Fault diagnosis method and system for ventilation system of hydraulic power plant

Abstract

The invention provides a hydraulic power plant ventilation system fault diagnosis method and system, and the method comprises the steps: obtaining a sensor historical data set containing various fault information in the operation process of a hydraulic power plant ventilation system; inputting the high-dimensional data set into an improved generative adversarial network (GAN) for data expansion, and performing feature dimension reduction on the expanded high-dimensional data set by using improved kernel principal component analysis (improved KPCA) to generate a low-dimensional data set; the method comprises the following steps: constructing a soft contrast learning framework SoftCLT model of a time sequence, using an elite reverse learning strategy, an integrated normal cloud model strategy and a Cauchy variation strategy to improve a mountaineering team optimization algorithm MTBO, and using the improved mountaineering team optimization algorithm MTBO to optimize hyper-parameters of the SoftCLT model based on a sample low-dimensional data set to obtain optimal model hyper-parameters; a ventilation system fault diagnosis model is obtained; and using the ventilation system fault diagnosis model to diagnose faults in the operation process of the hydraulic power plant ventilation system to obtain a corresponding diagnosis result.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fault diagnosis of ventilation systems of hydropower plants, and in particular to a method and system for fault diagnosis of ventilation systems of hydropower plants. BACKGROUND

[0002] With the increasing global energy demand and the increasing awareness of environmental protection, hydropower as a clean and renewable energy form has become increasingly important. The stable operation of hydropower plants is crucial to ensure power supply and grid safety. However, the ventilation system of a hydropower plant, as a key auxiliary system, directly affects the heat dissipation effect and operating efficiency of the hydraulic turbine, generator and other key equipment.

[0003] Generally, fault detection and diagnosis methods can be divided into three categories: model-based methods, knowledge-based methods and data-based methods. Model-based fault diagnosis methods are generally based on reaction mechanisms and require the establishment of a multi-dimensional, non-linear physical model of the entire system, which involves mechanical dynamics, fluid mechanics, thermodynamics, etc. Although this method can truly reflect the causal relationship between variables and faults, it is extremely challenging to establish a mechanism model of a hydropower plant ventilation system that considers fault diagnosis due to the complexity of the internal structure of the system and the difficulty in characterizing material properties. Knowledge-based methods rely on the professional knowledge of operators, which is very subjective. Data-based methods use historical data to identify the operating state of a hydropower plant ventilation system without the need for complex and accurate mathematical models, and are commonly used in fault detection research. Data-based methods can generally be divided into statistical methods and machine learning methods. Deep learning, as the dominant branch of machine learning, has strong data processing and feature learning capabilities and has been widely applied in the field of fault diagnosis.

[0004] Current fault diagnosis models cannot fully utilize the fault characteristics of hydropower plant ventilation system data, and the model parameters are not accurate enough, so the diagnosis accuracy needs to be improved. Existing technologies also use the Mountaineering Team-Based Optimization (MTBO) algorithm to optimize the diagnosis model, but it does not consider the shortcoming that the MTBO algorithm is prone to local optimization in the later iteration period. SUMMARY

[0005] The present application aims to at least partially solve one of the technical problems in the related art.

[0006] To this end, a first object of the present application is to provide a method for fault diagnosis of a ventilation system of a hydropower plant, thereby improving the utilization of fault information and effectively improving the fault diagnosis efficiency of the ventilation system of the hydropower plant during operation.

[0007] A second object of the present application is to provide a system for fault diagnosis of a ventilation system of a hydropower plant.

[0008] A third object of the present application is to provide an electronic device.

[0009] A fourth object of the present application is to provide a computer-readable storage medium.

[0010] A fifth object of the present application is to provide a computer program product.

[0011] To achieve the above objects, the first aspect of the present application provides a method for diagnosing faults of a ventilation system of a hydropower plant, comprising:

[0012] obtaining a sensor historical data set containing various fault information during operation of the ventilation system of the hydropower plant;

[0013] inputting the sensor historical data set into an improved generative adversarial network (GAN) for data augmentation, performing feature dimension reduction on the augmented high-dimensional data set using an improved kernel principal component analysis (improved KPCA), and generating a low-dimensional data set;

[0014] constructing a time series soft contrastive learning framework (SoftCLT) model, improving a mountain team optimization (MTBO) algorithm using an elite reverse learning strategy, an integrated normal cloud model strategy, and a Cauchy mutation strategy, and optimizing hyperparameters of the SoftCLT model based on the sample low-dimensional data set using the improved MTBO algorithm to obtain optimal model hyperparameters and obtain a ventilation system fault diagnosis model;

[0015] based on the low-dimensional data set, using the ventilation system fault diagnosis model to diagnose faults during operation of the ventilation system of the hydropower plant, and obtaining corresponding diagnosis results.

[0016] Optionally, the various fault information includes faults of air volume, air pressure, speed, temperature, current, external temperature, and humidity; and the fault types of the sensors include motor faults, fan faults, damper and auxiliary device faults.

[0017] Optionally, the improved GAN model generates samples using a two-stage model, including a variational autoencoder data augmentation stage and a deep attention mechanism optimization sample generation stage; both stages use a single-stage GAN as the backbone, input initial samples generated in the first stage into the second stage, and perform high-quality regeneration on the initial samples; and the inputting of the sensor historical data set into the improved GAN for data augmentation includes:

[0018] During the training process, the objective of the generator G is to generate real samples to deceive the discriminator D, and the final ideal state is D(G(z)) = 0.5, and the loss function formula of the GAN is:

[0019]

[0020] Wherein, V(D, G) represents the difference degree of generated samples and real samples, the cross-entropy loss function of binary classification includes And Two parts: Indicates that the fixed generator G trains the discriminator D, and the parameters of the discriminator D are updated by maximizing the cross-entropy loss V(D, G); Indicates that the fixed discriminator D trains the generator G;

[0021] The sensor historical data set is input into the variational autoencoder to obtain the first stage generated sample; the encoding network in the variational autoencoder VAE is composed of three fully connected layers, the first fully connected layer performs sampling operation on the input high-dimensional data, and the activation function adopts Relu; The other two fully connected layers respectively map the down-sampled samples into two low-dimensional parameter features, i.e. the mean and variance of the normal distribution; the loss function of the encoding network is KL divergence, which is used to measure the difference between the implicit vector distribution and the standard normal distribution, and the KL loss function is:

[0022]

[0023] Wherein, μ and σ 2 Respectively, the mean and variance output by the encoding network; F KL Indicates the calculation of the KL divergence of the two approximate distributions, which is used to represent the KL divergence between the data distribution and the standard normal distribution N(0, 1);

[0024] The decoding network in the variational autoencoder VAE is composed of two fully connected networks, the parameter settings of the two networks are opposite to the parameter settings of the encoding network, the input low-dimensional hidden variables are amplified in features, and finally the same size of the generated data as the input data is generated, and its loss function is:

[0025]

[0026] The discriminant network is composed of 2 convolutional layers and 3 fully connected layers, the activation function between each network layer adopts Relu, the output layer outputs the true and false label through the Sigmoid activation function, and the loss function of the discriminant network is:

[0027]

[0028] Wherein, x~P data (x) represents that the real sample conforms to P data (x) distribution; z~P z (z) represents that the expectation conforms to P z (z) distribution;

[0029] The sample generated in the first stage is concatenated with the feature vector to the second stage to obtain a final optimized generated image, i.e., an expanded high-dimensional data set; in the second stage, an encoder is used to remove the multi-layer perceptron classification head in the original network model, and a vector originally used for additional classification prediction output is removed as the encoder of the generator in this stage to obtain image features Y, which is expressed as:

[0030]

[0031] y l '=MSA(LN(y l-1 ))+y l-1 ;l=1,2,…,L.

[0032] y l =MLP(LN(y l '))+y l ';l=1,2,…,L.

[0033] Y=y1+y2+y3+…+y L

[0034] wherein, is a trainable linear projection parameter, E pos ∈R (N+1)×K is a position encoding, and LN(·) is a layer normalization function; the input y l-1 of each encoder block of the ViT encoder is the output of the previous encoder block, and the output y l of the current encoder block is calculated through multi-head self-attention and MLP, and after obtaining the sample features Y, the feature vector is concatenated to the image features to obtain a new feature vector Y', which is expressed as:

[0035]

[0036] The feature vector Y' is input into a series of convolutional decoders to obtain a second-stage generated sample, and then the sample is discriminated for authenticity by a discriminator to supervise the image generation quality; the loss function of the second stage is composed of a generator loss function and a discriminator loss function:

[0037]

[0038] wherein, is a second-stage generator loss function, is a second-stage discriminator loss function, is a deep attention multi-modal similarity model (DAMSM) loss function, For the feature matching loss function, λ1 and λ2 are the respective weight parameters of the DAMSM loss and the feature matching loss in the second stage generator.

[0039] Optionally, the improved kernel principal component analysis (improved KPCA) is used to reduce the dimensionality of the expanded high-dimensional data set, and a low-dimensional data set is generated.

[0040] The time series of the input signal is defined as x a (a = 1, 2,..., m) and linearly independent variables y a (a = 1, 2,..., n), and n ≤ m; singular value decomposition is used to obtain x a The expression is:

[0041] U = [x1, x2,..., x m ]

[0042] V = [y1, y2,..., y n ]

[0043] U = CV

[0044] In the formula, U is the input data matrix, C is an n × n transformation matrix, and V is the output observation matrix.

[0045] Further:

[0046] R x = BR y B T

[0047] Since the autocorrelation matrix R x is a real symmetric matrix, the autocorrelation matrix R y can be converted to a diagonal matrix composed of m positive real eigenvalues ξ i (i = 1, 2,..., m) in R x , and B is an n-dimensional vector, i.e.:

[0048] R y = [E(y i ) 2 ] = [ξ i ]

[0049] ξ1>ξ2>...>ξ m

[0050] In the autocorrelation matrix R y , the m eigenvalues correspond to m eigenvectors, and the m eigenvectors form m subspaces; in order to maintain the original information in the m eigenvectors, m satisfies:

[0051]

[0052] Setting up a nonlinear mapping function and the original feature set W = [w1, w2, …, wn]T s ] T Transforming into a kernel feature set where s is the number of samples; the covariance matrix F of the kernel feature set is expressed as:

[0053]

[0054] Through the eigenvalue λ i and eigenvector ζ i of the matrix F, the following expression is obtained:

[0055]

[0056] For the inner product operation in the high-dimensional space of KPCA, the Gaussian radial basis function is selected as the kernel function:

[0057]

[0058] Selecting q eigenvectors corresponding to the q largest eigenvalues as the column vectors of the transformation matrix γ, and the result of KPCA dimensionality reduction is:

[0059] R = γ T · W'

[0060] In order to measure the interval size between samples of the same class and samples of different classes, the intra-class cohesion degree is defined, and its expression is:

[0061]

[0062] where ε is the number of the i-th class of samples, σa is the standard deviation of the a-th dimensional component of the class; and βia is the a-th dimensional component of the j-th sample in the i-th class.

[0063] At the same time, the dispersion degree between different classes of samples is expressed as:

[0064]

[0065] where ε is the number of the i-th class of samples, σa is the standard deviation of the a-th dimensional component of the class; and βia is the a-th dimensional component of the j-th sample in the i-th class. aij The separability degree of the sample classes is:

[0066]

[0067] When the value of J a is smaller, the separability degree of the sample classes is higher, and the contribution degree to the sample classification is greater; in order to illustrate the correlation degree between the extracted eigenvectors and the sample class vectors, its expression is defined as: ​

[0068]

[0069] where τ is a correlation coefficient, ζ i is the extracted feature vector, ζ s is the sample category vector in the original feature set.

[0070] According to the correlation coefficient τ, the feature vector with high contribution to sample classification is retained, and the interference of invalid samples is reduced; if the absolute value of τ tends to 1, it means that the correlation between the two vectors is stronger; if the absolute value of τ tends to 0, it means that the correlation is weaker.

[0071] Optionally, the implementation process of the time series soft contrast learning framework SoftCLT model includes:

[0072] Input the sample low-dimensional data set into the time series soft contrast learning framework SoftCLT model;

[0073] Given a batch of K time series X = {x1,......,x k}, the goal is to learn a nonlinear embedding function f θ that maps time series to representation vectors where T c is the sequence length, H is the input feature dimension, and M is the embedding feature dimension.

[0074] Put the four parameters of sharpness, loss contribution, iteration number and learning rate into the improved mountain-climbing optimization algorithm for optimization, and then perform the subsequent steps.

[0075] Use the minimum-maximum normalized distance metric D1(·,·), and use the sigmoid function β v (a) = 1 / (1+exp(-a)) to define the soft assignment of a pair of data indices (e,e') for instance contrastive loss:

[0076] W I (e,e') = 2α·β v (-τ I ·D1(x e ,x e' ))

[0077] where τ I is a hyperparameter that controls the sharpness, and α is an upper bound in the range [0,1] that distinguishes between pairs of the same time series and pairs of close different time series; when α = 1, the pair with a distance of 0 and the pair with the same time series are assigned a value of 1.

[0078] Let r e,t = r e+2K,t and For two augmented embedding vectors of x at time stamp t e The softmax probability of relative similarity among all similarities when computing loss is defined as:

[0079]

[0080] where dot product is used as similarity measure For two augmented embedding vectors of x at time stamp t e The soft instance contrastive loss at time stamp t is defined as:

[0081]

[0082] where, The first term in corresponds to the loss of positive pairs, and the second term corresponds to the loss of other pairs weighted by soft assignment W I (e,e') ;

[0083] When time stamps are closer, the assignment is close to 1, and when they are farther apart, the assignment is close to 0. Define a soft assignment of a pair of time stamps (t, t') for time contrastive loss:

[0084]

[0085] where, is a hyper-parameter to control the sharpness, and by adjusting to control the degree of soft assignment;

[0086] Let and For two augmented embedding vectors of x at time stamp t e The softmax probability of relative similarity among all similarities when computing loss is defined as:

[0087]

[0088] The soft time contrastive loss at time stamp t is defined as:

[0089]

[0090] where, The first term in corresponds to the loss of positive pairs, and the second term corresponds to the loss of other pairs weighted by soft assignment ;

[0091] The final loss of SoftCLT is the combination of soft instance and time contrastive loss:

[0092]

[0093] where λ cis a hyperparameter controlling each loss contribution;

[0094] The final output is obtained using the full connection layer to obtain the fault diagnosis result of the ventilation system of the hydropower plant.

[0095] Optionally, the elite reverse learning strategy, integrated normal cloud model strategy and Cauchy mutation strategy are used to improve the mountain team optimization algorithm MTBO, and the hyperparameters of the SoftCLT model are optimized based on the sample low-dimensional data set using the improved mountain team optimization algorithm MTBO, to obtain the best model hyperparameters, and obtain the ventilation system fault diagnosis model, including:

[0096] The objective function of the mountain team optimization algorithm MTBO is set as the ventilation system fault diagnosis accuracy, and the related parameters are initialized, including: population size, dimension size, sharpness, loss contribution, iteration number and learning rate;

[0097] The elite reverse learning strategy is used to replace the original random initialization in the MTBO algorithm, the reverse solution of the current solution is calculated to increase the diversity of the population, and the optimal solution is selected from the current solution and the reverse solution as a new generation of individuals; The reverse solution of the elite is defined as: assuming that the extreme point corresponding to the individual in the population is the elite individual Its reverse solution is defined as:

[0098]

[0099] Wherein, L' is a d-dimensional vector on (0, 1), Alpha j = min(X i,j ), beta j = min(X i,j ), alpha j , beta j are dynamic boundaries, which overcome the shortcomings of fixed boundaries that are difficult to preserve search experience, so that the elite reverse solution can search in a narrow space and is not easy to fall into local optimum; If the dynamic boundary operation makes Cross the boundary and become an infeasible solution, reset it using a random generation method, as follows:

[0100]

[0101] According to the objective function, the fitness value of the fault data is calculated, and the optimal solution of the ventilation system fault diagnosis accuracy is obtained according to the calculated fault data fitness.

[0102] The integrated normal cloud model strategy is used to replace the original mountain team algorithm position updating formula, which is represented by a mathematical formula as follows:

[0103] X = [x1, x2, ..., x Nd =Gnc(Ex,En,He,Nd)

[0104] Where Ex is the expected value, En is the entropy value, He is the hyperentropy value, and Nd is the expected number of cloud droplets;

[0105] The new formula for calculating the assisted climbing phase is:

[0106] position = Gnc(position) best ,En,He,dim)

[0107]

[0108]

[0109] Among them, Position best The optimal position for the current population; dim sets the dimension for the algorithm; λ e =0.3, θ=2, t″ is the current iteration number, and T″ is the maximum iteration number;

[0110] In the disaster threat phase of the MTBO algorithm, the position update formula for the phase solution is set as follows:

[0111]

[0112] in, To obtain the new position, X Avalanche This represents the location of other solutions when a local optimum occurs randomly.

[0113] During the solution process, when a local optimum occurs, the position update formula for each solution is as follows:

[0114]

[0115] Among them, X Team The average position of all solutions;

[0116] The random member replacement phase of the original mountaineering team optimization algorithm is replaced with a Cauchy mutation strategy. The position update formula for this phase using the Cauchy mutation operator is as follows:

[0117]

[0118] Where cauchy(0,1) is the standard Cauchy distribution function;

[0119] The current solution is continuously updated through the MTBO optimization process, and the global optimal solution is updated synchronously.

[0120] It is judged whether the maximum iteration number is reached, if yes, the optimal solution is output, and the hyperparameters of the SoftCLT are extracted therefrom, otherwise, the fault data fitness value is calculated according to the objective function, and the subsequent steps are performed.

[0121] To achieve the above purpose, the second aspect of the present application proposes a hydroelectric power plant ventilation system fault diagnosis system, comprising:

[0122] A data acquisition module is configured to acquire a sensor historical data set containing various fault information during operation of the ventilation system of the hydroelectric power plant.

[0123] A data processing module is configured to input the sensor historical data set into an improved generative adversarial network (GAN) for data augmentation, perform feature dimension reduction on the augmented high-dimensional data set by using an improved kernel principal component analysis (improved KPCA), and generate a low-dimensional data set.

[0124] A model construction and optimization module is configured to construct a time series soft contrast learning (SoftCLT) model, improve a mountain team optimization (MTBO) algorithm by using an elite reverse learning strategy, an integrated normal cloud model strategy, and a Cauchy mutation strategy, optimize hyperparameters of the SoftCLT model based on the sample low-dimensional data set by using the improved MTBO algorithm, obtain optimal model hyperparameters, and obtain a ventilation system fault diagnosis model.

[0125] A fault diagnosis module is configured to diagnose faults in the operation of the ventilation system of the hydroelectric power plant by using the ventilation system fault diagnosis model based on the low-dimensional data set, and obtain corresponding diagnosis results.

[0126] To achieve the above purpose, the third aspect of the present application proposes an electronic device, comprising a processor and a memory in communication connection with the processor.

[0127] The memory stores computer execution instructions.

[0128] The processor executes the computer execution instructions stored in the memory to implement the method of any one of the first aspect.

[0129] To achieve the above purpose, the fourth aspect of the present application proposes a computer readable storage medium, wherein the computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by a processor to implement the method of any one of the first aspect.

[0130] To achieve the above purpose, the fifth aspect of the present application proposes a computer program product, which is executed by a processor to implement the method of any one of the first aspect.

[0131] The embodiments of the present application provide at least the following beneficial effects:

[0132] The improved generative adversarial network (GAN) is used to expand the unbalanced data to achieve data set balance in the present application. The ratio of normal service to fault state in the data of the operation state monitoring of the ventilation system of the hydropower plant is seriously out of balance, and the normal samples are more and the fault samples are less, which leads to the problems of unsatisfactory training effect of the ventilation system fault diagnosis model and low diagnosis accuracy. The improved generative adversarial network (GAN) is used to solve the problem of unbalanced sample data.

[0133] The improved mountain-climbing team optimization algorithm is used for parameter optimization in the present application. In order to avoid the problem of the mountain-climbing team optimization algorithm falling into local optimum, an improved strategy is proposed, which uses the elite reverse learning strategy to replace the random initialization of the population position in the original mountain-climbing team optimization algorithm. The integrated normal cloud model strategy and Cauchy mutation strategy are introduced in the optimization process to enhance the disturbance of the optimal solution position update and increase the diversity of the population, which helps to jump out of the local optimum.

[0134] In addition, the time series soft contrast learning framework (SoftCLT) model is established in the present application, and the improved mountain-climbing team optimization algorithm is used to optimize the SoftCLT model, which significantly improves the accuracy of model fault diagnosis.

[0135] The additional aspects and advantages of the present application will be partially given in the following description, partially become obvious from the following description, or be known by the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0136] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, with reference to the following drawings, in which:

[0137] Figure 1 A flowchart of a hydropower plant ventilation system fault diagnosis method provided by an embodiment of the present application;

[0138] Figure 2 A flowchart of a hydropower plant ventilation system fault diagnosis method provided by an embodiment of the present application;

[0139] Figure 3 An improved VDGAN model architecture provided by an embodiment of the present application;

[0140] Figure 4 A flowchart of an improved MTBO algorithm provided by an embodiment of the present application. DETAILED DESCRIPTION

[0141] Embodiments of the present application are described below in detail, examples of which are shown in the accompanying drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application.

[0142] To solve the problems in the prior art, the embodiments of the present application provide a hydroelectric power plant ventilation system fault diagnosis method, Figure 1 and Figure 2 A flowchart of a hydroelectric power plant ventilation system fault diagnosis method provided by the embodiments is shown in the figure. As shown in Figure 1 and Figure 2 , the method comprises the following steps:

[0143] Step 1, obtaining a sensor historical data set containing various fault information in the running process of the hydroelectric power plant ventilation system.

[0144] In the implementation process of the present application, first, the historical data of various sensors in the hydroelectric power plant ventilation system need to be obtained. These data sets contain a plurality of key indicators related to the operation of the ventilation system, and the changes of these indicators when faults occur. Specifically, these data sets cover the following aspects:

[0145] Air volume fault: the air volume of the ventilation system does not reach the expected value, which may be caused by fan failure, air door blockage or control system abnormality.

[0146] Air pressure fault: abnormal fluctuation of air flow pressure in the ventilation system, which may be caused by fan performance decline or pipeline blockage.

[0147] Speed fault: the fan speed does not reach the set range, which may be caused by motor failure or fan component problem.

[0148] Temperature fault: the temperature sensor in the system reads abnormal value, which may be caused by environmental change or system failure.

[0149] Current fault: abnormal fluctuation of motor current, which may indicate that the motor has a fault or abnormal load change.

[0150] External temperature and humidity fault: changes in external environmental temperature or humidity have a certain impact on system performance, and if the sensor fails, it may affect the stability of the ventilation system.

[0151] Motor fault: the motor, as the core component of the hydroelectric power plant ventilation system, its fault will directly affect the running state of the whole ventilation system. The sensor fault of the motor is manifested as abnormal fluctuation of current, speed and other parameters.

[0152] Fan failure: The key role of the fan in the ventilation system makes the fan failure cause obvious abnormalities in parameters such as air volume and air pressure. Fan sensor failure often accompanies changes in the fan operating state, affecting air volume and air pressure.

[0153] Damper and auxiliary device failure: Damper failure may cause abnormal regulation of air volume flow, affecting the normal operation of the system. The failure of auxiliary devices such as electric valves may also cause similar operation abnormalities.

[0154] The historical data set of these sensors is obtained to comprehensively understand the variation of various parameters of the ventilation system under different failure conditions, providing basic data support for subsequent fault diagnosis and prediction. In order to ensure the integrity and accuracy of the data, the historical data set should include sensor data under different seasons and different operating conditions, so as to analyze the influence of different factors on the performance of the ventilation system.

[0155] Step 2: Input the sensor historical data set into the improved generative adversarial network (GAN) for data augmentation, and use the improved kernel principal component analysis (improved KPCA) to reduce the dimension of the augmented high-dimensional data set, generating a low-dimensional data set.

[0156] Due to the difficulty in obtaining large amounts of hydroelectric power plant ventilation system failure data in actual work, the imbalance between normal data and failure data seriously affects the training of deep models, resulting in low diagnostic accuracy, etc. Therefore, the present application proposes a VDGAN model for few-shot parameters combining variational autoencoder and deep attention mechanism, as shown in Figure 3 The specific implementation steps are as follows:

[0157] (1) The VDGAN model proposed in the present application generates samples in two stages, namely the variational autoencoder data enhancement stage and the deep attention mechanism optimized sample generation stage; both stages use a single-level GAN as the backbone, and the initial samples generated in the first stage are input into the second stage for high-quality regeneration of the initial samples.

[0158] (2) During training, the goal of the generator G is to generate as many real samples as possible to deceive the discriminator D, and the ideal state is D(G(z)) = 0.5, so the loss function formula of GAN is:

[0159] where V(D, G) represents the difference between the generated samples and the real samples, and the cross-entropy loss function of binary classification includes and Two parts: D(G(z)) represents training the discriminator D by fixing the generator G, and updating the parameters of the discriminator D by maximizing the cross-entropy loss V(D, G); D denotes the training of the generator G, the generator wants to minimize the cross-entropy function under the condition that the discriminator maximizes the true and false sample cross-entropy loss V(D, G).

[0160] (3) The encoding network in the variational autoencoder VAE is composed of three fully connected layers. The first fully connected layer performs a sampling operation on the input high-dimensional data, and the activation function is Relu. The other two fully connected layers respectively map the down-sampled samples into two low-dimensional parameter features, i.e. the mean and variance of the normal distribution. The loss function of the encoding network is the KL divergence, which is used to measure the difference between the implicit vector distribution and the standard normal distribution. The KL loss function is:

[0161]

[0162] Wherein, μ and σ 2 are the mean and variance of the encoding network output; F KL denotes the calculation of the KL divergence of two approximate distributions, which is used to represent the KL divergence between the data distribution obtained in the application and the standard normal distribution N(0, 1);

[0163] In the embodiment of the application, the decoding network (generating network) in the variational autoencoder VAE is composed of two fully connected networks. The parameter settings of the two networks are opposite to the parameter settings of the encoding network. The input low-dimensional hidden variables are amplified in features, and finally the same size of the generated data as the input data is generated. The Relu activation function is used between the two network layers, and the Sigmoid activation function is used in the final output layer. The optimization goal of the generating network not only needs to ensure the minimization of the feature loss between the generated sample and the true sample, but also needs to ensure that the generated data can deceive the discrimination network. The loss function of the generating network is:

[0164]

[0165] In the embodiment of the application, the discrimination network is composed of 2 convolutional layers and 3 fully connected layers. The activation function between each network layer adopts Relu, and the output layer outputs the true and false label through the Sigmoid activation function. The loss function of the discrimination network is:

[0166]

[0167] Wherein, x ~ P data (x) denotes that the real sample is subject to P data (x) distribution; z ~ P z (z) denotes that the expectation is subject to P z (z) distribution;

[0168] (4) The sample is input to the second stage through the high-quality new sample obtained in the first stage, and the final optimized generated image is obtained.

[0169] In the embodiments of the present application, the input of the second stage is the sample generated by the first stage and the feature vector In this stage, the encoder is used to remove the multi-layer perceptron classification head in the original network model, and the vector originally used for additional classification prediction output is removed. The encoder of the generator in this stage is obtained, and the image feature Y is expressed as:

[0170]

[0171] y l '=MSA(LN(y l-1 ))+y l-1 ;l=1,2,…,L.

[0172] y l =MLP(LN(y l '))+y l ';l=1,2,…,L.

[0173] Y=y1+y2+y3+…+y L

[0174] Wherein: is a trainable linear projection parameter, E pos ∈R (N+1)×K is a position encoding, LN(·) is a layer normalization function; the input y l-1 of each encoder block of the ViT encoder is the output of the previous encoder block, and the output y l of the current encoder block is calculated through multi-head self-attention and MLP, and the sample feature Y is obtained. After that, the feature vector Y' is spliced into the image feature to obtain a new feature vector Y', and the expression is:

[0175]

[0176] The feature vector Y' is input into a series of convolution decoding to obtain the second stage generated sample, and then the discriminator discriminates the authenticity of the sample to supervise the image generation quality.

[0177] The generator and the discriminator are alternately trained, the generator loss and the discriminator loss of the data enhancement stage are optimized through the variational autoencoder data enhancement stage and the deep attention mechanism, and the sample meeting the text condition is obtained; the loss function of the DFA-GAN model is the second stage loss function, and the loss function of this stage is composed of the generator loss function and the discriminator loss function, as follows:

[0178]

[0179] Wherein: is a second stage generator loss function, is a second stage discriminator loss function, is a deep attention multi-modal similarity model (DAMSM) loss function, is a feature matching loss function, and λ1, λ2 are the respective weight parameters of the DAMSM loss and the feature matching loss for the second stage generator.

[0180] It should be noted that the second stage generator re-enters the The purpose is to avoid the instability of the adversarial generation network in the training process, so that the generator can better capture the similarity between the intermediate layer feature representations, and help the generator to generate more real and coherent images.

[0181] Further, the improved kernel principal component analysis (improved KPCA) is used to reduce the dimension of the expanded high-dimensional data set and eliminate redundant data to generate a low-dimensional data set. The specific steps are as follows:

[0182] (1) Define a time series of input signals as x a (a=1,2,…,m) and linearly independent variables y a (a=1,2,…,n), and n≤m. x a can be obtained by singular value decomposition, and its expression is:

[0183] U=[x1,x2,...,x m ]

[0184] V=[y1,y2,...,y n ]

[0185] U=CV

[0186] In the formula, U is the input data matrix, C is the n×n transformation matrix, and V is the output observation matrix.

[0187] Further,

[0188] R x =BR y B T

[0189] Since the autocorrelation matrix R x is a real symmetric matrix, the autocorrelation matrix R y can be converted into a diagonal matrix composed of m positive real eigenvalues ξ i (i=1,2,…,m) in R x , and B is an n-dimensional vector, i.e.:

[0190]

[0191] ξ1>ξ2>…>ξ m

[0192] In the autocorrelation matrix R y , m eigenvalues correspond to m eigenvectors, and the m eigenvectors constitute m subspaces; in order to maintain the original information in the m eigenvectors, m should satisfy:

[0193]

[0194] From the above formula, it can be seen that the eigenvalue ξ i compresses the information contained in the original signal, so the eigenvalue and the eigenvector can be used to classify the fault modes.

[0195] (2) Establish a nonlinear mapping function and convert the original feature set W = [w1, w2, …, w s ] T into a kernel feature set where s is the number of samples; the covariance matrix F of the kernel feature set is expressed as:

[0196]

[0197] Further, the eigenvalue λ i and the eigenvector ζ i of the matrix F are obtained, and the following expression is obtained:

[0198]

[0199] For the inner product operation in the high-dimensional space of KPCA, the Gaussian radial basis function is selected as the kernel function:

[0200]

[0201] Then, q eigenvectors corresponding to the q largest eigenvalues are selected as the column vectors of the transformation matrix γ, and the result of KPCA dimension reduction is:

[0202] R = γ T · W'

[0203] In order to measure the interval size between the same samples and different samples, the application defines the intra-class cohesion degree, which is expressed as:

[0204]

[0205] In the formula, ε is the number of the i-th class sample, is the standard deviation of the a-th dimensional component of the class.

[0206] At the same time, the dispersion degree expression between different types of samples is:

[0207]

[0208] In the formula, ε is the number of the i-th type of sample, β aij is the a-th dimensional component of the j-th sample in the i-th type, and the sample class separability is:

[0209]

[0210] When the value of J a is smaller, the sample class separability is higher, and the contribution to the sample classification is greater. In order to illustrate the correlation degree between the extracted feature vector and the sample class vector, an embodiment of the application defines its expression as:

[0211]

[0212] In the formula, τ is the correlation degree coefficient, ζ i is the extracted feature vector, and ζ s is the sample class vector in the original feature set.

[0213] (3) According to the correlation degree coefficient τ, the feature vector with high contribution to the sample classification is retained, and the interference of invalid samples is reduced; if the absolute value of τ tends to 1, it indicates that the correlation between the two vectors is stronger; if the absolute value of τ tends to 0, it indicates that the correlation is weaker. Therefore, according to the value of τ, the feature vector with high contribution to the sample classification can be retained, the interference of invalid samples can be reduced, and the accuracy of subsequent fault classification can be improved.

[0214] Step 3, constructing a time series soft contrast learning framework SoftCLT model, using an elite reverse learning strategy, an integrated normal cloud model strategy and a Cauchy mutation strategy to improve the mountain team optimization algorithm MTBO, and based on the sample low-dimensional data set, using the improved mountain team optimization algorithm MTBO to optimize the hyperparameters of the SoftCLT model, obtaining the best model hyperparameters, obtaining the ventilation system fault diagnosis model.

[0215] In an embodiment of the application, the implementation process of constructing the time series soft contrast learning framework SoftCLT model includes:

[0216] (1) Input the sample low-dimensional data set into the time series soft contrast learning framework SoftCLT model.

[0217] (2) Given a batch of K time series X = {x1,......,x k}, the goal is to learn a nonlinear embedding function f θMapping time series to represent vector In the formula T c H is the sequence length, H is the input feature dimension, and M is the embedding feature dimension.

[0218] (3) The four parameters of sharpness, loss contribution, number of iterations and learning rate are put into the improved mountaineering team optimization algorithm for optimization and subsequent steps.

[0219] (4) Use the minimum-maximum normalized distance metric D1(·,·) and the sigmoid function β. v (a) = 1 / (1 + exp(-a)) defines a pair of soft assignments for data indices (e, e') used for instance contrast loss:

[0220] W I (e,e')=2α·β v (-τ I ·D1(x e ,x e' ))

[0221] Where, τ I It is a hyperparameter that controls sharpness. α is the upper bound in the range [0,1] that distinguishes between pairs of the same time series and pairs of different time series that are close to each other. When α = 1, pairs with a distance of 0 and pairs with the same time series are assigned a value of 1.

[0222] (5) Let r e,t =r e+2K,t and x at timestamp t e The two augmented embedding vectors define the softmax probability of relative similarity among all similarities when calculating the loss as:

[0223]

[0224] Among them, the dot product is used as a similarity measure. Timestamp t at time x e The soft instance contrast loss is defined as:

[0225]

[0226] in, The first term corresponds to the direct loss, and the second term corresponds to the soft assignment W. I (e,e') is the loss of other weighted pairs.

[0227] (6) As timestamps get closer, the assignment is close to 1; as they get further apart, the assignment is close to 0. In this embodiment, a pair of timestamps (t, t') is further defined as a soft assignment for time comparison loss:

[0228]

[0229] where, is a hyper-parameter controlling the sharpness, by adjusting to control the degree of soft assignment.

[0230] (7) Let and be two augmented embedding vectors of x e at time stamp t, we define the softmax probability of relative similarity among all similarities when computing the loss as:

[0231]

[0232] Then the soft time-contrastive loss at time stamp t is defined as:

[0233]

[0234] where, the first term in the above equation corresponds to the loss of positive pairs, and the second term corresponds to the loss of other pairs weighted by the soft assignment .

[0235] (8) The final loss of SoftCLT is the combination of soft instance and time-contrastive losses:

[0236]

[0237] where λ c is a hyper-parameter controlling the contribution of each loss;

[0238] (9) The final output is obtained using a fully connected layer to obtain the fault diagnosis result of the ventilation system of the hydropower plant.

[0239] In the embodiment of the application, the elite reverse learning strategy, the integrated normal cloud model strategy and the Cauchy mutation strategy are used to improve the mountain team optimization algorithm MTBO, and the low-dimensional data set is used to optimize the hyperparameters of the SoftCLT model based on the improved mountain team optimization algorithm MTBO, the best model hyperparameters are obtained, and the implementation steps of the ventilation system fault diagnosis model are referred to Figure 4 as follows:

[0240] (1) Set the objective function of the mountain team optimization algorithm MTBO as the ventilation system fault diagnosis accuracy and initialize the related parameters, including: population size, dimension size, sharpness, loss contribution, iteration number and learning rate.

[0241] (2) The elite reverse learning strategy is used to replace the original random initialization in the MTBO algorithm. The reverse solution of the current solution is calculated to increase the diversity of the population, and the optimal solution is selected from the current solution and the reverse solution as a new generation of individuals. The elite reverse solution is defined as follows: assuming that the extreme point corresponding to the individual in the population is an elite individual its reverse solution can be defined as:

[0242]

[0243] wherein L' is a d-dimensional vector on (0, 1), α j = min(X i,j ), β j = min(X i,j ), α j , β j are dynamic boundaries. The dynamic boundaries overcome the shortcomings of fixed boundaries in preserving search experience, so that the elite reverse solution can search in a narrow space and is not easy to fall into local optimum. If the dynamic boundary operation makes the boundary beyond the boundary becomes an infeasible solution, the random generation method is used for resetting, and the resetting method is as follows:

[0244]

[0245] (3) The fitness value of the fault data is calculated according to the objective function, and the optimal solution of the fault diagnosis accuracy of the ventilation system is obtained according to the calculated fault data fitness.

[0246] (4) The integrated normal cloud model strategy is used to replace the original climbing team position updating formula in the mountain climbing algorithm, and this process is represented by a mathematical formula as follows:

[0247] X = [x1, x2,..., x Nd ] = Gnc(Ex, En, He, Nd)

[0248] wherein Ex is the expected value, En is the entropy value, He is the hyper-entropy value, and Nd is the expected number of cloud droplets.

[0249] The new climbing team stage calculation formula is:

[0250] position = Gnc(position best , En, He, dim)

[0251]

[0252] wherein Position best is the optimal position of the current population; dim is the dimension set by the algorithm; and λ e= 0.3, 0 = 2, t" is the current iteration number, T" is the maximum iteration number.

[0253] (5) In the disaster threat stage of the MTBO algorithm, the occurrence of an avalanche or a fall off a cliff may threaten the lives of mountaineers, thereby preventing the mountaineering team from reaching the summit, that is, forcing the algorithm to fall into a local optimal state; in the algorithm, in order to avoid falling into a local optimum and moving towards the best solution position, the position update formula of the solution in this stage is:

[0254]

[0255] wherein, is the new position obtained, X Avalanche is the position of other solutions when a local optimal state occurs randomly.

[0256] (6) In the solving process, when a local optimum occurs, the entire algorithm will try to rescue and avoid each solution from falling into the position update formula as follows:

[0257]

[0258] wherein, X Team is the average position of all solutions.

[0259] (7) The Cauchy mutation strategy is used to replace the random replacement member stage of the original mountaineering team optimization algorithm. In the random replacement solution stage, a new solution is generated in the form of a normal distribution random number. The position update formula of the Cauchy mutation operator used in this stage is:

[0260]

[0261] wherein, cauchy(0, 1) is a standard Cauchy distribution function.

[0262] (8) The current solution is updated continuously through the MTBO optimization process, and the global optimal solution is updated synchronously.

[0263] (9) It is judged whether the maximum iteration number is reached, if yes, the optimal solution is output, and the hyperparameters of the SoftCLT are extracted therefrom, otherwise, step (3) is returned, the fault data fitness value is calculated according to the objective function, and subsequent steps are performed.

[0264] Step 4, based on the low-dimensional data set, the ventilation system fault diagnosis model is used to diagnose the faults in the operation process of the ventilation system of the hydropower plant, and the corresponding diagnosis result is obtained.

[0265] In the embodiment of the present application, based on these low-dimensional data, the trained ventilation system fault diagnosis model is applied to diagnose the faults of the ventilation system of the hydropower plant, thereby obtaining the corresponding diagnosis result.

[0266] Based on the diagnosis results, the operation and maintenance personnel can quickly take the following measures:

[0267] Fault repair: according to the fault type and severity, take appropriate repair measures to ensure that the system resumes normal operation as soon as possible; preventive maintenance: in the case of no fault, predict the fault trend and running state of the system in advance to reduce system downtime and prolong equipment life; real-time monitoring and alarm: deploy real-time monitoring modules in the ventilation system, and the system performs real-time diagnosis through continuous sensor data to automatically trigger alarms for timely response by operation and maintenance personnel.

[0268] In addition, the embodiments of the present application also store and display the original data of the operation process of the ventilation system of the hydropower plant and the corresponding diagnosis results, and at the same time, issue an alarm and display the fault type when a fault occurs.

[0269] In order to realize the above-mentioned embodiments, the present application also proposes a fault diagnosis system for a ventilation system of a hydropower plant, which comprises:

[0270] A data acquisition module for acquiring a sensor historical data set containing various fault information during the operation process of the ventilation system of the hydropower plant;

[0271] A data processing module for inputting the sensor historical data set into an improved generative adversarial network (GAN) for data augmentation, using an improved kernel principal component analysis (improved KPCA) to reduce the dimensionality of the augmented high-dimensional data set, and generating a low-dimensional data set;

[0272] A model construction and optimization module for constructing a time series soft contrast learning (SoftCLT) model, using an elite back learning strategy, an integrated normal cloud model strategy, and a Cauchy mutation strategy to improve the mountain team optimization (MTBO) algorithm, and based on the sample low-dimensional data set, using the improved mountain team optimization (MTBO) algorithm to optimize the hyperparameters of the SoftCLT model to obtain the best model hyperparameters, and obtaining a ventilation system fault diagnosis model;

[0273] A fault diagnosis module for diagnosing the faults in the operation process of the ventilation system of the hydropower plant based on the low-dimensional data set using the ventilation system fault diagnosis model, and obtaining the corresponding diagnosis results.

[0274] Regarding the system in the above-mentioned embodiments, the specific manner in which each module performs operations has been described in detail in the embodiments related to the method, and will not be described in detail here.

[0275] To achieve the above-mentioned embodiments, the present application further provides an electronic device, comprising: a processor, and a memory connected with the processor in communication; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory to realize the method provided by the foregoing embodiments.

[0276] To achieve the above-mentioned embodiments, the present application further provides a computer readable storage medium, wherein the computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by a processor to realize the method provided by the foregoing embodiments.

[0277] To achieve the above-mentioned embodiments, the present application further provides a computer program product, comprising a computer program, wherein the computer program is executed by a processor to realize the method provided by the foregoing embodiments.

[0278] The collection, storage, use, processing, transmission, provision and disclosure of user personal information involved in the present application comply with relevant laws and regulations and do not violate public order and good customs.

[0279] It should be noted that the personal information from the user should be collected for legal and reasonable purposes, and should not be shared or sold outside these legal uses. In addition, such collection / sharing should be carried out after the user's informed consent is received, including but not limited to informing the user to read the user agreement / user notice before the user uses the function, and signing the agreement / authorization including authorization of relevant user information. In addition, any necessary steps should be taken to protect and ensure access to such personal information data, and to ensure that other people with access to personal information data comply with their privacy policy and processes.

[0280] It should be understood that the various forms of flow shown above can be reordered, added or deleted steps. For example, the steps described in the present application can be executed in parallel, sequentially or in different order, as long as the desired results of the technical solutions of the present application can be achieved, which are not limited herein.

[0281] The above detailed description does not constitute a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A method for diagnosing a fault of a ventilation system of a hydroelectric power plant, characterized by, The method comprises the following steps: obtain a sensor historical data set containing various fault information in the operation process of the ventilation system of the hydropower plant; input the sensor historical data set into an improved generative adversarial network (GAN) for data augmentation, perform feature dimension reduction on the augmented high-dimensional data set by using improved kernel principal component analysis (improved KPCA), and generate a low-dimensional data set; construct a time series soft contrast learning framework (SoftCLT) model, improve the mountain climbing team optimization (MTBO) algorithm by using an elite reverse learning strategy, an integrated normal cloud model strategy, and a Cauchy mutation strategy, optimize the hyperparameters of the SoftCLT model based on the sample low-dimensional data set by using the improved MTBO algorithm, obtain the best model hyperparameters, and obtain a ventilation system fault diagnosis model; based on the low-dimensional data set, diagnose the faults in the operation process of the ventilation system of the hydropower plant by using the ventilation system fault diagnosis model, and obtain corresponding diagnosis results.

2. The method of claim 1, wherein, The various fault information includes faults of air volume, air pressure, rotating speed, temperature, current, external temperature, and humidity; and the fault types of the sensors include motor faults, fan faults, and faults of air doors and auxiliary devices.

3. The method of claim 2, wherein, The improved GAN model generates samples in two stages, namely a variational autoencoder data enhancement stage and a deep attention mechanism sample generation stage; both stages use a single-level GAN as the main part, input the initial samples generated in the first stage into the second stage, and regenerate the initial samples with high quality; the inputting of the sensor historical data set into the improved GAN for data augmentation comprises the following steps: During the training process, the objective of the generator G is to generate real samples to deceive the discriminator D, and the final ideal state is D(G(z))=0.5, and the loss function formula of the GAN is as follows: Wherein, V(D, G) represents the difference degree of the generated sample and the real sample, the cross-entropy loss function of binary classification includes And Two parts: The fixed generator G trains the discriminator D, and the parameters of the discriminator D are updated by maximizing the cross-entropy loss V(D, G); The fixed discriminator D trains the generator G; input the sensor historical data set into the variational autoencoder to obtain the samples generated in the first stage; the encoding network in the variational autoencoder (VAE) comprises three fully connected layers, the first fully connected layer performs a sampling operation on the input high-dimensional data, and the activation function adopts Relu; the other two fully connected layers respectively map the down-sampled samples into two low-dimensional parameter features, namely the mean and variance of the normal distribution; the loss function of the encoding network is the KL divergence, which is used to measure the difference between the implicit vector distribution and the standard normal distribution, and the KL loss function is as follows: where μ and σ 2 are the mean and variance of the encoding network output, respectively; F KL denotes the computation of the KL divergence of two approximated distributions, used to express the KL divergence between the data distribution and the standard normal distribution N(0, 1); the decoding network in the variational autoencoder (VAE) comprises two fully connected networks, the parameter settings of the two networks are opposite to those of the encoding network, the input low-dimensional hidden variables are amplified in features, and finally the same size generated data as the input data are generated, and the loss function thereof is as follows: the discriminator network comprises two convolutional layers and three fully connected layers, the activation function between each network layer adopts Relu, the output layer outputs a true or false label through a Sigmoid activation function, and the loss function of the discriminator network is as follows: where x ~ P data (x) indicates that the true sample follows P data (x) distribution; z ~ P z (z) indicates that the expectation follows P z (z) distribution; The sample generated in the first stage and the feature vector are input to the second stage to obtain a final optimized generated image, i.e., an expanded high-dimensional data set; in the second stage, an encoder is used to remove the multi-layer perceptron classification head in the original network model, and a vector originally used for additional classification and prediction of output results is removed, serving as the encoder of the generator in this stage to obtain image features Y, expressed as: y l ' = MSA(LN(y l-1 )) + y l-1 ; l = 1, 2,..., L. y l = MLP(LN(y l ')) + y l ' ; i = 1,2,..., L. Y = y1 + y2 + y3 +... + y L wherein, is trainable linear projection parameter, E pos ∈R (N+1)×K is position encoding, LN(·) is layer normalization function; input y of each encoder block of ViT encoder l-1 is output of the previous encoder block, output y of the current encoder block is calculated by multi-head self-attention and MLP l , after obtaining sample feature Y, feature vector is spliced into image feature to obtain new feature vector Y', expression is: The feature vector Y' is input into a series of convolutional decoding to obtain a second stage generated sample, and then the discriminator discriminates the authenticity of the sample to supervise the image generation quality; the loss function of the second stage is composed of the generator loss function and the discriminator loss function: wherein, is a second stage generator loss function, is a second stage discriminator loss function, is a deep attention multi-modal similarity model (DAMSM) loss function, is a feature matching loss function, and λ1, λ2 are the respective weight parameters of the DAMSM loss and the feature matching loss for the second stage generator.

4. The method of claim 3, wherein, The improved kernel principal component analysis (improved KPCA) is used to reduce the dimensionality of the expanded high-dimensional data set, and a low-dimensional data set is generated, including: Let the time series of the input signal be denoted by x a (a = 1, 2,..., m) and linearly independent variables y a (a = 1, 2,..., n), and n < m; x is obtained by singular value decomposition a The expression is: U = [x1, x2,..., x m ] V = [y1, y2,..., y n ] U = CV In the formula, U is an input data matrix, C is an n*n transformation matrix, and V is an output observation matrix. Further has: R x = BR y B T Since the autocorrelation matrix R x is a real symmetric matrix, the autocorrelation matrix R y can be transformed into R x with m positive real eigenvalues ξ i (i = 1, 2, …, m) and a diagonal matrix B of size n x n, i.e.: R y = [E(y i ) 2 ] = [ξ i ] ξ1>ξ2>...>ξ m In the autocorrelation matrix R y m eigenvalues correspond to m eigenvectors, and the m eigenvectors constitute m subspaces; in order to maintain the original information in the m eigenvectors, m satisfies: Setting up a non-linear mapping function and the original feature set W = [w1, w2, …, wn]T s ] T Transforming into a kernel feature set where s is the number of samples; and the covariance matrix F of the kernel feature set is expressed as: By the eigenvalues λ i and eigenvectors ζ i of the matrix F, the following expression is obtained: For the inner product operation in high dimensional space, Gaussian radial basis function is chosen as the kernel function: Select q eigenvectors corresponding to the q largest eigenvalues as column vectors of the transformation matrix γ, and the KPCA dimensionality reduction result is: R = γ T • W' In order to measure the interval size between the same samples and different samples, the intra-class aggregation degree is defined, and its expression is: where ε is the number of the ith class sample, is the standard deviation of the athdimensional component of the class. At the same time, the dispersion degree expression between different samples is: In the formula, ε is the number of the i-th sample, β aij is the a-th dimensional component of the j-th sample in the i-th class, and the separability of the sample classes is: When J a The smaller the value, the higher the degree of separability of the sample category, and the greater the contribution to the sample classification. In order to illustrate the correlation between the extracted feature vector and the sample category vector, the expression is defined as: where τ is a correlation degree coefficient, ζ i is the extracted feature vector, ζ s is the sample class vector in the original feature set; According to the correlation coefficient τ, the feature vectors with high classification contribution degree are retained, and the interference of invalid samples is reduced; if the absolute value of τ tends to 1, it means that the correlation between the two vectors is stronger; if the absolute value of τ tends to 0, it means that the correlation is weaker.

5. The method of claim 4, wherein, The implementation process of the soft contrast learning framework SoftCLT model of time series includes: The low-dimensional data set is input into the SoftCLT model of time series; Given a set of K time series X = {x 1, ..., x k}, the goal is to learn a nonlinear embedding function f θ that maps time series to representation vectors where T c is the sequence length, H is the input feature dimension, and M is the embedding feature dimension; Put the four parameters of sharpness, loss contribution, iteration number and learning rate into the improved climbing team optimization algorithm for optimization, and then perform the subsequent steps; Using the min-max normalized distance metric D1(·,·), using the sigmoid function β v (a) = 1 / (1 + exp(-a)) defines a soft assignment of a pair of data indices (e, e') for the instance contrast loss: W I (e,e') = 2a-β v (-τ I ·D1(x e ,x e' )) where τ I is a hyperparameter that controls the sharpness, and a is an upper bound in the range [0, 1] that distinguishes between identical time series pairs and close time series pairs; when a = 1, pairs with distance 0 and pairs with identical time series are assigned a value of 1; Let r e,t = r e+2K,t and be the two augmented embedding vectors of x e at time stamp t, the softmax probabilities of relative similarities among all similarities when computing the loss are defined as: where the dot product is used as the similarity measure timestamp t at time x e The soft instance contrast loss is defined as: wherein, The first term in (1) corresponds to the loss of a direct hit, the second term corresponds to the loss of a soft assignment W I the loss of other pairs of (e, e') weighted When the timestamps are closer and closer, the assignment is close to 1, and when they are farther and farther away, the assignment is close to 0. Define the soft assignment of a pair of timestamps (t, t') for time contrast loss: wherein, is a hyperparameter that controls the sharpness, by adjusting the degree of soft assignment. Let and be the two augmented embedding vectors of x e at time t, the softmax probabilities of relative similarities among all similarities when computing the loss are defined as: Then define the soft time contrast loss at time t as: wherein, The first term in the sum corresponds to the loss of a direct pair, the second term corresponds to the loss of a soft assignment to the other pairs weighted. The final loss of SoftCLT is the combination of soft instance and time contrast loss: where λ c is a hyperparameter controlling each loss contribution; The final output is obtained by using the full connection layer, and the fault diagnosis result of the ventilation system of the hydropower plant is obtained.

6. The method of claim 5, wherein, The elite reverse learning strategy, integrated normal cloud model strategy and Cauchy mutation strategy are used to improve the mountain team optimization algorithm MTBO, and the hyperparameters of the SoftCLT model are optimized based on the low-dimensional data set using the improved mountain team optimization algorithm MTBO to obtain the best model hyperparameters, and the ventilation system fault diagnosis model is obtained, including: The objective function of the mountain team optimization algorithm MTBO is set to the ventilation system fault diagnosis accuracy, and the related parameters are initialized, including: population size, dimension size, sharpness, loss contribution, iteration number and learning rate; The elite reverse learning strategy is used to replace the original random initialization in the MTBO algorithm, the reverse solution of the current solution is calculated to increase the diversity of the population, and the optimal solution is selected from the current solution and the reverse solution as a new generation of individuals; the elite reverse solution is defined as: assuming that the extreme point corresponding to the individual in the population is an elite individual its reverse solution is defined as: where L' is a d-dimensional vector on (0, 1), α j = min(X i,j ), β j = min(X i,j ), α j , β j are dynamic boundaries, which overcome the shortcoming of fixed boundaries in preserving search experience, so that the elite reverse solution can search in a narrow space and is not easy to fall into local optimum; if the dynamic boundary operation makes the boundary become infeasible, it is reset by a random generation method as follows: According to the objective function, the fault data fitness value is calculated, and the optimal solution of the ventilation system fault diagnosis accuracy is obtained according to the calculated fault data fitness value; The integrated normal cloud model strategy is used to replace the auxiliary mountain position update formula in the original mountain team algorithm, and this process is represented by a mathematical formula as follows: X = [x1, x2,..., x Nd ] = Gnc(Ex, En, He, Nd) Wherein, Ex is the expected value, En is the entropy value, He is the hyper-entropy value, and Nd is the expected number of cloud droplets. The new auxiliary mountain stage calculation formula is: position = Gnc(position best , En, He, dim) where Position best is the current population optimal position; dim is the algorithm setting dimension; λ e = 0.3, θ = 2, t" is the current iteration number, and T" is the maximum iteration number. In the disaster threat stage of the MTBO algorithm, the position update formula of the stage solution is set as: wherein, X is the new position obtained Avalanche X is the position of other solutions when the local optimal state occurs randomly. In the solving process, when local optimization occurs, the position update formula of each solution is as follows: where X Team is the average position of all solutions; The Cauchy mutation strategy is used to replace the random replacement member stage of the original climbing team optimization algorithm, and the position update formula of the Cauchy mutation operator is used in the stage: Wherein, cauchy(0, 1) is a standard Cauchy distribution function; The current solution is updated through the MTBO optimization process, and the global optimal solution is also updated synchronously. It is judged whether the maximum iteration number is reached, if yes, the optimal solution is output, and the hyperparameters of the SoftCLT are extracted therefrom, otherwise, the fault data fitness value is calculated according to the objective function, and the subsequent steps are performed.

7. A hydroelectric power plant ventilation system fault diagnosis system characterized by, It comprises: A data acquisition module is configured to acquire a sensor historical data set containing various fault information in the operation process of the ventilation system of the hydropower plant; A data processing module is configured to input the sensor historical data set into an improved generative adversarial network (GAN) for data augmentation, and use an improved kernel principal component analysis (improved KPCA) to perform feature dimension reduction on the augmented high-dimensional data set to generate a low-dimensional data set; A model construction and optimization module is configured to construct a time series soft contrast learning (SoftCLT) model, improve a mountain team optimization (MTBO) algorithm using an elite reverse learning strategy, an integrated normal cloud model strategy and a Cauchy mutation strategy, and optimize hyperparameters of the SoftCLT model based on the sample low-dimensional data set using the improved MTBO to obtain optimal model hyperparameters and obtain a ventilation system fault diagnosis model; A fault diagnosis module is configured to diagnose faults in the operation process of the ventilation system of the hydropower plant based on the low-dimensional data set using the ventilation system fault diagnosis model to obtain corresponding diagnosis results.

8. An electronic device, characterized in that, It comprises: A processor and a memory in communication with the processor; The memory stores computer execution instructions; The processor executes the computer execution instructions stored in the memory to implement the method of any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by the processor to implement the method of any one of claims 1-6.

10. A computer program product, characterised in that, It comprises a computer program which, when executed by the processor, implements the method of any one of claims 1-6.

Images