A method for monitoring intermittent processes based on multi-directional concurrent local weighted projection regression
Through the multi-directional concurrent local weighted projection regression method, the problem of local structural change detection in intermittent process monitoring is solved, and the refined monitoring and fault detection of intermittent processes are realized, which improves detection efficiency and accuracy.
Patent Information
- Application Number
- CN202410643416.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-23
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-05-23
AI Technical Summary
The prior art is difficult to effectively monitor local structural changes in intermittent processes, resulting in low fault detection efficiency and failure to fully consider product quality impacts. Neural network-based methods lack local data structure capture and information retention.
A batch process monitoring method based on multi-directional concurrent local weighted projection regression is adopted, and a phase division is performed through local mass correlation characteristics and time series, a multi-directional concurrent local weighted projection regression model is established, and a regression structure is used for quality variable prediction and state monitoring.
The refined monitoring of intermittent processes is realized, complete monitoring indicators related to quality and process are established, the accuracy and efficiency of fault detection are improved, and important local information is retained.
Smart Images

Figure CN118625754B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of intermittent process fault detection, and more specifically to an intermittent process monitoring method based on multi-directional concurrent local weighted projection regression. Background Art
[0002] As one of the most important production methods in modern processing industry, intermittent process is widely used in many fields such as biopharmaceuticals, fine chemicals, food processing, semiconductor production, etc. due to its advantages of high value, multiple varieties and small batch production. Intermittent process is different from continuous process. Intermittent process has a wide range of multi-stage characteristics. In view of the multi-stage characteristics of intermittent process, it is necessary to establish a multi-stage model that can capture its inherent behavior. Each stage has specific local behavior. Multi-stage modeling is conducive to the characterization of process characteristics and improves process monitoring effect.
[0003] The existing multi-stage processing methods still have the following problems: First, in the process of stage division, the stages need to be identified by clustering methods, and complex post-processing is required. Second, each search for the optimal sharding point requires traversing all possible sharding points, and the order and sequence length have a high time complexity. Third, the final product quality of the process is not taken into account when dividing the stages. Therefore, reasonable stage division can effectively improve the effect of batch process monitoring and product quality prediction.
[0004] In some cases, appropriate fluctuations in process variables do not affect product quality and should not trigger disturbance alarms. However, some unsupervised methods only consider anomalies in the entire process space without using output variables. Supervised methods use quality data to extract relevant information and are suitable for quality-related process monitoring.
[0005] Intermittent process data has strong nonlinearity. In order to deal with nonlinear problems, most methods use kernel functions to project nonlinear process data into high-dimensional kernel space. However, the kernel function method cannot maintain the characteristics of the original data. In addition, during the training phase of the model, the kernel matrix constructed based on a large amount of data is used for feature decomposition, which increases the amount of calculation and affects the detection efficiency.
[0006] Artificial neural networks have more advantages in extracting features from nonlinear data. Traditional artificial neural network-based process fault detection methods do not consider the impact of monitoring on product quality. In addition, most neural network-based methods are global end-to-end and do not capture local data structures. During feature compression, important local neighbor information is lost, which is not conducive to monitoring.
[0007] Therefore, how to monitor changes in the local structure of process data and implement fault detection by monitoring changes in local information is a problem that technical personnel in this field urgently need to solve. Summary of the invention
[0008] In view of this, the present invention provides an intermittent process monitoring method based on multi-directional concurrent local weighted projection regression. The method fully considers the nonlinear characteristics of the intermittent process, divides the stages according to local quality-related characteristics and time series, and uses a multi-directional concurrent local weighted projection regression model for modeling in each stage, thereby establishing complete quality-related and process-related monitoring indicators, and using the regression structure to predict quality variables.
[0009] In order to achieve the above object, the present invention adopts the following technical solution:
[0010] A batch process monitoring method based on multi-directional concurrent local weighted projection regression, comprising:
[0011] S1: Obtain intermittent process variable data and perform standardization;
[0012] S2: Divide the entire process into stages according to the operating characteristics and time sequence of the intermittent process;
[0013] S3: For the divided stages, a multi-directional concurrent local weighted projection regression model is established respectively;
[0014] S4: Refine the monitoring statistical data, calculate the global statistics and obtain the control limit, and determine whether the global statistics exceed the control limit. If it exceeds the control limit, it means a fault has occurred, otherwise it is normal.
[0015] Preferably, S1 specifically includes:
[0016] Collect the intermittent process and form three-dimensional data X(I×J×K), where I, J and K represent the batch, variable and sampling time of the data respectively;
[0017] The collected three-dimensional data X(I×J×K) is expanded into X(I×JK) along the batch direction and standardized:
[0018]
[0019] in, represents the standardized data, mean(X(I×J)) and std(X(I×J)) represent the mean and variance of X(I×J), respectively.
[0020] Preferably, S2 specifically includes:
[0021] S2 specifically includes:
[0022] Determine the standardized data The initial stage preset values are used as the input values of the Gaussian mixture model, and cluster analysis is performed through the Gaussian mixture model. The standardized data The probability density corresponding to each cluster is
[0023]
[0024] Where i = 1, 2, ..., m, m is the number of standardized data, j = 1, 2, ..., B, B is the number of Gaussian models; ∑ j and μ j represents the jth cluster C j The variance matrix and mean vector of , T represents the transpose;
[0025] The probability density function of Represented by B Gaussian models:
[0026]
[0027] In the above formula, w j represents the mixing weight of the jth component;
[0028] Based on the probability density function And through Bayesian inference method, we get The probability density of the data belonging to stage b is:
[0029]
[0030] The standardized data is analyzed by the maximum value of the probability density Divide into stages:
[0031]
[0032] in, express The maximum probability density belonging to stage b.
[0033] Preferably, S2 also includes regularizing the divided stages, which specifically includes the following steps:
[0034] For intermittent processes, the batch with the closest time length to the normal batch is selected as the reference batch. For batches of unequal lengths, G = (g1, g2, ..., g i ,…,g m ) and reference batch H = (h1,h2,…,h j ,…,h n ), n represents the number of reference batch samples, m represents the number of unequal length batch samples, calculate gi and h j The distance d(g i ,h j ):
[0035] d(g i ,h j ) = g i -h j2
[0036] By g i and h j Calculate the shortest distance D and the optimal path W:
[0037]
[0038] In the above formula, (g i (l),h j (l)) indicates batches of unequal length g i and reference batch h j The lth point of , max(m,n)≤L≤m+n;
[0039] W = argmin[D(m,n)];
[0040] The optimal path W = {w1,w2,…,w L},w l =(i(l),j(l)) is the sequence to be searched in the m×n grid according to the shortest distance D(m,n);
[0041] The constraint allows a point in the grid (g m (l),h n (l)) can only pass through or reach point (g m (l-1),h n (l)), point (g m (l),h n (l-1)) or point (g m (l-1),h n (l-1));
[0042] Compute the path distance matrix:
[0043]
[0044] Among them, d(g m (l),h n (l)) indicates batches of unequal length g m and reference batch h n The distance of the lth point, D(g m (l-1),h n (l)) indicates batches of unequal length g mThe l-1th point and the reference batch h n The shortest distance of the lth point, D(g m (l),h n (l-1)) indicates batches of unequal length g m The lth point and the reference batch h n The shortest distance of the l-1th point, D(g m (l-1),h n (l-1)) indicates batches of unequal length g m The l-1th point and the reference batch h n The shortest distance to the l-1th point;
[0045] Repeat the above process to regularize each stage.
[0046] Preferably, S3 specifically includes:
[0047] Assume that the nonlinear function y α =f(x α )+e,x α and α are input variables and output variables respectively, e represents noise with a mean of 0;
[0048] Calculate the weight ψ through the Gaussian kernel α :
[0049]
[0050] Where D is a positive semidefinite matrix that determines the size and shape of the local model, x c is the data center of the local model, T represents transposition;
[0051] Based on the weight ψ α Compute the weighted mean for a multidirectional concurrent locally weighted projection regression model:
[0052]
[0053] in, and are the weighted means of the input and output variables of the local weighted projection regression model, respectively, and N is the number of sampling points;
[0054] Based on the normal sample (x δ ,y δ ),δ=1,2,...,Δ,Δ is the length of the sample, the weighted mean of the input and output variables of the local weighted projection regression model is used to calculate the input matrix X k And the output matrix Y k ;
[0055] Based on the input matrix X kAnd the output matrix Y k Construct the intermediate matrix S k , and based on the intermediate matrix S k and the input matrix X k Calculate the feature output matrix F k : F k =X k S k .
[0056] Preferably, the monitoring statistical data are refined into a predictable output subspace, an unpredictable output residual subspace, an unpredictable output principle subspace, an output irrelevant main subspace and an output irrelevant residual subspace.
[0057] Preferably, S4 specifically includes:
[0058] Through the feature output matrix F k and the mass-related load matrix Q k Calculate predicted values And the predicted value Perform singular value decomposition:
[0059]
[0060] in, represents the value after orthogonality, is the value of orthogonal decomposition;
[0061] Scoring Matrix The calculation formula is:
[0062]
[0063] Where, X k is the input matrix, S k is the intermediate matrix;
[0064] The unpredictable output is And use l y The principal components are used for PCA calculation:
[0065]
[0066] In the formula, represents the principal component matrix of the unpredictable output space, represents the unpredictable output space score matrix, represents the unpredictable output spatial residual matrix;
[0067] Input matrix X k And the output matrix Y k Perform singular value decomposition:
[0068]
[0069]
[0070] In the formula, and All represent load matrices, For X k Zhongyu The associated score matrix, Represents X k Medium and predicted Y k Irrelevant variables, Represents Y k Cannot be X k The predicted variables, Represents the input matrix X k The residual matrix of
[0071] Compute monitoring statistics for the predictable output subspace:
[0072]
[0073] In the formula, ω is the number of sampling points, x mz The input matrix X k The row vector of yes The row vector of is the predictable output subspace statistic at time k, is the intermediate weighting matrix;
[0074] Monitoring statistics of the unpredictable output main subspace:
[0075]
[0076] In the formula, is the monitoring statistic of the unpredictable output main subspace at time k, for The row vector of y mz is the output matrix Y k The row vector of
[0077] Compute monitoring statistics for the unpredictable output residual subspace:
[0078]
[0079] In the formula, is the monitoring statistic of the unpredictable output residual subspace at time k;
[0080] Compute monitoring statistics for the output-independent main subspace:
[0081]
[0082] In the formula, Output the monitoring statistics of the irrelevant main subspace at time k, for The row vector of yes The row vector of
[0083] For each data point (x k ,y k ), the global statistics calculation formula of the multi-directional concurrent local weighted projection regression model is:
[0084]
[0085]
[0086]
[0087]
[0088] In the formula, is the predictable output subspace statistic, For the statistics of the unpredictable output main subspace, SPE y is the unpredictable output residual subspace statistic, To output the monitoring statistics of the irrelevant main subspace, ψ k Calculate weights locally.
[0089] Preferably, the input matrix X k And the output matrix Y k The specific calculation process is:
[0090]
[0091]
[0092]
[0093]
[0094] Where T represents transpose, x n and n are normal samples, and are the local weighted values of the input and output variables of the local weighted projection regression model, respectively.
[0095] Preferably, the intermediate matrix construction process is:
[0096]
[0097] In the above formula, s1=u1, P k =[p1,…,p A ] and U k =[u1,…,u A ] are the input matrix, the load matrix and the weight matrix after output matrix training respectively.
[0098] Calculate the intermediate matrix S k =[s1,s2,…,s A ], A is the number of local models established.
[0099] The advantages and technical effects of the present invention are:
[0100] (1) Based on the local quality-related characteristics and time series, a time-sequential quality-related stage division model is proposed.
[0101] (2) A multi-directional concurrent local weighted projection regression model is used to model the nonlinear stage, ensuring the effective extraction of information.
[0102] (3) A complete set of quality-related and process-related monitoring indicators have been established to conduct detailed monitoring of the operating status of the intermittent process. BRIEF DESCRIPTION OF THE DRAWINGS
[0103] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0104] Figure 1 A flow chart of an intermittent process monitoring method based on multi-directional concurrent local weighted projection regression provided by the present invention;
[0105] Figure 2 is a flow chart of data preprocessing and chronological division of quality-related stages of the present invention;
[0106] Figure 3 It is a process monitoring diagram of the multi-directional concurrent local weighted projection regression algorithm of the present invention;
[0107] Figure 4a Aeration rate failure in penicillin fermentation process Statistical graphs;
[0108] Figure 4b Aeration rate failure in penicillin fermentation process Statistical graphs;
[0109] Figure 4c Aeration rate failure in penicillin fermentation process Statistical graphs;
[0110] Figure 4d SPE for aeration rate failure in penicillin fermentation process x Statistical chart. DETAILED DESCRIPTION
[0111] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0112] The present invention discloses an intermittent process monitoring method based on multi-directional concurrent local weighted projection regression, such as Figure 1 As shown, including:
[0113] S1: Obtain intermittent process variable data and perform standardization;
[0114] S2: Divide the entire process into stages according to the operating characteristics and time sequence of the intermittent process;
[0115] S3: For the divided stages, a multi-directional concurrent local weighted projection regression model is established respectively;
[0116] S4: Refine the monitoring statistical data, calculate the statistics and obtain the control limit, and determine whether the statistics exceed the control limit. If it exceeds the control limit, it means a fault has occurred, otherwise it is normal.
[0117] S1-S4 are executed in a loop until all samples are monitored.
[0118] Since the penicillin fermentation process is a typical intermittent process, the present invention describes the specific implementation method in conjunction with the penicillin fermentation process.
[0119] The penicillin production process is a typical multi-stage intermittent process, which is widely used in process monitoring and quality prediction. Pensim2.0 provides a benchmark platform for batch monitoring and fault diagnosis of penicillin production process. By simulating the penicillin fermentation process with Pensim2.0, a series of parameters such as microbial concentration, carbon dioxide, and pH value are obtained, which solves the problem of real-time acquisition of variable data. The simulation platform is used to generate 20 batches of normal operating data, where the reaction time of each batch is 400h and the sampling time is 1h. Eleven main process variables are selected as monitoring variables, among which penicillin concentration is the product quality variable, as shown in Table 1.
[0120] Table 1 Main variables of penicillin fermentation process
[0121]
[0122] (1) Collect penicillin fermentation process data X(20×11×400), expand the collected three-dimensional data X(20×11×400) into X(20×4400) along the batch direction, and standardize it, as follows: Figure 2 shown.
[0123]
[0124] in, represents the standardized data, mean(X(I×J)) and std(X(I×J)) represent the mean and variance of X(I×J), respectively;
[0125] (2) Standardized data R represents the matrix, l and z represent the number of rows and columns of the matrix respectively, and the k-means method is used to determine the standardized data The initial stage preset value is used as the input value of the Gaussian mixture model, and cluster analysis is performed through the Gaussian mixture model. The probability density corresponding to each cluster is The conditional probability density belonging to the jth cluster is shown in the following formula (2):
[0126]
[0127] In the above formula, i=1,2,...,m;j=1,2,...,B;∑ j and μ j represents the jth cluster C j The variance matrix and mean vector of . Therefore, The probability density function of can be expressed as B Gaussian models, as shown in formula (3):
[0128]
[0129] In the above formula, w j represents the mixing weight of the jth component;
[0130] By using Bayesian inference method The probability density of the data belonging to stage b is:
[0131]
[0132] Therefore, the stage can be obtained by formula (5).
[0133]
[0134] in, express The maximum probability density belonging to stage b;
[0135] The standardized data It is divided into B stages.
[0136] Since intermittent production cannot achieve ideal repeated production, the stage lengths of different batches will be different. The stage division is based on the Gaussian mixture model. In view of the asynchronous data of each stage, the DTW algorithm is used to translate, compress and stretch the data to synchronize the trajectory of the sub-stage.
[0137] For intermittent processes, the batch with the longest time closest to the normal batch is selected as the reference batch. For each stage of unequal length G = (g1, g2, ..., g i ,…,g m ) and reference batch H = (h1,h2,…,h j ,…,h n ), n represents the number of reference batch samples, m represents the number of unequal length batch samples, then g i and h j The distance can be obtained by formula (7).
[0138] d(g i ,h j ) = g i -h j2 (6)
[0139] The shortest distance D(m,n) and the optimal path W are shown in equations (7)-(8):
[0140]
[0141] W=argmin[D(m,n)](8)
[0142] In the above formula, (gi (l),h j (l)) indicates batches of unequal length g i and reference batch h j The lth point of , max(m,n)≤L≤m+n;
[0143] The optimal path W = {w1,w2,…,w L},w l =(i(l),j(l)) is the sequence searched according to the shortest distance D(m,n) in the m×n grid.
[0144] The constraint allows a point in the grid (g m (l),h n (l)) can only be reached through or at (g m (l-1),h n (l)),(g m (l),h n (l-1)) or (g m (l-1),h n (l-1)). Therefore, the path distance matrix is obtained by formula (9):
[0145]
[0146] Among them, d(g m (l),h n (l)) indicates batches of unequal length g m and reference batch h n The distance of the lth point, D(g m (l-1),h n (l)) indicates batches of unequal length g m The l-1th point and the reference batch h n The shortest distance of the lth point, D(g m (l),h n (l-1)) indicates batches of unequal length g m The lth point and the reference batch h n The shortest distance of the l-1th point, D(g m (l-1),h n (l-1)) indicates batches of unequal length g m The l-1th point and the reference batch h n The shortest distance to the l-1th point;
[0147] Therefore, asymmetric synchronization is achieved, and all trajectories will have the same duration and be synchronized with each other. The above process is repeated to regularize each stage.
[0148] (3) For each stage after division, there is strong nonlinearity between the data. The multi-directional concurrent local weighted regression algorithm uses several local linear models to approximate the nonlinear function f. The present invention introduces the nonlinear function y α =f(x α )+e,x α and α are the input variable and output variable respectively, and e represents the noise with mean 0. In order to measure the locality of the data point, the Gaussian kernel is used to calculate the weight ψ α As shown in formula (10):
[0149]
[0150] D is a positive semidefinite matrix, which is used to determine the size and shape of the local model. Then the weighted mean of the current local model is as shown in formula (11):
[0151]
[0152] In the above formula, N is the number of sampling points. The main task in nonlinear function approximation is to determine the number of local models, calculate the regression model of each local model, and automatically adjust the receptive field of the local model.
[0153] For each local model, local weighted partial least squares is used to determine the parameters. When the first training data sample (x1, y1) is input, a new receptive field is created. Then, the Gaussian kernel weights of all existing receptive fields are calculated. Then, each local model is updated with the corresponding weights.
[0154] In this way, a complex nonlinear problem becomes a weighted combination of several simple linear problems. δ ,y δ ),δ=1,2,...,ΔAfter local weighted projection regression training, the training model has K local linear models RF k To ensure that the input and output mean is zero, it is necessary to subtract the weighted mean from the training data samples. and As shown in (12).
[0155]
[0156] Therefore, the input and output matrices are and Feature output matrix F k It cannot be directly calculated from the input matrix, so the intermediate matrix is constructed as shown in formula (13):
[0157]
[0158] In the above formula, s1=u1, P k =[p1,…,p A ] and U k =[u1,…,u A ] are the load matrix and weight matrix after input and output matrix training, so we get the matrix S k =[s1,s2,…,s A ], F k Calculated by formula (14):
[0159] F k =X k S k (14)
[0160] (4) In order to further refine the statistical analysis space and realize a complete monitoring scheme for quality-related data and process-related data, the monitoring statistics are refined into predictable output subspace, unpredictable output residual subspace, unpredictable output principle subspace, output-irrelevant main subspace, and output-irrelevant residual subspace.
[0161] According to formula (14) and the partial least squares model, we can get And perform singular value decomposition on it, as shown in formula (15):
[0162]
[0163] In the formula, represents the value after orthogonality, It can be calculated by formula (16):
[0164]
[0165] In the formula, The unpredictable output is And use l y The principal components are used for PCA calculation, as shown in formula (17):
[0166]
[0167] In the formula, represents the principal component matrix of the unpredictable output space, represents the unpredictable output space score matrix, represents the unpredictable output space residual matrix.
[0168] Input matrix X k And the output matrix Y k It can be decomposed into formula (18):
[0169]
[0170] In the above formula, and All represent load matrices, For X k Zhongyu The associated score matrix. Represents X k Medium and predicted Y k Irrelevant variables, Represents Y k Cannot be X k The predicted variables, Represents the input matrix X k The residual matrix of .
[0171] Obtaining the statistical indicators of the training data in the offline modeling process, the monitoring statistics of the output subspace can be predicted and calculated according to formula (19):
[0172]
[0173] In the formula, ω is the number of sampling points, The input matrix X k The row vector of yes The row vector of is the predictable output subspace statistic at time k, is the intermediate weighting matrix;
[0174] The monitoring statistics of the unpredictable output main subspace are calculated according to formula (20):
[0175]
[0176] The monitoring statistic of the unpredictable output residual subspace is calculated according to formula (21):
[0177]
[0178] In the formula, is the monitoring statistic of the unpredictable output main subspace at time k, for The row vector of the unpredictable output space residual matrix y mz is the output matrix Y k The row vector of .
[0179] The monitoring statistics of the output irrelevant main subspace are calculated according to formula (22):
[0180]
[0181] The monitoring statistic of the output irrelevant residual subspace is calculated according to formula (23):
[0182]
[0183] In the formula, for The row vector of yes The row vector of
[0184]
[0185] For each data point (x k ,y k ), the global statistics of the multi-directional concurrent local weighted projection regression model are calculated by equations (24) to (27):
[0186]
[0187]
[0188]
[0189]
[0190] The control limits are obtained using kernel density estimation.
[0191] According to step (1), the data of the penicillin fermentation process under real-time working conditions are collected and standardized. According to step (2), the stages are divided, and each stage is modeled using the improved algorithm. Then, the statistic T of the new sample is calculated. c 2 , T x 2 , T y 2 and SPE x ; and determine whether the statistic exceeds the control limit. If it exceeds the control limit, it indicates that a fault has occurred. Otherwise, it is normal and returns to (1) to continue monitoring the next sample. The process monitoring diagram of the multi-directional concurrent local weighted projection regression of the present invention is shown in FIG. Figure 3 shown.
[0192] The above steps are the application of the present invention to obtain key information by performing nonlinear model approximation in each stage after dividing the typical intermittent process into stages. The proposed model is used for fault monitoring of the penicillin fermentation process when the stirring power generates a step signal with an amplitude of 5 between 100-300 sampling points. Figures 4a-4dThe model proposed in the present invention can detect abnormal working conditions timely and accurately, because the intermittent process monitoring method of multi-directional concurrent local weighted projection regression can effectively describe the multi-stage characteristics of the time series and extract local quality-related features.
[0193] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0194] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A batch process monitoring method based on multi-directional concurrent local weighted projection regression, characterized in that: include: S1: Obtain intermittent process variable data and perform standardization; S2: Divide the entire process into stages according to the operating characteristics and time sequence of the intermittent process; S3: For the divided stages, a multi-directional concurrent local weighted projection regression model is established respectively; S4: Refine monitoring statistical data, calculate global statistics and obtain control limits, and determine whether the global statistics exceed the control limits. If they exceed the control limits, it indicates a fault, otherwise it is normal; S2 specifically includes: Determine the standardized data The initial stage preset values are used as the input values of the Gaussian mixture model, and cluster analysis is performed through the Gaussian mixture model. The standardized data The probability density corresponding to each cluster is Where i = 1, 2, ..., m, m is the number of standardized data, j = 1, 2, ..., B, B is the number of Gaussian models; Σ j and μ j represents the jth cluster C j The variance matrix and mean vector of , T represents the transpose; The probability density function of Represented by B Gaussian models: In the above formula, w j represents the mixing weight of the jth component; Based on the probability density function And through Bayesian inference method, we get The probability density of the data belonging to stage b is: The standardized data is analyzed by the maximum value of the probability density Divide into stages: in, express The maximum probability density belonging to stage b; S2 also includes regularizing the divided stages, which specifically includes the following steps: For intermittent processes, the batch with the closest time length to the normal batch is selected as the reference batch. For batches of unequal lengths, G = (g1, g2, ..., g i ,…,g m ) and reference batch H = (h1,h2,…,h j ,…,h n ), n represents the number of reference batch samples, m represents the number of unequal length batch samples, calculate g i and h j The distance d(g i ,h j ): d(g i ,h j )=||g i -h j ||2 By g i and h j Calculate the shortest distance D and the optimal path W: In the above formula, (g i (l),h j (l)) indicates batches of unequal length g i and reference batch h j The lth point of max(m,n)≤L≤m+n; W = argmin[D(m,n)]; The optimal path W = {w1,w2,…,w L },w l =(i(l),j(l)) is the sequence to be searched in the m×n grid according to the shortest distance D(m,n); The constraint allows a point in the grid (g m (l),h n (l)) can only pass through or reach point (g m (l-1),h n (l)), point (g m (l),h n (l-1)) or point (g m (l-1),h n (l-1)); Compute the path distance matrix: Among them, d(g m (l),h n (l)) indicates batches of unequal length g m and reference batch h n The distance of the lth point, D(g m (l-1),h n (l)) indicates batches of unequal length g m The l-1th point and the reference batch h n The shortest distance of the lth point, D(g m (l),h n (l-1)) indicates batches of unequal length g m The lth point and the reference batch h n The shortest distance of the l-1th point, D(g m (l-1),h n (l-1)) indicates batches of unequal length g m The l-1th point and the reference batch h n The shortest distance to the l-1th point; Repeat the above process to regularize each stage; S3 specifically includes: Assume that the nonlinear function y α =f(x α )+e,x α and α are input variables and output variables respectively, e represents noise with a mean of 0; Calculate the weight ψ through the Gaussian kernel α : Where D is a positive semidefinite matrix that determines the size and shape of the local model, x c is the data center of the local model, T represents transposition; Based on the weight ψ α Compute the weighted mean for a multidirectional concurrent locally weighted projection regression model: in, and are the weighted means of the input and output variables of the local weighted projection regression model, respectively, and N is the number of sampling points; Based on the normal sample (x δ ,y δ ),δ=1,2,...,Δ,Δ is the length of the sample, the weighted mean of the input and output variables of the local weighted projection regression model is used to calculate the input matrix X k And the output matrix Y k ; Based on the input matrix X k And the output matrix Y k Construct the intermediate matrix S k , and based on the intermediate matrix S k and the input matrix X k Calculate the feature output matrix F k :F k =X k S k ; Refining the monitoring statistical data into a predictable output subspace, an unpredictable output residual subspace, an unpredictable output principle subspace, an output irrelevant main subspace and an output irrelevant residual subspace; S4 specifically includes: Through the feature output matrix F k and the mass-related load matrix Q k Calculate predicted values And the predicted value Perform singular value decomposition: in, represents the value after orthogonality, is the value of orthogonal decomposition; Scoring Matrix The calculation formula is: In the formula, X k is the input matrix, S k is the intermediate matrix; The unpredictable output is And use l y The principal components are used for PCA calculation: In the formula, represents the principal component matrix of the unpredictable output space, represents the unpredictable output space score matrix, represents the unpredictable output spatial residual matrix; Input matrix X k And the output matrix Y k Perform singular value decomposition: In the formula, and All represent load matrices, For X k Zhongyu The associated score matrix, Represents X k Medium and predicted Y k Irrelevant variables, Represents Y k Cannot be X k The predicted variables, Represents the input matrix X k The residual matrix of Compute monitoring statistics for the predictable output subspace: In the formula, ω is the number of sampling points, x mz The input matrix X k The row vector of yes The row vector of is the predictable output subspace statistic at time k, is the intermediate weighting matrix; Monitoring statistics of the unpredictable output main subspace: In the formula, is the monitoring statistic of the unpredictable output main subspace at time k, for The row vector of the unpredictable output space residual matrix y mz is the output matrix Y k The row vector of Compute monitoring statistics for the unpredictable output residual subspace: In the formula, is the monitoring statistic of the unpredictable output residual subspace at time k; Compute monitoring statistics for the output-independent main subspace: In the formula, Output the monitoring statistics of the irrelevant main subspace at time k, for The row vector of the input matrix X k The row vector of the residual matrix is yes The row vector of For each data point (x k ,y k ), the global statistics calculation formula of the multi-directional concurrent local weighted projection regression model is: In the formula, is the predictable output subspace statistic, For the statistics of the unpredictable output main subspace, SPE y is the unpredictable output residual subspace statistic, To output the monitoring statistics of the irrelevant main subspace, ψ k Calculate weights locally; Input matrix X k And the output matrix Y k The specific calculation process is: Where T represents transpose, x n and n are normal samples, and are the local weighted values of the input and output variables of the local weighted projection regression model respectively; The intermediate matrix construction process is: In the above formula, s1=u1, P k =[p1,…,p A ] and U k =[u1,…,u A ] are the input matrix, the load matrix and the weight matrix after output matrix training respectively. Calculate the intermediate matrix S k =[s1,s2,…,s A ], A is the number of local models established.
2. The intermittent process monitoring method based on multi-directional concurrent local weighted projection regression according to claim 1 is characterized in that: S1 specifically includes: Collect the intermittent process and form three-dimensional data X(I×J×K), where I, J and K represent the batch, variable and sampling time of the data respectively; The collected three-dimensional data X(I×J×K) is expanded into X(I×JK) along the batch direction and standardized: in, represents the standardized data, mean(X(I×J)) and std(X(I×J)) represent the mean and variance of X(I×J), respectively.
