Ctrf m ; (7) According to step S1.2.3 slices, calculate the statistical control limits of each interval to determine whether the slices need to be merged, and merge the slices with similar control limits to obtain 4. The abnormality monitoring method according to claim 1, characterized in that: The S2 is specifically: S2.1: Use the manifold learning algorithm to extract the manifold structure of each mode, specifically: S2.1.1: Analyze the slow features TSFA in S1 slices, and use the manifold learning algorithm-neighborhood preserving embedding NPE to construct the objective function L=min(XA-WXA) T (XA-WXA) obtains the manifold structure of each original slice and obtains the mapping matrix A, as follows: (1) Use the knn nearest neighbor algorithm to find the nearest neighbor matrix W for the first slice of data i ; (2) Use KL divergence to calculate the similarity of each slice, and modify the neighbor matrix W of the slice to avoid redundant calculation; (3) Calculate the mapping matrix A of each slice i ; S2.1.

2. Each slice that has been trained The manifold structure model is used to reduce the dimension of the original data and obtain the reduced-dimensional data Y that retains the original geometry and local features of the modal data. i =X i A i .

5. The abnormality monitoring method according to claim 1, characterized in that: The S3 is specifically: S3.

1. Construct an autoregressive model AR to extract the reduced dimension data Y with manifold structure i Dynamic latent variables in ; S3.

2. After extracting the dynamic latent variables, the data of each slice is decomposed using the traditional principal component analysis PCA and the statistical control limits of each slice are calculated Ctrl i .

6. The abnormality monitoring method according to claim 1, characterized in that: The S4 is specifically: S4.

1. The control limits of each slice are fused using Bayesian to construct a multimodal dynamic manifold learning MM-DML model. S4.

2. For the newly collected monitoring data, the system operation status is located according to TSFA, and the statistics corresponding to the state slice model to which it belongs are calculated. By comparing the statistics with the control limit, it is determined whether an abnormal operating condition occurs.

Images