Industrial scene working condition clustering method, system, device and storage medium based on Markov random field model

Abstract

A method, system, device and storage medium for clustering industrial scene working conditions based on a Markov random field model, belonging to the technical field of industrial data analysis. The method of the present invention includes step S100, preliminarily screening out features that can affect the changes in the operating conditions of a power plant, and preliminarily processing them; step S200, performing noise reduction processing on the features that affect the changes in the operating conditions of the power plant after the preliminary processing; step S300, performing dimensionality reduction processing on the feature data that affects the changes in the operating conditions of the power plant after the noise reduction processing; step S400, clustering the data after the dimensionality reduction processing to obtain the operating conditions. The system of the present invention includes a preliminary screening module, a noise suppression module, a feature dimensionality reduction module and a working condition clustering module. The present invention further improves the efficiency of processing and analyzing massive data of the unit, and realizes the high efficiency and economy that cannot be achieved when manually processing massive unit data.

Description

Technical Field

[0001] The present invention relates to a method, system, device and storage medium for constructing working conditions in an industrial scenario, and belongs to the technical field of industrial data analysis. Background Art

[0002] In the field of industrial big data analysis, actual industrial signals often come from many production system sensors, and most of these data change over time. The operating condition is defined as the possible states of the high-dimensional time series signal. For example, the operating state of a thermal power plant can be roughly divided into the startup segment, the operation segment, the stop segment, the abnormal segment, etc. Operating condition mining and identification uses machine learning algorithms to mine the operating condition information hidden in high-dimensional data, which has certain guiding significance for actual production and is also an important pre-processing method for many subsequent algorithms (prediction, classification).

[0003] The existing working condition clustering algorithms mainly include methods based on distance measurement and hierarchical clustering based on data probability distribution. The main problem of clustering algorithms based on distance measurement is that no matter what distance measurement and clustering algorithm are used, the original signal needs to be cut, and the length of the subsequences after cutting must be consistent. Too long subsequences may contain more than one hidden working condition, while too short subsequences may contain fewer data features, which will have an adverse effect on the results of the clustering algorithm, and artificial segmentation of the sequence is not conducive to retaining the time consistency of the original time series. The hierarchical clustering method based on data distribution intuitively determines the segmentation points for clustering according to the distribution characteristics of the data in certain dimensions. The advantage is that the algorithm is intuitive and highly interpretable. The disadvantage is that the features used for hierarchical clustering are often category features with a very concentrated distribution. Unstable or volatile features cannot be used in this type of method, and its use is relatively limited.

[0004] Therefore, there is an urgent need for a working condition clustering method in industrial scenarios to explore potential working conditions in industrial scenarios. Summary of the invention

[0005] In order to realize the potential working condition mining in industrial scenarios with high precision and high efficiency, the present invention provides a brief overview of the present invention below to provide a basic understanding of certain aspects of the present invention. It should be understood that this overview is not an exhaustive overview of the present invention. It is not intended to determine the key or important parts of the present invention, nor is it intended to limit the scope of the present invention.

[0006] The present invention provides an industrial scene working condition clustering method based on a Markov random field model, comprising the following steps:

[0007] Step S100: preliminarily screen out features that can affect the change of the operating condition of the power plant from among the numerous features of the power plant, and preliminarily process these features that can affect the change of the operating condition of the power plant;

[0008] Step S200, performing noise reduction processing on the characteristics of changes in the operating conditions of the power plant that have been initially processed in step S100;

[0009] Step S300, performing dimensionality reduction processing on the characteristic data affecting the change of the operating condition of the power plant after the noise reduction processing in step S200;

[0010] Step S400: clustering the data after dimensionality reduction processing to obtain operating conditions.

[0011] Further: in step S100, among all the existing measurement point data of the coal-fired boiler power plant, data with a time period of 5 seconds and a total time length of 4 months are collected, and features that affect changes in boiler operating conditions are selected.

[0012] Further: In step S200, the method for performing noise reduction processing on the characteristics of the change in the operating conditions of the power plant is: firstly correcting the abnormal data, and then suppressing the high-frequency noise, specifically in the following manner:

[0013] How to handle abnormal data:

[0014] Custom time window value w, from X m Starting from the first data, we take the subsequent w data, and then continuously translate the window forward. Each time, we calculate the standard deviation σ and mean μ of the data in the window. Data outside the range of μ±3σ is considered abnormal data. The average of the 6 data before and after the abnormal data is used as the replacement value to replace the abnormal data. Then, we continuously translate the window forward, and replace the abnormal data each time until all the data in the feature are processed.

[0015] The noise processing method is:

[0016] Custom time window value w, from X m The first data is taken to obtain a total of w data, and then the window is continuously shifted forward. Each time, the data in the window is Gaussian smoothed, that is, the data in the center of the window is weighted averaged according to the Gaussian distribution probability coefficient, the center data of each window is modified, and finally X is used. m The first w / 2 and last w / 2 data are removed.

[0017] The high-frequency noise of the input data is suppressed by Gaussian filtering, wavelet denoising algorithm or digital low-pass filter, and the data anomaly of the input data is corrected by wavelet soft threshold denoising method.

[0018] Further: In step S300, the specific method of performing dimensionality reduction processing on the characteristic data affecting the change of the operating condition of the power plant is as follows:

[0019] Step S310, calculate the correlation coefficient of any two characteristic data that affect the change of the operating conditions of the power plant. When the correlation coefficient is greater than 0.95, remove the features with smaller variance in the two features, that is, remove the motor speed feedback signal of each coal mill, the primary wind speed of the air-powder mixture, the temperature of the air-powder mixture, the mixed air volume characteristics, and the secondary air layer damper opening characteristics at the upper and lower ends of the combustion chamber. At this time, the data feature is reduced to X∈R T×k ;

[0020] Step S320, X∈R T×k The data is reduced in dimension by principal component analysis and singular value decomposition algorithm, and the eigenvector set corresponding to the eigenvalues ​​with the first 95% of the eigenvalue weight is taken as the data after dimensionality reduction, denoted as X∈R T×n .

[0021] Further: in step S400, the data X∈R after the dimension reduction processing of the characteristic data affecting the power plant operating condition change in step S300 is T×n As the input of the working condition clustering, clustering is performed to obtain the operating condition. The specific method includes:

[0022] Step S410: Input data X∈R T×n Initialize and customize clustering parameters to form data in a specific form. Generate clustering result P according to the customized clustering parameters. Define Gaussian inverse covariance matrix Θ with Toeplitz matrix properties for clustering result P. Randomly initialize Toeplitz matrices U and Z according to Gaussian inverse covariance matrix Θ as parameters for subsequent calculations.

[0023] Step S420: According to the clustering result P and the parameter Θ, a value that can characterize the clustering result is obtained. When the change of the characterization value is less than a certain threshold (the threshold is set according to the specific requirements for characterizing the clustering result), or the number of iterations reaches the maximum number of iterations, the iteration is stopped. The clustering result at this time is the final output result.

[0024] Further: it also includes step S430, iteratively updating the existing clusters and their parameters according to the overall cluster optimization goal, and the parameter update is divided into two steps. First, P is updated when Θ remains unchanged. When Θ remains unchanged, the optimization goal is rewritten, and the optimization goal is solved by a dynamic programming algorithm, and P is updated; secondly, the cluster Θ is updated when P remains unchanged. When P remains unchanged, the optimization goal is rewritten, and the optimization goal is solved by an alternating direction multiplier algorithm, and Θ, U, and Z are updated.

[0025] Further: In step S410, data X∈R T×n The method for initializing and forming data in a specific form includes the following steps:

[0026] Step S411: The input data is in the form of a multivariate time series with a dimension of n and a length of T, which can be expressed as x orig ∈R T ×n , x orig =[x1,x2,x3...x T ], where x t ∈R 1×n

[0027] Customize the number of clusters K, the window length w for subsequence segmentation, the maximum number of iterations max_iter, and the regularization coefficients β and π according to actual needs;

[0028] Step S412: x t Mapped to an array X consisting of subsequences t , for each time point t, X t is a subsequence of length w that ends at time t:

[0029] X t =[x t-w+1 , ..., x t ]

[0030] Step S413: Perform random initial clustering according to the user-defined clustering number K, and the result is expressed as:

[0031]

[0032] Step S414: define the Gaussian inverse covariance matrix θ for each cluster i ∈R nw×nw , the matrix is ​​defined as the Toeplitz matrix, expressed as:

[0033]

[0034] Among them A (0) , A (1) , ..., A (w-1) ∈R n×n , represents the correlation between n features (i.e. n sensors) at a certain moment, and Represents the correlation between feature i at time t and feature j at time t+1. Since the model assumes that clustering has time-invariant characteristics, the matrix Θ representing the correlation between time 0 and time 1 0，1 and any matrix Θ representing the time t and t+1 t，t+1 are equal, so Θ i The matrix has the characteristics of a Toeplitz matrix;

[0035] Step S415: for each θ i Randomly initialize the Toeplitz matrices U and Z as model parameters, where U, Z∈Rnw ×nw ;

[0036] Therefore, the parameter initialization module maps the input data into a subsequence array X t , initialize K clusters P, generate an nw×nw Toeplitz matrix Θ for each cluster, and randomly generate the Toeplitz matrices Z and U corresponding to each Θ as model parameters.

[0037] Further: In step S420, the Gaussian inverse covariance matrix Θ optimization problem complies with the following three conditions:

[0038] (1) Assume that X i Satisfy the multivariate Gaussian distribution and minimize its maximum likelihood function value;

[0039] (2) Add a penalty term to encourage consecutive time points to be classified into the same cluster;

[0040] (3) To avoid overfitting, we hope that Θ is as sparse as possible;

[0041] Based on the above three points, we can get the optimization goal of the problem:

[0042]

[0043] Where ll(X i , Θ i ) is the maximum likelihood term, and its expression is:

[0044]

[0045] ll(X i , Θ i ) i Belongs to cluster P i Medium X t The mean of

[0046] Therefore, the clustering result characterization value expression is:

[0047]

[0048] Further: In step S430, the existing clusters and their parameters are iteratively updated according to the overall cluster optimization goal, the purpose of which is to:

[0049]

[0050] For P = {P1, ..., P k} and Θ={Θ1, ..., Θ k} to update;

[0051] The update is divided into two steps. The first step is to update P = {P1, ..., P k}; Step 2: Update clustering Θ = {Θ1, ..., Θ k}.

[0052] Update first step:

[0053] When Θ remains unchanged, the optimization objective is:

[0054]

[0055] Solving this optimization problem is equivalent to a trade-off between maximizing the log-likelihood and making the clustering as continuous as possible in time. The importance of temporal continuity is controlled by the parameter β.

[0056] This optimization problem can be solved using a dynamic programming algorithm. The directed graph of the dynamic programming algorithm is shown in the figure below.

[0057] From the perspective of dynamic programming algorithm, this optimization problem is equivalent to: solving the shortest path from 1 to T in the directed graph, where the value of the node (i, j) in the graph is when X i The negative log-likelihood value of cluster j, j∈1,...,K is assigned, and when a node has an edge connected to different clusters, the weight of the edge is set to β. The algorithm complexity is O(KT).

[0058] Therefore, through the above method, the optimization problem can be solved by using the dynamic programming algorithm when Θ remains unchanged. k} to update.

[0059] Update Step 2:

[0060] When P remains unchanged, the maximum likelihood term can be rewritten as:

[0061]

[0062]

[0063] Since θ i The sparsity of , and the Toeplitz condition must be satisfied, the optimization objective can be rewritten as:

[0064]

[0065]

[0066] To solve this optimization problem, it needs to be repeated at least K times each time, and the overall algorithm complexity is relatively high. In order to reduce the computational complexity, the alternating direction multiplier algorithm (ADMM) with parallel computing characteristics is used to solve this optimization problem.

[0067] Using the ADMM algorithm, introducing the consistency variable Z and the dual variable U, the optimization problem can be rewritten as:

[0068]

[0069]

[0070] The corresponding augmented Lagrangian function is:

[0071]

[0072] Where ρ>0 is the penalty coefficient of the ADMM algorithm.

[0073] The ADMM algorithm performs the following three steps repeatedly until convergence, and the stopping condition is set when the residual of Θ and Z at the end of two iterations is less than the threshold:

[0074] (a)

[0075] (b)

[0076] (c)U k+1 :=U k +(Θ k+1 -Z k+1 ),

[0077] In the above iterative steps, steps (a) and (b) belong to the optimization functions of Θ and Z, and solving the optimization functions can obtain new Θ and Z. According to the new Θ and Z, a new U can be obtained through step (c). The updating method of Θ and Z is described as follows.

[0078] Θ update: The update of Θ parameters is independent of Z and U variables and can be written as follows:

[0079]

[0080] This optimization problem has a known analytical solution:

[0081]

[0082] Among them, QDQ T is the matrix ρ(Z k -U k )-The eigenvalue decomposition result of S. is the maximum likelihood estimate of the sample covariance matrix under the multivariate normal distribution.

[0083] Z update: The Z parameter update can be written as follows:

[0084]

[0085] Defining variables Indicates the lth occurrence of A in the Z matrix (m) The block position is the element at (i, j), which can represent the position of any element in the Z matrix. Values, set the values ​​of these positions in the Z matrix to:

[0086]

[0087] make

[0088] The following closed-form solution of Z can be obtained:

[0089]

[0090] Therefore, by updating the first step and the second step, we can get a new P = {P1, ..., P k} and Θ={Θ1, ..., Θ k}, and pass the updated results to the clustering assignment submodule.

[0091] An industrial scene working condition clustering system based on a Markov random field model, comprising a preliminary screening module, a noise suppression module, a feature dimension reduction module and a working condition clustering module;

[0092] The preliminary screening module is used to preliminarily screen out the features that can affect the changes in the operating conditions of the power plant from the numerous features of the power plant, and perform preliminary processing on these features;

[0093] The noise suppression module is used to screen abnormal data in the power plant operation data and correct the noise;

[0094] The feature dimension reduction module is used to perform dimension reduction processing on the feature data that affects the change of the operating condition of the power plant after the noise reduction processing;

[0095] The operating condition clustering module clusters the data after dimensionality reduction processing to obtain the operating condition.

[0096] Further: the working condition clustering module includes a parameter initialization submodule, a clustering assignment submodule and a parameter updating submodule;

[0097] The parameter initialization submodule is used to initialize the input data and customize the clustering parameters to form data in a specific form for subsequent clustering operations;

[0098] The clustering value assignment submodule obtains a value that can characterize the clustering result according to the clustering result parameters;

[0099] The parameter updating submodule iteratively updates the existing clusters and their parameters according to the overall cluster optimization objective.

[0100] A device for clustering industrial scene working conditions based on a Markov random field model comprises a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the steps of the aforementioned method for clustering industrial scene working conditions based on a Markov random field model are implemented.

[0101] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the aforementioned industrial scene working condition clustering method based on a Markov random field model.

[0102] The present invention has the following beneficial effects:

[0103] 1. The present invention can smoothly realize the clustering of steady-state load conditions of industrial production units, and extract the steady-state conditions of massive data of the units by screening, suppressing, reducing dimensions and clustering industrial data, thus providing effective and valuable data support for subsequent calculation of unit performance indicators and diagnosis and analysis of equipment operation status;

[0104] 2. The present invention further improves the efficiency of processing and analyzing massive data of the unit, and achieves high efficiency and economy that cannot be achieved when manually processing massive data of the unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] Figure 1 It is a clustering system diagram of industrial scene working conditions based on Markov random field model;

[0106] Figure 2 It is a flow chart of the industrial scenario working condition clustering method based on the Markov random field model;

[0107] Figure 3 It is a directed graph of dynamic programming algorithm. DETAILED DESCRIPTION

[0108] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is described below by the specific embodiments shown in the accompanying drawings. However, it should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention. Specific implementation method one:

[0110] The purpose of this embodiment is to provide an industrial scenario working condition clustering method and system based on the Markov random field model. The real-time operating data of the power plant is obtained by arranging sensors at various locations in the power plant, and the characteristic data that can affect the changes in the boiler working conditions are selected as much as possible through manual screening. The data is processed through data cleaning, noise suppression, feature dimensionality reduction and other methods, and then the data is clustered, and each category is regarded as a different working condition. This embodiment is mainly divided into a preliminary screening module; a noise suppression module; a feature dimensionality reduction module; a working condition clustering module;

[0111] Initial screening module

[0112] The function of this module is to preliminarily select the features that can affect the changes in the operating conditions of the power plant from the many features of the power plant, and perform preliminary processing on these features.

[0113] Among the data collected by all existing measuring points in a coal-fired boiler power plant for a certain time period (e.g., 5 seconds) and a total time length of 2 months, select the features that can affect the change of boiler operating conditions. Features include total coal supply, total water supply, total air supply, boiler load, coal mill data, and combustion chamber data. Coal mill data includes the motor speed feedback signal, coal supply feedback signal, air-powder primary wind speed, air-powder mixture temperature, and mixed air volume of each coal mill. The combustion chamber data includes the damper opening of the four corners of the sofa air layer, the primary air layer at the upper and lower ends, and the secondary air layer, for a total of 106 features.

[0114] The selected characteristic data are preliminarily processed, including obtaining the water-coal ratio according to the ratio of the total water supply and the total coal supply, obtaining the air-coal ratio according to the ratio of the total air supply and the total coal supply, and taking the average of the four-corner damper opening data of each air layer in the combustion chamber to obtain the average damper opening of each air layer. Then the total water supply, total air supply, and the four-corner damper opening of each air layer are eliminated. Save as preliminary screening data. The preliminary screening data dimensions are 51.

[0115] The preliminary screening data is saved as the output of this module and as the input of the noise suppression module.

[0116] Noise Suppression Module

[0117] The function of this module is to reduce the noise in the input data. There are various types of noise in the actual power plant operation data, which can be divided into data anomalies caused by the fluctuation of industrial sensors themselves and accidental distortion of industrial sensors.

[0118] This module filters out abnormal data based on the standard deviation and replaces the abnormal data with the mean of the data around the abnormal data; Gaussian filtering method is used to correct the noise.

[0119] The noise suppressed data is used as the output of this module and as the input of the feature dimension reduction module.

[0120] Feature Dimensionality Reduction Module

[0121] The function of this module is to reduce the dimension of the input data to reduce the calculation difficulty and save calculation time.

[0122] This module removes redundant features that show positive linear correlation through correlation coefficient method and variance selection method, and can remove the motor speed feedback signal, primary wind speed of air-powder mixture, temperature of air-powder mixture, and mixed air volume features of each coal mill. It can also remove the secondary air layer damper opening features at the upper and lower ends of the combustion chamber, reducing the dimensions to 19; and then reduce the dimensions of the features through principal component analysis (PCA) and singular value decomposition (SVD) algorithms, reducing the dimensions to 8. The features of the data processed by PCA and SVD no longer have a definite physical meaning.

[0123] The data after feature dimension reduction is used as the output of this module and as the input of the working condition clustering algorithm.

[0124] Working condition clustering module

[0125] The function of this module is to cluster the input data. According to the clustering results, the timestamp of each category is obtained, and the boiler operation data can be divided into different working conditions.

[0126] This module is divided into parameter initialization submodule, clustering assignment submodule and parameter update submodule.

[0127] The parameter initialization submodule is used to initialize the input data and customize the clustering parameters to form a specific form of data for subsequent clustering operations. According to the customized clustering parameters, the clustering result P is generated; the Gaussian inverse covariance matrix Θ with the properties of the Toeplitz matrix is ​​defined for the cluster P; the Toeplitz matrices U and Z are randomly initialized according to Θ as parameters for subsequent calculations.

[0128] The clustering assignment submodule is used to obtain a value that can characterize the clustering result based on the clustering result P and the parameter Θ. When the change in the characterization value is less than a certain threshold, or the number of iterations reaches the maximum number of iterations, the iteration is stopped, and the clustering result at this time is the final output result.

[0129] The function of the parameter update submodule is to iteratively update the existing clusters and their parameters according to the overall cluster optimization goal. The parameter update is divided into two steps. First, P is updated when Θ remains unchanged, and then the cluster Θ is updated when P remains unchanged. When Θ remains unchanged, the optimization goal is rewritten, the dynamic programming algorithm is used to solve the optimization goal, and P is updated. When P remains unchanged, the optimization goal is rewritten, the alternating direction multiplier algorithm (ADMM) is used to solve the optimization goal, and Θ, U, and Z are updated. Each update of the clustering results and parameters is an iteration. The updated clustering results and parameters are passed to the cluster assignment submodule. Specific implementation method 2:

[0131] This implementation method is to obtain the real-time operation data of the power plant by arranging sensors at various locations in the power plant, and perform basic data processing through data cleaning, feature dimension reduction and other methods, and then input the data into the operating condition clustering module to cluster the boiler data, and each category is regarded as a different operating condition. The process is as follows Figure 1 and Figure 2 As shown, where:

[0132] 1. Preliminary screening module

[0133] The function of the preliminary screening module is to preliminarily screen out the features that can affect the changes in the operating conditions of the power plant from the many features of the power plant, and perform preliminary processing on these features.

[0134] Among the data collected by all existing measuring points in coal-fired boiler power plants for a certain time period (e.g., 5 seconds) and a total time length of 4 months, select the features that can affect the changes in boiler operating conditions. Features include total coal feed, total water feed, total air feed, boiler load, coal mill data, and combustion chamber data. Coal mill data includes the motor speed feedback signal, coal feed feedback signal, air-powder primary wind speed, air-powder mixture temperature, and mixed air volume of each coal mill. The combustion chamber data includes the damper openings of the four corners of the sofa air layer, the primary air layer at the upper and lower ends, and the secondary air layer, for a total of 106 features.

[0135] The selected characteristic data are preliminarily processed, including obtaining the water-coal ratio according to the ratio of the total water supply and the total coal supply, obtaining the air-coal ratio according to the ratio of the total air supply and the total coal supply, and taking the average of the four-corner damper opening data of each wind layer in the combustion chamber to obtain the average damper opening of each wind layer. Then the total water supply, total air supply, and the four-corner damper opening of each wind layer are eliminated. Save as preliminary screening data. The preliminary screening data format is X = {X1,...,X M}, X∈R T , that is, the data has M-dimensional features, and each dimension feature is a time series with a length of T. Among them, M = 51, and T is about 2 million.

[0136] The preliminary screening data is used as the output of this module and as the input of the noise suppression module.

[0137] 2. Noise Suppression Module

[0138] The function of the noise suppression module is to reduce the noise in the input data. There are various types of noise in the actual power plant operation data. The noise can be divided into data anomalies caused by the fluctuation of the industrial sensor itself and the accidental distortion of the industrial sensor. Therefore, it is necessary to reduce the noise of the input data.

[0139] The input data has M = 51 dimensional features, each dimensional feature is a time series of length T, and any dimensional feature data is recorded as X m .

[0140] The method for handling abnormal data is:

[0141] Custom time window value w (optional w = 100), from X m Starting from the first data, we take the subsequent w data, and then continuously shift the window forward. Each time, we calculate the standard deviation σ and mean μ of the data in the window. Data outside the range of μ±3σ is considered abnormal data, and the average of the 6 data before and after the abnormal data is used as the replacement value to replace the abnormal data. Then, we continuously shift the window forward, and replace the abnormal data each time. This is until all the data in the feature are processed.

[0142] The noise processing method is:

[0143] Custom time window value w (optional w = 100), from X m The first data is taken to obtain a total of w data, and then the window is continuously shifted forward. Each time, the data in the window is Gaussian smoothed, that is, the data in the center of the window is weighted averaged according to the Gaussian distribution probability coefficient, and the center data of each window is modified. Finally, X m The first w / 2 and last w / 2 data are removed.

[0144] Through the above method, the abnormal data is first corrected, and then the high-frequency noise is suppressed. The noise-suppressed data has M-dimensional features, and the length of each dimension feature is still recorded as T. The processed data is saved as the output of this module and as the input of the feature dimensionality reduction module.

[0145] 3. Feature Dimensionality Reduction Module

[0146] The function of the feature dimension reduction module is to reduce the dimension of the input data to reduce the calculation difficulty and save calculation time.

[0147] The specific method of dimensionality reduction is as follows:

[0148] Step 1: Calculate the correlation coefficient of any two feature data. When the correlation coefficient is greater than 0.95, remove the feature with smaller variance in the two features. Through this step, the motor speed feedback signal, the primary wind speed of the air-powder mixture, the temperature of the air-powder mixture, and the mixed air volume characteristics of each coal mill can be removed. The secondary air layer damper opening characteristics at the upper and lower ends of the combustion chamber can also be removed. At this time, the data feature is reduced to k = 19, the length is T, and it is recorded as X∈R T×k .

[0149] Step 2: X∈R T×k The data is reduced in dimension using principal component analysis (PCA) and singular value decomposition (SVD) algorithms. PCA and SVD algorithms are mature and common algorithms in the field of data mining. The eigenvector set corresponding to the eigenvalues ​​with the top 95% of the eigenvalue weights is taken as the reduced-dimensional data. At this time, the data features are reduced to n = 8, with a length of T, denoted as X∈R T×n The features of data processed by PCA and SVD no longer have exact physical meanings.

[0150] The dimension of the data after feature dimension reduction is n=8 and the length is T. The data after dimension reduction is the output of this module and serves as the input of the working condition clustering algorithm.

[0151] 4. Working Condition Clustering Module

[0152] The function of the operating condition clustering module is to cluster the input data. According to the clustering results, the timestamp of each category is obtained, and the boiler operation data can be divided into different operating conditions.

[0153] The input data dimension of this module is n=8, the length is T, and it is recorded as X∈R T×n Each dimension feature has no exact physical meaning. All processes in this module are pure mathematical operations.

[0154] The working condition clustering module can be divided into parameter initialization submodule, clustering assignment submodule and parameter updating submodule.

[0155] Parameter initialization submodule:

[0156] The input data of the parameter initialization submodule is the output data of the feature dimension reduction module. Its function is to initialize the input data and form data in a specific form for subsequent calculations.

[0157] The input data is in the form of a multivariate time series with dimension n and length T, which can be expressed as x orig ∈R T×n :

[0158] x orig =[x1,x2,x3...x T ]

[0159] where xt ∈R 1×n .

[0160] Step 1: You need to customize the number of clusters K, the window length w for subsequence segmentation, the maximum number of iterations max_iter, and the regularization coefficients β and ρ according to actual needs.

[0161] Step 2: x t Mapped to an array X consisting of subsequences t For each time point t, X t To terminate at time t, the length of the subsequence is w. The purpose of this step is to encourage consecutive time points to belong to the same cluster:

[0162] X t =[x t-w+1 , ..., x t ]

[0163] Step 3: Perform random initial clustering according to the custom cluster number K. The result is expressed as

[0164] Step 4: Define the Gaussian inverse covariance matrix Θ for each cluster i ∈R nw×nw , the matrix is ​​defined as the Toeplitz matrix, expressed as:

[0165]

[0166] Among them A (0) , A (1) , ..., A (w-1) ∈R n×n , represents the correlation between n features (i.e. n sensors) at a certain moment, and Represents the correlation between feature i at time t and feature j at time t+1. Since the model assumes that clustering has time-invariant characteristics, the matrix Θ representing the correlation between time 0 and time 1 0，1 and any matrix Θ representing the time t and t+1 t，t+1 are equal, so Θ i The matrix has the characteristics of a Toeplitz matrix (a symmetric matrix with equal elements on each diagonal).

[0167] Step 5: for each Θ i Randomly initialize the Toeplitz matrices U and Z as model parameters, where U, Z∈R nw ×nw .

[0168] Therefore, the parameter initialization module maps the input data into a subsequence array X t, initialize K clusters P, generate the nw×nw Toeplitz matrix Θ of each cluster, and randomly generate the Toeplitz matrices Z and U corresponding to each Θ as model parameters. The data is passed to the cluster assignment module.

[0169] Clustering assignment submodule:

[0170] The clustering assignment submodule is used to obtain a value that can characterize the clustering result based on the clustering result P and the parameter Θ. When the change of the characterization value is less than a certain threshold, or the number of iterations reaches the maximum number of iterations max_iter, the iteration is stopped. At this time, the clustering result P = {P1, ..., P k} is the final output result.

[0171] Θ optimization problem follows the following three conditions: 1. Assume that X i Satisfy the multivariate Gaussian distribution and minimize its maximum likelihood function value; 2. Add penalty terms to encourage continuous time points to be classified into the same cluster; 3. Avoid overfitting and hope that Θ is as sparse as possible. Based on the above three points, the optimization goal of the problem can be obtained:

[0172]

[0173] Where ll(X i , Θ i ) is the maximum likelihood term, and its expression is:

[0174]

[0175] ll(X i , Θ i ) i Belongs to cluster P i Medium X t The mean of .

[0176] Therefore, the clustering result characterization value expression is:

[0177]

[0178] Parameter update submodule:

[0179] The parameter update submodule is used according to the optimization goal:

[0180]

[0181] For P = {P1, ..., P k} and Θ={Θ1, ..., Θ k} to update.

[0182] The update is divided into two steps: 1. Update P = {P1, ..., P k}; 2 Update clustering Θ={Θ1,...,Θ k}.

[0183] Update step 1:

[0184] When Θ remains unchanged, the optimization objective is:

[0185]

[0186] Solving this optimization problem is equivalent to a trade-off between maximizing the log-likelihood and making the clustering as continuous as possible in time. The importance of temporal continuity is controlled by the parameter β.

[0187] This optimization problem can be solved by using a dynamic programming algorithm. The directed graph of the dynamic programming algorithm is as follows: Figure 3 shown.

[0188] From the perspective of dynamic programming algorithm, this optimization problem is equivalent to: solving the shortest path from 1 to T in the directed graph, where the value of the node (i, j) in the graph is when X i The negative log-likelihood values ​​of clusters i, j∈1,…,K are assigned, and when a node has an edge connected to different clusters, the weight of the edge is set to β. The algorithm complexity is O(KT).

[0189] Therefore, through the above method, the dynamic programming algorithm can be used to solve the optimization problem when Θ remains unchanged. k} to update.

[0190] Update step 2:

[0191] When P remains unchanged, the maximum likelihood term can be rewritten as:

[0192]

[0193]

[0194] Since θ i The sparsity of , and the Toeplitz condition must be satisfied, the optimization objective can be rewritten as:

[0195]

[0196]

[0197] To solve this optimization problem, it needs to be repeated at least K times each time, and the overall algorithm complexity is relatively high. In order to reduce the computational complexity, the alternating direction multiplier algorithm (ADMM) with parallel computing characteristics is used to solve this optimization problem.

[0198] Using the ADMM algorithm, introducing the consistency variable Z and the dual variable U, the optimization problem can be rewritten as:

[0199]

[0200]

[0201] The corresponding augmented Lagrangian function is:

[0202]

[0203] Where ρ>0 is the penalty coefficient of the ADMM algorithm.

[0204] The ADMM algorithm performs the following three steps repeatedly until convergence, and the stopping condition is set when the residual of Θ and Z at the end of two iterations is less than the threshold:

[0205] (a)

[0206] (b)

[0207] (c)U k+1 :=U k +(Θ k+1 -Z k+1 ),

[0208] In the above iterative steps, steps (a) and (b) belong to the optimization functions of Θ and Z, and solving the optimization functions can obtain new Θ and Z. According to the new Θ and Z, a new U can be obtained through step (c). The updating method of Θ and Z is described as follows.

[0209] Θ update: The update of Θ parameters is independent of Z and U variables and can be written as follows:

[0210]

[0211] This optimization problem has a known analytical solution:

[0212]

[0213] Among them, QDQ T is the matrix ρ(Z k -U k )-The eigenvalue decomposition result of S.

[0214] in is the maximum likelihood estimate of the sample covariance matrix under the multivariate normal distribution.

[0215] Z update: The Z parameter update can be written as follows:

[0216]

[0217] Defining variables Indicates the lth occurrence of A in the Z matrix (m) The block position is the element at (i, j), which can represent the position of any element in the Z matrix. Values, set the values ​​of these positions in the Z matrix to:

[0218]

[0219] make

[0220] The following closed-form solution of Z can be obtained:

[0221]

[0222] Therefore, through updating step 1 and updating step 2, a new P = {P1, ..., P k} and Θ={Θ1, ..., Θ k}, and pass the updated results to the clustering assignment submodule. Specific implementation method three:

[0224] Those skilled in the art can provide the system and method mentioned in the above embodiments as a method, a system, or a computer program product. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware, and the modules can also be reorganized according to the computer logic structure. Moreover, the present embodiment can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0225] The flowchart or block diagram of the method, device (system), and computer program product according to the present embodiment is described. It should be understood that each process or box in the flowchart or block diagram, as well as the combination of the processes or boxes in the flowchart or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the process in the flowchart. Figure 1 A process or multiple processes or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0226] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes or boxes Figure 1 A function specified in one or more boxes.

[0227] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0228] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0229] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0230] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0231] This implementation is only an exemplary description of this patent and does not limit its protection scope. Those skilled in the art may also make partial changes to it, as long as they do not exceed the spirit of this patent, they are all within the protection scope of this patent.

Claims

1. An industrial scene working condition clustering method based on Markov random field model, characterized in that: The following steps are involved: Step S100: preliminarily screen out features that can affect the change of the operating condition of the power plant from among the numerous features of the power plant, and preliminarily process these features that can affect the change of the operating condition of the power plant; Step S200, performing noise reduction processing on the characteristics of changes in the operating conditions of the power plant that have been initially processed in step S100; Step S300, performing dimensionality reduction processing on the characteristic data affecting the change of the operating condition of the power plant after the noise reduction processing in step S200, the specific method is as follows: Step S310, calculate the correlation coefficient of any two characteristic data that affect the change of the operating conditions of the power plant. When the correlation coefficient is greater than 0.95, remove the features with smaller variance in the two features, that is, remove the motor speed feedback signal of each coal mill, the primary wind speed of the air-powder mixture, the temperature of the air-powder mixture, the mixed air volume characteristics, and the secondary air layer damper opening characteristics at the upper and lower ends of the combustion chamber. At this time, the data characteristics are reduced to ; Step S320: The data is reduced in dimension through principal component analysis and singular value decomposition algorithm, and the eigenvector set corresponding to the eigenvalues ​​with the first 95% of the eigenvalue weight is taken as the data after dimensionality reduction, which is recorded as ; Step S400: clustering the data after the dimension reduction process to obtain the operating conditions. The specific method includes: Step S410: Input data Initialize and customize clustering parameters to form data in a specific form. Generate clustering result P based on customized clustering parameters, and define Gaussian inverse covariance matrix with Toeplitz matrix properties for clustering result P , according to the Gaussian inverse covariance matrix Randomly initialize the Toeplitz matrices U and Z as parameters for subsequent calculations; Step S420: Based on the clustering result P and parameters , get the value that can represent the clustering result. When the change of the representation value is less than a certain threshold, or the number of iterations reaches the maximum number of iterations, the iteration is stopped, and the clustering result at this time is the final output result; The step S430 is also included, in which the existing clusters and their parameters are iteratively updated according to the overall cluster optimization goal. The parameter update is divided into two steps. First, Update P without changing Under the condition that P remains unchanged, rewrite the optimization objective, use the dynamic programming algorithm to solve the optimization objective, and update P; then update the clustering under the condition that P remains unchanged , under the condition that P remains unchanged, rewrite the optimization objective and use the alternating direction multiplier algorithm to solve the optimization objective. , U, and Z are updated, where the specific method of using the alternating direction multiplier algorithm is: By using the alternating direction multiplier algorithm and introducing the consistency variable Z and the dual variable U, the optimization problem can be rewritten as: ； The corresponding augmented Lagrangian function is: ； in , is the penalty coefficient of the alternating direction multiplier algorithm; The alternating direction multiplier algorithm performs the following three steps repeatedly until convergence, and the stopping condition is set to The resulting residual of Z at the end of two iterations is less than the threshold: Step a. ; Step b: ; Step c. ; In the above iterative steps, steps a and b belong to and Z optimization function, the solution optimization function can get a new and Z, according to the new and Z, we can get new U, The update methods of and Z are as follows: renew: The update of parameters is independent of the Z and U variables and can be written as follows: ； This optimization problem has a known analytical solution: ； in For the matrix The eigenvalue decomposition result of , is the maximum likelihood estimate of the sample covariance matrix under the multivariate normal distribution; Z update: The Z parameter update can be written as follows: ； Defining variables Indicates that in the Z matrix Occurrence The block location is The element at which this variable can represent the position of any element in the Z matrix. For each group Values, set the values ​​of these positions in the Z matrix to: ； make ; The following closed-form solution of Z can be obtained: ； Therefore, by updating the first step and updating the second step, we can get the new and , cluster the update results.

2. The industrial scene working condition clustering method based on the Markov random field model according to claim 1 is characterized by: In step S100, from all the existing measurement point data of the coal-fired boiler power plant, data with a time period of 5 seconds and a total time length of 4 months are collected, and features that affect changes in boiler operating conditions are selected.

3. The industrial scene working condition clustering method based on the Markov random field model according to claim 1 is characterized in that: In step S200, the method for performing noise reduction processing on the characteristics of the change in the operating conditions of the power plant is: firstly, the abnormal data is corrected, and then the high-frequency noise is suppressed. The specific method is: How to handle abnormal data: Custom time window value w, from Starting from the first data, we take the next w data, and then move the window forward continuously, and calculate the standard deviation of the data in the window each time. and mean , appears in Data outside the range is considered abnormal data. The average of the six data before and after the abnormal data is used as the replacement value to replace the abnormal data. Then the window is continuously shifted forward, and the abnormal data is replaced each time until all the data in the feature is processed; The noise processing method is: Custom time window value w, from The first data is used to obtain the subsequent w data, and then the window is continuously shifted forward. Each time, the data in the window is Gaussian smoothed, that is, the data in the center of the window is weighted averaged according to the Gaussian distribution probability coefficient, and the center data of each window is modified. The front and finally Data removed.

4. An industrial scene working condition clustering system based on a Markov random field model is implemented based on an industrial scene working condition clustering method based on a Markov random field model according to any one of claims 1 to 3, characterized in that: It includes preliminary screening module, noise suppression module, feature dimension reduction module and working condition clustering module; The preliminary screening module is used to preliminarily screen out the features that can affect the changes in the operating conditions of the power plant from the numerous features of the power plant, and perform preliminary processing on these features; The noise suppression module is used to screen abnormal data in the power plant operation data and correct the noise; The feature dimension reduction module is used to perform dimension reduction processing on the feature data that affects the change of the operating condition of the power plant after the noise reduction processing; The operating condition clustering module clusters the data after dimensionality reduction processing to obtain the operating condition.

5. The industrial scene working condition clustering system based on the Markov random field model according to claim 4 is characterized in that: The working condition clustering module includes a parameter initialization submodule, a clustering assignment submodule and a parameter updating submodule; The parameter initialization submodule is used to initialize the input data and customize the clustering parameters to form data in a specific form for subsequent clustering operations; The clustering value assignment submodule obtains a value that can characterize the clustering result according to the clustering result parameters; The parameter updating submodule iteratively updates the existing clusters and their parameters according to the overall cluster optimization objective.

6. An industrial scene working condition clustering device based on a Markov random field model, characterized in that: It comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of an industrial scene working condition clustering method based on a Markov random field model as described in any one of claims 1 to 3 when executing the computer program.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for clustering industrial scene working conditions based on a Markov random field model as described in any one of claims 1 to 3 is implemented.

Images