Data mining-based boiler water-cooled wall fault diagnosis method and data mining-based boiler water-cooled wall fault diagnosis equipment

Abstract

The invention discloses a boiler water-cooled wall fault diagnosis method and device based on data mining, and the method comprises the steps: collecting the fault data of a boiler water-cooled wall, and carrying out the noise reduction of the data through an LLTSA (Linear Local Tangent Space Arrangement) algorithm; carrying out modal decomposition on the denoised data by adopting a characteristic modal decomposition algorithm FMD, and extracting different modal components; and constructing a Moirai model, optimizing a component weight in the Moirai model by adopting a RBMO optimization algorithm improved by an adaptive disturbance strategy, and carrying out fault probability distribution prediction on the boiler water-cooled wall on the modal component by utilizing the optimized Moirai model. According to the invention, more accurate and efficient boiler water-cooled wall fault diagnosis can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to a fault diagnosis method and device, in particular to a boiler water wall fault diagnosis method and device based on data mining. BACKGROUND

[0002] The boiler water wall is a key component in power plants, industrial boilers and other equipment, which is in a harsh working condition of high temperature, high pressure and corrosive medium for a long time. It is easy to cause pipe wall thinning, cracking, leakage and other faults due to factors such as thermal stress, wear and corrosion. Once a fault occurs, it not only affects the efficiency of the boiler, but also may cause safety accidents and result in significant economic losses. The boiler water wall fault diagnosis mainly relies on periodic inspection by technical personnel or simple threshold judgment in the past, which is difficult to achieve accurate diagnosis. Artificial analysis is low in efficiency and cannot meet the real-time monitoring demand, especially in complex working conditions, which is easy to miss detection or misjudgment. During the operation of the boiler, the data collected by the sensor is often affected by mechanical vibration, electromagnetic interference and environmental noise, resulting in signal distortion. The traditional filtering method (such as mean filtering and wavelet denoising) cannot effectively separate the noise and the real fault characteristics. The water wall fault signal usually contains multiple frequency components (such as high-frequency noise and low-frequency leakage characteristics), and the traditional method (such as Fourier transform) cannot adaptively decompose the signal mode, resulting in insufficient extraction of fault characteristics. The existing fault diagnosis model (such as support vector machine and neural network) relies on fixed weight distribution, which is difficult to adapt to the change of data distribution under different working conditions, and the prediction accuracy is unstable. SUMMARY

[0003] The purpose of the present application is to provide a more accurate and efficient boiler water wall fault diagnosis method and device based on data mining.

[0004] Technical scheme: The boiler water wall fault diagnosis method based on data mining of the present application comprises:

[0005] (1) Collecting boiler water wall fault data, and using linear local tangent space alignment algorithm LLTSA to perform denoising processing on the data;

[0006] (2) Using feature mode decomposition algorithm FMD to perform mode decomposition on the denoised data, and extracting different modal components;

[0007] (3) Constructing a Moirai model to predict the fault probability distribution of the boiler water wall for the modal components.

[0008] Further, step (1) comprises: collecting boiler water wall fault data to form a data set, which contains the monitoring data of the boiler water wall operation in previous years, including the temperature, pressure and flow in the water wall pipe; and the monitoring data in a specific period of time in the data set is described in the form of a set of formulas:

[0009]

[0010] Among them, V represents the monitoring data set, V ij It represents the sampling value of the i-th monitoring quantity at the j-th sampling point, m is the dimension of the monitoring data, and n is the number of sampling points.

[0011] Furthermore, step (1) further includes:

[0012] Find a transformation matrix A so that the transformation matrix A can transform R D The monitoring dataset V in space is mapped to R d The spatial data set Y=[y1,…,y N ], N is the number of sampling points, that is:

[0013] Y=A T VH N ,d <D

[0014] Among them, D is the original dimension of the monitoring data, d is the target dimension after dimensionality reduction, T is the transpose of the matrix, and the central matrix H is N =I-ee T / N, I is the identity matrix, e is an N-dimensional column vector with all elements set to 1;

[0015] The linear local tangent space arrangement algorithm LLTSA includes the following steps:

[0016] Step 1: Get a sample dataset;

[0017] Step 2: To achieve the noise reduction effect, set A PCA The sample data set is mapped by PCA to obtain the mapped data set X = [x1, x2…x N ];

[0018] Step 3: Use K-nearest neighbor method to find data point X i k similar points X of the same type j , X i is the coordinate of the i-th sample point in the PCA space, and j ranges from 1 to k;

[0019] Step 4: Extract local information and calculate the i H k The d largest eigenvalues ​​correspond to the local low-dimensional space matrix M composed of d eigenvectors, and calculate W i =H k (iV i V i T ); where H k =I-ee T / k,i=1,…,N,M iis the right singular corresponding to the dth largest singular value of X i H k is the right singular corresponding to the dth largest singular value of X i is a local weight matrix, V i is a local tangent space basis

[0020] Step 5: Construct the matrix B by local accumulation as follows:

[0021] B(I i ,I i )←B(I i ,I i )+W i W i T

[0022] where B=0 and I i ={i i ,…i k} denote the index set of the k nearest neighbors of X i ;

[0023] Calculate the eigenvalues and eigenvectors of the generalized eigenproblem:

[0024] XH N BH N X T α=λXH N X T α

[0025] The eigenvectors corresponding to the eigenvalues λ1, λ2, …, λ n (λ1<λ2…< n ) are α1, α2, …, α d , respectively. Then A LLTSA =(α1, α2, …, α d ); d is the smallest eigenvalue and n is the number of eigenvalues;

[0026] Step 6: Obtain the transformation matrix A=A PCA ×A LLTSA . Then the sample data of the original data in low dimension is:

[0027] Y' = A T XH N

[0028] where Y' is the low-dimensional data sample to be obtained.

[0029] Further, step (2) comprises:

[0030] (201)First, initialize the denoised boiler water wall fault data, set the input decomposition mode number n and the filter length L, initialize the FIR filter through M Hanning windows, and initialize the iteration i = 1;

[0031] Then use Obtain the filtered signal, that is, the decomposition modal component, where represents the i-th modal component, x represents the denoised boiler water wall fault time series data, * represents convolution operation, represents the i-th signal frequency, m = 1, 2, …, M;

[0032] Finally, through the boiler water wall fault data, estimate the modal period as the point where the autocorrelation spectrum reaches the local maximum after zero crossing to update the filter coefficients; complete one iteration and set i = i + 1;

[0033] (202)Taking the correlation kurtosis as the objective function, the constraint problem of the FMD algorithm is converted into the maximum problem of the correlation kurtosis CK:

[0034] argmax{CK M (u k )}

[0035]

[0036] Where CK M (u k ) represents the correlation kurtosis value of the k-th decomposition mode; N represents the number of iterations; M represents the shift order; u k (n) represents the k-th modal component; T s represents the input period measured using the number of samples; f k represents the k-th FIR filter; l represents the weighted sum position of the historical data points of the input signal x;

[0037] Determine whether the current iteration number reaches the maximum iteration number: if yes, execute the next step; otherwise, continue iteration;

[0038] (203)Construct a correlation matrix to select two adjacent modal components with the largest correlation coefficient, calculate the correlation kurtosis of the modal components according to the estimated boiler water wall fault data, select the modal component with larger correlation kurtosis as the optimal modal component of FMD, and set M = M-1;

[0039] The correlation matrix of two modal components u p and u q is represented by the following formula:

[0040]

[0041] where CC pq represents the modal component u p and the correlation matrix of u q ; and are the mean values of u p and u q , respectively;

[0042] The correlation kurtosis value of the modal component is calculated by the following formula:

[0043]

[0044] It is determined whether the mode number M reaches the preset mode number K: if yes, the K optimal modal components are output; otherwise, the modal component decomposition is performed again.

[0045] Further, the kurtosis value updating formula of the modal component is as follows:

[0046]

[0047] where K represents the kurtosis value of the modal component; N represents the number of sampling points; x(n) represents the modal component; X rms represents the root mean square value of the modal component x(n).

[0048] Further, step (3) comprises:

[0049] A data set of N time series is established where is a target time series containing d yi variables and T i time steps; each time series has a set of covariates

[0050] The goal is to predict the distribution by learning the model f θ : Maximize the log-likelihood to predict the distribution

[0051] where p(D) is the data distribution of the sampled time series (Y, Z); p(T|D) is the task distribution defining the review window; Y t-1:T =(y t-1，t ,...,y t-1 ) and the content length l and the prediction period Y t:t+h =(y t ,…,y t+h-1 ), h is the prediction length;

[0052] The discarded layer and attention head index, the attention score between (i, m) queries and the attention score between (j, n) keys are calculated as follows, j represents a summation index variable, and n represents the number of sampling points:

[0053]

[0054] wherein E ij,mn is the degree of attention from one time point (i, m) to another time point (j, n); are the respective query and key vectors, d h represents the nth modal component, is a rotation matrix; are two different attention weights; is a binary attention bias component indicating whether the condition is true; i represents a time index, and m represents a variable index; A ij,mn represents a normalized attention score;

[0055] Using a mixed parameter distribution, the mixed distribution probability density function of c components is:

[0056]

[0057] wherein ω i is the weight of the ith component, is the probability density function of the ith component,

[0058] Further, step (3) further comprises: using the Red-Billed Blue Magpie Optimization algorithm RBMO to optimize the component weights in the Moirai model.

[0059] Further, the implementation steps of the Red-Billed Blue Magpie Optimization algorithm RBMO include:

[0060] Step 1: Random initialization

[0061]

[0062] wherein X represents a search position, n represents a population size, and dim represents the dimension of the problem to be solved;

[0063] x i,j =(ub-lb)×Rand1+lb

[0064] wherein ub represents the upper limit of the model component weight, lb represents the lower limit of the model component weight, and Rand1 represents a random number from 0 to 1;

[0065] Step 2: Find food

[0066]

[0067] where X i (t+1) represents the i-th new search factor position, X i represents the i-th randomly selected individual, t represents the current iteration number, p represents 10-n search factors randomly selected from all searched individuals, X m represents the m-th randomly selected individual, X rs represents the search factor randomly selected in the current iteration, q represents 2-5 search factors randomly selected from all searched individuals; Rand2 and Rand3 represent random numbers from 0 to 1;

[0068] Step 3: Attacking the prey

[0069]

[0070] where X food (t) represents the position of the food, Randn represents a random number for generating a standard normal distribution; T represents the total iteration number;

[0071] Step 4: Storing the food

[0072]

[0073] where, represents the fitness value of the i-th component weight before position update, represents the fitness value of the i-th component weight after position update.

[0074] Further, the individual with better fitness value is disturbed in the food storage stage, and the adaptive disturbance function formula is as follows:

[0075] Δp id = p id + η i

[0076] where Δp id is the disturbed correction position of p id ; p id is the optimal position of individual i in the t-th iteration; η i ∈(-ζ d ,ζ d ) is the correction value of individual i in the t-th iteration process, and the calculation formula of ζ d is as follows:

[0077]

[0078] where c is an adjustment constant; is the average value of the optimal target position searched in the t-1-th iteration; c iis a proportionality coefficient; Ar is a search region width; t is a current iteration number.

[0079] Based on the same inventive concept, a boiler water wall fault diagnosis device based on data mining provided by the present application comprises a processor and a memory, the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the steps of the above-mentioned boiler water wall fault diagnosis method are realized.

[0080] Beneficial effects: Compared with the prior art, the present application has the following obvious advantages:

[0081] 1. The LLTSA algorithm is used to retain the local structure of data while effectively removing noise interference, and the FMD algorithm is used to accurately extract multi-modal fault features, thereby significantly improving the fault recognition rate and solving the problem of insufficient feature extraction of traditional methods.

[0082] 2. The improved RBMO algorithm is used to dynamically optimize the Moirai model weight, avoid local optimization, ensure that the model can stably predict under different working conditions, and greatly improve the generalization ability and robustness.

[0083] 3. The fully automatic processing process realizes rapid diagnosis and early warning, is compatible with the industrial Internet of Things platform, provides real-time support for boiler safe operation and preventive maintenance, and reduces operation and maintenance costs. BRIEF DESCRIPTION OF DRAWINGS

[0084] Figure 1 is a flow chart of a boiler water wall fault diagnosis method based on data mining provided by an embodiment of the present application;

[0085] Figure 2 is a flow chart of a linear local tangent space arrangement algorithm in an embodiment of the present application;

[0086] Figure 3 is a flow chart of an improved Red-Billed Blue Magpie optimization algorithm in an embodiment of the present application. DETAILED DESCRIPTION

[0087] The present application will be further described below in conjunction with the drawings.

[0088] Embodiment 1

[0089] As shown in Figure 1 Embodiment 1 provides a boiler water wall fault diagnosis method based on data mining, comprising the following steps:

[0090] (1) Collecting boiler water wall fault data, using a linear local tangent space arrangement algorithm (LLTSA) to perform noise reduction processing on the data, thereby reducing the influence of noise on fault diagnosis;

[0091] (101)Firstly, the PLC system of the boiler collects sensor monitoring data through RS485, ModBus and other transmission links and stores it locally. Then, the PLC system packs the collected data into data frames at regular intervals and sends it to the remote boiler Internet of Things platform through the defined Ethernet interface every 10s. The platform is responsible for data analysis, storage and statistics of the received data. Finally, the required detection data is exported from the platform interface to form a data set, which contains the historical monitoring data of the boiler water-cooled wall, including temperature, pressure, flow and other sensor parameters in the water-cooled wall pipe. The monitoring data of a certain period in the data set can be described in the form of a set in the formula:

[0092]

[0093] Wherein, V represents the monitoring data set, V ij represents the sampling value of the ith monitoring quantity at the jth sampling point, m is the dimension of the monitoring data, and n is the number of sampling points.

[0094] (102)The sensor is easily disturbed in the vibration and dust environment, and the sensor itself records noise, so that the monitored data deviates from the true value and produces noise. The boiler water-cooled wall monitoring data has nonlinear characteristics, and the noise types are various, therefore, the application proposes to use linear local tangent space arrangement algorithm LLTSA to clean the data noise.

[0095] LLTSA algorithm uses local information near the data point for noise reduction, which can better preserve the local characteristics of the data and avoid information loss caused by global smoothing. By using the tangent space near the data point to represent the local geometry, these local tangent spaces are aligned in the low-dimensional space linearly mapped from the original high-dimensional space. Find a conversion matrix A, so that the conversion matrix A can map the monitoring data set V in the R D space to the data set Y in the R d space, that is: N N is the number of sampling points (the number of columns n of the matrix V), that is:

[0096] Y=A T VH N ,d<D

[0097] Wherein, D is the original dimension of the monitoring data (i.e. the number of monitoring variables collected by each sampling point), d represents the target dimension after dimension reduction (i.e. the dimension of the low-dimensional space to which the original data is mapped from the high-dimensional space by the linear local tangent space arrangement algorithm), T represents the transpose of the matrix, and the center matrix H N =I-ee T / N, I is the identity matrix, e is an N-dimensional column vector with all elements equal to 1; Y is the underlying d-dimensional nonlinear manifold of V.

[0098] like Figure 2 As shown, the linear local tangent space arrangement algorithm LLTSA includes the following steps:

[0099] Step 1: Get a sample dataset;

[0100] Step 2: To achieve the noise reduction effect, set A PCA The sample data set is mapped by PCA to obtain the mapped data set X = [x1, x2…x N ];

[0101] Step 3: Use K-nearest neighbor method (KNN) to find data point X i k similar points X of the same type j , X i is the coordinate of the i-th sample point in the PCA space, and j ranges from 1 to k;

[0102] Step 4: Extract local information and calculate the i H k The d largest eigenvalues ​​correspond to the local low-dimensional space matrix M composed of d eigenvectors, and calculate W i =H k (iV i V i T ); where H k =I-ee T / k,i=1,…,N,M i Is with X i H k The right singular value corresponding to the d-th largest singular value of i is the local weight matrix, reflecting the difference between the sample point and the local tangent space; V i is the local tangent space basis (corresponding to the first d largest singular values).

[0103] Step 5: Construct matrix B by local accumulation:

[0104] B(I i ,I i )←B(I i ,I i )+W i W i T

[0105] Among them, initialize B = 0, I i ={i i ,…i k} represents X ian index set of k nearest neighbors of the point;

[0106] Eigenvalues and eigenvectors of the generalized eigenproblem are calculated:

[0107] XH N BH N X T alpha = lambda XH N X T alpha

[0108] corresponding to eigenvalues lambda1, lambda2, …, lambda n (lambda1 < lambda2 < … < lambda n ) are alpha1, alpha2, …, alpha d , then A LLTSA = (alpha1, alpha2, …, alpha d ); d is the minimum eigenvalue, and n is the number of eigenvalues;

[0109] Step 6: Obtain the conversion matrix A = A PCA x A LLTSA , then the sample data of the original data in the low dimension are:

[0110] Y' = A T XH N

[0111] Where Y' is the low-dimensional data sample to be solved.

[0112] (2) The characteristic modal decomposition algorithm FMD is used for modal decomposition on the denoised data to extract different modal components;

[0113] The boiler water wall monitoring data contains multiple frequency components, and different frequency components may correspond to different fault types, so the characteristic modal decomposition algorithm FMD is used for signal decomposition on the denoised boiler water wall fault data to extract different modal components, and the running state of the boiler water wall is understood more deeply, so as to improve the efficiency and accuracy of fault diagnosis, and provide support for the safe operation and preventive maintenance of the boiler. Specifically:

[0114] (201) The FMD algorithm has the advantages of adaptive filter, correlation kurtosis optimization, mode selection and interpretability, and can effectively decompose the signal into multiple modal components and retain the characteristics of different frequency components. In the process of processing the denoised data, first, the denoised boiler water wall fault data (which are multi-dimensional physical quantity data arranged in time sequence collected by long-term monitoring of the water wall running state by the sensor. These data directly reflect the health status of the water wall and are the basis for fault diagnosis) are initialized, the input decomposition mode number n and the filter length L are set, the FIR filter is initialized through M Hanning windows, and the iteration i = 1 is initialized;

[0115] Then use Obtain the filtered signal, that is, decompose the modal components, where represents the i-th modal component, x represents the denoised boiler water wall fault time series data (i.e. the original signal to be decomposed), * represents the convolution operation, represents the frequency of the i-th signal, m=1,2,…,M;

[0116] Finally, the modal period is estimated based on the boiler water wall fault data as the point where the autocorrelation spectrum reaches a local maximum after crossing the zero point to update the filter coefficient; one iteration is completed and i=i+1 is set;

[0117] (202) Taking the correlation kurtosis as the objective function, the constraint problem of the FMD algorithm is transformed into the maximum value problem of the correlation kurtosis CK:

[0118] argmax{CK M (u k )}

[0119]

[0120]

[0121] Among them, CK M (u k ) represents the correlation kurtosis value of the kth decomposition mode; N represents the number of iterations; M represents the shift order; u k (n) represents the kth modal component; T s represents the input period measured using the number of samples; f k represents the kth FIR filter; l represents the weighted summation position of the historical data points of the input signal x;

[0122] Determine whether the current number of iterations has reached the maximum number of iterations: if so, proceed to the next step; otherwise, continue iterating;

[0123] (203) A correlation matrix is ​​constructed to select two adjacent modal components with the largest correlation coefficients. The correlation kurtosis of the modal components is calculated based on the estimated boiler water wall fault data. The modal component with the larger correlation kurtosis is selected as the optimal modal component of FMD, and M = M-1 is set;

[0124] The two modal components u are expressed by the following formula p and u q The correlation matrix is:

[0125]

[0126] Among them, CC pq represents the modal component u pand u q the correlation matrix of u and the average of u p and u q

[0127] The correlation kurtosis value of the modal component is calculated by the following formula:

[0128]

[0129] Determine whether the mode number M reaches the preset mode number K: if yes, output the K optimal modal components; otherwise, re-perform the modal component decomposition.

[0130] Further, the kurtosis value updating formula of the modal component is as follows:

[0131]

[0132] Wherein, K represents the kurtosis value of the modal component; N represents the sampling point number; x(n) represents the modal component; X rms represents the root mean square value of the modal component x(n).

[0133] (3) Construct the Moirai model, and use the red-billed blue magpie optimization algorithm RBMO improved by the adaptive disturbance strategy to optimize the component weight in the Moirai model, and use the optimized Moirai model to predict the fault probability distribution of the boiler water wall.

[0134] In order to more effectively predict the fault probability distribution of the boiler water wall and improve the prediction accuracy, the Moirai model is used in the present application, and the modal component obtained by LLTSA noise reduction and FMD decomposition is used as the input of the Moirai model, which is used to predict the fault probability distribution of the boiler water wall. Further, the present application proposes to use the red-billed blue magpie optimization algorithm RBMO to optimize the component weight in the Moirai model, so as to improve the prediction accuracy of the Moirai model. The improved RBMO algorithm is used to disturb the individual with better fitness value, so as to avoid falling into local optimum, and the model can more accurately predict the fault probability of the boiler water wall in the prediction process. Specifically:

[0135] (301) Moirai is a basic model for time series prediction, which can predict a wide range of time series. A dataset of N time series is established where is the target time series containing d yi variables and T i time steps; each time series has a set of covariates

[0136] The goal is to learn the model f​θ : Maximizing log-likelihood to predict distribution

[0137] where p(D) is the data distribution of the sampled time series (Y, Z); p(T|D) is the task distribution defining the review window; Y t-1:T = (y t-1，t ,...,y t-1 ) and the content length l and the prediction period, Y t:t+h = (y t ,…,y t+h-1 ), h is the prediction length;

[0138] The attention score between the (i, m) query and the (j, n) key is calculated as follows, where j represents the summation index variable (a temporary variable that traverses all sampling points (from 1 to n) when calculating the kurtosis value), and n represents the number of sampling points (i.e., the time series length or the total number of data points of the modal component signal):

[0139]

[0140] where E ij,mn is the degree of attention from one time point (i, m) to another time point (j, n); are the respective query and key vectors, d h represents the nth modal component (i.e., one of the decomposed signal components), is a rotation matrix; are two different attention weights; is a binary attention bias component indicating whether the condition is true or not; i represents the time index, and m represents the variable index; A ij,mn represents the normalized attention score;

[0141] To achieve the goal of flexibly adapting to different distribution target variables while ensuring that the operations of sampling and evaluating the loss function remain simple, a hybrid parameter distribution is used, and the probability density function of the mixed distribution of c components is:

[0142]

[0143] where ω i is the weight of the ith component, is the probability density function of the ith component,

[0144] The RBMO establishes a mathematical model of the RBMO by simulating the search, pursuit, attack of prey, and food storage behaviors of the blue jay.

[0145] As Figure 3 shown, the implementation steps of the Red-billed Blue Magpie Optimization (RBMO) algorithm include:

[0146] Step 1: Random initialization

[0147]

[0148] where X represents the search position, n represents the population size, and dim represents the dimension of the problem being solved.

[0149] x i,j = (ub - lb) × Rand1 + lb

[0150] where ub represents the upper limit of the model component weight, lb represents the lower limit of the model component weight, and Rand1 represents a random number from 0 to 1.

[0151] Step 2: Finding food. In the process of finding food, red-billed blue magpies usually act in small groups (2 to 5) or in groups (more than 10) to improve search efficiency. They use various ways such as jumping on the ground, walking or searching for food resources in trees. This adaptability and flexibility allows red-billed blue magpies to adopt various hunting strategies according to environmental conditions and available resources, ensuring adequate food supply.

[0152]

[0153] where X i (t+1) represents the i-th new search factor position, X i represents the i-th randomly selected individual, t represents the current iteration number, p represents the random selection of 10-n search factors from all searched individuals, X m represents the m-th randomly selected individual, X rs represents the randomly selected search factor in the current iteration, q represents the random selection of 2-5 search factors from all searched individuals; Rand2 and Rand3 represent random numbers from 0 to 1.

[0154] Step 3: Attacking prey. Red-billed blue magpies show high hunting proficiency and cooperation when chasing prey. They use tactics such as rapid pecking, jumping to catch prey or flying to catch insects. In small group actions, the main target is usually small prey or plants.

[0155]

[0156] where X food (t) represents the position of the food, Randn represents a random number for generating a standard normal distribution; T represents the total number of iterations;

[0157] Fourth step: storing food. In addition to searching for and attacking food, the red-billed blue magpie also stores excess food in tree holes or other hidden places for future consumption, ensuring a stable food supply in times of food shortage.

[0158]

[0159] wherein, represents the fitness value before the position update of the i-th component weight, represents the fitness value after the position update of the i-th component weight.

[0160] (303) In the algorithm optimization process, after a certain number of iterations, the search area of the entire population will automatically converge to the optimal solution set. In order to accelerate the convergence speed of the search area and improve the accuracy of the optimal solution, the adaptive disturbance strategy is proposed, and the individuals with better fitness values are disturbed in the storage food stage. The adaptive disturbance function formula is as follows:

[0161] Δp id = p id + η i

[0162] wherein, Δp id is the disturbance correction position of p id ; p id is the optimal position of individual i in the t-th iteration; η i ∈(-ζ d ,ζ d ) is the correction value of individual i in the t-th iteration process, and the calculation formula of ζ d is as follows:

[0163]

[0164] wherein, c is an adjustment constant; is the average value of the optimal target position searched in the t-1th iteration; c i is a proportional coefficient; Δr is the search area width; t is the current iteration number.

[0165] Example 2

[0166] Embodiment 2 provides a boiler water wall fault diagnosis device based on data mining, which can include a processor and a memory, the memory has computer instructions stored therein, the processor is configured to execute the computer instructions stored in the memory, when the computer instructions are executed by the processor, the boiler water wall fault diagnosis device implements the steps of the boiler water wall fault diagnosis method as described in Embodiment 1 above, and achieves the same technical effects as the boiler water wall fault diagnosis method.

[0167] The memory can include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The device can further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage media can be used for reading and writing non-removable, non-volatile media such as magnetic disks (e.g., conventional, floppy disks, and / or hard disks). The program / utility, having a set of program modules, can be stored in, for example, memory by way of example, such as an operating system, one or more application programs, other program modules, and program data, each or some combination thereof, can include implementation of a network environment. The program modules typically carry out the functions and / or methodologies of embodiments of the present application.

[0168] The processor performs various function applications and data processing by running the program stored in the memory, such as implementing the boiler water wall fault diagnosis method provided in Embodiment 1 of the present application.

Claims

1. A boiler water wall fault diagnosis method based on data mining, characterized in that: include: (1) Collect boiler water wall fault data and use the linear local tangent space arrangement algorithm (LLTSA) to reduce noise on the data; (2) The denoised data is subjected to modal decomposition using the eigenmode decomposition algorithm (FMD) to extract different modal components; (3) Construct the Moirai model to predict the failure probability distribution of the boiler water-cooled wall based on the modal components.

2. The boiler water wall fault diagnosis method based on data mining according to claim 1 is characterized in that: Step (1) includes: collecting boiler water wall fault data to form a data set, wherein the data set includes historical monitoring data of the boiler water wall operation, including the temperature, pressure, and flow rate in the water wall tube; the monitoring data within a specific period of time in the data set is described in the form of a set in the formula: Among them, V represents the monitoring data set, V ij It represents the sampling value of the i-th monitoring quantity at the j-th sampling point, m is the dimension of the monitoring data, and n is the number of sampling points.

3. The boiler water wall fault diagnosis method based on data mining according to claim 2 is characterized in that: Step (1) further includes: Find a transformation matrix A so that the transformation matrix A can transform R D The monitoring dataset V in space is mapped to R d The spatial data set Y=[y1,…,y N ], N is the number of sampling points, that is: Y=A T VH N ,d<D Among them, D is the original dimension of the monitoring data, d is the target dimension after dimensionality reduction, T is the transpose of the matrix, and the central matrix H is N =I-ee T / N, I is the identity matrix, e is an N-dimensional column vector with all elements set to 1; The linear local tangent space arrangement algorithm LLTSA includes the following steps: Step 1: Get the sample dataset; Step 2: To achieve the noise reduction effect, set A PCA The sample data set is mapped by PCA to obtain the mapped data set X = [x1, x2…x N ]; Step 3: Use K-nearest neighbor method to find data point X i k similar points X j , X i is the coordinate of the i-th sample point in the PCA space, and j ranges from 1 to k; Step 4: Extract local information and calculate the i H k The d largest eigenvalues ​​correspond to the local low-dimensional space matrix M composed of d eigenvectors, and calculate W i =H k (iV i V i T ); where H k =I-ee T / k,i=1,…,N,M i Is with X i H k The right singular value corresponding to the d-th largest singular value of i is the local weight matrix, V i Local tangent space basis; Step 5: Construct matrix B by local accumulation: B(I i ,I i )←B(I i ,I i )+W i W i T Among them, initialize B = 0, I i ={i i ,…i k } represents X i The index set of k nearest neighbor points; Compute the eigenvalues ​​and eigenvectors of a generalized eigenproblem: XH N BH N X T a=λXH N X T a and eigenvalues ​​λ1,λ2,…,λ n (λ1<λ2…<λ n ) The corresponding eigenvector solutions are α1,α2,…,α d , then A LLTSA =(α1,α2,…,α d ); d is the minimum eigenvalue, n is the number of eigenvalues; Step 6: Get the transformation matrix A=A PCA ×A LLTSA , then the sample data of the original data in low dimension is: Y'=A T XH N Among them, Y' is the low-dimensional data sample required.

4. The boiler water wall fault diagnosis method based on data mining according to claim 3 is characterized in that: Step (2) includes: (201) First, the denoised boiler water wall fault data is initialized, the number of input decomposition patterns n and the filter length L are set, the FIR filter is initialized through M Hanning windows, and the iteration i = 1 is initialized; Then use Obtain the filtered signal, that is, decompose the modal components, where represents the i-th modal component, x represents the denoised boiler water wall fault time series data, * represents the convolution operation, represents the frequency of the i-th signal, m=1,2,…,M; Finally, the modal period is estimated based on the boiler water wall fault data as the point where the autocorrelation spectrum reaches a local maximum after crossing the zero point to update the filter coefficient; one iteration is completed and i=i+1 is set; (202) Taking the correlation kurtosis as the objective function, the constraint problem of the FMD algorithm is transformed into the maximum value problem of the correlation kurtosis CK: Among them, CK M (u k ) represents the correlation kurtosis value of the kth decomposition mode; N represents the number of iterations; M represents the shift order; u k (n) represents the kth modal component; T s represents the input period measured using the number of samples; f k represents the kth FIR filter; l represents the weighted summation position of the historical data points of the input signal x; Determine whether the current number of iterations has reached the maximum number of iterations: if so, proceed to the next step; otherwise, continue iterating; (203) A correlation matrix is ​​constructed to select two adjacent modal components with the largest correlation coefficients. The correlation kurtosis of the modal components is calculated based on the estimated boiler water wall fault data. The modal component with the larger correlation kurtosis is selected as the optimal modal component of FMD, and M = M-1 is set; The two modal components u are expressed by the following formula p and u q The correlation matrix is: Among them, CC pq represents the modal component u p and u q The correlation matrix of and u p and u q The average value of The relevant kurtosis value of the modal component is calculated by the following formula: Determine whether the mode number M reaches the preset mode number K: if so, output K optimal modal components; otherwise, re-perform modal component decomposition.

5. The boiler water wall fault diagnosis method based on data mining according to claim 4 is characterized in that: The updating formula of the kurtosis value of the modal component is as follows: Where K represents the kurtosis value of the modal component; N represents the number of sampling points; x(n) represents the modal component; X rms represents the RMS value of the modal component x(n).

6. The boiler water wall fault diagnosis method based on data mining according to claim 5 is characterized in that: Step (3) includes: Create a dataset of N time series in Is included d yi Variables and T i target time series of time steps; each time series has a set of covariates The goal is to learn the model Maximize the log-likelihood to predict the distribution Where p(D) is the data distribution of the sampled time series (Y, Z); p(T|D) is the task distribution that defines the look-back window; Y t-1:T =(y t-1，t ,...,y t-1 ) with content length l and prediction period, Y t:t+h =(y t ,…,y t+h-1 ), h is the prediction length; Dropping the layer and attention head index, the attention score between (i,m) queries and the attention score between (j,n) keys are calculated as follows, where j represents the summation index variable and n represents the number of sampling points: Among them, E ij,mn is the degree of attention a certain moment (i,m) pays to another moment (j,n); are the respective query and key vectors, d h represents the nth modal component, is the rotation matrix; are two different attention weights; is a binary attention bias component that indicates whether the condition is true; i represents the time index, m represents the variable index; A ij,mn represents the normalized attention score; Using a mixed parametric distribution, the probability density function of the mixture distribution of c components is: Among them, ω i is the weight of the i-th component, is the probability density function of the ith component, 7. The boiler water wall fault diagnosis method based on data mining according to any one of claims 1 to 6, characterized in that: Step (3) also includes: optimizing the component weights in the Moirai model using the Red-billed Blue Magpie Optimization algorithm (RBMO).

8. The boiler water wall fault diagnosis method based on data mining according to claim 7, characterized in that: The implementation steps of the Red-billed Blue Magpie Optimization Algorithm (RBMO) include: Step 1: Random Initialization Among them, X represents the search position, n represents the population size, and dim represents the dimension of the problem to be solved; x i,j =(ub-lb)×Rand1+lb Among them, ub represents the upper limit of the model component weight, lb represents the lower limit of the model component weight, and Rand1 represents a random number from 0 to 1; Step 2: Finding Food Among them, X i (t+1) represents the i-th new search factor position, X i represents the randomly selected i-th individual, t represents the current number of iterations, p represents the 10-n search factors randomly selected from all searched individuals, X m represents the randomly selected mth individual, X rs represents the search factor randomly selected in the current iteration, q represents the random selection of 2-5 search factors from all searched individuals; Rand2 and Rand3 represent random numbers from 0 to 1; Step 3: Attack the prey Among them, X food (t) indicates the location of the food, Randn represents the random number used to generate the standard normal distribution; T represents the total number of iterations; Step 4: Storing Food in, Represents the fitness value before the i-th component weight position is updated, Represents the fitness value after the i-th component weight position is updated.

9. The boiler water wall fault diagnosis method based on data mining according to claim 8, characterized in that: During the food storage phase, individuals with better fitness values ​​are perturbed. The adaptive perturbation function formula is as follows: Δp id =p id +n i Where Δp id For p id The disturbance correction position of p id is the optimal position of individual i in the tth iteration; η i ∈(-ζ d ,ζ d ) is the correction value of individual i in the t-th iteration process, ζ d The calculation formula is: Where c is the adjustment constant; is the average value of the optimal target position searched within t-1 iterations; c i is the proportional coefficient; Δr is the width of the search area; t is the current number of iterations.

10. A boiler water wall fault diagnosis device based on data mining, characterized in that: The boiler water wall fault diagnosis device comprises a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the boiler water wall fault diagnosis device implements the steps of the boiler water wall fault diagnosis method as described in any one of claims 1 to 9.

Images