Nuclear power equipment fault diagnosis method based on machine learning

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a nuclear power equipment fault diagnosis method based on machine learning. The method comprises the following steps: S1, data acquisition and labeling; s2, data expansion; s3, training a feature extraction model; s4, training a feature dimension reduction model; s5, training a classifier model; and S6, performing fault diagnosis on the nuclear power equipment. According to the method, the generative adversarial network of quantum state reinforcement learning is adopted for data expansion, the problem that the model generalization ability is poor due to insufficient training samples is solved, meanwhile, the feature extraction process of the neural network is optimized through the chaos optimization algorithm, and it is ensured that key features can be effectively extracted when complex high-dimensional data is processed. According to the method, feature dimension reduction is carried out by adopting an auto-encoder based on dynamic routing, information redundancy is reduced, the data processing efficiency is improved, finally, accurate classification of equipment fault states is realized through a high-order neural network of dynamic step optimization, and the accuracy and reliability of fault diagnosis are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and particularly relates to a nuclear power equipment fault diagnosis method based on machine learning. Background Art

[0002] During the operation of nuclear power equipment, various key equipment are in a complex working environment for a long time, facing various physical and electrical factors such as temperature, pressure, vibration, and electrical load. The stable operation of these equipment is directly related to the safety and production efficiency of nuclear power plants. Therefore, timely and accurate equipment fault diagnosis is crucial for ensuring the safety of nuclear power plants. However, the existing equipment fault diagnosis methods generally rely on expert experience or simple statistical analysis models, and it is difficult to handle the complex and changeable operating conditions of nuclear power equipment. Especially when the amount of equipment operation data is huge and there are complex non-linear relationships between various physical parameters, the diagnostic accuracy of traditional methods often fails to meet the actual requirements. Summary of the Invention

[0003] The purpose of the present invention is to provide a nuclear power equipment fault diagnosis method based on machine learning, aiming to solve the problem of insufficient fault diagnosis accuracy in the complex operating environment of nuclear power equipment. By using a generative adversarial network with quantum state enhanced learning for data augmentation, the problem of poor model generalization ability caused by insufficient training samples is overcome. At the same time, a chaotic optimization algorithm is used to optimize the feature extraction process of the neural network to ensure that key features can be effectively extracted when dealing with complex high-dimensional data. In addition, the present invention also uses an autoencoder based on dynamic routing for feature dimensionality reduction, reduces information redundancy, improves data processing efficiency, and finally realizes the accurate classification of equipment fault states through a high-order neural network optimized by dynamic step, improving the accuracy and reliability of fault diagnosis.

[0004] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0005] A nuclear power equipment fault diagnosis method based on machine learning:

[0006] S1. Data acquisition and annotation

[0007] The data collected from the nuclear power plant exists in vector form, including temperature, pressure, and flow measurement values, and the storage format of the data is structured JSON format;

[0008] S2. Data augmentation

[0009] Use a generative adversarial network based on quantum state reinforcement learning for sample generation, and then achieve data augmentation. The generative adversarial network based on quantum state reinforcement learning consists of two parts: a generator and a discriminator. The generator is responsible for generating high-quality data samples, and the discriminator evaluates the authenticity of the samples. The generator utilizes the characteristics of quantum computing to enhance its learning and generation capabilities, using quantum noise and quantum entanglement as the sources of random variables to improve the diversity and complexity of the generated samples; the discriminator adopts a symmetric quantum network design to more accurately evaluate the authenticity of the samples, enabling the discriminator to more sensitively capture the subtle variations in the data when processing the generated data;

[0010] S3. Feature extraction model training

[0011] Input the augmented data into the feature extraction model for training the feature extraction model. Use a 6-layer fully connected neural network for feature extraction, and adopt a chaotic optimization algorithm to optimize the parameters of the neural network model to achieve the training of the neural network model. Adopt a self-regulating mechanism that allows the network to dynamically adjust its weight update strategy according to the error of the previous training cycle, and utilize the sensitivity in chaos theory to enhance the model's response ability to changes in initial conditions and parameters;

[0012] S4. Feature dimensionality reduction model training

[0013] Input the data after feature extraction into the feature dimensionality reduction model for training the feature dimensionality reduction model. Use an autoencoder algorithm based on dynamic routing for feature dimensionality reduction. The autoencoder consists of a symmetric encoder and decoder, and adopts a dynamic routing mechanism based on feature importance scoring to automatically determine the transmission paths of each feature in different layers of the encoder, enhancing the model's ability to reconstruct the original data while reducing information loss;

[0014] S5. Classifier model training

[0015] Input the data after dimensionality reduction into the classifier for training the classifier. Use a high-order neural network based on dynamic step optimization as the classification algorithm. Based on the high-order neural network, not only consider the influence of individual input features, but also consider the interaction between features, and capture complex non-linear relationships by using high-order terms to improve the generalization ability of the model;

[0016] S6. Nuclear power equipment fault diagnosis

[0017] Use the trained model to conduct fault diagnosis on new samples. The collected original data is input into the trained feature extraction and feature dimensionality reduction models for feature processing, and the processed features are input into the classifier model for training the classifier, and then obtain the classification results, including normal operation, abnormal warning, and fault status.

[0018] S1. Data collection and annotation: The attributes of the data include: temperature T a representing the temperature of the critical parts of the device, in degrees Celsius; pressure P a representing the pressure related to the main steam pipeline, in Pascals; flow rate F a representing the flow rate of the coolant or steam, in cubic meters per hour; vibration level V a representing the vibration level generated during the operation of the device, in meters per second squared; current I a representing the working current of the electrical equipment, in amperes; voltage U a representing the voltage during the operation of the device, in volts; power W a representing the power consumed by the device, in watts; bearing temperature B a representing the temperature of the critical bearings, in degrees Celsius; bearing wear index E a representing a quantitative index of the bearing wear condition, dimensionless; corrosion index C a representing a quantitative index of the corrosion state of the device, dimensionless.

[0019] S1. Data collection and annotation: The collected data is annotated. The annotation method is manual annotation, and the annotation categories include normal operation, abnormal warning, and fault status.

[0020] S2. Data augmentation: The training process of the generative adversarial network algorithm based on quantum state enhanced learning:

[0021] S201. Randomly initialize the parameters of the generator and discriminator. Let the generator be G c , and the discriminator be D c . The weights of the generator are The weights of the discriminator are The weights of the generator and discriminator are initialized randomly and follow a normal distribution with a mean of 0 and a standard deviation of the identity matrix;

[0022] S202. Before each iteration, use a quantum random number generator to generate a series of quantum states to ensure the diversity and randomness of the states, expressed as:

[0023]

[0024] In the formula, θ c is the angle of the qubit, φ c is the phase of the qubit. The angle of the qubit and the phase of the qubit are parameters in the quantum state generation process; U() is a quantum gate operation function, specifically the Hadamard gate, used to create a superposition state; is the initial state of n cz qubits, n czis the number of initial states of qubits;

[0025] S203. In the generation stage, the generator accepts a random variable and a quantum state as inputs and, through the internal neural network structure, generates new data samples, expressed as:

[0026]

[0027] where X gen，c is the generated data, z c is the random variable, ψ c is the quantum state input, are the weights of the generator, and G c () is the generator function;

[0028] The way the internal neural network structure of the generator processes the input quantum state and random variable is expressed as:

[0029]

[0030] where represents concatenating the information of the random variable and the quantum-encoded information, and QEncode() represents the encoding neural network function for quantum state encoding, are the bias parameters of the generator, Re() is the ReLU activation function; tanh() is the hyperbolic tangent function;

[0031] S204. In the discrimination stage, after the generated data samples are fed into the discriminator, the discriminator evaluates the similarity between these samples and the real samples, expressed as:

[0032]

[0033] where y gen，c is the discrimination result of the discriminator for the generated data, and y real，c is the discrimination result of the discriminator for the real data; D c () is the discriminator function; X real，c is the real data;

[0034] S205. According to the output of the discriminator, calculate the loss function of the generative adversarial network, and then update the weights of the generator and the discriminator through the backpropagation algorithm, expressed as:

[0035] L gan (G c , D c ) = log(D c (X real，c )) + log(1 - D c (G c (z c , ψ c)))

[0036]

[0037] Wherein, L gan () is the loss function of the generative adversarial network, and ← is the parameter update operation; η gan is the learning rate of the generative adversarial network; is the gradient of the loss function of the generative adversarial network with respect to the generator weight parameter; is the gradient of the loss function of the generative adversarial network with respect to the discriminator weight parameter, and η gan is set to 0.01;

[0038] S206. Repeat the above steps iteratively until the preset stop iteration condition is met, which indicates that the model training is completed.

[0039] The preset stop iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times; after the data augmentation model training is completed, use the trained data augmentation model to increase the number of samples. Suppose the original collected samples are 800, and the data augmentation model augments and generates 200 samples, then the augmented dataset contains 1000 samples.

[0040] S3. Feature extraction model training: The training process of the neural network algorithm based on chaotic optimization:

[0041] S301. Initialize the weights and biases of the neural network, and the initialization method is expressed as:

[0042] W p，0 = μ p + σ p ·N(0, I)

[0043] b p，0 = V p ·N(0, I)

[0044] Wherein, W p，0 is the initial value of the weights of the neural network; b p，0 is the initial value of the biases of the neural network; μ p is the mean offset of the weights of the neural network; V p is the mean offset of the weight biases of the neural network; σ p is the standard deviation of the weight initialization of the neural network; N(0, I) represents a normal distribution with a mean of 0 and a standard deviation of the identity matrix, and I is the identity matrix;

[0045] S302. During the training process, evaluate the output error of the neural network model and adjust the learning rate through a chaotic feedback mechanism to avoid overfitting or underfitting. The dynamic update method of self-adjusting the learning rate is expressed as:

[0046] η p，t = η p，t-1 · exp(-γ p · ΔE p，t-1 )

[0047] Where η p，t is the learning rate of the neural network at the t-th iteration; η p，t-1 is the learning rate of the neural network at the (t - 1)-th iteration; γ p is the adjustment factor of the neural network, controlling the sensitivity of the learning rate change; ΔE p，t-1 is the amount of error change of the neural network in the previous cycle;

[0048] According to the updated learning rate, the weights are updated, expressed as:

[0049]

[0050] Where W p，t is the weight of the neural network at the t-th iteration; W p，t-1 is the weight of the neural network at the (t - 1)-th iteration; E p is the error of the neural network, calculated from the output features of the last layer of the neural network by a preset Softmax function; is the symbol of partial derivative;

[0051] S303. During the training process, overfitting is prevented by adopting a regularization strategy to ensure the generalization ability of the model. Regularization sparsity is performed by dynamically adjusting the number of nodes in the hidden layer of the neural network, expressed as:

[0052] N p，t = N p，t-1 + sgn(ΔE p，t-1 )· β p

[0053] Where N p，t is the number of nodes in the hidden layer of the neural network at the t-th iteration; N p，t-1 is the number of nodes in the hidden layer of the neural network at the (t - 1)-th iteration; β p is the node adjustment step size of the hidden layer of the neural network; sgn() is the sign function of the error change, used to determine whether to increase or decrease the number of nodes, that is, for sgn(ΔE p，t-1 ), when ΔE p，t-1 is greater than zero, sgn(ΔE p，t-1 ) is 1, when ΔE p，t-1 is less than or equal to zero, sgn(ΔE p，t-1 ) is -1;

[0054] The node adjustment step size of the neural network hidden layer depends on the function of the network layer, and the calculation method is expressed as:

[0055]

[0056] In the formula, Int cs () is the rounding operation; α p is the influence coefficient for adjusting the adjustment step size; d p is the depth of the current network layer;

[0057] S304. Repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed.

[0058] μ p is set to 0.01, V p is set to 0.05, σ p is set to 0.001; d p takes the value of 6, α p is set to 2.5; the preset iteration stop condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0059] S4. Feature dimensionality reduction model training: The training process of the autoencoder algorithm based on dynamic routing:

[0060] S401. Initialize the weights and biases of the autoencoder. The initialization method is expressed as:

[0061]

[0062] In the formula, is the initial weight of the i-th layer of the autoencoder; is the initial bias of the i-th layer of the autoencoder; n r，i is the number of neurons in the i-th layer of the autoencoder; n r，i-1 is the number of neurons in the (i - 1)-th layer of the autoencoder; randn(n r，i , n r，i-1 ) generates standard normal distribution random numbers with a shape of n r，i ×n r,i-1 ; zeros(n r，i ) generates a zero vector with a length of n r，i ;

[0063] S402. The input data is passed through each layer of the encoder. Each layer makes a dynamic routing decision based on the output of the previous layer and the feature importance score of the current layer, and selectively passes the most informative features, which is expressed as:

[0064] z r，i = W r，i ·xr，i-1 +b r，i

[0065] a r，i = Re(z r，i )

[0066] x r，i = route(a r，i , imp(a r，i ))

[0067] In the formula, z r，i is the linear transformation output of the i-th layer of the autoencoder; a r，i is the output feature of the activation function of the autoencoder; W r，i is the weight of the i-th layer of the autoencoder; b r，i is the bias of the i-th layer of the autoencoder; Re() is the ReLU activation function; x r，i-1 is the input feature of the i-th layer of the autoencoder; route() represents the dynamic routing function based on feature importance; imp(a r，i ) calculates the importance scoring function of the activation function output feature of the autoencoder;

[0068] The dynamic routing function integrates the activation outputs of each feature by weighted average, where the weights are determined by the importance scores of the features, and the calculation method is expressed as:

[0069]

[0070] The calculation method of the importance scoring of the activation function output feature of the autoencoder is expressed as:

[0071]

[0072] In the formula, a r，i，j is the activation output of the j-th neuron in the i-th layer, a r，i，k is the activation output of the k-th neuron in the i-th layer; w r，imp，j is the weight corresponding to the feature importance, obtained through training, representing the importance of the j-th feature; w r，imp，k is the weight corresponding to the feature importance, obtained through training, representing the importance of the k-th feature; n r，i is the total number of neurons in the i-th layer;

[0073] S403. In the decoder stage, use the features from different encoding layers to perform data reconstruction and calculate the reconstruction error. Then, in the training process, the calculation method of the loss function of the autoencoder is expressed as:

[0074]

[0075] In the formula, Lr is the loss function of the autoencoder; m r is the number of samples in the current batch of inputs; is the j-th sample input to the autoencoder; is the j-th sample after reconstruction;

[0076] S404. Calculate the gradient of the loss function with respect to each parameter of the autoencoder through the backpropagation algorithm, and update the weights and biases of the autoencoder. The update method is expressed as:

[0077]

[0078] In the formula, ΔW r，i is the update increment of the weight of the i-th layer of the autoencoder; Δb r，i is the update increment of the bias of the i-th layer of the autoencoder; η r is the learning rate of the autoencoder; is the weight of the i-th layer of the updated autoencoder; is the bias of the i-th layer of the updated autoencoder; S r，i is the sparsity measure of the neurons in the i-th layer of the autoencoder; λ r，s is the sparsity pruning threshold; is the indicator function, which takes 1 when the condition in the parentheses holds, and 0 otherwise. λ r，s is set to 0.8, indicating that when the sparsity measure is higher than 80%, the weights will be pruned;

[0079] Improve the computational efficiency and generalization ability of the model by dynamically pruning redundant weights, that is, the weights corresponding to neurons with small activation values in most cases will be dynamically pruned, thereby simplifying the model structure. Specifically, the importance of neurons is judged by statistically calculating the sparsity of activation values. The calculation method of the sparsity measure of the neurons in the i-th layer of the autoencoder is expressed as:

[0080]

[0081] In the formula, S r，i is the sparsity measure of the neurons in the i-th layer; a r,i,j is the activation value of the j-th sample in the i-th layer; τ r is the sparsity threshold, τ r is set to 0.1;

[0082] S405. Repeat the above steps iteratively until the preset stop iteration condition is satisfied, which means the model training is completed. The preset stop iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0083] S5. Training of the classifier model: Training process of the high-order neural network algorithm based on dynamic step optimization:

[0084] S501. Initialize the weights and biases of the high-order neural network. Let the weight of the high-order neural network be W u , and the bias of the high-order neural network be b u . The initialization method is expressed as:

[0085]

[0086] where ∼ represents following a specific distribution, is the weight of the i-th layer of the high-order neural network, is the bias of the i-th layer of the high-order neural network;

[0087] S502. Process the feature vector after feature dimension reduction into high-order extended features and input them into the hidden layer of the high-order neural network for forward propagation of data, which is expressed as:

[0088]

[0089] where X u is the input vector of the high-order neural network, that is, the feature vector after feature dimension reduction; is the high-order extension of the input vector X u , including the original features, the squares of the features, and the cross-products between features, is the input to the first hidden layer of the high-order neural network; is the weight of the first layer of the high-order neural network, is the bias of the first layer of the high-order neural network;

[0090] The calculation method of the high-order extension of the input vector X u is expressed as:

[0091]

[0092] where represents the square of each input feature of the high-order neural network to capture the non-linear influence of a single variable; represents the product of all possible combinations of the i-th feature and the j-th input feature of the high-order neural network to detect the interaction between different features;

[0093] S503. Use the modified LeakyReLU activation function in the hidden layer of the high-order neural network to increase the model's ability to process complex data, which is expressed as:

[0094]

[0095] where is the activation output of the first hidden layer of the high-order neural network; γ u is the activation function parameter of the high-order neural network, enhancing or weakening the non-linear effect of the activation function; is the LeakyReLU activation function;

[0096] The activation function parameter of the high-order neural network is flexibly adjusted according to the size of the error relative to the target error, so as to promote the increase of the non-linear characteristics of the activation function when the error is large. The calculation method is expressed as:

[0097]

[0098] In the formula, β u is the adjustment sensitivity parameter, E u is the training error of the current high-order neural network, E target is the target error of the high-order neural network;

[0099] S504. The output of the hidden layer is passed through one or more output neurons, and the Softmax function is used for multi-class classification. The calculation method is expressed as:

[0100]

[0101] In the formula, Y u is the classification output of the high-order neural network, is the weight of the output layer of the high-order neural network, is the bias of the output layer of the high-order neural network, is the activation output of the last hidden layer of the high-order neural network; Soft() is the Softmax function;

[0102] S505. The error is calculated using the backpropagation algorithm, and the weights and biases of each layer are adjusted according to the error feedback. The calculation methods of the update amounts of the weights and biases of the high-order neural network are expressed as:

[0103]

[0104] In the formula, ΔW u is the update amount of the high-order neural network weight, Δb u is the update amount of the high-order neural network bias, α u is the learning rate of the high-order neural network; α() is the update learning rate function;

[0105] The update learning rate function is set in a dynamically adjusted manner, and its dynamic adjustment method is:

[0106] α(W u ) = α0 × exp(-β × H(W u ))

[0107] α(b u ) = α0 × exp(-β × H(b u ))

[0108] where α0 is the initial update factor, β is the update factor adjustment parameter, and H(W u ) is the historical update cumulative amount of the high-order neural network weights, and H(b u ) is the historical update cumulative amount of the high-order neural network biases;

[0109] The calculation methods of the historical update cumulative amount of the high-order neural network weights and the historical update cumulative amount of the high-order neural network biases are expressed as:

[0110]

[0111] where Krs is the number of past high-order neural network iterations considered, is the update increment of the high-order neural network weight parameters at the historical int(k)th time, is the update increment of the high-order neural network bias parameters at the historical int(k)th time, and int(k) represents the kth iteration;

[0112] The learning rate of the high-order neural network depends on the magnitude and change trend of the error E u to achieve more effective parameter updates, which is expressed as:

[0113]

[0114] In the formula, α base is the base learning rate, int(t) is the current iteration number, Tmax is the total number of iterations, and κ u is the learning rate update decay exponent;

[0115] S506. Repeat the above steps iteratively until the preset iteration stop condition is satisfied, which indicates that the model training is completed.

[0116] β u is set to 2, E target is set to 0.05; α0 is set to 0.1; α base is set to 0.01, Tmax is set to 1000, and κ u is set to 2; the preset iteration stop condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0117] The beneficial effects achieved by the present invention are:

[0118] 1. By adopting a generative adversarial network with quantum state enhanced learning, the present invention can still generate high-quality samples under the condition of limited training samples. The augmented data enhances the generalization ability of the model and improves the accurate discrimination of different operating states.

[0119] 2. By optimizing the fully connected neural network through a chaotic optimization algorithm, the model can more efficiently extract key features in device operations, effectively reducing the training instability caused by gradient problems and ensuring the comprehensiveness and reliability of feature extraction.

[0120] 3. The dynamic routing autoencoder optimizes the dimensionality reduction effect through an importance scoring mechanism during the feature dimensionality reduction process, can accurately retain the key features of device operation, while reducing unnecessary dimensions and improving the processing efficiency of subsequent classifiers.

[0121] 4. The high-order neural network optimized by dynamic step optimization can capture complex non-linear relationships in device operations, improve the prediction accuracy of the classifier through richer feature interaction information, and ensure the accuracy of device fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0122] Figure 1 is the flow chart of neural network training proposed by the present invention;

[0123] Figure 2 is the flow chart of autoencoder training proposed by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0124] The present invention will be described in detail below with reference to the drawings and specific embodiments.

[0125] The present invention proposes a nuclear power equipment fault diagnosis method based on machine learning, and the steps are as follows:

[0126] S1. Data collection and annotation

[0127] The present invention collects operation data from various devices in a nuclear power plant. The collected data exists in vector form and includes, but is not limited to, physical measurement values such as temperature, pressure, and flow rate. The storage format of the data is a structured JSON format. In one embodiment, the attributes of the data include: temperature T a representing the temperature of the key part of the device, with the unit of degree Celsius; pressure P a representing the pressure related to the main steam pipeline, with the unit of Pascal; flow rate F a representing the flow rate of coolant or steam, with the unit of cubic meters per hour; vibration level V a representing the vibration level generated by device operation, with the unit of meters per second squared; current I a representing the working current of electrical equipment, with the unit of ampere; voltage U aRepresents the voltage at which the device operates, in volts; power W a Represents the power consumed by the device, in watts; bearing temperature B a Represents the temperature of the critical bearing, in degrees Celsius; bearing wear index E a Represents a quantitative index of the bearing wear condition, dimensionless; corrosion index C a Represents a quantitative index of the device corrosion state, dimensionless.

[0128] It should be noted that this embodiment is only to illustrate a data format and type of the present invention. In actual applications, the attributes of data are usually more than 10, and the number of data attributes may reach dozens or even hundreds.

[0129] Furthermore, the collected data is labeled. The labeling method of the present invention is manual labeling. In one embodiment, the labeling categories include "normal operation", "abnormal warning", and "fault state".

[0130] S2. Data augmentation

[0131] It can be understood that in the task of the present invention, the acquisition, labeling, and preprocessing of training data are time-consuming and laborious, and insufficient training samples are likely to lead to poor generalization ability of the model and affect the accuracy of the model. The present invention uses a generative adversarial network based on quantum state reinforcement learning to generate samples, thereby realizing data augmentation. The generative adversarial network based on quantum state reinforcement learning consists of two parts: a generator and a discriminator. The generator is responsible for generating high-quality data samples, and the discriminator evaluates the authenticity of the samples. Based on the traditional generative adversarial network, the generator utilizes the characteristics of quantum computing to enhance its learning and generation ability, using quantum noise and quantum entanglement as the source of random variables to improve the diversity and complexity of the generated samples; the discriminator adopts a symmetric quantum network design, which can more precisely evaluate the authenticity of the samples, enabling the discriminator to more sensitively capture the subtle variations in the data when processing the generated data.

[0132] S3. Feature extraction model training

[0133] The expanded data is input into the feature extraction model for training the feature extraction model. In the present invention, a 6-layer fully connected neural network is used for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, problems such as gradient disappearance, gradient explosion, or getting stuck in local optimal solutions may occur, affecting the stability of training and the performance of the model. The present invention uses a chaotic optimization algorithm to optimize the parameters of the neural network model to achieve the training of the neural network model. Specifically, the present invention adopts a self-adjusting mechanism that allows the network to dynamically adjust its weight update strategy according to the error of the previous training cycle, and uses the sensitivity in chaos theory to enhance the model's response ability to changes in initial conditions and parameters.

[0134] S4. Feature dimensionality reduction model training

[0135] The data after feature extraction is input into the feature dimensionality reduction model for training the feature dimensionality reduction model. The present invention uses an autoencoder algorithm based on dynamic routing for feature dimensionality reduction. The autoencoder consists of a symmetric encoder and decoder. The present invention uses a dynamic routing mechanism based on feature importance scoring to automatically determine the transmission paths of each feature in different layers of the encoder, enhancing the model's ability to reconstruct the original data and reducing information loss at the same time.

[0136] S5. Classifier model training

[0137] The data after dimensionality reduction is input into the classifier for training the classifier. The present invention uses a high-order neural network based on dynamic step optimization as the classification algorithm. Based on the high-order neural network, not only the influence of individual input features is considered, but also the interaction between features is considered. By using high-order terms to capture complex non-linear relationships, the generalization ability of the model is improved.

[0138] S6. Nuclear power equipment fault diagnosis

[0139] The trained model is used to perform fault diagnosis on new samples. In one embodiment, the collected original data is input into the trained feature extraction and feature dimensionality reduction models for feature processing. Further, the processed features are input into the classifier model for training the classifier, and then the classification results are obtained. In this embodiment, the classification categories include "normal operation", "abnormal warning", and "fault status".

[0140] The present invention proposes a method for nuclear power equipment fault diagnosis based on machine learning, and the steps are as follows:

[0141] S1. Data collection and annotation

[0142] The present invention collects operation data from various devices in a nuclear power plant. The collected data exists in the form of vectors and includes, but is not limited to, physical measurement values such as temperature, pressure, and flow rate. The storage format of the data is a structured JSON format. In one embodiment, the attributes of the data include: temperature T a represents the temperature of the critical part of the device, in degrees Celsius; pressure P a represents the pressure related to the main steam pipeline, in Pascals; flow rate F a represents the flow rate of the coolant or steam, in cubic meters per hour; vibration level V a represents the vibration level generated during the operation of the device, in meters per second squared; current I a represents the operating current of the electrical device, in amperes; voltage U a represents the voltage of the device during operation, in volts; power W a represents the power consumed by the device, in watts; bearing temperature B a represents the temperature of the critical bearing, in degrees Celsius; bearing wear index E a represents a quantitative index of the bearing wear condition, unitless; corrosion index C a represents a quantitative index of the corrosion state of the device, unitless.

[0143] It should be noted that this embodiment is only to illustrate a data format and type of the present invention. In actual applications, the attributes of the data usually exceed 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0144] Furthermore, the collected data is labeled. The labeling method of the present invention is manual labeling. In one embodiment, the labeling categories include "normal operation", "abnormal warning", and "fault status".

[0145] S2. Data augmentation

[0146] It can be understood that in the task of the present invention, the acquisition, labeling, and preprocessing of training data are time-consuming and laborious, and insufficient training samples are likely to lead to poor generalization ability of the model and affect the accuracy of the model. The present invention uses a generative adversarial network based on quantum state reinforcement learning for sample generation, thereby realizing data augmentation. The generative adversarial network based on quantum state reinforcement learning consists of two parts: a generator and a discriminator. The generator is responsible for generating high-quality data samples, and the discriminator evaluates the authenticity of the samples. Based on the traditional generative adversarial network, the generator uses the characteristics of quantum computing to enhance its learning and generation ability, and uses quantum noise and quantum entanglement as the source of random variables to improve the diversity and complexity of the generated samples; the discriminator adopts a symmetric quantum network design, which can more accurately evaluate the authenticity of the samples, so that when the discriminator processes the generated data, it can more sensitively capture the tiny variations in the data.

[0147] Specifically, the training process of the generative adversarial network algorithm based on quantum state reinforcement learning is as follows:

[0148] S201. Randomly initialize the parameters of the generator and the discriminator. Let the generator be G c , and the discriminator be D c . The weights of the generator are . The weights of the discriminator are In one embodiment, the weights of the generator and the discriminator are initialized randomly and follow a normal distribution with a mean of 0 and a standard deviation of the identity matrix.

[0149] S202. Before each iteration, use a quantum random number generator to generate a series of quantum states to ensure the diversity and randomness of the states, expressed as:

[0150]

[0151] where θ c is the angle of the qubit, φ c is the phase of the qubit. The angle and phase of the qubit are parameters in the quantum state generation process; U() is a quantum gate operation function, specifically the Hadamard gate, which is used to create a superposition state; is the initial state of n cz qubits, and n cz is the number of the initial states of the qubits.

[0152] S203. In the generation stage, the generator takes a random variable and a quantum state as inputs and generates new data samples through an internal neural network structure, expressed as:

[0153]

[0154] where X gen，c is the generated data, z c is the random variable, ψ c is the quantum state input, are the weights of the generator, and G c () is the generator function.

[0155] In one embodiment, the way the internal neural network structure of the generator processes the input quantum state and random variable is expressed as:

[0156]

[0157] where denotes the concatenation of a random variable and quantum-encoded information, and QEncode() denotes the encoding neural network function for quantum state encoding. is the bias parameter of the generator, Re() is the ReLU activation function; tanh() is the hyperbolic tangent function.

[0158] S204. In the discrimination stage, after the generated data samples are fed into the discriminator, the discriminator evaluates the similarity between these samples and the real samples, which is expressed as:

[0159]

[0160] In the formula, y gen，c is the discrimination result of the discriminator for the generated data, and y real，c is the discrimination result of the discriminator for the real data; D c () is the discriminator function; x real，c is the real data.

[0161] S205. According to the output of the discriminator, calculate the loss function of the generative adversarial network, and then update the weights of the generator and the discriminator through the backpropagation algorithm, which is expressed as:

[0162] L gan (G c , D c ) = log(D c (X real，c )) + log(1 - D c (G c (z c , ψ c )))

[0163]

[0164] In the formula, L gan () is the loss function of the generative adversarial network, and ← is the parameter update operation; η gan is the learning rate of the generative adversarial network; is the gradient of the generative adversarial network loss function with respect to the generator weight parameter; is the gradient of the generative adversarial network loss function with respect to the discriminator weight parameter. Preferably, η gan is set to 0.01.

[0165] S206. Repeat the above steps iteratively until the preset stop iteration condition is satisfied, which means the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0166] After the data augmentation model training is completed, the trained data augmentation model is used to increase the number of samples. In one embodiment, assume that the original collected samples are 800, and the data augmentation model generates 200 samples through augmentation. Then the augmented dataset contains 1000 samples.

[0167] S3. Feature extraction model training

[0168] The augmented data is input into the feature extraction model for training the feature extraction model. The present invention uses a 6-layer fully connected neural network for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, problems such as gradient disappearance, gradient explosion, or getting stuck in local optimal solutions may occur, affecting the stability of training and the performance of the model. The present invention adopts a chaotic optimization algorithm to optimize the parameters of the neural network model to achieve the training of the neural network model. Specifically, the present invention adopts a self-regulating mechanism that allows the network to dynamically adjust its weight update strategy according to the error of the previous training cycle, and utilizes the sensitivity in chaos theory to enhance the model's response ability to changes in initial conditions and parameters.

[0169] Specifically, the training process of the neural network algorithm based on chaotic optimization is as follows:

[0170] S301. Initialize the weights and biases of the neural network. In one embodiment, the initialization method is expressed as:

[0171] W p，0 = μ p + σ p ·N(0, I)

[0172] b p，0 = V p ·N(0, I)

[0173] In the formula, W p，0 is the initial value of the weights of the neural network; b p，0 is the initial value of the biases of the neural network; μ p is the mean offset of the weights of the neural network; V p is the mean offset of the biases of the neural network; σ p is the standard deviation of the weights initialization of the neural network; N(0, I) represents a normal distribution with a mean of 0 and a standard deviation of the identity matrix, and I is the identity matrix. Preferably, μ p is set to 0.01, V p is set to 0.05, and σ p is set to 0.001.

[0174] S302. During the training process, evaluate the output error of the neural network model and adjust the learning rate through a chaotic feedback mechanism to avoid overfitting or underfitting. The dynamic update method of self-adjusting the learning rate is expressed as:

[0175] η p，t = η p,t-1 · exp(-γ p · ΔE p，t-1 )

[0176] In the formula, η p，t is the learning rate of the neural network at the t-th iteration; η p，t-1 is the learning rate of the neural network at the (t - 1)-th iteration; γ p is the adjustment factor of the neural network, controlling the sensitivity of the learning rate change; ΔE p，t-1 is the error change amount of the neural network in the previous cycle.

[0177] Furthermore, according to the updated learning rate, update the weights, which is expressed as:

[0178]

[0179] In the formula, W p，t is the weight of the neural network at the t-th iteration; W p，t-1 is the weight of the neural network at the (t - 1)-th iteration; E p is the error of the neural network, calculated by the preset Softmax function for the output features of the last layer of the neural network; is the partial derivative symbol.

[0180] S303. During the training process, prevent overfitting by adopting a regularization strategy to ensure the generalization ability of the model. Specifically, perform regularization sparsity by dynamically adjusting the number of nodes in the hidden layer of the neural network, which is expressed as:

[0181] N p，t = N p，t-1 + sgn(ΔE p，t-1 )· β p

[0182] In the formula, N p，t is the number of nodes in the hidden layer of the neural network at the t-th iteration; N p，t-1 is the number of nodes in the hidden layer of the neural network at the (t - 1)-th iteration; β p is the node adjustment step size of the hidden layer of the neural network; sgn() is the sign function of the error change, used to determine whether to increase or decrease the number of nodes, that is, for sgn(ΔE p，t-1 ), when ΔE p，t-1 is greater than zero, sgn(ΔE p，t-1 ) is 1, when ΔEp，t-1 When it is less than or equal to zero, sgn(ΔE p，t-1 ) is -1.

[0183] In one embodiment, the node adjustment step size of the neural network hidden layer depends on a function of the network hierarchy, and the calculation method is expressed as:

[0184]

[0185] In the formula, Int cs () is the rounding operation; α p is the influence coefficient for adjusting the adjustment step size; d p is the depth of the current network layer. Preferably, d p takes the value of 6, and α p is set to 2.5.

[0186] S304. Repeat the above steps iteratively until the preset stop iteration condition is met, which indicates that the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0187] S4. Feature dimensionality reduction model training

[0188] Input the data after feature extraction into the feature dimensionality reduction model for training the feature dimensionality reduction model. The present invention uses an autoencoder algorithm based on dynamic routing for feature dimensionality reduction. The autoencoder consists of a symmetric encoder and a decoder. The present invention uses a dynamic routing mechanism based on feature importance scoring to automatically determine the transfer path of each feature in different layers of the encoder, enhancing the model's ability to reconstruct the original data and reducing information loss at the same time.

[0189] Specifically, the training process of the autoencoder algorithm based on dynamic routing is as follows:

[0190] S401. Initialize the weights and biases of the autoencoder. In one embodiment, the initialization method is expressed as:

[0191]

[0192] In the formula, is the initial weight of the i-th layer of the autoencoder; is the initial bias of the i-th layer of the autoencoder; n r,i is the number of neurons in the i-th layer of the autoencoder; n r，i-1 is the number of neurons in the (i - 1)-th layer of the autoencoder; randn(n r，i , n r，i-1 ) generates a shape of n r，i ×n r，i-1Standard normal distribution random numbers; zeros(n r，i ) Generates a zero vector of length n r，i .

[0193] S402. The input data is passed through each layer of the encoder. Each layer makes a dynamic routing decision based on the output of the previous layer and the feature importance score of the current layer, and selectively passes the most informative features, expressed as:

[0194] z r，i = W r，i ·x r，i-1 + b r，i

[0195] a r，i = Re(z r，i )

[0196] x r，i = route(a r，i , imp(a r，i ))

[0197] In the formula, z r，i is the output of the linear transformation of the i-th layer of the autoencoder; a r，i is the output feature of the activation function of the autoencoder; W r，i is the weight of the i-th layer of the autoencoder; b r，i is the bias of the i-th layer of the autoencoder; Re() is the ReLU activation function; x r，i-1 is the input feature of the i-th layer of the autoencoder; route() represents the dynamic routing function based on feature importance; imp(a r，i ) calculates the importance scoring function of the output feature of the activation function of the autoencoder.

[0198] In one embodiment, the dynamic routing function integrates the activation outputs of each feature by weighted average, where the weights are determined by the importance scores of the features, and the calculation method is expressed as:

[0199]

[0200] Furthermore, the calculation method of the importance score of the output feature of the activation function of the autoencoder is expressed as:

[0201]

[0202] In the formula, a r，i，j is the activation output of the j-th neuron in the i-th layer, and a r，i，k is the activation output of the k-th neuron in the i-th layer; w r，imp，j is the weight corresponding to the feature importance, obtained through training, representing the importance of the j-th feature; wr，imp，k is the weight corresponding to the importance of the feature, obtained through training, representing the importance of the k-th feature; n r，i is the total number of neurons in the i-th layer.

[0203] S403. In the decoder stage, use the features from different encoding layers to perform data reconstruction and calculate the reconstruction error. Then, during the training process, the calculation method of the loss function of the autoencoder is expressed as:

[0204]

[0205] In the formula, L r is the loss function of the autoencoder; m r is the number of samples in the current batch input; is the j-th sample input to the autoencoder; is the j-th sample after reconstruction.

[0206] S404. Calculate the gradient of the loss function with respect to each parameter of the autoencoder through the backpropagation algorithm, and update the weights and biases of the autoencoder. The update method is expressed as:

[0207]

[0208] In the formula, ΔW r，i is the update increment of the weight of the i-th layer of the autoencoder; Δb r，i is the update increment of the bias of the i-th layer of the autoencoder; η r is the learning rate of the autoencoder; is the weight of the i-th layer of the updated autoencoder; is the bias of the i-th layer of the updated autoencoder; s r，i is the sparsity measure of the neurons in the i-th layer of the autoencoder; λ r，s is the sparsity pruning threshold; is the indicator function, which takes 1 when the condition in the parentheses holds, and 0 otherwise. Preferably, λ r，s is set to 0.8, indicating that when the sparsity measure is higher than 80%, the weights will be pruned.

[0209] In one embodiment, improve the computational efficiency and generalization ability of the model by dynamically pruning redundant weights, that is, the weights corresponding to the neurons with relatively small activation values in most cases will be dynamically pruned, thereby simplifying the model structure. Specifically, judge the importance of neurons by statistically calculating the sparsity of activation values. The calculation method of the sparsity measure of the neurons in the i-th layer of the autoencoder is expressed as:

[0210]

[0211] In the formula, S r，iis the sparsity measure of the i-th layer of neurons; a r，i，j is the activation value of the j-th sample in the i-th layer; τ r is the sparsity threshold. Preferably, τ r is set to 0.1.

[0212] S405. Repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed. In one embodiment, the preset iteration stop condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0213] S5. Classifier model training

[0214] Input the data after dimensionality reduction into the classifier for the training of the classifier. The present invention uses a high-order neural network based on dynamic step optimization as the classification algorithm. Based on the high-order neural network, not only the influence of individual input features is considered, but also the interaction between features is considered. By using high-order terms to capture complex non-linear relationships, the generalization ability of the model is improved.

[0215] Specifically, the training process of the high-order neural network algorithm based on dynamic step optimization is as follows:

[0216] S501. Initialize the weights and biases of the high-order neural network. Let the weights of the high-order neural network be W u and the biases of the high-order neural network be b u . In one embodiment, the initialization method is expressed as:

[0217]

[0218] where ~ means following a specific distribution, is the weight of the i-th layer of the high-order neural network, is the bias of the i-th layer of the high-order neural network.

[0219] S502. Process the feature vector after dimensionality reduction into high-order extended features and input them into the hidden layer of the high-order neural network for forward propagation of data, which is expressed as:

[0220]

[0221] where X n is the input vector of the high-order neural network, that is, the feature vector after dimensionality reduction; is the high-order extension of the input vector X u , including the original features, the squares of the features, and the cross-products between features. is the input of the first hidden layer of the high-order neural network; is the weight of the first layer of the high-order neural network, is the bias of the first layer of the high-order neural network.

[0222] In one embodiment, the input vector X u is calculated as the high-order expansion of:

[0223]

[0224] where represents the square of each input feature of the high-order neural network to capture the non-linear effect of a single variable; represents the product of all possible combinations of the i-th feature and the j-th input feature of the high-order neural network to detect the interaction between different features.

[0225] S503. Use the modified LeakyReLU activation function in the hidden layer of the high-order neural network to increase the model's ability to process complex data, which is expressed as:

[0226]

[0227] where is the activation output of the first hidden layer of the high-order neural network; γ u is the activation function parameter of the high-order neural network, which enhances or weakens the non-linear effect of the activation function; is the LeakyReLU activation function.

[0228] In one embodiment, the activation function parameter of the high-order neural network is flexibly adjusted according to the size of the error relative to the target error to promote the increase of the non-linear characteristics of the activation function when the error is large, and the calculation method is expressed as:

[0229]

[0230] where β u is the adjustment sensitivity parameter, E u is the training error of the current high-order neural network, E target is the target error of the high-order neural network. Preferably, β u is set to 2, and E target is set to 0.05.

[0231] S504. The output of the hidden layer is passed through one or more output neurons and the Softmax function is used for multi-class classification, and the calculation method is expressed as:

[0232]

[0233] where Y u is the classification output of the high-order neural network, is the weight of the output layer of the high-order neural network, is the bias of the output layer of the high-order neural network, is the activation output of the last hidden layer of the high-order neural network; Soft() is the Softmax function.

[0234] S505. Calculate the error using the backpropagation algorithm, and adjust the weights and biases of each layer according to the error feedback. The calculation method of the update amount of the weights and biases of the high-order neural network is expressed as:

[0235]

[0236] In the formula, ΔW u is the update amount of the weight of the high-order neural network, Δb u is the update amount of the bias of the high-order neural network, α u is the learning rate of the high-order neural network; α() is the update learning rate function.

[0237] In one embodiment, the update learning rate function is set in a dynamically adjusted manner, and its dynamic adjustment method is:

[0238] α(W u ) = α0 × exp(-β × H(W u ))

[0239] α(b u ) = α0 × exp(-β × H(b u ))

[0240] where α0 is the initial update factor, β is the update factor adjustment parameter, H(W u ) is the historical update cumulative amount of the weight of the high-order neural network, and H(b u ) is the historical update cumulative amount of the bias of the high-order neural network. Preferably, α0 is set to 0.1.

[0241] Furthermore, the calculation methods of the historical update cumulative amount of the weight of the high-order neural network and the historical update cumulative amount of the bias of the high-order neural network are expressed as:

[0242]

[0243] where Krs is the number of past iterations of the high-order neural network considered, is the update increment of the weight parameter of the high-order neural network at the historical int(k)th time, is the update increment of the bias parameter of the high-order neural network at the historical int(k)th time, and int(k) represents the kth iteration.

[0244] Furthermore, the learning rate of the high-order neural network depends on the error E uThe size and change trend are used to achieve more effective parameter updates, expressed as:

[0245]

[0246] In the formula, α base is the base learning rate, int(t) is the current iteration number, Tmax is the total number of iterations, and κ u is the learning rate update decay exponent. Preferably, α base is set to 0.01, Tmax is set to 1000, and κ u is set to 2.

[0247] S506. Repeat and iterate the above steps until the preset stop iteration condition is satisfied, which indicates that the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0248] S6. Fault diagnosis of nuclear power equipment

[0249] Use the trained model to perform fault diagnosis on new samples. In one embodiment, the collected original data is input into the trained feature extraction and feature dimensionality reduction model for feature processing. Further, the processed features are input into the classifier model for classifier training, and then the classification result is obtained. In this embodiment, the classification categories include "normal operation", "abnormal warning", and "fault status".

Claims

1. A nuclear power equipment fault diagnosis method based on machine learning, characterized in that: S1. Data acquisition and annotation The data collected from the nuclear power plant exists in vector form, including temperature, pressure, and flow measurement values, and the storage format of the data is structured JSON format; S2. Data augmentation Use a generative adversarial network based on quantum state reinforcement learning for sample generation, and then realize data augmentation. The generative adversarial network based on quantum state reinforcement learning consists of two parts: a generator and a discriminator. The generator is responsible for generating high-quality data samples, and the discriminator evaluates the authenticity of the samples. The generator uses the characteristics of quantum computing to enhance its learning and generation capabilities, and uses quantum noise and quantum entanglement as the source of random variables to improve the diversity and complexity of the generated samples; the discriminator adopts a symmetric quantum network design to more accurately evaluate the authenticity of the samples, so that the discriminator can more sensitively capture the tiny variations in the data when processing the generated data; S3. Feature extraction model training Input the augmented data into the feature extraction model for training the feature extraction model. Use a 6-layer fully connected neural network for feature extraction, and use a chaotic optimization algorithm to optimize the parameters of the neural network model to realize the training of the neural network model. Adopt a self-adjusting mechanism to allow the network to dynamically adjust its weight update strategy according to the error of the previous training cycle, and use the sensitivity in chaos theory to enhance the model's response ability to changes in initial conditions and parameters; S4. Feature dimensionality reduction model training Input the data after feature extraction into the feature dimensionality reduction model for training the feature dimensionality reduction model. Use an autoencoder algorithm based on dynamic routing for feature dimensionality reduction. The autoencoder consists of a symmetric encoder and decoder. Adopt a dynamic routing mechanism based on feature importance scoring to automatically determine the transmission path of each feature in different layers of the encoder, enhance the model's ability to reconstruct the original data, and at the same time reduce information loss; S5. Classifier model training Input the data after dimensionality reduction into the classifier for training the classifier. Use a high-order neural network based on dynamic step optimization as the classification algorithm. On the basis of the high-order neural network, not only consider the influence of individual input features, but also consider the interaction between features, and capture complex nonlinear relationships by using high-order terms to improve the generalization ability of the model; S6. Nuclear power equipment fault diagnosis Use the trained model to diagnose faults in new samples. The collected original data is input into the trained feature extraction and feature dimensionality reduction models for feature processing, and the processed features are input into the classifier model for training the classifier, and then the classification results are obtained, including normal operation, abnormal warning, and fault status.

2. The method for diagnosing faults of nuclear power equipment based on machine learning according to claim 1, characterized in that: S1. Data Acquisition and Annotation: The attributes of the data include: Temperature T a indicating the temperature of the key parts of the equipment, in degrees Celsius; Pressure P a indicating the pressure related to the main steam pipeline, in Pascals; Flow rate F a indicating the flow rate of the coolant or steam, in cubic meters per hour; Vibration level V a indicating the vibration level generated during the operation of the equipment, in meters per second squared; Current I a indicating the operating current of the electrical equipment, in Amperes; Voltage U a indicating the voltage during the operation of the equipment, in Volts; Power W a indicating the power consumed by the equipment, in Watts; Bearing temperature B a indicating the temperature of the key bearings, in degrees Celsius; Bearing wear index E a indicating the quantitative index of the bearing wear condition, dimensionless; Corrosion index C a indicating the quantitative index of the equipment corrosion state, dimensionless.

3. The method for diagnosing faults in nuclear power equipment based on machine learning according to claim 1, characterized in that: S1. Data acquisition and annotation: Annotate the collected data. The annotation method is manual annotation, and the annotation categories include normal operation, abnormal warning, and fault status.

4. The method for diagnosing faults of nuclear power equipment based on machine learning according to claim 1, characterized in that: S2. Data augmentation: Training process of the generative adversarial network algorithm based on quantum state reinforcement learning: S201. Randomly initialize the parameters of the generator and the discriminator. Let the generator be G c , and the discriminator be D c . The weights of the generator are . The weights of the discriminator are . The weights of the generator and the discriminator are initialized randomly and follow a normal distribution with a mean of 0 and a standard deviation of the identity matrix; S202. Before each iteration, use a quantum random number generator to generate a series of quantum states to ensure the diversity and randomness of the states, expressed as: where θ c is the angle of the qubit, φ c is the phase of the qubit, and the angle and phase of the qubit are parameters in the quantum state generation process; U( ) is the quantum gate operation function, specifically the Hadamard gate, which is used to create a superposition state; is the initial state of n cz qubits, and n cz is the number of the initial states of the qubits; S203. In the generation stage, the generator takes a random variable and a quantum state as inputs and, through its internal neural network structure, generates new data samples, expressed as: where X gen,c is the generated data, z c is a random variable, ψ c is the quantum state input, is the weight of the generator, G c ( ) is the generator function; The way the internal neural network structure of the generator processes the input quantum state and random variable is expressed as: In the formula, represents the concatenation of a random variable and the quantum-encoded information, and QEncode( ) represents the encoding neural network function for quantum state encoding. is the bias parameter of the generator, Re( ) is the ReLU activation function; tanh( ) is the hyperbolic tangent function. S204. In the discrimination stage, after the generated data samples are fed into the discriminator, the discriminator evaluates the similarity of these samples to the real samples, expressed as: where y gen，c is the discriminator's discrimination result for the generated data, and y real，c is the discriminator's discrimination result for the real data; D c () is the discriminator function; X real，c is the real data; S205. According to the output of the discriminator, calculate the loss function of the generative adversarial network, and then update the weights of the generator and the discriminator through the backpropagation algorithm, expressed as: L gan (G c ,D c ) = log(D c (X real，c )) + log(1 - D c (G c (z c ,ψ c ))) Where L gan ( ) is the loss function of the generative adversarial network, and ← is the parameter update operation; η gan is the learning rate of the generative adversarial network; is the gradient of the loss function of the generative adversarial network with respect to the generator weight parameters; is the gradient of the loss function of the generative adversarial network with respect to the discriminator weight parameters, η gan is set to 0.01; S206. Repeat the above steps iteratively until the preset stop iteration condition is met, which means the model training is completed.

5. The method for diagnosing faults of nuclear power equipment based on machine learning according to claim 4, characterized in that: The preset stop iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times; after the data augmentation model training is completed, use the trained data augmentation model to increase the number of samples. Suppose the original collected samples are 800, and the data augmentation model generates 200 samples through augmentation, then the augmented dataset contains 1000 samples.

6. The method for diagnosing faults of nuclear power equipment based on machine learning according to claim 1, wherein: S3. Feature extraction model training: The training process of the neural network algorithm based on chaotic optimization: S301. Initialize the weights and biases of the neural network, and the initialization method is expressed as: W p，0 = μ p + σ p ·N(0·I) b p，0 = v p · N(0, I) Where, W p，0 is the initial value of the weight of the neural network; b p，0 is the initial value of the bias of the neural network; μ p is the mean offset of the weight of the neural network; v p is the mean offset of the bias of the weight of the neural network; σ p is the standard deviation of the weight initialization of the neural network; N(0, I) represents a normal distribution with a mean of 0 and a standard deviation of the identity matrix, and I is the identity matrix; S302. During the training process, evaluate the output error of the neural network model, and adjust the learning rate through the chaotic feedback mechanism to avoid overfitting or underfitting. The dynamic update method of self-adjusting the learning rate is expressed as: η p，t = η p，t-1 · exp(-γ p · ΔE p,t-1 ) where η p，t is the learning rate of the neural network for the t-th iteration; η p，t-1 is the learning rate of the neural network for the (t - 1)-th iteration; γ p is the adjustment factor of the neural network, controlling the sensitivity of the change in the learning rate; ΔE p，t-1 is the amount of change in the error of the neural network in the previous cycle; According to the updated learning rate, update the weights, expressed as: where, W p，t is the weight of the neural network in the t-th iteration; W p，t-1 is the weight of the neural network in the (t-1)-th iteration; E p is the error of the neural network, which is calculated by a preset Softmax function for the output features of the last layer of the neural network; is the symbol of partial derivative; S303. During the training process, prevent overfitting by adopting a regularization strategy to ensure the generalization ability of the model. Regularization sparsity is achieved by dynamically adjusting the number of nodes in the hidden layer of the neural network, expressed as: N p，t = N p，t-1 + sgn(ΔE p，t-1 )·β p Where N p，t is the number of nodes in the hidden layer of the neural network at the t-th iteration; N p，t-1 is the number of nodes in the hidden layer of the neural network at the (t - 1)-th iteration; β p is the node adjustment step size of the neural network hidden layer; sgn( ) is the sign function of the error change, which is used to determine whether to increase or decrease the number of nodes, that is, for sgn(ΔE p，t-1 ), when ΔE p，t-1 is greater than zero, sgn(ΔE p，t-1 ) is 1, and when ΔE p，t-1 is less than or equal to zero, sgn(ΔE p，t-1 ) is -1; The step size of adjusting the nodes in the hidden layer of the neural network depends on the function of the network layer, and the calculation method is expressed as: where Int cs ( ) is the rounding operation; α p is the influence coefficient for adjusting the adjustment step size; d p is the depth of the current network layer; S304. Repeat the above steps iteratively until the preset stop iteration condition is met, which means the model training is completed.

7. The method for diagnosing faults in nuclear power equipment based on machine learning according to claim 6, characterized in that: μ p is set to 0.01, ν p is set to 0.05, σ p is set to 0.001; d p takes the value of 6, α p is set to 2.5; The preset stop iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

8. The method for diagnosing faults of nuclear power equipment based on machine learning according to claim 1, characterized in that: S4. Feature dimensionality reduction model training: The training process of the autoencoder algorithm based on dynamic routing: S401. Initialize the weights and biases of the autoencoder, and the initialization method is expressed as: In the formula, is the initial weight of the i-th layer of the autoencoder; is the initial bias of the i-th layer of the autoencoder; n r，i is the number of neurons in the i-th layer of the autoencoder; n r，i-1 is the number of neurons in the (i - 1)-th layer of the autoencoder; randn(n r，i , n r，i-1 ) generates standard normal distribution random numbers with a shape of n r，i ×n r，i-1 ; zeros(n r,i ) generates a zero vector with a length of n r,i ; S402. The input data is passed through each layer of the encoder. Each layer makes a dynamic routing decision based on the output of the previous layer and the feature importance score of the current layer, and selectively passes the most informative features, expressed as: z r，i = W r，i · x ri-1 + b r，i a r，i = Re(z r，i ) x r，i = rollte(a r，i , imp(a r，i )) where z r,i is the linear transformation output of the i-th layer of the autoencoder; a r,i is the output feature of the activation function of the autoencoder; W r，i is the weight of the i-th layer of the autoencoder; b r，i is the bias of the i-th layer of the autoencoder; Re( ) is the ReLU activation function; x r，i-1 is the input feature of the i-th layer of the autoencoder; ro ute( ) represents a dynamic routing function based on feature importance; imp(a r，i ) calculates the importance scoring function of the activation function output features of the autoencoder; The dynamic routing function integrates the activation outputs of each feature by weighted averaging, where the weights are determined by the importance scores of the features, and the calculation method is expressed as: The calculation method of the importance score of the activation function output feature of the autoencoder is expressed as: where a r，i，j is the activation output of the j-th neuron in the i-th layer, and a r，i，k is the activation output of the k-th neuron in the i-th layer; w r，imp，j is the weight corresponding to the importance of the feature, obtained through training, representing the importance of the j-th feature; w r，imp，k is the weight corresponding to the importance of the feature, obtained through training, representing the importance of the k-th feature; n r，i is the total number of neurons in the i-th layer; S403. In the decoder stage, use the features from different encoding layers to perform data reconstruction and calculate the reconstruction error. Then, during the training process, the calculation method of the loss function of the autoencoder is: where, L r is the loss function of the autoencoder; m r is the number of samples in the current batch input; is the j-th sample input to the autoencoder; is the j-th sample after reconstruction; S404. Calculate the gradients of the loss function with respect to each parameter of the autoencoder through the backpropagation algorithm, and update the weights and biases of the autoencoder. The update method is expressed as: where Δw r，i is the update increment of the weight of the i-th layer of the autoencoder; Δb r，i is the update increment of the bias of the i-th layer of the autoencoder; η r is the learning rate of the auto - encoder; is the weight of the i - th layer of the updated auto - encoder; is the bias of the i - th layer of the updated auto - encoder; S r，i is the sparsity measure of the neurons in the i - th layer of the auto - encoder; λ r，s is the sparsity pruning threshold; is an indicator function that takes 1 when the condition in the parentheses holds and 0 otherwise, and λ r，s is set to 0.8, indicating that when the sparsity metric is higher than 80%, the weights will be pruned; Improve the computational efficiency and generalization ability of the model by dynamically pruning redundant weights, that is, the weights corresponding to neurons with relatively small activation values in most cases will be dynamically pruned, thereby simplifying the model structure. Specifically, judge the importance of neurons by statistically measuring the sparsity of activation values. The calculation method of the sparsity measure of the i-th layer neurons of the autoencoder is expressed as: where S r，i is the sparsity measure of the i-th layer of neurons; a r，i，j is the activation value of the j-th sample at the i-th layer; τ r is the sparsity threshold, and τ r is set to 0.1; S405. Repeat the above steps iteratively until the preset stopping iteration condition is met, which means the model training is completed. The preset stopping iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

9. The method for diagnosing faults in nuclear power equipment based on machine learning according to claim 1, wherein: S5. Classifier model training: The training process of the high-order neural network algorithm based on dynamic step optimization: S501. Initialize the weights and biases of the high-order neural network. Let the weight of the high-order neural network be w u , and the bias of the high-order neural network be b u . The initialization method is expressed as: wherein, ~ indicates being subject to a specific distribution, is the weight of the i-th layer of the high-order neural network, is the bias of the i-th layer of the high-order neural network; S502. Process the feature vector after dimensionality reduction into high-order extended features and input them into the hidden layer of the high-order neural network for forward propagation of data, which is expressed as: Where X u is the input vector of the high-order neural network, that is, the feature vector after feature dimensionality reduction; is the high-order expansion of the input vector X u , including the original features, the squares of the features, and the cross-products between the features, is the input to the first hidden layer of the high-order neural network; is the weight of the first layer of the high-order neural network, is the bias of the first layer of the high-order neural network; Input vector X u The calculation method of the high-order extension of is expressed as: wherein, represents the square of each input feature of the high-order neural network to capture the non-linear effect of a single variable; represents the product of all possible combinations of the i-th feature and the j-th input feature of the high-order neural network to detect the interaction between different features; S503. Use the modified LeakyReLU activation function in the hidden layer of the high-order neural network to increase the model's ability to process complex data, which is expressed as: In the formula, is the activation output of the first hidden layer of the high-order neural network; γ u is the activation function parameter of the high-order neural network, enhancing or weakening the non-linear effect of the activation function; is the LeakyReLU activation function; The activation function parameters of the high-order neural network are flexibly adjusted according to the magnitude of the error relative to the target error to promote the increase of the non-linear characteristics of the activation function when the error is large. The calculation method is expressed as: where β u is the sensitivity adjustment parameter, and E u is the training error of the current high-order neural network, and E target is the target error of the high-order neural network; S504. The output of the hidden layer is passed through one or more output neurons, and the Softmax function is used for multi-class classification. The calculation method is expressed as: where Y u is the classification output of the high-order neural network, is the weight of the output layer of the high-order neural network, is the bias of the output layer of the high-order neural network, is the activation output of the last hidden layer of the high-order neural network; Soft( ) is the Softmax function; S505. Use the backpropagation algorithm to calculate the error, and adjust the weights and biases of each layer according to the error feedback. The calculation method of the update amount of the weights and biases of the high-order neural network is expressed as: where, ΔW u is the update amount of the high-order neural network weight, and Δb u is the update amount of the high-order neural network bias, and α u is the learning rate of the high-order neural network; α( ) is the update learning rate function; The update learning rate function is set in a dynamically adjusted manner, and its dynamic adjustment method is: α(W u ) = α0 × exp(-β × H(W u )) α(b u ) = α0 × exp(-β × H(b u )) Among them, α0 is the initial update factor, β is the update factor adjustment parameter, and H(W u ) is the historical update cumulative amount of the high-order neural network weights, and H(b u ) is the historical update cumulative amount of the high-order neural network biases; The calculation methods of the historical update cumulative amount of the high-order neural network weights and the historical update cumulative amount of the high-order neural network biases are expressed as: Among them, Krs is the number of past high-order neural network iterations considered, is the update increment of the high-order neural network weight parameter for the historical int(k) times, is the update increment of the high-order neural network bias parameter for the historical int(k) times, and int(k) represents the k-th iteration; The learning rate of the high-order neural network depends on the size and trend of change of the error E to achieve more effective parameter update, expressed as: u and is represented as follows: where α base is the base learning rate, int(t) is the current iteration number, Tmax is the total number of iterations, and κ u is the learning rate update decay exponent; S506. Repeat the above steps iteratively until the preset stopping iteration condition is met, which means the model training is completed.

10. The method for diagnosing faults of nuclear power equipment based on machine learning according to claim 9, characterized in that: β u Set to 2, E target Set to 0.05; α0 is set to 0.1; α base Set to 0.01, Tmax is set to 1000, κ u Set to 2; The preset stopping iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

Images