Turbine fault diagnosis method based on generative adversarial network

Through the fault diagnosis method based on the generative adversarial network, the generation of extended data and combined with dynamic niche neural optimization and manifold learning technology, the problems of slow iteration speed and low accuracy in traditional methods are solved, and efficient and accurate steam turbine fault diagnosis is achieved. , improves the health management capabilities of the equipment.

CN118965101APending Publication Date: 2024-11-15HARBIN ELECTRIC SCI & TECH CO LTD
View PDF 0 Cites 16 Cited by

Patent Information

Application Number
CN202410982483.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-22
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The iterative speed of traditional steam turbine fault diagnosis methods is slow, the accuracy of fault identification is low, the sample size is insufficient, the diversity is poor, the gradient disappears or explosions are susceptible to the influence of gradients, the processing of high-dimensional data is low efficiency, and the adaptability and real-time response capabilities are insufficient.

Method used

The fault diagnosis method based on the generative adversarial network is adopted, and the generative adversarial network algorithm of momentum factor is generated and expanded data is generated, combined with dynamic niche neural optimization algorithm and manifold learning, and the auto-coded neural network is selected and dimensionality reduction is reduced. The extreme learning machine is used to classify. The algorithm performs troubleshooting.

Benefits of technology

It improves the accuracy and efficiency of fault detection, enhances the generalization ability of the model, reduces the risk of overfitting, ensures stability and rapid response capabilities in complex data environments, reduces downtime, and extends equipment life.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118965101A_ABST
    Figure CN118965101A_ABST
Patent Text Reader

Abstract

The invention discloses a turbine fault diagnosis method based on a generative adversarial network, and belongs to the technical field of turbine fault recognition. The problems that in the prior art, a traditional turbine fault diagnosis method is low in iteration speed and low in fault recognition accuracy are solved. The method comprises the following steps: acquiring and marking steam turbine operation data output by a steam turbine sensor to obtain an original training set; constructing a sample generation model, inputting the original training set and training, and integrating to obtain training data; constructing a feature extraction model, and inputting training data for training to obtain an updated feature extraction model; constructing a feature dimension reduction model, and inputting the collected data in the updated feature extraction model for training to obtain an updated feature dimension reduction model; and constructing an extreme learning machine network, and inputting data in the updated feature dimension reduction model for training to obtain a steam turbine fault diagnosis model. The method effectively improves the fault classification speed and accuracy, and can be applied to turbine fault detection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a steam turbine fault diagnosis method, in particular to a steam turbine fault diagnosis method based on a generative adversarial network, and belongs to the technical field of steam turbine fault identification. Background Art

[0002] With the improvement of industrial automation and intelligence, the demand for stable operation and maintenance of major mechanical equipment such as steam turbines is increasing. Traditional fault diagnosis methods mainly rely on regular physical inspections and empirical judgments, which are not only time-consuming and labor-intensive, but also difficult to achieve early detection and prevention of faults. In addition, the above methods are usually incapable of processing complex data and responding to rapidly changing industrial environments. Therefore, it is particularly important to develop an intelligent diagnosis technology that can monitor the status of steam turbines in real time and quickly and accurately diagnose their potential faults.

[0003] In the prior art, a Chinese invention patent document with a publication (announcement) number of CN118070138A discloses a method, storage medium and device for diagnosing turbine rotor faults under unbalanced samples, including the following steps: clustering a minority of samples using an agglomerative hierarchical clustering algorithm; calculating the center distance between each subcluster and the majority class, comparing their center distances, selecting candidate subclusters and denoising using the K nearest neighbor algorithm; determining the oversampling size of each subcluster based on the size of the subcluster and the misclassification cost of the subcluster; synthesizing new instances in each minority subcluster using a minority class sample oversampling method with majority weights; and combining with a least squares support vector machine to perform small sample mixed unbalanced fault diagnosis to increase the interpretability of the model. The above method cleverly solves the turbine rotor fault problem. In order to solve the problem of unbalanced and heterogeneous data sampling modeling in high-end equipment such as systems, an explainable sampling model is provided. However, the traditional turbine fault diagnosis method still has the following problems: (1) The existing technology often leads to insufficient training and weak generalization ability of the turbine fault diagnosis model due to insufficient sample size and poor diversity; (2) The traditional neural network training is easily affected by gradient vanishing or exploding, which affects the stability of the turbine fault diagnosis model training and the final performance of the model; (3) The traditional method fails to effectively utilize the intrinsic structure of the data, resulting in low efficiency in processing high-dimensional data for turbine fault diagnosis and insufficient robustness in handling outliers; (4) The traditional algorithm has slow iteration, insufficient adaptability and real-time response capabilities, and is difficult to meet the high standards of real-time and accuracy required by industrial sites.

[0004] In summary, a steam turbine fault diagnosis method with high fault detection accuracy and efficiency is needed. Summary of the invention

[0005] A brief overview of the present invention is provided below in order to provide a basic understanding of certain aspects of the present invention. It should be understood that this overview is not an exhaustive overview of the present invention. It is not intended to identify key or important parts of the present invention, nor is it intended to limit the scope of the present invention. Its purpose is merely to present certain concepts in a simplified form as a prelude to a more detailed description discussed later.

[0006] In view of this, in order to solve the problems of slow iteration speed and low fault identification accuracy of traditional steam turbine fault diagnosis methods in the prior art, the present invention provides a steam turbine fault diagnosis method based on a generative adversarial network.

[0007] The technical solution is as follows: A steam turbine fault diagnosis method based on a generative adversarial network comprises the following steps:

[0008] S1. Collect and annotate the steam turbine operation data output by the steam turbine sensor to obtain the original training set;

[0009] S2. Use the momentum factor-based generative adversarial network algorithm to build a sample generation model, input the original training set and perform training, and integrate to obtain training data;

[0010] S3. construct a feature extraction model, input the training data into the feature extraction model, and use the dynamic niche neural optimization algorithm for training to obtain an updated feature extraction model;

[0011] S4. construct a feature dimensionality reduction model using an autoencoder neural network algorithm based on manifold learning, input the collected data in the updated feature extraction model into the feature dimensionality reduction model, and train it using a gradient descent method to obtain an updated feature dimensionality reduction model;

[0012] S5. Construct an extreme learning machine network, input the data in the updated feature dimension reduction model into the classifier of the extreme learning machine network, and use the extreme learning machine classification algorithm based on the nonlinear metric of the activation function for training to obtain the turbine fault diagnosis model.

[0013] Furthermore, the step S2 specifically includes the following steps:

[0014] S201. Construct a sample generation model, set the maximum number of iterations, initialize the parameters of the generator and the discriminator by randomly extracting from a Gaussian distribution, and obtain a parameter initialization result;

[0015] S202. According to the parameter initialization result, randomly select a batch of data, i.e., real data, from the original training set to perform sample generation model training;

[0016] S203. Generate false data through the generator according to the generator parameters and the input random noise;

[0017] S204. Mix the real data and the fake data and input them into the discriminator, distinguish the real data from the fake data through the discriminator, and update the discriminator parameters in the parameter initialization result according to the discriminator loss function to obtain updated discriminator parameters;

[0018] S205. Referring to the process of updating the discriminator parameters, the generator parameters in the parameter initialization result are updated according to the generator loss function to obtain updated generator parameters;

[0019] S206. Repeat steps S201-S205 for iterative training until the maximum number of iterations is reached, that is, the sample generation training is completed and the training data is obtained;

[0020] In S201, the momentum factor-based generative adversarial network algorithm includes a generator G and a discriminator D, the generator G is used to generate false data that is similar to real data but does not appear in the original steam turbine operation data, and the discriminator D is used to distinguish whether the input data is the original steam turbine operation data or false data;

[0021] The parameter initialization result is expressed as:

[0022]

[0023] Among them, θ g is the generator parameter, θ d are the discriminator parameters, represents a normal distribution with mean 0 and covariance as the identity matrix I;

[0024] In S202, during the random selection process, batch data The randomness of the selection is achieved through the sampling probability p(xe), that is, the real data sample xe is uniformly distributed in the original turbine operation data The sampling probability of being selected is p(xe);

[0025] The sampling probability p(xe) is expressed as:

[0026]

[0027] Among them, N is the original turbine operation data The total number of data points in ;

[0028] A batch of data randomly selected It is expressed as:

[0029]

[0030] Among them, xei is the i-th real data sample, i=1,2,...,m, m is the batch size;

[0031] In S203, the false data sample It is expressed as:

[0032]

[0033] Where z is the input random noise;

[0034] In S204, the discriminator loss function L d It is expressed as:

[0035] L d =L d,real +L d,fake

[0036]

[0037] Among them, L d,real is the loss of real data, L d,fake is the loss of generated data, is the i-th false data sample; the process of updating the discriminator parameters is expressed as:

[0038]

[0039] Among them, θ d ' is the updated discriminator parameter, λ d Update the learning rate for the discriminator, is the discriminator gradient parameter;

[0040] In S205, the generator loss function L g It is expressed as:

[0041]

[0042] Among them, L M is the momentum factor;

[0043] Momentum Factor L M It is expressed as:

[0044]

[0045] Where M is the momentum factor metric, f() is the feature extraction function, ∥∥ M represents the distance metric obtained through distance metric learning;

[0046] The distance metric of data sample features is expressed as:

[0047]

[0048] Among them, Me is a symmetric positive definite matrix, which represents the importance and relevance of different dimensions in the feature space, and T is the matrix transpose;

[0049] Calculate the generator gradient using the chain rule

[0050] Generator Gradients It is expressed as:

[0051]

[0052] in, Denotes the generator loss function L g The partial derivative with respect to the discriminator output, Represents the discriminator output to generate data The partial derivative of Indicates the generated data For the generator parameter θ g The partial derivative of

[0053] The process of updating the generator parameters is expressed as:

[0054]

[0055] Among them, θ g ' is the updated generator parameter, λ g Update the learning rate for the generator, is the generator gradient parameter.

[0056] Furthermore, the step S3 specifically includes the following steps:

[0057] S301. Construct a feature extraction model based on a neural network, input training data, set the maximum number of iterations, and generate an initial population in the initialization phase;

[0058] S302. In the dynamic evaluation phase, the performance of the neural network parameters is evaluated using a fitness function according to the attractor properties of the dynamic trajectory;

[0059] S303. Define the neural network iteration process and calculate the loss function of the neural network;

[0060] S304. The individual adjusts its position in the ecological niche according to its fitness, and obtains an individual after position adjustment;

[0061] S305. Perform crossover and update operations to obtain new individuals;

[0062] S306. Using mutation mechanism based on nonlinear dynamics theory to perform mutation and mutation operations, so that individuals can make random jumps in a larger range;

[0063] S307. Perform global dynamic updates on all populations according to the theory of ecological dynamic systems;

[0064] S308. Repeat steps S301-S307 for iterative training until the maximum number of iterations is reached, indicating that the feature extraction model training is completed and an updated feature extraction model is obtained;

[0065] In S301, P (0) is the initial population, individual p i That is, the neural network parameters, which include the weight matrix and bias vector of the neural network;

[0066] Individual p i It is expressed as:

[0067]

[0068] Among them, Nu is the population size, W i is the i-th weight matrix, b i is the i-th bias vector, Represents direct concatenation of vectors;

[0069] In S302, the evaluation process is represented as follows:

[0070] fit(p i )=α Ts ·Acc(p i )+β Ts Stab(p i )

[0071] Among them, fit() is the fitness function, Acc() represents the accuracy on the labeled training data, which is calculated by the output of the preset Softmax function, Stab() represents dynamic stability, α Ts is the performance hyperparameter, β Ts is a stability hyperparameter;

[0072] Dynamic stability of neural network parameters Stab(p i ) is expressed as:

[0073]

[0074] Among them, δ sta is the attenuation coefficient, is the loss function of the neural network, w ij For individual p i The jth weight parameter in , Mu is the total number of weight parameters;

[0075] In S303, the regularization strength of the t-th iteration is defined as κ(t), and the regularization strength of the t+1-th iteration is defined as κ(t+1);

[0076] The iterative process is expressed as:

[0077] κ(t+1)=κ(t)·exp(-α pe g(t)+β pe ·(1-L(t)))

[0078] g(t)=(f(p (t) )-f(p (t-1) ))·exp(-γ pe ·Ds(p (t) ))

[0079]

[0080] Among them, α pe is the fitness gradient hyperparameter, β pe is the learning progress hyperparameter, g(t) is the fitness gradient, L(t) is the learning progress, γ pe is the fractal dimension adjustment factor, Ds() is the fractal dimension calculation function;

[0081] Loss Function of Neural Network It is expressed as:

[0082]

[0083] Among them, R(W) is the L2 norm of the weight vector W, which is used to calculate the weight w i The sum of the squares of is the cross entropy loss;

[0084] In S304, the individual position adjustment process is expressed as:

[0085]

[0086] Among them, f avg is the average fitness of the current population, η vd is the interaction adjustment factor, p best is the parameter of the current optimal individual, γ vd To update the step size hyperparameter, λ vd is the learning strategy hyperparameter, Δp i is the individual after position adjustment;

[0087] In S305, the crossover and update operation process is represented as follows:

[0088]

[0089] Among them, p new For new individuals, is a randomly generated crossover factor, p parent1 is the first randomly selected individual among the individuals after position adjustment, p parent2 is the second individual randomly selected from the individuals after position adjustment;

[0090] In S306, the variation and mutation operation process is represented as follows:

[0091] p′ i =p i +∈ bg ·randn() ·mask()

[0092] mask()=rand(size(p i ))<μ bg

[0093] Among them, ∈ bg is the mutation intensity, μ bg is the mutation probability, rand() represents the generation of a normally distributed random number, and mask() represents the generation of a random number with the same distribution as individual p i Random binary masks of the same size control parameter variation, size() is the length measurement function, p′ i is the individual after random jump;

[0094] In S307, the global dynamic update process is represented as follows:

[0095] P (t+1) = sort(P (t) ,f)

[0096] Among them, sort() represents the function of sorting the population according to the fitness function f(), P (t) is the population of the tth iteration, P (t+1) is the population of the t+1th iteration.

[0097] Furthermore, the step S4 specifically includes the following steps:

[0098] S401. Construct a feature dimensionality reduction model, input the data in the updated feature extraction model, and initialize the weight matrices of the encoder and decoder of the feature dimensionality reduction model;

[0099] S402. Perform manifold structure sampling on the input data of the feature dimensionality reduction model by using the k-nearest neighbor algorithm to obtain the average distance of the k-nearest neighbors to estimate the local manifold structure;

[0100] S403. Performing random mask generation on the input data of the feature dimensionality reduction model to obtain random mask and masked input data;

[0101] S404. compressing the masked input data through an encoder to obtain a low-dimensional potential representation of the input data;

[0102] S405. In the latent space, a manifold constraint loss is obtained according to the average distance of the k nearest neighbors, and the low-dimensional latent representation is adjusted by the manifold constraint loss;

[0103] S406. Reconstruct the input data through the decoder according to the low-dimensional potential representation to obtain the reconstruction error and the manifold constraint error, and then update the neural network parameters through back propagation to calculate the total loss function;

[0104] S407. Optimizing the parameters of the autoencoder neural network using a gradient descent method;

[0105] S408. Repeat steps S401-S407 to iterate until the maximum number of iterations is met, that is, the feature dimensionality reduction model training is completed, and an updated feature dimensionality reduction model is obtained;

[0106] In S401, the initialization method is random initialization;

[0107] The initialization process is expressed as:

[0108]

[0109] Among them, W e is the weight matrix of the encoder, W d is the weight matrix of the decoder; b e is the bias vector of the encoder, b d is the bias vector of the decoder, σ 2 represents the standard deviation of initialization;

[0110] In S402, the k-nearest neighbor method is used to analyze the input data xr i Perform manifold structure sampling and calculate the input data xr i The average distance of the k nearest neighbors;

[0111] The manifold structure sampling process is expressed as:

[0112]

[0113] Among them, NN k (xr i ) is the input data xr i The k nearest neighbors of ∥xr i -x j ∥ 2Represents the first data point xr i and the second data point xr j The square of the Euclidean distance between them;

[0114] In S403, for the input data xr i , generate a random mask m i ;

[0115] Random mask m i It is expressed as:

[0116] m i = Bernoulli(p)

[0117] Among them, the random mask m i Each mask element of takes the value 0 or 1 independently, Bernoulli(p) represents Bernoulli distribution, and p is the probability that the mask element is 1;

[0118] In S404, the low-dimensional potential representation of the input data includes a first data point xr i The low-dimensional latent representation z i and the second data point xr j The low-dimensional latent representation z j ;

[0119] The first data point xr i The low-dimensional latent representation z i It is expressed as:

[0120] z i =Sig(W e ·(1+γ csd β csd )·(xr i ⊙m i )+b e )

[0121] The second data point xr j The low-dimensional latent representation z j It is expressed as:

[0122] z j =Sig(W e ·(1+γ csd β csd )·(xr j ⊙m i )+b e )

[0123] Among them, ⊙ represents the corresponding multiplication of elements, Sig() is the Sigmoid activation function, β csd is the dynamic coding adjustment coefficient, γ csd is the preset scaling factor;

[0124] In S405, the manifold constraint loss L manifold It is expressed as:

[0125]

[0126] In S406, the total loss function Lur is expressed as:

[0127]

[0128]

[0129] in, is the reconstructed input data, λ kc is the regularization parameter of the manifold constraint, represents the reconstruction error, λ kc L manifold represents the manifold constraint error;

[0130] In S407, the autoencoder neural network parameter optimization process is expressed as:

[0131]

[0132] Among them, α gk Learning rate optimized for neural network parameters.

[0133] Furthermore, the S5 specifically includes the following steps:

[0134] S501. Construct an extreme learning machine network, input the data in the updated feature dimensionality reduction model as classifier input data, randomly initialize the weight matrix and bias vector from the input layer to the hidden layer in the extreme learning machine network, and obtain the randomly initialized weight matrix and bias vector;

[0135] S502. Adopting an activation method based on Green's function to adjust the randomly initialized weight matrix and bias vector according to the local structure of the classifier input data for adaptive activation, and obtaining an activated output matrix;

[0136] S503. The hidden layers of the extreme learning machine network are all adaptively activated using the activation method based on the Green function to obtain the output of the i-th layer of the multi-layer structure;

[0137] S504. The output of the last hidden layer is calculated based on the output of the i-th layer of the multi-layer structure, the output layer is trained using the output of the last hidden layer, the weight matrix of the output layer is calculated using the least squares method, and the turbine fault diagnosis model is obtained by integration;

[0138] In S501, the process of randomly initializing the weight matrix and the bias vector is expressed as:

[0139] W (1) =randn(n',m')

[0140] b (1) =randn(m',1)

[0141] Among them, W (1) is the weight matrix from the input layer to the first hidden layer after random initialization, and its dimension is n'×m'; n' is the number of input features, m' is the number of neurons in the hidden layer, and b (1) is the bias vector from the input layer to the first hidden layer after random initialization, with a dimension of m'×1, and randn() represents drawing random numbers from a standard normal distribution;

[0142] In S502, the output matrix after activation is expressed as:

[0143] H=green(W (1) xv+b (1) )

[0144] Among them, x is the classifier input data matrix, green() represents the activation function based on Green's function;

[0145] The activation function green() based on Green's function is expressed as:

[0146]

[0147] Among them, G() is the Green's function, fr() is the input source function;

[0148] In S503, for a multi-layer structure, the output H of the i-th layer (i) It is expressed as:

[0149] H (i) = green(W (i) H (i-1) +b (i) )

[0150] Among them, W (i) is the weight matrix of the i-th layer, b (i) is the bias vector, H (i-1) is the output of the previous layer of the i-th layer;

[0151] In S504, the weight matrix W of the output layer (out) It is expressed as:

[0152] W (out) =(H (last)T H (last) +λI) -1 H (last)TTs

[0153] Among them, H (last) is the output of the last hidden layer, Ts is the target output matrix, and λ is the regularization parameter.

[0154] The beneficial effects of the present invention are as follows: the present invention adopts a generative adversarial network based on a momentum factor, generates augmented data through a generator and uses a discriminator to determine the authenticity, thereby improving the diversity and coverage of the data set, introducing a momentum factor in the training process to ensure the stability of the generation process, and preventing the generator from deviating in the early stages of training. In addition, the data augmentation technology improves the generalization ability of subsequent models by increasing the diversity of samples, and reduces the risk of overfitting due to insufficient samples; the present invention adopts a three-layer fully connected neural network and a dynamic niche neural optimization algorithm, the latter of which is based on the niche theory to optimize the neural network parameter configuration, ensure the optimal performance of the neural network when facing complex data, and avoid the problem of gradient disappearance or explosion. The feature extraction technology adjusts the parameters through the optimization algorithm, enhances the stability and performance of the neural network, and ensures that it can perform well in a variable data environment. performance in an environment; the present invention adopts an autoencoder neural network based on manifold learning, uses an encoder and a decoder to effectively compress and reconstruct high-dimensional data, enhances the processing capability of noise and outliers through a mask transformation mechanism, and retains the intrinsic structure of the data. The feature dimensionality reduction technology improves the processing efficiency by accurately mapping and reconstructing the data, and at the same time enhances the model's ability to identify abnormal data; the present invention adopts an extreme learning machine classification algorithm based on nonlinear metrics of activation functions, supports fast weight setting without iteration, effectively adapts to complex nonlinear data distribution, and improves the speed and accuracy of fault classification; therefore, the present invention improves the accuracy and efficiency of fault detection, reduces downtime, and extends equipment life by integrating advanced data processing, feature extraction, and machine learning technologies, and can provide scientific data support for the health management of steam turbines. BRIEF DESCRIPTION OF THE DRAWINGS

[0155] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0156] Figure 1 It is a flow chart of a steam turbine fault diagnosis method based on generative adversarial network;

[0157] Figure 2 A schematic diagram of an embodiment flow chart of feature extraction model training;

[0158] Figure 3 Schematic diagram of an embodiment of a feature dimensionality reduction model;

[0159] Figure 4 The present invention is a flow chart of an embodiment of a method for diagnosing steam turbine faults based on a generative adversarial network. DETAILED DESCRIPTION

[0160] In order to make the technical solutions and advantages of the embodiments of the present invention more clearly understood, the exemplary embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than an exhaustive list of all the embodiments. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0161] refer to Figure 1-Figure 4 The present embodiment is described in detail. A method for diagnosing a steam turbine fault based on a generative adversarial network comprises the following steps:

[0162] S1. Collect and annotate the steam turbine operation data output by the steam turbine sensor to obtain the original training set;

[0163] Specifically: In this embodiment, the steam turbine is equipped with multiple sensors, which record the operating status of the steam turbine in real time. The attributes of the steam turbine operating data include timestamp ax, speed ay (in rpm), pressure az (in bar), temperature aw (in °C), vibration frequency av (in Hz), vibration amplitude au (in mm), power at (in kW), flow as (in m 3 / h), current ar (unit is A), voltage aq (unit is V);

[0164] In practical applications, the attributes of steam turbine operation data are usually more than 10, and the number of attributes may reach dozens or even hundreds;

[0165] In this embodiment, the steam turbine operation data is annotated manually, and the annotated categories include normal, bearing damage, blade corrosion, seal looseness and other faults, a total of 5 categories of fault diagnosis results;

[0166] S2. Use the momentum factor-based generative adversarial network algorithm to build a sample generation model, input the original training set and perform training, and integrate to obtain training data;

[0167] S3. construct a feature extraction model, input the training data into the feature extraction model, and use the dynamic niche neural optimization algorithm for training to obtain an updated feature extraction model;

[0168] S4. construct a feature dimensionality reduction model using an autoencoder neural network algorithm based on manifold learning, input the collected data in the updated feature extraction model into the feature dimensionality reduction model, and train it using a gradient descent method to obtain an updated feature dimensionality reduction model;

[0169] S5. Construct an extreme learning machine network, input the data in the updated feature dimension reduction model into the classifier of the extreme learning machine network, and use the extreme learning machine classification algorithm based on the nonlinear metric of the activation function for training to obtain the turbine fault diagnosis model.

[0170] Furthermore, the step S2 specifically includes the following steps:

[0171] S201. Construct a sample generation model, set the maximum number of iterations, initialize the parameters of the generator and the discriminator by randomly extracting from a Gaussian distribution, and obtain a parameter initialization result;

[0172] S202. According to the parameter initialization result, randomly select a batch of data, i.e., real data, from the original training set to perform sample generation model training;

[0173] S203. Generate false data through the generator according to the generator parameters and the input random noise;

[0174] S204. Mix the real data and the fake data and input them into the discriminator, distinguish the real data from the fake data through the discriminator, and update the discriminator parameters in the parameter initialization result according to the discriminator loss function to obtain updated discriminator parameters;

[0175] S205. Referring to the process of updating the discriminator parameters, the generator parameters in the parameter initialization result are updated according to the generator loss function to obtain updated generator parameters;

[0176] S206. Repeat steps S201-S205 for iterative training until the maximum number of iterations is reached, that is, the sample generation training is completed and the training data is obtained;

[0177] Specifically: In this embodiment, the maximum number of iterations is set to 1000;

[0178] In S201, the momentum factor-based generative adversarial network algorithm includes a generator G and a discriminator D, the generator G is used to generate false data that is similar to real data but does not appear in the original steam turbine operation data, and the discriminator D is used to distinguish whether the input data is the original steam turbine operation data or false data;

[0179] The parameter initialization result is expressed as:

[0180]

[0181] Among them, θ g is the generator parameter, θ d are the discriminator parameters, represents a normal distribution with mean 0 and covariance as the identity matrix I;

[0182] In S202, during the random selection process, batch data The randomness of the selection is achieved through the sampling probability p(xe), that is, the real data sample xe is uniformly distributed in the original turbine operation data The sampling probability of being selected is p(xe);

[0183] The sampling probability p(xe) is expressed as:

[0184]

[0185] Among them, N is the original turbine operation data The total number of data points in ;

[0186] A randomly selected batch of data B is expressed as:

[0187]

[0188] Among them, xe i is the i-th real data sample, i=1,2,...,m, m is the batch size;

[0189] In S203, the false data sample It is expressed as:

[0190]

[0191] Where z is the input random noise;

[0192] In S204, the discriminator loss function L d It is expressed as:

[0193] L d =L d,real +L d,fake

[0194]

[0195] Among them, L d,real is the loss of real data, L d,fake is the loss of generated data, x i is a real data sample, is the i-th false data sample;

[0196] The process of updating the discriminator parameters is expressed as:

[0197]

[0198] Among them, θ d' is the updated discriminator parameter, λ d Update the learning rate for the discriminator, λ d Set to 0.01, is the discriminator gradient parameter;

[0199] In S205, the generator loss function L g It is expressed as:

[0200]

[0201] Among them, L M is the momentum factor;

[0202] Momentum Factor L M It is expressed as:

[0203]

[0204] Where M is the momentum factor metric, f() is the feature extraction function, ∥∥ M represents the distance metric obtained through distance metric learning;

[0205] The distance metric of the data sample features is expressed as:

[0206]

[0207] Among them, Me is a symmetric positive definite matrix, which represents the importance and relevance of different dimensions in the feature space, and T is the matrix transpose;

[0208] Calculate the generator gradient using the chain rule

[0209] Generator Gradients It is expressed as:

[0210]

[0211] in, Denotes the generator loss function L g The partial derivative with respect to the discriminator output, Represents the discriminator output to generate data The partial derivative of Indicates the generated data For the generator parameter θ g The partial derivative of

[0212] The process of updating the generator parameters is expressed as:

[0213]

[0214] Among them, θ g ' is the updated generator parameter, λg Update the learning rate for the generator, is the generator gradient parameter;

[0215] Specifically, the collection, acquisition, labeling and preprocessing of turbine operation data are time-consuming and labor-intensive, and insufficient samples in the original training set can easily lead to poor model generalization ability, while affecting the accuracy of the model. The present invention uses a generative adversarial network algorithm based on a momentum factor to generate samples, thereby achieving data expansion. The use of a momentum factor in the training process of the generator helps to maintain the stability of the generation process and prevent extreme deviations in the generator output in the early stages of training. The momentum factor can also ensure that the expanded data can better reflect the feature distribution under the actual equipment operating state. Under the guidance of the discriminator, adjusting the generator parameters can improve the authenticity of the false data, making it more difficult for the discriminator to identify it.

[0216] Furthermore, the step S3 specifically includes the following steps:

[0217] S301. Construct a feature extraction model based on a neural network, input training data, set the maximum number of iterations, and generate an initial population in the initialization phase;

[0218] S302. In the dynamic evaluation phase, the performance of the neural network parameters is evaluated using a fitness function according to the attractor properties of the dynamic trajectory;

[0219] S303. Define the neural network iteration process and calculate the loss function of the neural network;

[0220] S304. The individual adjusts its position in the ecological niche according to its fitness, and obtains an individual after position adjustment;

[0221] S305. Perform crossover and update operations to obtain new individuals;

[0222] S306. Using mutation mechanism based on nonlinear dynamics theory to perform mutation and mutation operations, so that individuals can make random jumps in a larger range;

[0223] S307. Perform global dynamic updates on all populations according to the theory of ecological dynamic systems;

[0224] S308. Repeat steps S301-S307 for iterative training until the maximum number of iterations is reached, indicating that the feature extraction model training is completed and an updated feature extraction model is obtained;

[0225] Specifically: In this embodiment, the maximum number of iterations is set to 1000;

[0226] In S301, P (0) is the initial population, individual p iThat is, the neural network parameters, which include the weight matrix and bias vector of the neural network;

[0227] Individual p i It is expressed as:

[0228]

[0229] Among them, Nu is the population size, W i is the i-th weight matrix, b i is the i-th bias vector, Represents direct concatenation of vectors;

[0230] In S302, the evaluation process is represented as follows:

[0231] fit(p i )=α Ts ·Acc(p i )+β Ts Stab(p i )

[0232] Among them, fit() is the fitness function, Acc() represents the accuracy on the labeled training data, which is calculated by the output of the preset Softmax function, Stab() represents dynamic stability, α Ts is the performance hyperparameter, β Ts is a stability hyperparameter, α Ts Set to 0.7, β Ts Set to 0.3;

[0233] Dynamic stability of neural network parameters Stab(p i ) is expressed as:

[0234]

[0235] Among them, δ sta is the attenuation coefficient, δ sta Set to 0.05, is the loss function of the neural network, w ij For individual p i The jth weight parameter in , Mu is the total number of weight parameters;

[0236] In S303, the regularization strength of the t-th iteration is defined as κ(t), and the regularization strength of the t+1-th iteration is defined as κ(t+1);

[0237] The iterative process is expressed as:

[0238] κ(t+1)=κ(t)·exp(-α pe g(t)+β pe ·(1-L(t)))

[0239] g(t)=(f(p (t) )-f(p (t-1) ))·exp(-γ pe ·Ds(p (t) ))

[0240]

[0241] Among them, α pe is the fitness gradient hyperparameter, α pe Set to 0.3, β pe Set to 0.7, β pe is the learning progress hyperparameter, g(t) is the fitness gradient, which measures the rate of change of the fitness function over time, L(t) is the learning progress, which represents the improvement of the performance of the neural network on the training data set at each iteration, and γ pe is the fractal dimension adjustment factor, Ds() is the fractal dimension calculation function;

[0242] Loss Function of Neural Network It is expressed as:

[0243]

[0244] Among them, R(W) is the L2 norm of the weight vector W, which is used to calculate the weight w i The sum of the squares of is the cross entropy loss;

[0245] In the above S304, during the individual position adjustment process, excellent individuals occupy the dominant ecological niche, and weaker individuals explore new ecological niches or are eliminated;

[0246] The individual position adjustment process is expressed as:

[0247]

[0248] Among them, f avg is the average fitness of the current population, η vd is the interaction adjustment factor, η vd Set to 0.1, p best is the parameter of the current optimal individual, γ vd To update the step size hyperparameter, γ vd Set to 0.1, λ vd is the learning strategy hyperparameter, λ vd Set to 0.5, Δp i is the individual after position adjustment;

[0249] In S305, the crossover and update operation process is represented as follows:

[0250]

[0251] Among them, p new A new individual may inherit the characteristics of two or more excellent individuals. is a randomly generated crossover factor, p parent1 is the first randomly selected individual among the individuals after position adjustment, p parent2 is the second individual randomly selected from the individuals after position adjustment;

[0252] In S306, the variation and mutation operation process is represented as follows:

[0253] p′ i =p i +∈ bg rand() mask()

[0254] mask()=rand(size(p i ))<μ bg

[0255] Among them, ∈ bg is the mutation strength, ∈ bg Set to 0.01, μ bg is the mutation probability, μ bg Set to 0.1, rand() means generating a normally distributed random number, and mask() means generating a random number with the individual p i Random binary masks of the same size control parameter variation, size() is the length measurement function, p′ i is the individual after random jump;

[0256] In S307, the global dynamic update process is represented as follows:

[0257] P (t+1) = sort(P (t) ,f)

[0258] Among them, sort() represents the function of sorting the population according to the fitness function f(). When sorting, individuals with high fitness are retained and individuals with low fitness are eliminated. (t) is the population of the tth iteration, P (t+1) is the population of the t+1th iteration;

[0259] Specifically, in this embodiment, Ds(p) is defined as the local fractal dimension at the parameter vector p, and a small radius ∈ is defined around the parameter vector p. br The local area of ​​the sphere, calculate the number of individuals N(∈ br );

[0260] Then the local fractal dimension Ds(p) at the parameter vector point p is expressed as:

[0261]

[0262] Among them, ∈ br Set to 0.0001;

[0263] The expanded data is input into the feature extraction model to train the feature extraction model. The present invention adopts a 3-layer fully connected neural network for feature extraction. In the prior art, some schemes use neural networks for feature extraction. In some neural network structures, gradient vanishing, gradient explosion or falling into local optimal solutions may be encountered, affecting the stability of training and the performance of the model. The present invention adopts a feature extraction model constructed by a neural network based on a dynamic niche neural optimization algorithm. Inspired by the niche theory in ecology, each species seeks its optimal position in the ecosystem to achieve an advantage in the competition for survival. The dynamic niche neural optimization algorithm aims to find the optimal configuration of neural network parameters so that the network performs best on a given task.

[0264] Furthermore, the step S4 specifically includes the following steps:

[0265] S401. Construct a feature dimensionality reduction model, input the data in the updated feature extraction model, and initialize the weight matrices of the encoder and decoder of the feature dimensionality reduction model;

[0266] S402. Perform manifold structure sampling on the input data of the feature dimensionality reduction model by using the k-nearest neighbor algorithm to obtain the average distance of the k-nearest neighbors to estimate the local manifold structure;

[0267] S403. Performing random mask generation on the input data of the feature dimensionality reduction model to obtain random mask and masked input data;

[0268] S404. compressing the masked input data through an encoder to obtain a low-dimensional potential representation of the input data;

[0269] S405. In the latent space, a manifold constraint loss is obtained according to the average distance of the k nearest neighbors, and the low-dimensional latent representation is adjusted by the manifold constraint loss;

[0270] S406. Reconstruct the input data through the decoder according to the low-dimensional potential representation to obtain the reconstruction error and the manifold constraint error, and then update the neural network parameters through back propagation to calculate the total loss function;

[0271] S407. Optimizing the parameters of the autoencoder neural network using a gradient descent method;

[0272] S408. Repeat steps S401-S407 to iterate until the maximum number of iterations is met, that is, the feature dimensionality reduction model training is completed, and an updated feature dimensionality reduction model is obtained;

[0273] Specifically: In this embodiment, the maximum number of iterations is set to 1000;

[0274] In S401, the initialization method is random initialization;

[0275] The initialization process is expressed as:

[0276]

[0277] Among them, W e is the weight matrix of the encoder, W d is the weight matrix of the decoder; b e is the bias vector of the encoder, b d is the bias vector of the decoder, σ 2 represents the standard deviation of initialization;

[0278] In S402, the k-nearest neighbor method is used to analyze the input data xr i Perform manifold structure sampling and calculate the input data xr i The average distance of the k nearest neighbors;

[0279] The manifold structure sampling process is expressed as:

[0280]

[0281] Among them, NN k (xr i ) is the input data xr i The k nearest neighbors of ∥xr i -x j ∥ 2 Represents the first data point xr i and the second data point xr j The square of the Euclidean distance between them;

[0282] In S403, for the input data xr i , generate a random mask m i ;

[0283] Random mask m i It is expressed as:

[0284] m i = Bernoulli(p)

[0285] Among them, the random mask m iEach mask element of takes the value 0 or 1 independently, Bernoulli(p) represents Bernoulli distribution, and p is the probability that the mask element is 1;

[0286] The Bernoulli distribution is a discrete distribution used to simulate a random variable with one of two possible results. In this embodiment, the Bernoulli random variable m ij The generation method is expressed as:

[0287]

[0288] Among them, U is a random number drawn from the uniform distribution Uniform(0,1), pr is the preset threshold parameter, indicating that m ij =1, pr is set to 0.45, if U is less than p, then the Bernoulli random variable m ij is set to 1 if U is greater than or equal to p, then the Bernoulli random variable m ij is set to 0;

[0289] In S404, the low-dimensional potential representation of the input data includes a first data point xr i The low-dimensional latent representation z i and the second data point xr j The low-dimensional latent representation z j ;

[0290] The first data point xr i The low-dimensional latent representation z i It is expressed as:

[0291] z i =Sig(W e ·(1+γ csd β csd )·(xr i ⊙m i )+b e )

[0292] The second data point xr j The low-dimensional latent representation z j It is expressed as:

[0293] z j =Sig(W e ·(1+γ csd β csd )·(xr j ⊙m i )+b e )

[0294] Among them, ⊙ represents the corresponding multiplication of elements, Sig() is the Sigmoid activation function, β csd is the dynamic coding adjustment coefficient, γcsd is the preset scaling factor, γ csd Set to 0.03;

[0295] The dynamic coding adjustment coefficient is dynamically adjusted based on the statistical characteristics of the data in the updated feature extraction model, and the skewness S(xr i ) and kurtosis K(xr i );

[0296] Skewness S(xr i ) is expressed as:

[0297]

[0298] Kurtosis K(xr i ) is expressed as:

[0299]

[0300] Among them, μ pd For xr i The mean value, σ pd For xr i The standard deviation of

[0301] Dynamic coding adjustment coefficient β csd It is expressed as:

[0302]

[0303] Among them, tanh() is the hyperbolic tangent function;

[0304] In S405, the manifold constraint loss L manifold It is expressed as:

[0305]

[0306] In S406, the total loss function Lur is expressed as:

[0307]

[0308]

[0309] in, is the reconstructed input data, λ kc is the regularization parameter of the manifold constraint, represents the reconstruction error, λ kc L manifold represents the manifold constraint error;

[0310] In S407, the autoencoder neural network parameter optimization process is expressed as:

[0311]

[0312] Among them, α gk is the learning rate for optimizing neural network parameters, α gk Set to 0.04;

[0313] Specifically, the present invention inputs the data after feature extraction, that is, the data in the updated feature extraction model, into the feature dimensionality reduction model to train the feature dimensionality reduction model, and adopts an autoencoder neural network algorithm based on manifold learning to perform feature dimensionality reduction. The autoencoder neural network based on manifold learning includes an encoder and a decoder. The function of the encoder is to map high-dimensional input features to a low-dimensional latent space, and the decoder attempts to reconstruct the original input from this low-dimensional latent space. The manifold learning principle is adopted to enable the neural network to better understand and retain the intrinsic structure of the data. At the same time, a mask transformation mechanism is adopted to enhance the robustness of the model to noise and outliers in the input data.

[0314] Furthermore, the S5 specifically includes the following steps:

[0315] S501. Construct an extreme learning machine network, input the data in the updated feature dimensionality reduction model as classifier input data, randomly initialize the weight matrix and bias vector from the input layer to the hidden layer in the extreme learning machine network, and obtain the randomly initialized weight matrix and bias vector;

[0316] S502. Adopting an activation method based on Green's function to adjust the randomly initialized weight matrix and bias vector according to the local structure of the classifier input data for adaptive activation, and obtaining an activated output matrix;

[0317] S503. The hidden layers of the extreme learning machine network are all adaptively activated using the activation method based on the Green function to obtain the output of the i-th layer of the multi-layer structure;

[0318] S504. The output of the last hidden layer is calculated based on the output of the i-th layer of the multi-layer structure, the output layer is trained using the output of the last hidden layer, the weight matrix of the output layer is calculated using the least squares method, and the turbine fault diagnosis model is obtained by integration;

[0319] In S501, the process of randomly initializing the weight matrix and the bias vector is expressed as:

[0320] W (1) =randn(n',m')

[0321] b (1) =randn(m',1)

[0322] Among them, W (1)is the weight matrix from the input layer to the first hidden layer after random initialization, and its dimension is n'×m'; n' is the number of input features, m' is the number of neurons in the hidden layer, and b (1) is the bias vector from the input layer to the first hidden layer after random initialization, with a dimension of m'×1, and randn() represents drawing random numbers from a standard normal distribution;

[0323] In S502, the output matrix after activation is expressed as:

[0324] H=green(W (1) xv+b (1) )

[0325] Among them, x is the classifier input data matrix, green() represents the activation function based on Green's function;

[0326] The activation function green() based on Green's function is expressed as:

[0327]

[0328] Among them, G() is the Green's function, which is the input source function;

[0329] In S503, for a multi-layer structure, the output H of the i-th layer (i) It is expressed as:

[0330] H (i) = green(W (i) H (i-1) +b (i) )

[0331] Among them, W (i) is the weight matrix of the i-th layer, b (i) is the bias vector, H (i-1) is the output of the previous layer of the i-th layer;

[0332] In S504, the weight matrix W of the output layer (out) It is expressed as:

[0333] W (out) =(H (last)T H (last) +λI) -1 H (last)T Ts

[0334] Among them, H (last) is the output of the last hidden layer, Ts is the target output matrix, and λ is the regularization parameter;

[0335] Specifically, in step S5, the dimension-reduced data is input into a classifier for classifier training. The present invention adopts an extreme learning machine classification algorithm based on nonlinear metrics of activation functions. The extreme learning machine is a feedforward neural network with a single hidden layer, which is fast and does not require iterative weight adjustment. The present invention improves its structure and increases the ability to support multi-layer hidden structures to adapt to more complex nonlinear data distributions. After the output layer training is completed, the entire extreme learning machine classification algorithm based on nonlinear metrics of activation functions is trained. The turbine fault diagnosis model after training is deployed for real-time fault diagnosis, which can quickly respond to and accurately identify potential faults of various turbines.

[0336] By using the trained turbine fault diagnosis model to process new samples, the potential fault status of the turbine can be quickly identified and classified. The newly collected turbine operation data is first processed by the trained feature extraction model, that is, the updated feature extraction model to extract key features. The data after key feature extraction is processed by the data dimension reduction model to obtain the reduced features. The reduced features are input into the classifier model, and finally the turbine fault diagnosis model outputs the fault diagnosis results. The fault diagnosis results include: normal, bearing damage, blade corrosion, seal looseness and other faults, a total of 5 categories.

[0337] Although the present invention has been described according to a limited number of embodiments, it will be apparent to those skilled in the art, with the benefit of the above description, that other embodiments may be envisioned within the scope of the invention thus described. In addition, it should be noted that the language used in this specification is selected primarily for readability and didactic purposes, rather than for explaining or defining the subject matter of the present invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the present invention is illustrative, not restrictive, with respect to the scope of the present invention, which is defined by the appended claims.

Claims

1. A steam turbine fault diagnosis method based on generative adversarial network, characterized in that: The following steps are involved: S1. Collect and annotate the steam turbine operation data output by the steam turbine sensor to obtain the original training set; S2. Use the momentum factor-based generative adversarial network algorithm to build a sample generation model, input the original training set and perform training, and integrate to obtain training data; S3. construct a feature extraction model, input the training data into the feature extraction model, and use the dynamic niche neural optimization algorithm for training to obtain an updated feature extraction model; S4. construct a feature dimensionality reduction model using an autoencoder neural network algorithm based on manifold learning, input the collected data in the updated feature extraction model into the feature dimensionality reduction model, and train it using a gradient descent method to obtain an updated feature dimensionality reduction model; S5. Construct an extreme learning machine network, input the data in the updated feature dimension reduction model into the classifier of the extreme learning machine network, and use the extreme learning machine classification algorithm based on the nonlinear metric of the activation function for training to obtain the turbine fault diagnosis model.

2. A steam turbine fault diagnosis method based on a generative adversarial network according to claim 1, characterized in that: The S2 specifically includes the following steps: S201. Construct a sample generation model, set the maximum number of iterations, initialize the parameters of the generator and the discriminator by randomly extracting from a Gaussian distribution, and obtain a parameter initialization result; S202. According to the parameter initialization result, randomly select a batch of data, i.e., real data, from the original training set to perform sample generation model training; S203. Generate false data through the generator according to the generator parameters and the input random noise; S204. Mix the real data and the fake data and input them into the discriminator, distinguish the real data from the fake data through the discriminator, and update the discriminator parameters in the parameter initialization result according to the discriminator loss function to obtain updated discriminator parameters; S205. Referring to the process of updating the discriminator parameters, the generator parameters in the parameter initialization result are updated according to the generator loss function to obtain updated generator parameters; S206. Repeat steps S201-S205 for iterative training until the maximum number of iterations is reached, that is, the sample generation training is completed and the training data is obtained; In S201, the momentum factor-based generative adversarial network algorithm includes a generator G and a discriminator D, the generator G is used to generate false data that is similar to real data but does not appear in the original steam turbine operation data, and the discriminator D is used to distinguish whether the input data is the original steam turbine operation data or false data; The parameter initialization result is expressed as: Among them, θ g is the generator parameter, θ d are the discriminator parameters, represents a normal distribution with mean 0 and covariance as the identity matrix I; In S202, during the random selection process, batch data The randomness of the selection is achieved through the sampling probability p(xe), that is, the real data sample xe is uniformly distributed in the original turbine operation data The sampling probability of being selected is p(xe); The sampling probability p(xe) is expressed as: Among them, N is the original turbine operation data The total number of data points in ; A batch of data randomly selected It is expressed as: Among them, xe i is the i-th real data sample, i=1,2,...,m, m is the batch size; In S203, the false data sample It is expressed as: Where z is the input random noise; In S204, the discriminator loss function L d It is expressed as: L d =L d,real +L d,fake Among them, L d,real is the loss of real data, L d,fake is the loss of generated data, x i is a real data sample, is the i-th false data sample; The process of updating the discriminator parameters is expressed as: Among them, θ d ' is the updated discriminator parameter, λ d Update the learning rate for the discriminator, is the discriminator gradient parameter; In S205, the generator loss function L g It is expressed as: Among them, L M is the momentum factor; Momentum Factor L M It is expressed as: Where M is the momentum factor metric, f() is the feature extraction function, ∥∥ M represents the distance metric obtained through distance metric learning; The distance metric of data sample features is expressed as: Among them, Me is a symmetric positive definite matrix, which represents the importance and relevance of different dimensions in the feature space, and T is the matrix transpose; Calculate the generator gradient using the chain rule Generator Gradients It is expressed as: in, Denotes the generator loss function L g The partial derivative with respect to the discriminator output, Represents the discriminator output to generate data The partial derivative of Indicates the generated data For the generator parameter θ g The partial derivative of The process of updating the generator parameters is expressed as: Among them, θ g ' is the updated generator parameter, λ g Update the learning rate for the generator, is the generator gradient parameter.

3. A steam turbine fault diagnosis method based on generative adversarial network according to claim 2, characterized in that: The S3 specifically includes the following steps: S301. Construct a feature extraction model based on a neural network, input training data, set the maximum number of iterations, and generate an initial population in the initialization phase; S302. In the dynamic evaluation phase, the performance of the neural network parameters is evaluated using a fitness function according to the attractor properties of the dynamic trajectory; S303. Define the neural network iteration process and calculate the loss function of the neural network; S304. The individual adjusts its position in the ecological niche according to its fitness, and obtains an individual after position adjustment; S305. Perform crossover and update operations to obtain new individuals; S306. Using mutation mechanism based on nonlinear dynamics theory to perform mutation and mutation operations, so that individuals can make random jumps in a larger range; S307. Perform global dynamic updates on all populations according to the theory of ecological dynamic systems; S308. Repeat steps S301-S307 for iterative training until the maximum number of iterations is reached, indicating that the feature extraction model training is completed and an updated feature extraction model is obtained; In S301, P (0) is the initial population, individual p i That is, the neural network parameters, which include the weight matrix and bias vector of the neural network; Individual p i It is expressed as: Among them, Nu is the population size, W i is the i-th weight matrix, b i is the i-th bias vector, Represents direct concatenation of vectors; In S302, the evaluation process is represented as follows: fit(p i )=a Ts ·Acc(p i )+b Ts ·Stab(p i ) Among them, fit() is the fitness function, Acc() represents the accuracy on the labeled training data, which is calculated by the output of the preset Softmax function, Stab() represents dynamic stability, α Ts is the performance hyperparameter, β Ts is a stability hyperparameter; Dynamic stability of neural network parameters Stab(p i ) is expressed as: Among them, δ sta is the attenuation coefficient, is the loss function of the neural network, w ij For individual p i The jth weight parameter in , Mu is the total number of weight parameters; In S303, the regularization strength of the t-th iteration is defined as κ(t), and the regularization strength of the t+1-th iteration is defined as κ(t+1); The iterative process is expressed as: κ(t+1)=κ(t)·exp(-α pe ·g(t)+β pe ·(1-L(t))) g(t)=(f(p (t) )-f(p (t-1) ))·exp(-γ pe ·Ds(p (t) )) Among them, α pe is the fitness gradient hyperparameter, β pe is the learning progress hyperparameter, g(t) is the fitness gradient, L(t) is the learning progress, γ pe is the fractal dimension adjustment factor, Ds() is the fractal dimension calculation function; Loss Function of Neural Network It is expressed as: Among them, R(W) is the L2 norm of the weight vector W, which is used to calculate the weight w i The sum of the squares of is the cross entropy loss; In S304, the individual position adjustment process is expressed as: Among them, f avg is the average fitness of the current population, η vd is the interaction adjustment factor, p best is the parameter of the current optimal individual, γ vd To update the step size hyperparameter, λ vd is the learning strategy hyperparameter, Δp i is the individual after position adjustment; In S305, the crossover and update operation process is represented as follows: p new =ζ kg ·p parent1 +(1-z kg )·p parent2 Among them, p new For a new individual, kg is a randomly generated crossover factor, ζ kg ∈(0,1), p parent1 is the first randomly selected individual among the individuals after position adjustment, p parent2 is the second individual randomly selected from the individuals after position adjustment; In S306, the variation and mutation operation process is represented as follows: pα i =p i +∈ bg ·rand()·mask() mask()=rand(size(p i ))<μ bg Among them, ∈ bg is the mutation intensity, μ bg is the mutation probability, rand() represents the generation of a normally distributed random number, and mask() represents the generation of a random number with the same distribution as individual p i Random binary masks of the same size control parameter variation, size() is the length measurement function, p′ i is the individual after random jump; In S307, the global dynamic update process is represented as follows: P (t+1) =sort(P (t) ,f) Among them, sort() represents the function of sorting the population according to the fitness function f(), P (t) is the population of the tth iteration, P (t+1) is the population of the t+1th iteration.

4. A steam turbine fault diagnosis method based on generative adversarial network according to claim 3, characterized in that: The S4 specifically includes the following steps: S401. Construct a feature dimensionality reduction model, input the data in the updated feature extraction model, and initialize the weight matrices of the encoder and decoder of the feature dimensionality reduction model; S402. Perform manifold structure sampling on the input data of the feature dimensionality reduction model by using the k-nearest neighbor algorithm to obtain the average distance of the k-nearest neighbors to estimate the local manifold structure; S403. Performing random mask generation on the input data of the feature dimensionality reduction model to obtain random mask and masked input data; S404. compressing the masked input data through an encoder to obtain a low-dimensional potential representation of the input data; S405. In the latent space, a manifold constraint loss is obtained according to the average distance of the k nearest neighbors, and the low-dimensional latent representation is adjusted by the manifold constraint loss; S406. Reconstruct the input data through the decoder according to the low-dimensional potential representation to obtain the reconstruction error and the manifold constraint error, and then update the neural network parameters through back propagation to calculate the total loss function; S407. Optimizing the parameters of the autoencoder neural network using a gradient descent method; S408. Repeat steps S401-S407 to iterate until the maximum number of iterations is met, that is, the feature dimensionality reduction model training is completed, and an updated feature dimensionality reduction model is obtained; In S401, the initialization method is random initialization; The initialization process is expressed as: Among them, W e is the weight matrix of the encoder, W d is the weight matrix of the decoder; b e is the bias vector of the encoder, b d is the bias vector of the decoder, σ 2 represents the standard deviation of initialization; In S402, the k-nearest neighbor method is used to analyze the input data xr i Perform manifold structure sampling and calculate the input data xr i The average distance of the k nearest neighbors; The manifold structure sampling process is expressed as: Among them, NN k (xr i ) is the input data xr i The k nearest neighbors of ∥xr i -x j ∥ 2 Represents the first data point xr i and the second data point xr j The square of the Euclidean distance between them; In S403, for the input data xr i , generate a random mask m i ; Random mask m i It is expressed as: m i =Bernoulli(p) Among them, the random mask m i Each mask element of takes the value 0 or 1 independently, Bernoulli(p) represents Bernoulli distribution, and p is the probability that the mask element is 1; In S404, the low-dimensional potential representation of the input data includes a first data point xr i The low-dimensional latent representation z i and the second data point xr j The low-dimensional latent representation z j ; The first data point xr i The low-dimensional latent representation z i It is expressed as: z i =Sig(W e ·(1+γ csd β csd )·(xr i ⊙m i )+b e ) The second data point xr j The low-dimensional latent representation z j It is expressed as: z j =Sig(W e ·(1+γ csd β csd )·(xr j ⊙m i )+b e ) Among them, ⊙ represents the corresponding multiplication of elements, Sig() is the Sigmoid activation function, β csd is the dynamic coding adjustment coefficient, γ csd is the preset scaling factor; In S405, the manifold constraint loss L manifold It is expressed as: In S406, the total loss function Lur is expressed as: in, is the reconstructed input data, λ kc is the regularization parameter of the manifold constraint, represents the reconstruction error, λ kc L manifold represents the manifold constraint error; In S407, the autoencoder neural network parameter optimization process is expressed as: Among them, α gk Learning rate optimized for neural network parameters.

5. A steam turbine fault diagnosis method based on generative adversarial network according to claim 4, characterized in that: The S5 specifically includes the following steps: S501. Construct an extreme learning machine network, input the data in the updated feature dimensionality reduction model as classifier input data, randomly initialize the weight matrix and bias vector from the input layer to the hidden layer in the extreme learning machine network, and obtain the randomly initialized weight matrix and bias vector; S502. Adopting an activation method based on Green's function to adjust the randomly initialized weight matrix and bias vector according to the local structure of the classifier input data for adaptive activation, and obtaining an activated output matrix; S503. The hidden layers of the extreme learning machine network are all adaptively activated using the activation method based on the Green function to obtain the output of the i-th layer of the multi-layer structure; S504. The output of the last hidden layer is calculated based on the output of the i-th layer of the multi-layer structure, the output layer is trained using the output of the last hidden layer, the weight matrix of the output layer is calculated using the least squares method, and the turbine fault diagnosis model is obtained by integration; In S501, the process of randomly initializing the weight matrix and the bias vector is expressed as: W (1) =randn(n',m') b (1) =randn(m',1) Among them, W (1) is the weight matrix from the input layer to the first hidden layer after random initialization, and its dimension is n'×m'; n' is the number of input features, m' is the number of neurons in the hidden layer, and b (1) is the bias vector from the input layer to the first hidden layer after random initialization, with a dimension of m'×1, and randn() represents drawing random numbers from a standard normal distribution; In S502, the output matrix after activation is expressed as: H=green(W (1) xv+b (1) ) Among them, x is the classifier input data matrix, green() represents the activation function based on Green's function; The activation function green() based on Green's function is expressed as: Among them, G() is the Green's function, which is the input source function; In S503, for a multi-layer structure, the output H of the i-th layer (i) It is expressed as: H (i) =green(W (i) H (i-1) +b (i) ) Among them, W (i) is the weight matrix of the i-th layer, b (i) is the bias vector, H (i-1) is the output of the previous layer of the i-th layer; In S504, the weight matrix W of the output layer (out) It is expressed as: W (out) =(H (last)T H (last) +λI) -1 H (last)T Ts Among them, H (last) is the output of the last hidden layer, Ts is the target output matrix, and λ is the regularization parameter.

Citation Information

Cited By

  • State prediction-oriented all-solid waste pervious concrete detection data processing method

    CN119272136A

  • Market dishonesty information knowledge graph entity identification method

    CN119272775A

  • Method and system for identifying operation state of compact five-return type biomass steam boiler

    CN119337260A

  • Geological mineral resource prediction model training method based on artificial intelligence

    CN119378642A

  • Training method of geological mineral resources prediction model based on artificial intelligence

    CN119378642B