Multi-feature coupling fault prediction method for power generation steam turbine

Abstract

The invention provides a multi-feature coupling fault prediction method for a power generation steam turbine. According to the method, trend extraction is carried out on fault related parameters by adopting median regression empirical mode decomposition (MREMD). The method comprises the following steps: firstly, performing MREMD decomposition on related parameters to obtain a plurality of IMF components and residual components; selecting indexes representing IMF signal quality, such as relative smoothness, information entropy, power spectrum entropy and the like; a CRITIC weight analysis method is adopted, the weight of each index is determined by analyzing the comparison strength and conflict between the indexes, and data fusion and comprehensive evaluation of each index are achieved. And selecting the IMF component with a relatively high comprehensive evaluation value and the residual component of each parameter to reconstruct a trend term of each parameter. According to the method, noise, interference and sensor measuring point abnormal values can be effectively removed, trend information of original signals is reserved, high-quality feature vectors are provided for follow-up model training, the model can capture time sequence information of data more comprehensively, and the convergence speed and prediction accuracy of the model are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault prediction for marine power generation steam turbines, and particularly to a multi-feature coupling fault prediction method for power generation steam turbines. Background Technique

[0002] As a core component in a thermoelectric system for converting thermal energy into mechanical energy, the operating state of a power generation steam turbine is directly related to the operating stability and safety of the entire system. By mining the operating data before a turbine fault, its current operating state can be quantitatively evaluated, and future parameter fluctuations can be predicted, thereby realizing system fault diagnosis and prediction.

[0003] Some existing fault prediction methods have certain limitations. For example, the operating parameters of a power generation steam turbine are highly coupled, making it difficult to predict fault characteristic parameters and unable to effectively predict the development trend of fault parameters, thus failing to trigger system warnings in a timely manner and take corresponding emergency measures. Traditional prediction methods are difficult to accurately extract data features when dealing with complex time-series data and multi-feature coupling problems, resulting in low prediction accuracy. In the field of multi-feature prediction, although some models combined with deep learning have achieved certain results, there is still room for optimization, such as the subjectivity of model input feature selection, slow convergence speed, and overfitting problems. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the purpose of the present invention is to provide a multi-feature coupling fault prediction method for power generation steam turbines to solve the problems raised in the above background technique. The present invention solves the problems of highly coupled operating parameters of power generation steam turbines and difficult prediction of fault characteristic parameters, and realizes the effective prediction of the development trend of fault parameters through fluctuations in fault-related parameters, providing a reliable reference for system warning and emergency handling.

[0005] To achieve the above purpose, the present invention is realized through the following technical solutions: A multi-feature coupling fault prediction method for power generation steam turbines includes using median regression empirical mode decomposition (MREMD) to extract the trends of fault-related parameters. First, perform MREMD decomposition on the relevant parameters to obtain several IMF components and a residual component;

[0006] Select indicators such as relative smoothness, information entropy, and power spectrum entropy to characterize the quality of IMF signals, which are used to characterize the relative strength of the volatility of each component, the amount of information contained, and the information value contained in the power spectrum, etc.;

[0007] Adopt the CRITIC weight analysis method to determine the weights of each indicator through the analysis of the comparison intensity (standard deviation) and conflict (correlation) between indicators, realizing data fusion and comprehensive evaluation of each indicator;

[0008] Perform k-means clustering on each comprehensive evaluation value, discard the IMF components of the class with lower comprehensive evaluation value, and select the IMF components with higher comprehensive evaluation value and the residual components of each parameter to reconstruct the trend terms of each parameter. This can remove noise, interference, and outliers at sensor measurement points while retaining the trend information of the original signal as much as possible. The processed result is used as the training set of the neural network.

[0009] Furthermore, it also includes the MREMD decomposition process:

[0010] Step 1: For the time series s 0 (t) studied, extend both ends of s(t) to between the adjacent two extreme points of the extended time series through an autoregressive (AR) model:

[0011]

[0012] where: s is the mean point; φ 0 , φ 1 , … φ p are p + 1 real numbers; μ t (p + 1, p + 2, …, N) is a white noise sequence with zero mean;

[0013] Step 2: There are k extreme points in the extended s 0 (t). Calculate the mean of adjacent extreme points to obtain the sequence where i = 1, 2…k - 1, and then perform cubic spline interpolation on the mean sequence to obtain:

[0014]

[0015] where: m 1,0 (t), s 0 (t) - m 1,0 (t) is the signal mean sequence of s 0 (t), and h 1,0 (t) is the 1st order signal component of s 0 (t);

[0016] Step 3: Repeat Step 1 and Step 2 for iterative calculation on h 1,0 (t). Assume that after l iterations, h 1,l (t) satisfies the following termination condition, then stop the iteration at this time. The termination condition is shown in Equation (3):

[0017]

[0018] where: σ * and σ i are the signal components h 1，t-1(t) after the (l - 1)th iteration and the signal component h after the lth iteration respectively.1,l Standard deviation of the mean point sequence of (t); s z The z-th extreme point of the extended signal sequence; P is the conditional probability; θ 0 and θ l Are the ratios of the mean points of the initial signal and the signal after the l-th iteration to xMσ, and the calculation formula is as shown in (4):

[0019]

[0020] In the formula: sm1,0, sm2,0, sm3,0 are the mean points of the initial signal s without iteration 0 (t); sm 1,l, sm2,l, sm k,l are the mean points of the signal component h 1,l-1 (t) after the (l - 1)-th iteration;

[0021] Step four, h 1,l (t) is the 1st order IMF component, x 0 (t) - IMF 1 The first order residual signal R 1 of, take R 1 as the original signal and repeat steps one to three until the residual component becomes a monotonic function or no new IMF component can be separated, and the calculation process is as shown in (5)

[0022]

[0023] In the formula: n is the maximum number of IMF components that can be decomposed; Rn is the n-th order residual signal of the original signal s 0 (t), called the residual component

[0024] After the above decomposition, the original signal s 0 (t) can be expressed as the sum of all IMF components and the residual component, and the expression is as shown in (6):

[0025]

[0026] Furthermore, the calculation formulas for smoothness, information entropy, and power spectrum entropy are:

[0027]

[0028] In the formula: y (i) is the i-th value of the original data; y’ (i) is the i-th value of the IMF component; n is the length of the original data, p(x i ) is the probability density function of the IMF component in the original signal; p(x’ i ) is the probability density function of the power spectrum of the IMF component in the power spectrum of the original data.

[0029] Furthermore, the calculation steps of the comprehensive evaluation value are as follows:

[0030] S1. Assume that a certain signal has m IMF components after decomposition. Taking its relative smoothness, information entropy, and power spectrum entropy as evaluation indicators, an m×3 data matrix X can be constructed as shown in Equation (10);

[0031]

[0032] In the formula: X m1 is the relative smoothness of the m-th IMF component; X m2 is the information entropy of the m-th IMF component; X m3 is the power spectrum entropy of the m-th IMF component;

[0033] S2. Standardize the data. PH, H(x), and psdE(x) are negative indicators, and the standardization process is as shown in Equation (11);

[0034]

[0035] S3. Calculate the information carrying capacity, and use the standard deviation to represent the volatility between indicators, as shown in Equation (12) specifically;

[0036]

[0037] In the formula: is the mean value of each indicator data;

[0038] S4. Use the magnitude of the correlation coefficient to represent the conflict between indicators, and the calculation steps are as shown in Equation (13) and Equation (14);

[0039]

[0040] In the formula: R is the correlation matrix of indicators; r ij is the correlation coefficient between the i-th indicator and the j-th indicator; n is the number of evaluation indicators. In this paper, 3 evaluation indicators are set, so n = 3;

[0041] S5. Let the information amount be C j As shown in Equation (9), the weight calculation steps of each evaluation indicator are as shown in Equation (15) and Equation (16);

[0042] C j = b j × A j (15)

[0043]

[0044] In the formula: w j is the weight of the j-th evaluation indicator;

[0045] S6. The comprehensive evaluation value of each IMF component is calculated as shown in Equation (17);

[0046]

[0047] where: s i is the comprehensive evaluation value of the i-th IMF component.

[0048] Furthermore, the steps of the given K-means clustering are as follows:

[0049] Step 1: First, randomly select 2 initial clustering centers ρ i from the comprehensive evaluation s, and calculate the Euclidean distance between the remaining data objects and the clustering centers, as shown in Equation (18):

[0050] d(s,2) = |s - ρ i | i = 1, 2 (18)

[0051] Step 2: Find the clustering center ρ i closest to the target data object, and assign the data object to the cluster corresponding to the clustering center ρ i ;

[0052] Step 3: Calculate the average value of the data objects in each cluster as the new clustering center, and perform the next iteration until the clustering center no longer changes or reaches the maximum number of iterations and then stop.

[0053] Furthermore, it also includes the construction of a prediction model: constructing a Convolutional-Bidirectional Long Short-Term Memory (CNN-BiLSTM) prediction model, which is composed of a CNN module, a BiLSTM module, and an output module, and the training set selected is the processed training set.

[0054] Furthermore, the CNN module is composed of a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The convolutional layer extracts data features through convolutional calculations, the pooling layer further extracts features and reduces the network complexity, the fully connected layer realizes the combination of temporal features, and finally the output layer outputs the result. Since the selected data set is the processed data set, and in order to better retain the main features and make the model have a certain robustness to the position change of the target area containing feature information in the data set, the pooling layer selected here is the mean pooling;

[0055] The BiLSTM module introduces a reverse LSTM structure on the basis of the traditional forward LSTM, which can obtain the information from the back to the front of the signal and is conducive to extracting bidirectional temporal features. This module designs three types of mechanism gates, namely an input gate, a forget gate, and an output gate, for controlling the introduction or removal of information in the cell state;

[0056] The output module consists of the fully connected layer of the CNN and the output layer shared by the CNN and the BiLSTM module. The combination of spatial features and temporal features is achieved through the fully connected layer.

[0057] Furthermore, it also includes model hyperparameter optimization: The sparrow search algorithm (SSA) is used to optimize the hyperparameters of the CNN-BiLSTM model. Inspired by the foraging behavior and anti-predation behavior of sparrows, the sparrow search algorithm has the advantages of strong optimization ability and fast convergence speed. Through this algorithm, global optimization is carried out on hyperparameters such as the number of hidden units of the model, the kernel size of the convolutional layer, the number of iterations, and the initial learning rate, so as to improve the convergence speed and prediction accuracy of the model.

[0058] Furthermore, the calculation formula of the convolutional layer of the CNN module is:

[0059]

[0060] In the formula: represents the convolution calculation; F represents the input data of the convolutional layer; w c represents the weight parameter of the convolution kernel; C, H f , W e are the number of channels, height, and width of the convolution kernel respectively.

[0061] Furthermore, the method for introducing or removing information in the cell state of the BiLSTM module is as follows: Assume the cell state is C and the hidden unit is h. The calculation steps are as follows:

[0062] Step 1: The retention degree of historical information is controlled by the Sigmoid function of the forget gate, as shown in Equation (20);

[0063] f t = σ(W f × [h t-1 , x t + b f ) (20)

[0064] In the formula: W f is the weight matrix and bias of b f and f t ;

[0065] Step 2: The retention degree of current information is controlled by the Sigmoid function of the input gate, as shown in Equation (21);

[0066] i t = σ(W i × [h t-1 , x t + b i ) (21)

[0067] In the formula: W i and bi The weight matrix and bias for i t ;

[0068] Step 3: Establish a new candidate vector through the Tanh function, as shown in Equation (22);

[0069]

[0070] In the formula: W c is the weight matrix of Tanh, and b c is the bias term;

[0071] Step 4: Combine the current information and historical information to update the current state as shown in Equation (23);

[0072]

[0073] Step 5: Output the latest information through the Sigmoid function of the output gate, as shown in Equations (24) and (25);

[0074] o t = δ(W 0 × [h t-1 , x t + b 0 ) (24)

[0075] h t = o t × Tanh(C t ) (25)

[0076] In the formula: o t is the output result of the output gate, and h t is the output result of the hidden unit;

[0077] Output of the output module: Output of the fully connected layer, as shown in Equation (26):

[0078] o d = f d (o p × W d + b d ) (26)

[0079] In the formula: W d is the weight matrix of the fully connected layer, b d is the bias, and f d is the activation function, generally one of ReLu, Tanh, and Sig-moid;

[0080] The result of the fully connected layer is output by the output layer, as shown in Equation (27):

[0081] o d = fo (o d ×W o +b o ) (27)

[0082] Where: W o is the weight matrix of the output layer, b o is the bias, and f o is the activation function.

[0083] Advantages of the present invention:

[0084] 1. The multi-feature coupling fault prediction method for the power generation steam turbine extracts the operation trends of fault-related parameters through the MREMD decomposition and screening reconstruction method, which can effectively remove noise, interference, and abnormal values of sensor measurement points, retain the trend information of the original signal, and provide high-quality feature vectors for subsequent model training.

[0085] 2. The multi-feature coupling fault prediction method for the power generation steam turbine replaces the LSTM in the traditional CNN-LSTM model with BiLSTM, which can extract reverse time series features, enabling the model to capture the time series information of data more comprehensively and improving the prediction accuracy.

[0086] 3. The multi-feature coupling fault prediction method for the power generation steam turbine uses the SSA algorithm to optimize the hyperparameters of the CNN-BiLSTM model, avoiding the problem of model overfitting, and improving the convergence speed and prediction accuracy of the model. Verified by actual cases, the MREMD-CNN-BiLSTM model constructed by the present invention can, through the changes of fault-related parameters, ignore the short-term fluctuations of the system, and effectively predict the development trend of fault parameters, with the mean absolute error of 0.0071 and the root mean square error of 0.0074, meeting the requirements of engineering prediction. Description of the Drawings

[0087] Figure 1 is the flowchart of trend extraction based on MREMD of the present invention;

[0088] Figure 2 is the structure diagram of the CNN-BiLSTM model in the present invention;

[0089] Figure 3 is the structure diagram of the bidirectional long short-term memory model (BiLSTM) in the present invention;

[0090] Figure 4 is the flowchart of optimizing the CNN-BiLSTM model by the SSA algorithm;

[0091] Figure 5 are the IMF components and the residual component R obtained by decomposing the pressure before the quick closing valve of the steam turbine;

[0092] Figure 6The IMF components and the residual component R obtained by decomposing the pressure after the governing stage;

[0093] Figure 7 The IMF components and the residual component R obtained by decomposing the main condenser vacuum;

[0094] Figure 8 The actual data and the trend term of the steam pressure before the quick closing valve;

[0095] Figure 9 The actual data and the trend term of the pressure after the governing stage;

[0096] Figure 10 The actual data and the trend term of the main condenser vacuum;

[0097] Figure 11 The change in fitness during the iterative process;

[0098] Figure 12 The actual data, the trend term and the predicted data of the steam turbine speed training set;

[0099] Figure 13 The actual data, the trend term and the predicted data of the steam turbine speed verification set;

[0100] Figure 14 The change in the relative error of the predicted data over time. Specific implementation manners

[0101] To make the technical means, creative features, achieved purposes and functions of the present invention easy to understand, the present invention will be further described below in conjunction with specific implementation manners.

[0102] Please refer to Figures 1 to 14 , the present invention provides the following technical solution: A multi-feature coupling fault prediction method for a power generation steam turbine, which extracts the operation trends of fault-related parameters through the method of MREMD decomposition and screening reconstruction, can effectively remove noise, interference and abnormal values of sensor measurement points, retain the trend information of the original signal, and provide high-quality feature vectors for subsequent model training; replacing the LSTM in the traditional CNN-LSTM model with BiLSTM can extract reverse time series features, enabling the model to capture the time series information of the data more comprehensively and improving the prediction accuracy.

[0103] This embodiment also takes the speed fluctuation of a certain type of ship power generation steam turbine as an example to verify the fault prediction method proposed by the present invention. The specific process is as follows:

[0104] (1) MREMD decomposition

[0105] Select the steam pressure before the turbine quick - closing valve, the pressure after the governing stage, and the main condenser vacuum as the fault - related parameters of the turbine speed fluctuation, and perform the MREMD decomposition of equations (1)-(6) respectively. The results are shown in Figure

[0106] Select the steam pressure before the turbine quick - closing valve, the pressure after the governing stage, and the main condenser vacuum as the fault - related parameters of the turbine speed fluctuation, and perform the MREMD decomposition of equations (1)-(6) respectively. The results are as Figures 5 - 7 shown.

[0107] (2) Quality index calculation

[0108] Calculate the quality index and comprehensive evaluation value of the IMF components of each parameter according to equations (7)-(18). The results are shown in Table 1.

[0109] Table 1 Relative smoothness, information entropy, power spectrum entropy and comprehensive evaluation value of the IMF components of each parameter

[0110]

[0111] (3) Trend term reconstruction

[0112] As can be seen from Table 1, for the steam pressure before the quick - closing valve, the IMF5, IMF6 and IMF7 components have higher comprehensive evaluation values; for the pressure after the governing stage, the IMF4, IMF5 and IMF6 components have higher comprehensive evaluation values; for the main condenser vacuum, the IMF4 and IMF5 components have higher comprehensive evaluation values; reconstruct the components with higher comprehensive evaluation values and the residual components of each parameter into the trend terms of each parameter. The comparison with the original signals is as Figures 8 - 10 shown.

[0113] (4) Hyperparameter optimization

[0114] Use the SSA algorithm to optimize four hyperparameters of the CNN - BiLSTM prediction model constructed in Section 2, namely the number of hidden units, the kernel size of the convolutional layer, the number of iterations, and the initial learning rate. The population size and the number of iterations are taken as 10, the discoverer ratio is 0.7, and the scout ratio is selected as 0.2. The fitness curve of the training process is as Figure 11 shown.

[0115] After 10 iterations of the SSA algorithm, the optimal results are as follows: the number of model hidden units is 75, the convolutional kernel size is [1,1], the number of iterations is 50, and the initial learning rate is 0.0046.

[0116] (5) Model training and validation

[0117] As Figures 8 - 10The various parameter trend terms shown are used as the input vector of the feature set, and the steam turbine speed is used as the output vector of the feature set. The first 90% is taken as the training set, and the last 10% is taken as the validation set to train the constructed CNN-BiLSTM prediction model. The actual data, trend terms, and prediction results of the training set are as Figure 12 shown, and the actual data, trend terms, and prediction results of the validation set are as Figure 13 shown, and the relative error of the prediction results is as Figure 14 shown.

[0118] It can be seen from Figure 14 that as the prediction time increases, the prediction error also gradually increases, but the prediction trend is the same as the change trend of the original data. After calculation, the mean absolute error of the prediction results is 0.0071, and the root mean square error is 0.0074, meeting the requirements of engineering prediction.

[0119] The above shows and describes the basic principles, main features, and advantages of the present invention. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic features of the present invention.

[0120] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative manner of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for predicting multi-feature coupling faults of a power generation steam turbine, characterized in that: It includes using median regression empirical mode decomposition (MREMD) to extract the trend of fault related parameters. First, the related parameters are decomposed by MREMD to obtain several IMF components and residual components; Select indicators such as relative smoothness, information entropy, and power spectrum entropy to characterize the quality of IMF signals, which are used to characterize the relative strength of the volatility of each component, the amount of information contained, and the value of information contained in the power spectrum; The CRITIC weight analysis method is used to determine the weight of each indicator by analyzing the comparison strength (standard deviation) and conflict (correlation) between indicators, so as to achieve data fusion and comprehensive evaluation of each indicator; K-means clustering is performed on each comprehensive evaluation value, and the IMF components with lower comprehensive evaluation values ​​are discarded. The IMF components with higher comprehensive evaluation values ​​and the residual components of each parameter are selected to reconstruct the trend items of each parameter. This can remove noise, interference and abnormal values ​​of sensor measurement points while retaining the trend information of the original signal as much as possible. The processed results are used as the training set of the neural network.

2. A method for predicting multi-feature coupling faults of a power generation steam turbine according to claim 1, characterized in that: It also includes the MREMD decomposition process: Step 1: For the time series s0(t) under study, extend both ends of s(t) to the point where the left and right endpoints of the original series are between two adjacent extreme points of the extended time series through the autoregressive (AR) model: Where: s is the mean point; φ0, φ1, …φ p is p+1 real numbers; μ t (p+1,p+2,…,N) is a white noise sequence with zero mean; Step 2: After the extension, s0(t) has k extreme points, and the average value sequence of adjacent extreme points is calculated. Where i = 1, 2...k-1, and then the mean sequence is interpolated by cubic spline: Where: m 1,0 (t),s0(t)-m 1,0 (t) is the signal mean sequence of s0(t), h 1,0 (t) is the first-order signal component of s0(t); Step 3: h 1,0 (t) Repeat steps 1 and 2 for iterative calculation. Suppose that after l iterations h 1,l (t) satisfies the following termination condition, then the iteration is stopped. The termination condition is shown in formula (3): Where: * and σ i are the signal components h after the l-1th iteration respectively. 1，t-1(t) and the signal component h after the lth iteration 1,l (t) is the standard deviation of the mean point sequence; s z is the zth extreme point of the extended sequence; P is the conditional probability; θ0 and θ l is the ratio of the initial signal and the signal mean point after the lth iteration to xMσ, and the calculation formula is shown in (4): Where: sm1,0, sm2,0, sm3,0 are the mean points of the initial signal s0(t) when no iteration is performed; sm 1,l, sm 2,l, smk,l are the signal components h after the l-1th iteration 1,l-1 The mean point of (t); Step 4: h 1,l (t) is the first-order IMF component, and the first-order residual signal R1 of x0(t)-IMF1 is taken as the original signal. Repeat steps 1 to 3 until the residual component becomes a monotonic function or no new IMF component can be separated. The calculation process is shown in (5) Where: n is the maximum number of IMF components that can be decomposed; Rn is the n-order residual signal of the original signal s0(t), called the residual component After the above decomposition, the original signal s0(t) can be expressed as the sum of all IMF components and residual components, as shown in (6):

3. A method for predicting multi-feature coupling faults of a power generation steam turbine according to claim 2, characterized in that: The calculation formulas for smoothness, information entropy, and power spectrum entropy are: Where: y (i) is the i-th value of the original data; y' (i) is the ith value of the IMF component; n is the length of the original data, p(x i ) is the probability density function of the IMF component in the original signal; p(x' i ) is the probability density function of the IMF component power spectrum to the original data power spectrum.

4. A method for predicting multi-feature coupling faults of a power generation steam turbine according to claim 1, characterized in that: The calculation steps of the comprehensive evaluation value are as follows: S1. Assuming that a signal has m IMF components after decomposition, its relative smoothness, information entropy, and power spectrum entropy are used as evaluation indicators to construct an m×3 data matrix X, as shown in formula (10); Where: X m1 is the relative smoothness of the mth IMF component; X m2 is the information entropy of the mth IMF component; X m3 is the power spectrum entropy of the mth IMF component; S2. Standardize the data. PH, H(x), and psdE(x) are negative indicators. The standardization process is shown in formula (11); S3. Calculate the information carrying capacity and use the standard deviation to represent the volatility between indicators, as shown in formula (12); Where: is the mean value of each indicator data; S4. Use the size of the correlation coefficient to represent the conflict between indicators. The calculation steps are shown in formula (13) and formula (14); Where: R is the correlation matrix of the indicator; r ij is the correlation coefficient between the i-th indicator and the j-th indicator; n is the number of evaluation indicators. This paper sets 3 evaluation indicators, so n = 3; S5. Let the amount of information be C j As shown in formula (9), the weight calculation steps of each evaluation index are shown in formula (15) and formula (16); C j =b j ×A j (15) Where: w j is the weight of the jth evaluation index; S6, the comprehensive evaluation value of each IMF component is calculated as shown in formula (17); Where: s i is the comprehensive evaluation value of the ith IMF component.

5. A method for predicting multi-feature coupling faults of a power generation steam turbine according to claim 4, characterized in that: The given K-means clustering, the calculation steps are as follows: Step 1: First, randomly select two initial cluster centers ρ from the comprehensive evaluation s i , calculate the Euclidean distance between the remaining data objects and the cluster center, as shown in formula (18): d(s,2)=|s-ρ i | i=1.2 (18) Step 2: Find the cluster center ρ closest to the target data object i , and assign the data objects to the cluster centers ρ i The corresponding cluster; Step 3: Calculate the average value of the data objects in each cluster as the new cluster center and perform the next iteration until the cluster center no longer changes or the maximum number of iterations is reached.

6. A method for predicting multi-feature coupling faults of a power generation steam turbine according to claim 1, characterized in that: It also includes prediction model construction: building a convolutional-bidirectional long short-term memory (CNN-BiLSTM) prediction model, which consists of a CNN module, a BiLSTM module and an output module. The training set is the processed training set.

7. A method for predicting multi-feature coupling faults of a power generation steam turbine according to claim 6, characterized in that: The CNN module consists of a convolutional layer, a pooling layer, a fully connected layer and an output layer. The convolutional layer extracts data features through convolution calculations, the pooling layer further extracts features and reduces network complexity, the fully connected layer realizes the combination of temporal features, and finally the output layer outputs the results. Since the selected data set is a processed data set, and in order to better retain the main features, the model has a certain robustness to the position change of the target area containing feature information in the data set, the pooling layer here uses mean pooling; The BiLSTM module introduces a reverse LSTM structure based on the traditional forward LSTM, which can obtain information from the back to the front of the signal, which is conducive to extracting bidirectional time series features. The module designs three types of mechanism gates: input gate, forget gate and output gate, which are used to control the introduction or removal of information in the cell state; The output module consists of a fully connected layer of CNN and an output layer shared by CNN and BiLSTM modules, and the combination of spatial features and temporal features is achieved through the fully connected layer.

8. A method for predicting multi-feature coupling faults of a power generation steam turbine according to claim 7, characterized in that: It also includes model hyperparameter optimization: the sparrow search algorithm (SSA) is used to optimize the hyperparameters of the CNN-BiLSTM model. The sparrow search algorithm is inspired by the foraging and anti-predation behaviors of sparrows, and has the advantages of strong optimization ability and fast convergence speed. Through this algorithm, the model's number of hidden units, the kernel size of the convolutional layer, the number of iterations, the initial learning rate and other hyperparameters are globally optimized to improve the model's convergence speed and prediction accuracy.

9. A method for predicting multi-feature coupling faults of a power generation steam turbine according to claim 8, characterized in that: The calculation formula of the convolution layer of the CNN module is: Where: represents the convolution calculation; F represents the input data of the convolution layer; w c Represents the weight parameters of the convolution kernel; C, H f , W e are the number of channels, height and width of the convolution kernel respectively.

10. A method for predicting multi-feature coupling faults of a power generation steam turbine according to claim 9, characterized in that: The BiLSTM module controls the introduction or removal of information in the cell state as follows: Assuming the cell state is C and the hidden unit is h, the calculation steps are as follows: Step 1: The Sigmoid function of the forget gate controls the degree of retention of historical information, as shown in formula (20); f t =σ(W f ×[h t-1 ,x t ]+b f ) (20) Where: W f for b f and f t The weight matrix and bias of Step 2: The Sigmoid function of the input gate controls the degree of retention of the current information, as shown in formula (21); i t =σ(W i ×[h t-1 ,x t ]+b i ) (21) Where: W i and b i for i t The weight matrix and bias of Step 3: Create a new candidate vector through the Tanh function, as shown in formula (22); Where: W c is the Tanh weight matrix, b c is the bias term; Step 4: Merge the current information and historical information and update the current state as shown in formula (23); Step 5: The Sigmoid function of the output gate outputs the latest information, as shown in equations (24) and (25); oh t =δ(W0×[h t-1 ,x t ]+b0) (24) h t =o t ×Tanh(C t ) (25) In the formula: o t is the output result of the output gate, h t Output results for hidden units; Output of the output module: The output of the fully connected layer, as shown in formula (26): o d =f d (o p ×W d +b d ) (26) Where: W d is the weight matrix of the fully connected layer, b d is the bias, f d It is the activation function, which is generally ReLu, Tanh and Sigmoid; The result of the fully connected layer is output by the output layer, as shown in formula (27): o d =f o (o d ×W o +b o ) (27) Where: W o is the weight matrix of the output layer, b o is the bias, f o is the activation function.

Images