A high-voltage heater online monitoring method based on parallel GMM-LCKSVD
The clustering and monitoring of high-pressure heater data through parallel GMM and LCKSVD dictionary learning algorithms solves the problem of difficulty in mining the essential characteristics of data in the existing technology, realizes efficient fault diagnosis and monitoring, and improves the economicality of power plant operation and Security.
Patent Information
- Application Number
- CN202210676562.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-15
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2042-06-15
AI Technical Summary
The prior art is difficult to mine essential characteristics from high-pressure heater operation data, resulting in poor results in high-pressure heater status monitoring and fault warning, especially when the data distribution modes are diversified during long-term wide load operation of high-pressure heaters.
The parallel GMM algorithm is used for data clustering, and the model is established in combination with the LCKSVD dictionary learning algorithm, and the high-pressure heater is monitored and fault diagnosis is performed online, and fault parameters are located using reconstruction methods.
Real-time monitoring and fault warning of the operating status of high-pressure heaters is achieved, the monitoring effect and fault isolation are improved, and economic and safety risks in power plant operation are reduced.
Smart Images

Figure CN115130564B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of high-voltage heater fault diagnosis, and in particular to an online monitoring method for a high-voltage heater based on parallel GMM-LCKSVD. Background Art
[0002] As one of the core systems of a coal-fired power plant, the safe and stable operation of the steam-water system directly determines whether the entire unit can generate electricity normally. The high-pressure heater is an important equipment in the steam-water system, which can improve the thermal efficiency of the power plant and reduce coal consumption.
[0003] If the high-pressure heater cannot operate normally, the output of the generator will generally be reduced by about 10%, and the unit's heat consumption and power coal consumption will also increase. At the same time, the failure of the high-pressure heater may also cause water hammer in the turbine, increase the expansion difference of the turbine, reduce the safety and stability of boiler operation, and even threaten personal safety in severe cases. Therefore, the status monitoring and fault warning of the high-pressure heater are of great significance to the safety and economic operation of the unit.
[0004] In the past two decades, with the development of computer storage and communication technology, my country's thermal power units have built highly automated systems such as distributed control systems (DCS), plant-level information management systems (MIS) and plant-level information monitoring systems (SIS).
[0005] On the one hand, these systems improve the automation level of the entire power plant, and on the other hand, they also store a large amount of historical operating data of the unit equipment, which provides a solid foundation for the online monitoring and fault warning of the unit equipment.
[0006] In recent years, scholars have studied data-driven modeling techniques such as PCA and deep learning to mine the relationships between operating parameters from massive historical data of high-pressure booster operations and establish fault diagnosis models by processing and analyzing these relationships.
[0007] With the continuous improvement of unit parameters, as well as the increasing frequency of deep peak regulation and equipment start-up and shutdown, real-time status monitoring and fault warning of equipment running in the unit have become an important part of the construction of smart power plants.
[0008] However, these methods do not obtain sparse representation of data, making it difficult to mine the essential characteristics of the data.
[0009] At the same time, since thermal power units frequently participate in deep peak regulation, the high-voltage heater will operate in a wide load range for a long time, which will cause the historical data of the high-voltage heater operation to often show many different distribution patterns. At this time, if the samples are not processed and all historical data are directly used to establish a process monitoring model, ideal results will not be achieved.
[0010] Therefore, in-depth research on the distribution characteristics of high-pressure heater operation data and fully extracting the essential characteristics of the data are of great practical significance for improving the monitoring effect of the model. Summary of the invention
[0011] The technical problem to be solved by the present invention is as follows: In order to solve the problem that it is difficult to mine the essential characteristics of the high-voltage heater operation data in the prior art, and to meet the needs of real-time status monitoring and fault warning of the high-voltage heater, the present invention proposes a new online monitoring method and fault diagnosis method for the high-voltage heater based on parallel GMM and label consistent dictionary learning algorithm.
[0012] Technical solution of the present invention: The present invention discloses a high-voltage heater online monitoring method based on parallel GMM-LCKSVD, comprising the following steps:
[0013] S1. Use the parallel GMM algorithm to quickly cluster the original data and divide the data into multiple data sets;
[0014] S2. Use LCKSVD dictionary learning algorithm to establish a model and perform online monitoring of the high-voltage heater;
[0015] S3. After the fault is detected, the fault isolation method based on reconstruction is used to locate the parameters where the fault occurred.
[0016] Furthermore, the S1 specifically includes the following steps:
[0017] SA1. Cluster the data set using the parallel GMM algorithm:
[0018] Assume that the data sample set Y = {y1, y2, ...y m}, where: m represents the number of samples, sample y i ={y i1 ,y i2 ,...y in}, subscript i represents the sample number, and n represents the number of parameters;
[0019] Randomly divide the data sample set Y into N independent parts and specify the number of clusters K;
[0020] for and use and To update the Gaussian mixture model parameters of each part;
[0021] Where: m t represents the number of samples in the tth part;
[0022] l g,i The calculation method is:
[0023] Where: g is the weight of each Gaussian part; ∑ g is the covariance; μ g For expectations;
[0024] SA2. Use OPENMP to parallelize the calculation of the N-part mixed model parameters. After the calculation is completed, use To update the parameters of the entire GMM model;
[0025] SA3. Repeat the above calculation process until the number of iterations is greater than the preset maximum number of iterations or the error of the likelihood function is less than the given error;
[0026] SA4. After the calculation is completed, the posterior probability of each sample is calculated using the Bayesian formula, and the maximum posterior probability is defined as the clustering label of the sample.
[0027] Furthermore, the specific steps of S2 are as follows:
[0028] SB1. Establish LCKSVD dictionary learning model:
[0029] According to the clustering results of the parallel GMM algorithm, the sample label matrix H of the data sample set Y is constructed, and then Y is standardized and transposed, and the LCKSVD dictionary learning objective function about the data sample set Y is constructed to obtain the dictionary D;
[0030] The objective function of LCKSVD dictionary learning for data sample set Y is:
[0031]
[0032] Where: D is the dictionary matrix; X is the sparse matrix; T is the number of non-zero elements in X; A and W are the matrices to be calculated;
[0033] Q is a k×n dimensional matrix. If y i and d k The labels of are the same, then the corresponding element of Q is 1; otherwise, the element of Q is 0; k is the number of dictionaries;
[0034] SB2. Calculate the statistical indicators of the model and its threshold δ 2 :
[0035] After obtaining the dictionary D, for each sample y i ,according to Calculate its square prediction error SPE and get the square prediction error set SPE of the entire training data set list ;
[0036] Using the kernel density estimation method, the SPE list Perform nonparametric estimation and take the value corresponding to 99% of the probability density function area as the threshold δ of the model 2 ;
[0037] SB3. Online monitoring of real-time data:
[0038] For real-time data sample y new , using the FISTA algorithm to solve The optimization problem is solved and the sparse matrix X is obtained. Then according to Calculate its squared prediction error SPE;
[0039] If SPE>δ 2 It is considered that a fault has occurred, and the process goes to step S3;
[0040] If SPE≤δ 2 It is considered that no fault occurs, and the step S2 is looped.
[0041] Furthermore, the specific steps of S3 are as follows:
[0042] Sc1: Initialization parameter set S = {1, 2, ... n}, SPE set SPES = {}, fault parameter set fs = {};
[0043] Sc2: For Initialize β, where β i =0,
[0044] Where: β represents the direction vector whose kth element is 0 and other elements are all 1;
[0045] Solution , and add the obtained SPE to the SPES set;
[0046] Sc3: Calculate the minimum value SPE of SPES min and its corresponding parameter subscript i in Sc2, adding i to the fault parameter set fs;
[0047] Fruit SPE min >δ 2Or the number of elements in fs is less than n, use S=S-fs to update the parameter set, clear the error set SPES, and return to Sc2 for the next calculation;
[0048] If SPE min <δ 2 Or if the number of elements in fs is equal to n, then go to Sc4;
[0049] Sc4: Output fault parameter set fs.
[0050] Beneficial effects: Compared with the prior art, the present invention has the following advantages:
[0051] 1. The system and method provided by the present invention can quickly perform sample clustering on the historical data of the high-voltage heater, and use the dictionary learning algorithm to fully extract the essential characteristics of the data, so as to realize the real-time monitoring of the operating status of the high-voltage heater and the fault warning. After the fault occurs, if the correct fault direction and fault amplitude are found, the SPE value of the whole system can be less than the threshold value δ of the model. 2 ; For sample y i For each parameter of y, we can assume that one of the parameters has failed. If the assumption is correct, then change this parameter from y i After removing the remaining m-1 parameters, the calculated SPE value will inevitably be less than the threshold δ 2 Based on the above analysis, the fault isolation problem can be transformed into The FISTA algorithm is used for rapid solution, which improves the economy and safety of power plant operation.
[0052] 2. The system and method provided by the present invention solve the problem of fault isolation after the dictionary learning model detects a fault, and can provide reasonable guidance and decision-making for subsequent maintenance work. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 It is a flow chart of the high-pressure heater online monitoring method based on parallel GMM-LCKSVD of the present invention;
[0054] Figure 2 It is a scatter diagram of the distribution of the high-pressure heater end difference and outlet temperature parameters in a specific embodiment of the present invention;
[0055] Figure 3 A diagram showing the monitoring results of the FGMM algorithm for high pressure heater leakage fault in a specific embodiment of the present invention;
[0056] Figure 4 The monitoring result diagram of the GMM-LCKSVD algorithm for high-pressure heater leakage fault in the specific implementation mode of the present invention;
[0057] Figure 5The monitoring result diagram of the FGMM algorithm for the short circuit fault of the high-pressure heater inlet and outlet water chambers in the specific implementation mode of the present invention;
[0058] Figure 6 A diagram showing the monitoring results of the GMM-LCKSVD algorithm for a short circuit fault in the high-pressure water heater inlet and outlet compartments in a specific implementation manner of the present invention. DETAILED DESCRIPTION
[0059] In order to deepen the understanding of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings. The embodiments are only used to explain the present invention and do not constitute a limitation on the protection scope of the present invention.
[0060] A case analysis is carried out on the No. 1 high-pressure heater in the thermal system of a 600MW power plant. The 14 measurement points shown in Table 1 are selected as modeling parameters. The sampling interval is 10s. 2000 samples are collected for model training and testing respectively, and the test samples are divided into two groups, A and B.
[0061] Among them, the first 750 samples of Group A are normal operation data, and the last 250 are fault data of high-pressure heater leakage; the first 750 samples of Group B are normal operation samples, and the last 250 are fault data of high-pressure heater inlet and outlet water chamber short circuit.
[0062] Figure 2 The results of clustering the high-pressure heater end difference and outlet temperature using the parallel GMM algorithm are given. It can be seen that along the load change direction, the parallel GMM algorithm divides the data into three categories, which is completely consistent with the actual distribution characteristics of the data.
[0063] Table 1 List of measurement points and statistical characteristics
[0064]
[0065] The GMM-LCKSVD method of the present invention is used to establish a process monitoring model, and the monitoring results are compared with the FGMM model. The monitoring results of the two algorithms are as follows: Figure 3-Figure 6 As shown, the solid line in the figure represents the limit calculated for each sample, and the dotted line represents the threshold δ of the model 2 .
[0066] Table 2 shows the fault diagnosis results.
[0067] Table 2 High-pressure booster fault isolation results
[0068]
[0069] The results show that:
[0070] 1. By Figure 3 and Figure 5It can be seen that after the fault occurs, the sample limit BIP of the FGMM model is lower than the model threshold δ 2 , indicating that the algorithm failed to monitor the occurrence of the fault; it can also be seen that before the 250th fault sample, there are a large number of sample limits in the FGMM model that are higher than the model threshold, indicating that the FGMM algorithm regards normal samples as fault data at this time, and the monitoring error rate is very high.
[0071] 2. By Figure 4 and Figure 6 It can be seen that the GMM-LCKSVD model further extracts data features by using a dictionary learning algorithm.
[0072] Therefore, after the fault occurs, the sample limit SPE is higher than the model threshold δ 2 , which shows that the algorithm can monitor the occurrence of faults very well, can monitor high-pressure booster faults in real time, and the monitoring error rate is very low.
[0073] On this basis, the reconstruction-based fault isolation method can accurately locate the fault parameters, providing a new and reliable method for online monitoring and fault diagnosis of high-voltage power converters.
[0074] The present invention will be further described below in conjunction with the accompanying drawings.
[0075] The high-voltage heater online monitoring method based on parallel GMM-LCKSVD includes the following steps:
[0076] (1) Use the parallel GMM algorithm to cluster the data set Y:
[0077] Assume that the data sample set Y = {y1, y2, ...y m}, where: m represents the number of samples, sample y i ={y i1 ,y i2 ,...y in}, subscript i represents the sample number, n represents the number of variables in the sample; randomly divide the sample Y into N independent parts, and specify the number of clusters K. use and To update the Gaussian mixture model parameters of each part. Where: m t represents the number of samples in the tth part; l g,i The calculation method is: Where: g is the weight of each Gaussian part; ∑ g is the covariance; μ g Use OPENMP to parallelize the N parts. When all the N parts are calculated, use To update the parameters of the entire GMM model. Repeat the above calculation process until the number of iterations is greater than the maximum number of iterations or the error of the likelihood function is less than the given error. After the calculation is completed, the posterior probability of each sample is calculated using the Bayesian formula, and the maximum posterior probability is defined as the cluster label of the sample.
[0078] (2) Establish LCKSVD dictionary learning model:
[0079] First, the sample label matrix H of the training data Y is constructed according to the clustering results of the parallel GMM algorithm. Then Y is standardized and transposed, and the LCKSVD dictionary learning objective function for the training data Y is constructed:
[0080]
[0081] Where: D is the dictionary matrix; X is a sparse matrix; T is the number of non-zero elements in X; Q is a k×n dimensional matrix. If y i and d k The labels of are the same, then the corresponding element of Q is 1, otherwise the element of Q is 0; k is the number of dictionaries.
[0082] In order to use the KSVD algorithm to solve the parameters of LCKSVD dictionary learning, we can use Then the objective function can be written as:
[0083]
[0084] It can be seen that after the transformation, the objective function has only two unknown parameters and the form is exactly the same as the objective function of the KSVD algorithm. However, when solving the KSVD algorithm, the initial value of the dictionary D must be given first. To this end, we can construct The ridge regression model is obtained, and the initialization value of the A matrix is A = QX T (XX T +λI) -1 ;
[0085] Similarly, a ridge regression model is constructed for W, and the initial value of the W matrix is W = HX T (XX T +λI) -1 ;
[0086] After obtaining the initial values of A and W, we can obtain the initial value of the dictionary D and use the KSVD algorithm to solve the objective function to obtain the dictionary D.
[0087] (3) Calculate the statistical indicators of the model and its threshold δ 2 :
[0088] After obtaining the dictionary D, for each sample y i , according to ||y i -Dx i || F 2 Calculate its square prediction error SPE and get the square prediction error set SPE of the entire training data set list , using kernel density estimation to estimate SPE list Perform nonparametric estimation and use the value corresponding to 99% of the probability density function area as the threshold δ of the model 2 ;
[0089] (4) Fault monitoring of real-time data:
[0090] For real-time data sample y new , using the FISTA algorithm to solve The optimization problem is obtained and the sparse matrix X is obtained. Then according to ||y i -Dx i || F 2 Calculate its SPE, if SPE>δ 2 If yes, it is considered that a fault has occurred, otherwise the sample is considered normal;
[0091] (5) Fault diagnosis is performed on fault data using a reconstruction-based method. The specific steps are as follows:
[0092] Step 1: Initialize parameter set S = {1, 2, ... n}, SPE set SPES = {}, fault parameter set fs = {};
[0093] Step 2: For Initialize the direction vector β, where β i =0, Solving using the FISTA algorithm The optimization problem is solved and the obtained SPE is added to the SPES set;
[0094] Step 3: Calculate the minimum value SPE of SPES min And its corresponding parameter index i in step 2, add i to the fault parameter set fs, if SPE min <δ 2 Or if the number of elements of fs is equal to n, go to step 4, otherwise use S=S-fs to update the parameter set, clear the error set SPES, and return to step 2 for the next calculation;
[0095] Step 4: Output the fault parameter set fs.
[0096] The specific implementation steps of the online fault diagnosis method provided by the present invention are introduced below by taking the No. 1 high pressure heater reheater in the thermal system of a 600MW power plant as an example.
[0097] Combination Figure 1 The process shown, the steps are as follows:
[0098] 1. Collect a large amount of historical data on equipment operation through sensors installed in the on-site high-voltage heater;
[0099] 2. Use the parallel GMM algorithm to cluster the original data and divide the data into multiple groups of variable data sets;
[0100] 3. Use the LCKSVD dictionary learning algorithm to build a model, and use the trained LCKSVD model to monitor the real-time data of the high-voltage heater;
[0101] 4. After the fault is detected, a reconstruction-based fault isolation method is used to locate the parameters where the fault occurred.
[0102] It is further explained that any modification or equivalent substitution of the technical solution of the present invention by relevant technical personnel without departing from the purpose and scope of the technical solution of the present invention should be included in the scope of the claims of the present invention.
Claims
1. A method for online monitoring of a high-voltage heater based on parallel GMM-LCKSVD, characterized in that: The steps include: S1. Use the parallel GMM algorithm to quickly cluster the original data and divide the data into multiple data sets; S2. Use LCKSVD dictionary learning algorithm to establish a model and perform online monitoring of the high-voltage heater; The specific steps of S2 are as follows: SB1. Establish LCKSVD dictionary learning model; SB2. Calculate the statistical indicators of the model and its threshold δ 2 ; After obtaining the dictionary D, for each sample y i ,according to Calculate its square prediction error SPE and get the square prediction error set SPE of the entire training data set list ; Using the kernel density estimation method, the SPE list Perform nonparametric estimation and take the value corresponding to 99% of the probability density function area as the threshold δ of the model 2 ; SB3. Online monitoring of real-time data: For real-time data sample y new , using the FISTA algorithm to solve The optimization problem is solved and the sparse matrix X is obtained. Then according to Calculate its squared prediction error SPE; If SPE>δ 2 It is considered that a fault has occurred, and the process goes to step S3; If SPE≤δ 2 It is considered that no fault occurs, and the step S2 is looped; S3. After the fault is detected, the fault isolation method based on reconstruction is used to locate the parameters where the fault occurred; The specific steps of S3 are as follows: Sc1: Initialization parameter set S = {1, 2, ... n}, SPE set SPES = {}, fault parameter set fs = {}; Sc2: For Initialize β, where β i =0, Where: β represents the direction vector whose kth element is 0 and other elements are all 1; Solution , and add the obtained SPE to the SPES set; Sc3: Calculate the minimum value SPE of SPES min and its corresponding parameter subscript i in Sc2, adding i to the fault parameter set fs; Fruit SPE min >δ 2 Or the number of elements of fs is less than n, use S=S-fs to update the parameter set, clear the error set SPES, and return to Sc2 for the next calculation; If SPE min <δ 2 Or if the number of elements in fs is equal to n, then go to Sc4; Sc4: Output fault parameter set fs.
2. The method for online monitoring of a high-voltage heater based on parallel GMM-LCKSVD according to claim 1, characterized in that: The S1 specifically includes the following steps: SA1. Cluster the data set using the parallel GMM algorithm: Assume that the data sample set Y = {y1, y2, ...y m }, where: m represents the number of samples, sample y i ={y i1 ,y i2 ,...y in }, subscript i represents the sample number, and n represents the number of parameters; Randomly divide the data sample set Y into N independent parts and specify the number of clusters K; for and use and To update the Gaussian mixture model parameters of each part; Where: m t represents the number of samples in the tth part; l g,i The calculation method is: Where: g is the weight of each Gaussian part; ∑ g is the covariance; μ g For expectations; SA2. Use OPENMP to parallelize the calculation of the N-part mixed model parameters. After the calculation is completed, use To update the parameters of the entire GMM model; SA3. Repeat the above calculation process until the number of iterations is greater than the preset maximum number of iterations or the error of the likelihood function is less than the given error; SA4. After the calculation is completed, the posterior probability of each sample is calculated using the Bayesian formula, and the maximum posterior probability is defined as the clustering label of the sample.
3. The method for online monitoring of a high-voltage heater based on parallel GMM-LCKSVD according to claim 1, characterized in that: The LCKSVD dictionary learning model is established, including constructing a sample label matrix H of the data sample set Y according to the clustering results of the parallel GMM algorithm, then standardizing and transposing Y, and constructing an LCKSVD dictionary learning objective function for the data sample set Y, thereby obtaining a dictionary D; The objective function of LCKSVD dictionary learning for data sample set Y is: st||x||0≤T Where: D is the dictionary matrix; X is the sparse matrix; T is the number of non-zero elements in X; A and W are the matrices to be calculated; Q is a k×n dimensional matrix. If y i and d k The labels of are the same, then the corresponding element of Q is 1; otherwise, the element of Q is 0; k is the number of dictionaries.
