Artificial intelligence-based steam generator state real-time monitoring method

Abstract

The steam generator state real-time monitoring method based on artificial intelligence belongs to the field of artificial intelligence and comprises the following steps: S1, data acquisition and labeling; S2, sample generation is performed by using a quantum generative adversarial network based on random projection embedding to realize data expansion; S3, the expanded data is input into a feature extraction model to perform training of the feature extraction model, and a five-layer fully connected neural network is used for feature extraction; S4, the feature-extracted data is input into a feature dimension reduction model to perform training of the feature dimension reduction model, and a self-encoding neural network algorithm based on local preserving projection is used to realize feature dimension reduction; S5, the dimension-reduced data is input into a classifier to perform training of the classifier model; and S6, steam generator state recognition and monitoring are performed.The steam generator state real-time monitoring method based on artificial intelligence can solve the problems of insufficient sample quantity and lack of data diversity and enhances the robustness of the model when the model has noise or fuzzy classification boundary data.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method for real-time monitoring of the status of a steam generator based on artificial intelligence. Background Technology

[0002] With the increasing automation and intelligence of industry, the demand for monitoring the efficiency and safe operation of energy equipment such as steam generators is growing. As a key piece of equipment for energy conversion, steam generators play a central role in many industrial processes, and real-time monitoring of their status is crucial for ensuring production safety, improving energy efficiency, and reducing environmental impact. However, traditional monitoring methods rely on limited sensor data and manual monitoring, often struggling to accurately capture the complex dynamic changes of equipment in real time, especially when facing large-scale data processing and real-time response requirements, exhibiting significant limitations. Furthermore, conventional data processing methods are insufficient in handling inadequate samples, lack of data diversity, and complex nonlinear data structures, which limits the effectiveness of monitoring systems in anomaly prediction and fault diagnosis.

[0003] To address the aforementioned shortcomings, Chinese invention patent application CN202311255147.0 proposes a method for detecting unmanned aerial vehicles (UAVs), a machine learning model training method, and related equipment. The method includes: acquiring an initial image; using a target detection model to detect at least one first region in the initial image; obtaining detection results for target objects in the first region; determining the first region containing target objects whose detection results meet preset requirements as a second region; performing binary classification on the second region using a classification model to obtain a second classification label for target objects in the second region; filtering out the second regions containing the second classification label of target objects as UAV categories; and identifying the target objects in the filtered second regions as UAVs. This application first filters the initial image based on a target detection model, then performs accurate classification using a classification model, and performs a second filtering. The combination of these two filtering steps can effectively reduce the false alarm rate of UAV detection. Chinese invention patent application number CN202110250990.4 proposes a method, apparatus, storage medium, and electronic device for processing machine learning models. The method includes: acquiring a configuration file and processing data, wherein the configuration file includes at least one processing procedure of at least one machine learning model; and calling an executor corresponding to each processing procedure according to the configuration file to obtain the processing result of each processing procedure. The executor executes the corresponding processing procedure based on the machine learning model and processing data corresponding to each processing procedure in the configuration file. In this embodiment, by setting executors for each processing procedure of the machine learning model, the execution of each processing procedure is realized, and the execution result of each processing procedure is obtained. This eliminates the need to write program code, reduces the implementation difficulty of the machine learning model, and improves the learning and development efficiency of the machine learning model. Chinese invention patent application number CN202011607378.X proposes a training method, server, and computer device for a machine learning model. The method includes: generating a homomorphic key pair, wherein the homomorphic key pair includes a public key and a private key; sending the public key to multiple data owner servers, so that the multiple data owner servers encrypt their own data using the public key, and sending the encrypted ciphertext data to the machine learning server, wherein the machine learning server trains and tests the model based on the ciphertext data provided by the multiple data owner servers to generate a ciphertext machine learning model; and receiving the ciphertext machine learning model sent by the machine learning server, and decrypting it according to the private key to form a plaintext machine learning model, thereby protecting the data provided by the data owner servers and protecting the plaintext machine learning model, with high security.

[0004] The above technical solution has a certain degree of innovation, but the following problems still need to be further addressed:

[0005] 1. In traditional methods, data samples are often limited by the actual collection situation, which often leads to insufficient sample size, lack of sufficient diversity and representativeness, and limits the generalization ability of the model.

[0006] 2. Traditional feature extraction methods may suffer from gradient vanishing or gradient exploding due to structural or algorithmic limitations, or get stuck in local optima due to insufficient algorithm optimization strategies, affecting the training efficiency and performance of the model.

[0007] 3. Traditional feature reduction techniques may not be able to effectively preserve the main features and local structure of data when processing complex and nonlinear data, resulting in the loss of important information in the reduced data and affecting the effect of subsequent processing.

[0008] 4. Traditional classification methods often lack sufficient robustness when dealing with noisy or unclear classification boundaries, resulting in unstable classification results or low accuracy, especially in complex monitoring scenarios. Summary of the Invention

[0009] The purpose of this invention is to propose a real-time monitoring method for the status of a steam generator based on artificial intelligence. This method can solve the problems of insufficient sample size and lack of data diversity in traditional methods, effectively avoid common problems such as gradient vanishing, exploding, or getting trapped in local optima, and thus enhance the robustness of the model when dealing with noisy or fuzzy classification boundary data.

[0010] The technical solution adopted in this invention is:

[0011] The method for real-time monitoring of the status of a steam generator based on artificial intelligence includes the following steps:

[0012] S1. Data Acquisition and Labeling: The collected data comes from the process flow data of the biomass steam generator, as well as the control status data and system operation status of the induced draft fan, blower, feeder, and circulating pump. The data is collected in real time through sensors and monitoring systems and stored in structured JSON format. The collected data is manually labeled.

[0013] S2. Data augmentation: Samples are generated using a quantum generative adversarial network based on random projection embedding, thereby achieving data augmentation;

[0014] S3. Feature Extraction Model Training: The expanded data is input into the feature extraction model for training. A 5-layer fully connected neural network is used for feature extraction.

[0015] S4: Feature Dimensionality Reduction Model Training: Input the data after feature extraction into the feature dimensionality reduction model to train the feature dimensionality reduction model;

[0016] S5. Classifier Model Training: Input the dimensionality-reduced data into the classifier to train the classifier model;

[0017] S6. Steam Generator Status Identification and Monitoring: The trained model processes new steam generator status data to achieve real-time monitoring and status classification. Specifically, the collected raw data is input into the trained feature extraction and feature dimensionality reduction model for feature processing, and then the processed features are input into the classifier model for classifier training to obtain the classification result.

[0018] Furthermore, in step S2, the acquisition, labeling, and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor model generalization ability and affect model accuracy. A quantum generative adversarial network based on random projection embedding is used to generate samples, thereby achieving data expansion. This quantum generative adversarial network based on random projection embedding maps high-dimensional data to low-dimensional quantum states through random projection, then utilizes quantum mechanisms to accelerate the training process of the generative adversarial network. The initialization process of quantum states is designed using random matrix theory to improve the randomness and coverage of data mapping, ensuring the diversity and practicality of the generated data.

[0019] The training process for the quantum generative adversarial network algorithm based on random projection embedding is as follows:

[0020] S201. The quantum generative adversarial network (GAN) first maps high-dimensional data to a low-dimensional quantum state space. Specifically, it uses a random projection embedding method to initialize the quantum state of the GAN. The definition method is as follows:

[0021] ;

[0022] In the formula, Represents the initial quantum state; Indicates the first The complex probability amplitude of each data point; Indicates the first The quantum ground state after mapping 1 data point; N represents the total number of data samples;

[0023] S202. During the training process of the quantum generator, quantum states with a specific distribution are generated through quantum gate operations, and the parameters of the quantum generator are adjusted to minimize the difference between the generated data and the actual data. The parameter update method of the quantum generator is expressed as follows:

[0024] ;

[0025] In the formula, The parameters represent the quantum generator; Indicates the first The parameters of the quantum generator in the next iteration; Indicates the first The parameters of the quantum generator in the next iteration; This represents the learning rate of the quantum generator; Represents the quantum information gain loss function; The quantum information gain loss function is expressed with respect to the parameters. The gradient; For the first The dynamic adjustment coefficient for each iteration;

[0026] The dynamic adjustment coefficient is adjusted dynamically based on the coefficient of variation of the data, and the adjustment method is expressed as follows:

[0027] ;

[0028] In the formula, It is the first hyperparameter to be adjusted; It is the second hyperparameter to be adjusted; The coefficient of variation;

[0029] The coefficient of variation is calculated as follows:

[0030] ;

[0031] in, and These are the standard deviation and mean of the current input batch of data, respectively.

[0032] S203. The quantum state output by the quantum generator, after undergoing an inverse quantum Fourier transform, is converted into synthetic data for machine learning, represented as follows:

[0033] ;

[0034] In the formula, This represents the generated synthetic data; Represents the inverse quantum Fourier transform function;

[0035] S204. During the training process of the quantum discriminator, the quantum discriminator constructs its loss function by comparing the quantum states of real data and synthetic data, and optimizes its parameters by minimizing the loss function. The parameter update process of the quantum discriminator is expressed as follows:

[0036] ;

[0037] In the formula, The parameters represent the quantum discriminator; Indicates the first The parameters of the quantum discriminator in the next iteration; Indicates the first The parameters of the quantum discriminator in the next iteration; This represents the learning rate of the quantum discriminator; The loss function of the quantum discriminator; The loss function of the quantum discriminator with respect to the parameters The gradient;

[0038] The loss function of the quantum discriminator is calculated as follows:

[0039] ;

[0040] In the formula, It is the first The labels of each sample; It is the first The quantum state of a sample;

[0041] S205. Repeat steps S201-S204 until the preset stopping iteration condition is met, which indicates that the model training is complete. After the data augmentation model training is completed, use the trained data to augment the model to increase the number of samples.

[0042] Furthermore, in step S3, during the feature extraction model training, the expanded data is input into the feature extraction model for training. A 5-layer fully connected neural network is used for feature extraction. In existing technologies, some schemes use neural networks for feature extraction. However, in certain neural network structures, problems such as vanishing gradients, exploding gradients, or getting trapped in local optima may occur, affecting the stability of training and the performance of the model. This invention uses a feature extraction-based fruit fly optimized neural network as the feature extraction model. The input layer receives multi-dimensional sensor data, and the hidden layer performs deep feature extraction through a multi-layer neural network. This invention employs an information bottleneck strategy to constrain the feature extraction process, ensuring that the information transmitted to the output layer is the most critical, thereby reducing the propagation of redundant information and effectively accelerating the learning process.

[0043] Specifically, a neural network algorithm based on feature extraction optimized for fruit flies is used, and its training process is as follows:

[0044] S301. Initialize the weights and biases of the neural network. The initialization method for weights and biases is expressed as follows:

[0045] ;

[0046] In the formula, For the neural network Layer weights; For the neural network Layer bias; This indicates that it follows a specific distribution; This represents the number of neurons in layer l-1. This indicates that the mean is 0 and the standard deviation is 0. The normal distribution is assumed, and for the first layer of neurons, the initial standard deviation of the normal distribution is set to 0.01;

[0047] S302. Data is passed layer by layer through the network. During the transmission from the hidden layer to the output layer, an information bottleneck layer is used to force the network to learn more useful features by limiting the amount of information. Specifically, the forward propagation process is represented as follows:

[0048] ;

[0049] In the formula, For the neural network The linear transformation output of the layer; For the neural network Layer output; For the neural network Layer activation noise; It is the ReLU activation function; The scaling factor for the information bottleneck; express Norm;

[0050] The intensity of the activation noise is dynamically adjusted based on the value of the loss function, and the adjustment method is expressed as follows:

[0051] ;

[0052] In the formula, It is the baseline noise figure; This is the loss function for the neural network; It is a noise sensitivity hyperparameter; It is a random noise vector generated from a standard normal distribution;

[0053] Scaling factor of information bottleneck The model is dynamically adjusted based on the entropy changes in the input data to increase its sensitivity to data variations. The calculation method is expressed as follows:

[0054] ;

[0055] In the formula, For target data Entropy; For the first The entropy of the layer output data; The target data is input to the information bottleneck layer. The target data represents the probability data that the prediction of the input data is the correct sample label, and is calculated by the preset Softmax function.

[0056] Target data The entropy is calculated as follows:

[0057] ;

[0058] In the formula, The entropy calculation function is used; all log values ​​in the text are without a base because: in computer programming languages, the log function calculates the natural logarithm by default, so the base is not specified. Moreover, in explaining the principles of machine learning models in this field, the base is usually omitted by default, because no matter what the base is, as long as any base is fixed, it will not affect the model training process and results.

[0059] S303. Use a loss function to evaluate the difference between the output layer's feature representation and the target features. The loss function is calculated as follows:

[0060] ;

[0061] In the formula, Let be the loss function of the neural network. This is the actual output; The predicted output is obtained by calculating the model's output using a preset Softmax function; For balance parameters;

[0062] S304. During the error backpropagation phase, the fruit fly optimization algorithm is used to evaluate the loss corresponding to each set of parameters, guiding the parameters to be updated in the direction of minimizing the loss. The parameter update method is expressed as follows:

[0063] ;

[0064] In the formula, For the first Layer The weight parameters of the neural network in the next iteration. For the first Layer The weight parameters of the neural network in the next iteration. For the first Layer The bias parameters of the neural network in the next iteration. For the first Layer The bias parameters of the neural network in the next iteration; This refers to the adjustment amount of the weight parameters derived from the fruit fly algorithm; This refers to the adjustment amount of the bias parameters derived from the fruit fly algorithm; The learning rate of the neural network;

[0065] S305. Repeat steps S301-S304 until the preset stopping iteration condition is met, which means that the model training is complete.

[0066] Furthermore, in step S4, during the feature dimensionality reduction model training, the data after feature extraction is input into the feature dimensionality reduction model for training. This invention employs a locally preserving projection-based autoencoder neural network to achieve feature dimensionality reduction. This locally preserving projection-based autoencoder neural network includes an encoder and a decoder. The encoder maps the high-dimensional feature space to a low-dimensional feature space, while the decoder attempts to recover the original data from this low-dimensional space, thus preserving the main features of the data while reducing dimensionality. This invention uses a feature-based nonlinear hash encoding strategy to preserve the local structure of the data during the encoding process and maps features to the hash space through a nonlinear function, enhancing the model's ability to handle nonlinear data structures.

[0067] Specifically, feature dimensionality reduction is achieved using an autoencoder neural network algorithm based on local preservation projection. The training process is as follows:

[0068] S401. Randomly initialize the weights and biases of the autoencoder neural network. The initialization method is expressed as follows:

[0069] ;

[0070] In the formula, These are the weight parameters of an autoencoder neural network. These are the bias parameters of the autoencoder neural network. It is the initial standard deviation of the autoencoder neural network; It is the input layer dimension of the autoencoder neural network, that is, the feature dimension after feature extraction; It is the hidden layer dimension of an autoencoder neural network; To generate The interval follows a standard normal distribution. generate The zero vector of dimension;

[0071] S402. The input feature data after dimensionality reduction is processed by the encoder to form an encoded low-dimensional representation. The forward propagation process is represented as follows:

[0072] ;

[0073] In the formula, This represents the encoded feature representation; Indicates attention weight; The dot product symbol; Use the Sigmoid activation function; It is a weight matrix Transpose of; It is the data after dimensionality reduction of the features input into the autoencoder neural network;

[0074] Attention weights, by weighting the input features, highlight the features that are more important to the current monitoring task. The calculation method is expressed as follows:

[0075] ;

[0076] In the formula, The weight matrix is ​​a learnable matrix; For the Softmax function;

[0077] S403. In the hidden layer of the autoencoder neural network, a local preservation projection strategy is adopted to optimize the output of the hidden layer neurons based on the adjacency relationship of the data, so as to preserve the local relationship between data points, as shown below:

[0078] ;

[0079] In the formula, It is a local preservation loss term used to constrain the training process of hidden layer neurons; It is the first The first sample and the first The similarity weights between samples; express Norm; It is the bandwidth parameter of the kernel function;

[0080] S404. Locally preserved low-dimensional features are mapped to a hash space through a nonlinear function. This improves the dimensionality reduction effect by enhancing the separability of the data and strengthens the feature representation capability of the model. This can be expressed as:

[0081] ;

[0082] In the formula, Features after hash encoding; It is a mapping matrix from the feature space to the hash space; It is the bias term in the mapping process; It is a sign function used to generate binary hash codes;

[0083] The calculation method is expressed as follows:

[0084] ;

[0085] In the formula, express transpose, It is a regularization coefficient used to ensure numerical stability; Represents the identity matrix;

[0086] S405, By calculating the output layer and the original input The reconstruction error between the two is used to update the parameters of the autoencoder neural network using gradient descent. and To optimize the performance of the entire network, the method for updating model parameters is as follows:

[0087] ;

[0088] In the formula, It is the total loss function; for Transpose of; It is the input data for reconstruction; It is a regularization parameter; It is the learning rate of the autoencoder neural network; It is the Sigmoid activation function; This indicates a parameter update operation;

[0089] Total loss function Substituting these values ​​into the update formulas for weights and biases, we obtain the update method for weights and biases, expressed as:

[0090] ;

[0091] In the formula, the gradient term and Let represent the partial derivatives of the reconstruction error with respect to the weights and biases, respectively. and This represents the partial derivative of the local preservation loss with respect to the weights and biases;

[0092] weight matrix The update is performed using gradient descent, and the update method is expressed as follows:

[0093] ;

[0094] In the formula, This represents the partial derivative of the total loss with respect to the attention weights; This represents the partial derivative of the attention weights with respect to the weight matrix;

[0095] S406. Repeat steps S401-S405 until the preset stopping iteration condition is met, which indicates that the model training is complete.

[0096] Furthermore, in the classifier model training step S5, the dimensionality-reduced data is input into the classifier for training. This invention employs a random forest algorithm based on sparse subspace learning as the classifier. This random forest algorithm utilizes multiple decision trees to construct an ensemble learning framework, enhancing the model's generalization ability and accuracy by building each tree in a randomly selected subspace. This invention, based on the random forest algorithm, employs fuzzy variable constraints, making the model more robust when processing data with noisy or fuzzy classification boundaries. By using fuzzy logic in the node splitting process of the decision tree, the feature selection and classification decision process is optimized, thereby better adapting to the complexities of steam generator condition monitoring.

[0097] Specifically, the random forest algorithm based on sparse subspace learning is used for training, and the process is as follows:

[0098] S501. Initialize the number of decision trees. and the number of features for each tree The number of features for each tree is initialized as follows:

[0099] ;

[0100] In the formula, The total number of data features after dimensionality reduction of the input features;

[0101] S502, Randomly select from the original feature set Each tree has 10 features, and samples are randomly selected from the training samples to form the training data for each tree. For each tree, decision nodes are recursively constructed. At each node, fuzzy logic is used to evaluate and select the contribution of each feature to the sample classification, allowing for partial overlap and uncertainty to increase the model's ability to handle complex and fuzzy data. The calculation method of information gain based on fuzzy logic when splitting nodes is expressed as follows:

[0102] ;

[0103] In the formula, Information gain based on fuzzy logic is used to select the splitting feature se and the splitting threshold te; The operation is to take the maximum value under the conditions of splitting feature se and splitting threshold te; The fuzzy entropy under the threshold te; For the first Each sample weight; nf is the total number of input samples in the current batch; is the membership function for categories belonging to the threshold te; For the sample Total membership degree;

[0104] Among them, fuzzy entropy represents fuzzy membership degree, and its calculation method is expressed as:

[0105] ;

[0106] In the formula, For a sample to belong to the threshold te Fuzzy probabilities of each category A collection of all categories;

[0107] S503. In each split of the decision tree, calculate the fuzzy membership degree, and select the splitting features and points based on the fuzzy membership degree to optimize the tree structure. The calculation method is expressed as follows:

[0108] ;

[0109] In the formula, It is a parameter that controls the steepness of membership; For the sample Kernel function at threshold te; To control the smoothness parameter of membership;

[0110] Kernel function The calculation method is expressed as follows:

[0111] ;

[0112] In the formula, The bandwidth parameter of the adaptive kernel function;

[0113] The bandwidth parameter of the adaptive kernel function is calculated as follows:

[0114] ;

[0115] In the formula, As a regulating factor; This is the median calculation function. Indicates sample Its neighborhood samples The median of the Euclidean distance between them, and, ;

[0116] S504. During the growth of the decision tree, each node is split until the stopping condition is met.

[0117] S505. Once all decision trees have stopped classifying, the model training is complete. The output of the trained model is then used to determine the final classification result through a voting process.

[0118] Furthermore, in step S202, the quantum information gain loss function The calculation is performed using the Wasserstein loss based on quantum states, and the calculation method is expressed as follows:

[0119] ;

[0120] In the formula, Expressing expectations, and These represent the actual and generated quantum states, respectively. It is an interpolation between the actual and generated quantum states; The discriminant function represents the quantum discriminator; It is the balance coefficient; for Norm.

[0121] Further, in step S203, the generated synthetic data Adjustments are made using a hybrid kernel density estimation method, which utilizes kernel functions to more accurately simulate the target data distribution, as follows:

[0122] ;

[0123] In the formula, Represents the mixing kernel density function; Indicates the first Mixed weights for each core; Indicates the first The bandwidth of each kernel function; Indicates the nuclear center; Representing data The One feature; Indicates the generated synthetic data Feature dimensions;

[0124] The calculation method for the hybrid kernel density function is expressed as follows:

[0125] ;

[0126] In the formula, and It is data The The mean and standard deviation of each feature; These are the position parameters of the Laplace kernel; It is to prevent small positive numbers with a denominator of zero.

[0127] Furthermore, in step S304, the calculation methods for the weight parameter adjustment and bias parameter adjustment obtained based on the fruit fly algorithm are expressed as follows:

[0128] ;

[0129] In the formula, Let be the search radius of the fruit fly algorithm; Generate random numbers from a standard normal distribution in the interval (0,1); It is a symbolic function; Represents the loss function Gradient of weights; Represents the loss function The gradient with respect to the bias.

[0130] Further, in step S502, the fuzzy probability The calculation method is expressed as follows:

[0131] ;

[0132] In the formula, It is a sample Belonging to the th value under the threshold te Fuzzy membership degree of each category; It is a sample Fuzzy membership degree under the threshold te.

[0133] The beneficial effects of this invention are:

[0134] 1. This invention achieves the generation and expansion of data samples through a quantum generative adversarial network based on random projection embedding, solving the problems of insufficient sample quantity and lack of data diversity in traditional methods.

[0135] 2. This invention utilizes a fully connected neural network combined with an information bottleneck strategy and a fruit fly optimization algorithm for feature extraction, effectively avoiding common problems such as gradient vanishing, exploding, or getting trapped in local optima.

[0136] 3. This invention utilizes an autoencoder neural network based on local preserving projection for feature dimensionality reduction. By preserving the local relationships between data points, it improves the model's ability to process nonlinear data.

[0137] 4. This invention utilizes a random forest algorithm based on sparse subspace learning combined with fuzzy logic to enhance the robustness of the model when processing noisy or fuzzy classification boundary data. Detailed Implementation

[0138] This invention is based on existing hardware. The monitored steam generator is a vertical single-turn spiral tube gas-fired steam generator, using natural gas as fuel, primarily for steam production. The burner is installed at the bottom, employing fully premixed low-NOx combustion technology. A copper structure with a metal mesh is located in the center of the burner, extending to the top of the steam generator. The steam generator body is designed as a single-turn spiral finned tube, with a compact and simple structure and a small water capacity (less than 30L). No additional monitoring or inspection is required, and the small water capacity provides advantages such as rapid heating and low heat loss during shutdown. The heat exchange tubes of the steam generator are single-turn spiral finned tubes. The spiral fins are fixed to the tubes using high-frequency welding, and after welding, the fins are flattened. Flue gas passes through baffles on the outside of the spiral finned tubes. The baffles evenly distribute the flue gas, slowing the flow rate and improving heat exchange efficiency, thereby reducing the exhaust gas temperature.

[0139] The principle of the steam generator is that the burner produces high-temperature flue gas, which passes through the spiral finned tube and is affected by the baffle plate, and the flue gas temperature drops to about 200°C. Then it enters the condenser for further cooling, and the exhaust gas temperature drops to fifty or sixty degrees Celsius before finally being discharged from the exhaust port.

[0140] A monitoring system is installed on the steam generator, which includes the following main modules:

[0141] 1. Data Acquisition and Transmission Module: Collects real-time operating data of the steam generator through multiple sensors for subsequent algorithm analysis;

[0142] 2. Data Processing and Augmentation Module: Generates more diverse data samples through data augmentation to ensure sufficient data during training to improve the model's generalization ability;

[0143] 3. Feature Extraction and Dimensionality Reduction Module: Preserves the local structural relationships of the data through feature extraction and dimensionality reduction;

[0144] 4. Status Classification Module: Identifies and classifies different states to achieve multi-state identification of the steam generator;

[0145] 5. Real-time monitoring and alarm module: Displays the status information of the steam generator in real time, automatically triggers alarms to notify operators in case of faults, and records all alarm information.

[0146] Based on this, an artificial intelligence-based real-time monitoring method for the status of steam generators was established, including the following steps:

[0147] S1. Data Acquisition and Labeling

[0148] The collected data comes from the process flow data of the biomass steam generator, as well as the control status data and system operation status of the induced draft fan, blower, feeder, and circulating pump. The data is collected in real time through sensors and monitoring systems and stored in a structured JSON format.

[0149] In this embodiment, the attributes of the collected steam generator status data include:

[0150] Ta represents steam temperature in degrees Celsius (°C); Pa represents steam pressure in Pascals (Pa); Fa represents fuel flow rate in kilograms per hour (kg / h); Aa represents air flow rate in cubic meters per hour (m³ / h); Va represents steam velocity in meters per second (m / s); Ea represents heat output in joules (J); Ca represents chemical composition percentage in percentage (%); Da represents dew point temperature in degrees Celsius (°C); Oa represents oxygen content in percentage (%); Ba represents combustion efficiency in percentage (%).

[0151] It should be noted that this embodiment is only to illustrate one data format and type of the present invention. In practical applications, the data usually has more than 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0152] The collected data is manually labeled; in this embodiment, the labeling categories include: normal working state, initial fault state, intermediate fault state, severe fault state, and work stoppage state, for a total of 5 categories.

[0153] S2, Data Expansion

[0154] In this invention, the acquisition, annotation, and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor model generalization ability and affect model accuracy. A quantum generative adversarial network based on random projection embedding is used to generate samples, thereby expanding the data. This network maps high-dimensional data to low-dimensional quantum states through random projection, and then uses quantum mechanisms to accelerate the training process. The initialization process of the quantum states is designed using random matrix theory to improve the randomness and coverage of data mapping, ensuring the diversity and practicality of the generated data.

[0155] The training process for the quantum generative adversarial network algorithm based on random projection embedding is as follows:

[0156] S201. The quantum generative adversarial network (GAN) first maps high-dimensional data to a low-dimensional quantum state space. Specifically, it uses a random projection embedding method to initialize the quantum state of the GAN. The definition method is as follows:

[0157] ;

[0158] In the formula, Represents the initial quantum state; Indicates the first The complex probability amplitude of each data point; Indicates the first The quantum ground state after mapping 1 data point; N represents the total number of data samples;

[0159] S202. During the training process of the quantum generator, quantum states with a specific distribution are generated through quantum gate operations, and the parameters of the quantum generator are adjusted to minimize the difference between the generated data and the actual data. The parameter update method of the quantum generator is expressed as follows:

[0160] ;

[0161] In the formula, The parameters represent the quantum generator; Indicates the first The parameters of the quantum generator in the next iteration; Indicates the first The parameters of the quantum generator in the next iteration; This represents the learning rate of the quantum generator; Represents the quantum information gain loss function; The quantum information gain loss function is expressed with respect to the parameters. The gradient; For the first The dynamic adjustment coefficient for each iteration; preferably, Set to 0.01.

[0162] The dynamic adjustment coefficient is adjusted dynamically based on the coefficient of variation of the data, and the adjustment method is expressed as follows:

[0163] ;

[0164] In the formula, It is the first hyperparameter to be adjusted; It is the second hyperparameter to be adjusted; The coefficient of variation; preferably, Set to 0.1, Set it to 0.5.

[0165] The coefficient of variation is calculated as follows:

[0166] ;

[0167] in, and These are the standard deviation and mean of the current input batch of data, respectively.

[0168] Quantum information gain loss function The calculation is performed using the Wasserstein loss based on quantum states, and the calculation method is expressed as follows:

[0169] ;

[0170] In the formula, Expressing expectations, and These represent the actual and generated quantum states, respectively. It is an interpolation between the actual and generated quantum states; The discriminant function represents the quantum discriminator; It is the balance coefficient; for Norm. Preferably, Set it to 0.3.

[0171] S203. The quantum state output by the quantum generator, after undergoing an inverse quantum Fourier transform, is converted into synthetic data for machine learning, represented as follows:

[0172] ;

[0173] In the formula, This represents the generated synthetic data; Represents the inverse quantum Fourier transform function;

[0174] Generated synthetic data Adjustments are made using a hybrid kernel density estimation method, which utilizes kernel functions to more accurately simulate the target data distribution, as follows:

[0175] ;

[0176] In the formula, Represents the mixing kernel density function; Indicates the first Mixed weights for each core; Indicates the first The bandwidth of each kernel function; Indicates the nuclear center; Representing data The One feature; Indicates the generated synthetic data Feature dimensions;

[0177] The calculation method for the hybrid kernel density function is expressed as follows:

[0178] ;

[0179] In the formula, and It is data The The mean and standard deviation of each feature; These are the position parameters of the Laplace kernel; This is to prevent small positive numbers with a denominator of zero. Preferably, Set to 0.01.

[0180] S204. During the training process of the quantum discriminator, the quantum discriminator constructs its loss function by comparing the quantum states of real data and synthetic data, and optimizes its parameters by minimizing the loss function. The parameter update process of the quantum discriminator is expressed as follows:

[0181] ;

[0182] In the formula, The parameters represent the quantum discriminator; Indicates the first The parameters of the quantum discriminator in the next iteration; Indicates the first The parameters of the quantum discriminator in the next iteration; This represents the learning rate of the quantum discriminator; The loss function of the quantum discriminator; The loss function of the quantum discriminator with respect to the parameters The gradient; preferably, Set to 0.01.

[0183] The loss function of the quantum discriminator is calculated as follows:

[0184] ;

[0185] In the formula, It is the first The labels of each sample; It is the first The quantum state of a sample;

[0186] S205. Repeat steps S201-S204 until a preset stopping iteration condition is met, which indicates that the model training is complete. In this embodiment, the preset stopping iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000.

[0187] After the data augmentation model is trained, the number of samples is increased using the trained data augmentation model. In this embodiment, the original collected samples are 800, and the data augmentation model generates 200 samples, so the augmented dataset contains 1000 samples.

[0188] S3, Feature Extraction Model Training

[0189] The augmented data is input into the feature extraction model for training, employing a 5-layer fully connected neural network for feature extraction. Existing technologies, such as neural networks, may encounter problems like vanishing gradients, exploding gradients, or getting trapped in local optima in certain neural network structures, affecting training stability and model performance. This invention uses a feature extraction-optimized neural network based on fruit fly modeling. The input layer receives multi-dimensional sensor data, and the hidden layer performs deep feature extraction through a multi-layer neural network. This invention employs an information bottleneck strategy to constrain the feature extraction process, ensuring that the information passed to the output layer is the most critical, thereby reducing the propagation of redundant information and effectively accelerating the learning process.

[0190] Specifically, a neural network algorithm based on feature extraction optimized for fruit flies is used, and its training process is as follows:

[0191] S301. Initialize the weights and biases of the neural network. The initialization method for weights and biases is expressed as follows:

[0192] ;

[0193] In the formula, For the neural network Layer weights; For the neural network Layer bias; This indicates that it follows a specific distribution; This represents the number of neurons in the previous layer; This indicates that the mean is 0 and the standard deviation is 0. The normal distribution is assumed, and for the first layer of neurons, the initial standard deviation of the normal distribution is set to 0.01;

[0194] S302. Data is passed layer by layer through the network. During the transmission from the hidden layer to the output layer, an information bottleneck layer is used to force the network to learn more useful features by limiting the amount of information. Specifically, the forward propagation process is represented as follows:

[0195] ;

[0196] In the formula, For the neural network The linear transformation output of the layer; For the neural network Layer output; For the neural network Layer activation noise; It is the ReLU activation function; The scaling factor for the information bottleneck; express Norm;

[0197] The intensity of the activation noise is dynamically adjusted based on the value of the loss function, and the adjustment method is expressed as follows:

[0198] ;

[0199] In the formula, It is the baseline noise figure; This is the loss function for the neural network; It is a noise sensitivity hyperparameter; It is a random noise vector generated from a standard normal distribution; preferably, Set to 0.002. Set to 2.

[0200] Scaling factor of information bottleneck The model is dynamically adjusted based on the entropy changes in the input data to increase its sensitivity to data variations. The calculation method is expressed as follows:

[0201] ;

[0202] In the formula, For target data Entropy; For the first The entropy of the layer output data; The target data to be input into the information bottleneck layer;

[0203] Target data The entropy is calculated as follows:

[0204] ;

[0205] In the formula, This is a function for calculating entropy.

[0206] S303. Use a loss function to evaluate the difference between the output layer's feature representation and the target features. The loss function is calculated as follows:

[0207] ;

[0208] In the formula, Let be the loss function of the neural network. This is the actual output; The predicted output is obtained by calculating the model's output using a preset Softmax function; For balance parameters; preferably, Set it to 0.3.

[0209] S304. During the error backpropagation phase, the fruit fly optimization algorithm is used to evaluate the loss corresponding to each set of parameters, guiding the parameters to be updated in the direction of minimizing the loss. The parameter update method is expressed as follows:

[0210] ;

[0211] In the formula, For the first Layer The weight parameters of the neural network in the next iteration. For the first Layer The weight parameters of the neural network in the next iteration. For the first Layer The bias parameters of the neural network in the next iteration. For the first Layer The bias parameters of the neural network in the next iteration; This refers to the adjustment amount of the weight parameters derived from the fruit fly algorithm; This refers to the adjustment amount of the bias parameters derived from the fruit fly algorithm; The learning rate of the neural network; preferably, Set to 0.01.

[0212] The calculation methods for the weight parameter adjustment and bias parameter adjustment derived from the fruit fly algorithm are expressed as follows:

[0213] ;

[0214] In the formula, Let be the search radius of the fruit fly algorithm; Generate random numbers from a standard normal distribution in the interval (0,1); It is a symbolic function; Represents the loss function Gradient of weights; Represents the loss function The gradient with respect to the bias.

[0215] S305. Repeat steps S301-S304 until a preset stopping iteration condition is met, which indicates that the model training is complete. In this embodiment, the preset stopping iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000.

[0216] S4: Feature Dimensionality Reduction Model Training

[0217] The extracted features are input into a feature dimensionality reduction model for training. This invention employs a locally preserving projection-based autoencoder neural network (AEN) for feature dimensionality reduction. This AEN includes an encoder and a decoder. The encoder maps the high-dimensional feature space to a low-dimensional feature space, while the decoder attempts to recover the original data from this low-dimensional space, thus preserving the main features of the data while reducing dimensionality. This invention also uses a feature-based nonlinear hash encoding strategy to preserve the local structure of the data during encoding and maps features to a hash space through a nonlinear function, enhancing the model's ability to handle nonlinear data structures.

[0218] Specifically, feature dimensionality reduction is achieved using an autoencoder neural network algorithm based on local preservation projection. The training process is as follows:

[0219] S401. Randomly initialize the weights and biases of the autoencoder neural network. The initialization method is expressed as follows:

[0220] ;

[0221] In the formula, These are the weight parameters of an autoencoder neural network. These are the bias parameters of the autoencoder neural network. It is the initial standard deviation of the autoencoder neural network; It is the input layer dimension of the autoencoder neural network, that is, the feature dimension after feature extraction; It is the hidden layer dimension of an autoencoder neural network; To generate The interval follows a standard normal distribution. generate The zero vector of dimension; preferably, Set it to 100.

[0222] S402. The input feature data after dimensionality reduction is processed by the encoder to form an encoded low-dimensional representation. The forward propagation process is represented as follows:

[0223] ;

[0224] In the formula, This represents the encoded feature representation; Indicates attention weight; The dot product symbol; Use the Sigmoid activation function; It is a weight matrix Transpose of; It is the data after dimensionality reduction of the features input into the autoencoder neural network;

[0225] Attention weights, by weighting the input features, highlight the features that are more important to the current monitoring task. The calculation method is expressed as follows:

[0226] ;

[0227] In the formula, The weight matrix is ​​a learnable matrix; For the Softmax function;

[0228] S403. In the hidden layer of the autoencoder neural network, a local preservation projection strategy is adopted to optimize the output of the hidden layer neurons based on the adjacency relationship of the data, so as to preserve the local relationship between data points, as shown below:

[0229] ;

[0230] In the formula, It is a local preservation loss term used to constrain the training process of hidden layer neurons; It is the first The first sample and the first The similarity weights between samples; express Norm; It is the bandwidth parameter of the kernel function; preferably, Set it to 0.1.

[0231] S404. Locally preserved low-dimensional features are mapped to a hash space through a nonlinear function. This improves the dimensionality reduction effect by enhancing the separability of the data and strengthens the feature representation capability of the model. This can be expressed as:

[0232] ;

[0233] In the formula, Features after hash encoding; It is a mapping matrix from the feature space to the hash space; It is the bias term in the mapping process; It is a sign function used to generate binary hash codes;

[0234] The calculation method is expressed as follows:

[0235] ;

[0236] In the formula, express transpose, It is a regularization coefficient used to ensure numerical stability; Represents the identity matrix;

[0237] S405, By calculating the output layer and the original input The reconstruction error between the two is used to update the parameters of the autoencoder neural network using gradient descent. and To optimize the performance of the entire network, the method for updating model parameters is as follows:

[0238] ;

[0239] In the formula, It is the total loss function; for Transpose of; It is the input data for reconstruction; It is a regularization parameter; It is the learning rate of the autoencoder neural network; It is the Sigmoid activation function; This indicates a parameter update operation;

[0240] Total loss function Substituting these values ​​into the update formulas for weights and biases, we obtain the update method for weights and biases, expressed as:

[0241] ;

[0242] In the formula, the gradient term and Let represent the partial derivatives of the reconstruction error with respect to the weights and biases, respectively. and This represents the partial derivative of the local preservation loss with respect to the weights and biases;

[0243] weight matrix The update is performed using gradient descent, and the update method is expressed as follows:

[0244] ;

[0245] In the formula, This represents the partial derivative of the total loss with respect to the attention weights; This represents the partial derivative of the attention weights with respect to the weight matrix;

[0246] S406. Repeat steps S401-S405 until a preset stopping iteration condition is met, which indicates that the model training is complete. In this embodiment, the preset stopping iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000.

[0247] S5, Classifier Model Training

[0248] The dimensionality-reduced data is input into a classifier for training the classifier model. This invention employs a random forest algorithm based on sparse subspace learning as the classifier. This random forest algorithm utilizes multiple decision trees to construct an ensemble learning framework, enhancing the model's generalization ability and accuracy by building each tree in a randomly selected subspace. Building upon the random forest algorithm, this invention employs fuzzy variable constraints, making the model more robust when handling noisy or fuzzy classification boundaries. By using fuzzy logic in the node splitting process of the decision tree, the feature selection and classification decision-making processes are optimized, thus better adapting to the complexities of steam generator condition monitoring.

[0249] Specifically, the random forest algorithm based on sparse subspace learning is used for training, and the process is as follows:

[0250] S501. Initialize the number of decision trees. and the number of features for each tree In this embodiment, the number of decision trees is initialized to 100, and the number of features per tree is initialized as follows:

[0251] ;

[0252] In the formula, The total number of data features after dimensionality reduction of the input features;

[0253] S502, Randomly select from the original feature set Each tree has 10 features, and samples are randomly selected from the training samples to form the training data for each tree. For each tree, decision nodes are recursively constructed. At each node, fuzzy logic is used to evaluate and select the contribution of each feature to the sample classification, allowing for partial overlap and uncertainty to increase the model's ability to handle complex and fuzzy data. The calculation method of information gain based on fuzzy logic when splitting nodes is expressed as follows:

[0254] ;

[0255] In the formula, Information gain based on fuzzy logic is used to select the splitting feature se and the splitting threshold te; The operation is to take the maximum value under the conditions of splitting feature se and splitting threshold te; The fuzzy entropy under the threshold te; For the first Each sample weight; nf is the total number of input samples in the current batch; is the membership function for categories belonging to the threshold te; For the sample Total membership degree;

[0256] Among them, fuzzy entropy represents fuzzy membership degree, and its calculation method is expressed as:

[0257] ;

[0258] In the formula, For a sample to belong to the threshold te Fuzzy probabilities of each category A collection of all categories;

[0259] Fuzzy probability The calculation method is expressed as follows:

[0260] ;

[0261] In the formula, It is a sample Belonging to the th value under the threshold te Fuzzy membership degree of each category; It is a sample Fuzzy membership degree under the threshold te.

[0262] S503. In each split of the decision tree, calculate the fuzzy membership degree, and select the splitting features and points based on the fuzzy membership degree to optimize the tree structure. The calculation method is expressed as follows:

[0263] ;

[0264] In the formula, It is a parameter that controls the steepness of membership; For the sample Kernel function at threshold te; To control the smoothness parameter of membership; preferably, Set to 1, Set to 3.14.

[0265] Kernel function The calculation method is expressed as follows:

[0266] ;

[0267] In the formula, The bandwidth parameter of the adaptive kernel function;

[0268] The bandwidth parameter of the adaptive kernel function is calculated as follows:

[0269] ;

[0270] In the formula, As a regulating factor; This is the median calculation function. Indicates sample Its neighborhood samples The median of the Euclidean distance between them, and, Preferably, Set it to 0.5.

[0271] S504. During the growth of the decision tree, each node is split until a stopping condition is met; in this embodiment, the stopping condition is that the number of samples under a certain node is less than a threshold. Or reach the maximum depth of the decision tree Preferably, Set to 5. Set it to 4.

[0272] S505. Once all decision trees have stopped classifying, the model training is complete. The output of the trained model is then used to determine the final classification result through a voting process.

[0273] S6. Steam generator status identification and monitoring

[0274] The trained model processes new steam generator status data to achieve real-time monitoring and status classification. Specifically, the collected raw data is input into the trained feature extraction and feature dimensionality reduction model for feature processing. The processed features are then input into the classifier model for classifier training, thereby obtaining the classification result. In this embodiment, the classification categories include: normal operating status, initial fault status, intermediate fault status, severe fault status, and work stoppage status, for a total of 5 categories.

Claims

1. A method for real-time monitoring of the state of a steam generator based on artificial intelligence, characterized by, Comprising the following steps: S1, data acquisition and labeling: the data collected from the process data of the biomass combustion steam generator, the control state data of the induced draft fan, the blower, the feeding, the circulating pump and the system operation state, the data is collected in real time through the sensor and the monitoring system, and is stored in the structured JSON format; the collected data is manually labeled; S2, data expansion: quantum generative adversarial network based on random projection embedding is used for sample generation, and then data expansion is realized; wherein, the training process of sample generation is as follows: S201、The quantum generative adversarial network first maps high-dimensional data to a low-dimensional quantum state space. Specifically, the random projection embedding method is used to initialize the quantum state of the quantum generative adversarial network, and the initial quantum state | ψ c > is defined as: where |ψ c > represents the initialized quantum state; a c,i represents the complex probability amplitude of the i-th data point; |x c,i > represents the mapped quantum ground state of the i-th data point; and N represents the total number of data samples. S202, in the training process of the quantum generator, quantum states with specific distribution are generated through quantum gate operation, and the parameters of the quantum generator are adjusted, so as to minimize the difference between the generated data and the actual data, and the parameter updating mode of the quantum generator is represented as: where θ c denotes a parameter of the quantum generator; denotes a parameter of the quantum generator of the t+1th iteration; denotes a parameter of the quantum generator of the tth iteration; η c denotes a learning rate of the quantum generator; c represents a quantum information gain loss function; represents a gradient of the quantum information gain loss function with respect to the parameter θ c of the parameter θ; δ c (t) is a dynamic adjustment coefficient of the tth iteration; Wherein, the dynamic adjustment coefficient is dynamically adjusted according to the coefficient of variation of data, and the adjustment mode is represented as: wherein κ cd is a first tuning hyperparameter; λ cd is a second tuning hyperparameter; CV c is a coefficient of variation; The calculation method of the coefficient of variation is represented as: where σ cd and μ cd are the standard deviation and mean of the current batch of input data, respectively. S203, the quantum state output by the quantum generator is converted into synthetic data for machine learning after inverse quantum Fourier transform, which is represented as: X synth,c = IQFT(|ψ c >) where X synth,c represents the generated synthetic data; IQFT() represents an inverse quantum Fourier transform function; In step S203, the generated synthetic data X synth,c The adjustment is made by a mixture kernel density estimation, which uses kernel functions to more accurately model the target data distribution, denoted as: where K adjusted,c represents the mixture kernel density function; β c,j represents the mixture weight of the jth kernel; λ c,j represents the bandwidth of the jth kernel function; X synth,R represents the kernel center; represents the jth feature of the data X synth,c ; Mrrepresents the feature dimension of the generated synthetic data X synth,c ; Wherein, the calculation method of the mixed kernel density function is represented as: where μ c,j and σ c,j are the mean and standard deviation of the jth feature of data X synth,c ; v c,j is the location parameter of the Laplace kernel; and ∈ cd is a small positive number to prevent the denominator from being zero. S204, in the training process of the quantum discriminator, the quantum discriminator constructs the loss function of the quantum discriminator by comparing the quantum states of the real data and the synthetic data, and realizes the optimization of the parameters by minimizing the loss function of the quantum discriminator, and the updating process of the parameters of the quantum discriminator is represented as: ωt+1= ωt- η∇L(ωt) c denotes the parameters of the quantum discriminator; denotes the parameters of the quantum discriminator at the t-th iteration; denotes the parameters of the quantum discriminator at the t+1-th iteration; γ c denotes the learning rate of the quantum discriminator; Dc a loss function representing the quantum discriminator; a gradient of the loss function representing the quantum discriminator with respect to the parameter ω c a loss function representing the quantum discriminator; Wherein, the calculation method of the loss function of the quantum discriminator is represented as: where y c,i is the label of the ith sample; |ψ c,i is the quantum state of the ith sample; S205, repeat steps S201-S204 until the preset stopping iteration condition is met, that is, the model training is completed; after the data expansion model training is completed, the trained data expansion model is used to increase the sample number; S3, feature extraction model training: the expanded data is input into the feature extraction model for feature extraction model training, and a 5-layer fully connected neural network is used for feature extraction; S4, feature dimension reduction model training: the feature extracted data is input into the feature dimension reduction model for feature dimension reduction model training; S5, classifier model training: the dimension reduced data is input into the classifier for classifier model training; S6, steam generator state recognition monitoring: the new steam generator state data is processed through the trained model to realize real-time monitoring and state classification, specifically, the collected original data is input into the trained feature extraction and feature dimension reduction model for feature processing, and then the processed features are input into the classifier model for classifier training, and finally the classification result is obtained.

2. The real-time monitoring method of the steam generator state based on artificial intelligence according to claim 1, wherein, In the feature extraction model training of step S3, the neural network algorithm based on feature extraction Drosophila optimization is adopted, and the training process is as follows: S301, initialize the weights and biases of the neural network, and the initialization mode of the weights and biases is represented as: wherein, is the weight of the l-th layer of the neural network; is the bias of the l-th layer of the neural network; ~ denotes being subject to a certain distribution; n p,l-1 is the number of neurons of the l-1-th layer; denotes a normal distribution with mean 0 and standard deviation of 1, and the initialized normal distribution standard deviation is set to 0.01 for the first layer neurons; S302, data is transmitted through network layers layer by layer, and in the transmission process from the hidden layer to the output layer, the information bottleneck layer is used to limit the amount of information to force the network to learn more useful features, specifically, the process of forward propagation is represented as: where, is the linear transformation output of the l-th layer of the neural network; is the output of the l-1-th layer of the neural network; is the activation noise of the l-th layer of the neural network; Re() is the ReLU activation function; β p is the scaling factor of the information bottleneck; |||| denotes the L2 norm; The intensity of the activation noise is dynamically adjusted according to the value of the loss function, and the adjustment manner is represented as: where ζ p is the baseline noise coefficient; p is the loss function for the neural network; γ p is the noise sensitivity hyperparameter; ∈ p is a random noise vector generated from a standard normal distribution; Scaling factor β of information bottleneck p Dynamic adjustment based on the change in entropy of the input data to increase the sensitivity of the model to changes in data, calculated as: In the formula, H(y) pr ) represents the target data y pr Entropy; y is the entropy of the output data of layer l-1; pr The target data is input to the information bottleneck layer. The target data represents the probability data that the prediction of the input data is the correct sample label, and is calculated by the preset Softmax function. Target data y pr The entropy of the target data y is calculated as H(y pr ) = -∑y pr logy pr In the formula, H() is an entropy value calculation function; S303, the difference between the feature expression of the output layer and the target feature is evaluated by using the loss function, and the calculation manner of the loss function is represented as: In the formula, p is a loss function of the neural network, y p is a true output; is a predicted output calculated by a preset Softmax function on the output of the model; λ p is a balance parameter; S304, in the error back propagation stage, the loss corresponding to each group of parameters is evaluated by using the fruit fly optimization algorithm, and the parameters are updated in the direction of the minimum loss, and the updating manner of the parameters is represented as: wherein, is the weight parameter of the neural network for the lth layer at the t+1th iteration, is the weight parameter of the neural network for the lth layer at the tth iteration, is the bias parameter of the neural network for the lth layer at the t+1th iteration, is the bias parameter of the neural network for the lth layer at the tth iteration; is the weight parameter adjustment amount based on the fruit fly algorithm; is the bias parameter adjustment amount based on the fruit fly algorithm; a p is the learning rate of the neural network; S305, repeat steps S301-S304 until the preset stop iteration condition is met, that is, the model training is completed.

3. The real-time monitoring method of the steam generator state based on artificial intelligence according to claim 1, wherein, In the feature dimension reduction model training of step S4, the feature dimension reduction is realized by using the self-encoding neural network algorithm based on local preserving projection, and the training process is as follows: S401, the weights and biases of the self-encoding neural network are randomly initialized, and the initialization manner is represented as: W r = σ gs · randns(d r , h r ) b r = zeros(h r , 1) where W r are the weight parameters of the autoencoder neural network, b r are the bias parameters of the autoencoder neural network, σ gs is the initialization standard deviation of the autoencoder neural network; d r is the input layer dimension of the autoencoder neural network, i.e., the feature dimension after feature extraction; h r is the hidden layer dimension of the autoencoder neural network; randns(d r ,h r ) generates a standard normal distribution in the interval (d r ,h r ); and zeros(h r ,1) generates a zero vector of dimension h r . S402, the input feature dimension reduction data is processed by the encoder part to form the low-dimensional representation after encoding, and the forward propagation process is represented as: In the formula, Z r denotes the encoded feature representation; A r denotes an attention weight; is a dot product symbol; Sig() is a Sigmoid activation function; is a transpose of a weight matrix W r ; X r is data after feature reduction input to the auto-encoding neural network; The attention weight highlights the features more important for the current monitoring task by weighting the input features, and the calculation manner is represented as: A r = softmax(K r · tanh(W r X r + b r )) In the formula, K r is a learnable weight matrix; softmax() is a Softmax function; S403, in the hidden layer of the self-encoding neural network, the local preserving projection strategy is adopted, the output of the hidden layer neuron is optimized according to the adjacency relationship of the data, so as to maintain the local relationship between the data points, and is represented as: wherein L r is a local keep loss term for constraining the training process of hidden layer neurons; is a similarity weight between the i-th sample and the j-th sample; ||| denotes the L2 norm; σ r is a bandwidth parameter of the kernel function; S404, the low-dimensional features of local preservation are mapped to the hash space through a nonlinear function, the dimension reduction effect is optimized by improving the separability of the data, and the feature expression ability of the model is enhanced, and is represented as: In the formula, H r is the hash-coded feature; V r is a mapping matrix from the feature space to the hash space; c r is a bias term in the mapping process; sgn() is the sign function to generate binary hash codes; V r The calculation method is represented as: In the formula, Z represents r transpose, γ r It is a regularization coefficient used to ensure numerical stability; I represents the unit matrix; S405, update the parameters W of the auto-encoder neural network by gradient descent method through calculating the reconstruction error between the output layer and the original input X r r and b r , the way to update the model parameters is represented as:​ where Δ r is the total loss function; is the transpose of W r ; is the reconstructed input data; λ r is the regularization parameter; α r is the learning rate of the auto-encoding neural network; and Sig() is the Sigmoid activation function. <--- represents the parameter updating operation; The total loss function Δ r Substituting the update formula of the weight and the bias into the update formula of the weight and the bias, the update mode of the weight and the bias is obtained, which is represented as: where the gradient terms and denote the partial derivatives of the reconstruction error with respect to the weights and biases, respectively, and denote the partial derivatives of the local preservation loss with respect to the weights and biases, respectively; The weight matrix K r The updating is performed in a gradient descent manner, and the updating manner is represented as: wherein denotes the partial derivative of the total loss with respect to the attention weight; denotes the partial derivative of the attention weight with respect to the weight matrix; S406, repeat steps S401-S405 until the preset stop iteration condition is met, that is, the model training is completed.

4. The real-time monitoring method of the steam generator state based on artificial intelligence according to claim 1, wherein, In the classifier model training of step S5, the random forest algorithm based on sparse subspace learning is used for training, and the process is as follows: S501, initialize the number of decision trees N u and the number of features F for each tree u The number of features for each tree is initialized as In the formula, d u is the total number of data features after input feature dimension reduction; S502、Randomly extract F features from the original feature set u S502、Randomly extract F features from the original feature set For each tree, recursively construct decision nodes, at each node, use fuzzy logic to evaluate and select the contribution of each feature to the classification of samples, allow the existence of partial overlap and uncertainty to increase the ability of the model to handle complex and fuzzy data, and the calculation method of information gain based on fuzzy logic at node splitting is represented as: where G u (se,te) is the information gain based on fuzzy logic for selecting split feature se and split threshold te; is the max operation for split feature se and split threshold te; H u (te) is the fuzzy entropy at threshold te; w ui is the weight of the i-th sample; nf is the total number of input samples in the current batch; is the membership function for the class at threshold te; is the total membership of the sample . Wherein, the fuzzy entropy represents the fuzzy membership degree, and the calculation manner is represented as: In the formula, P k is the fuzzy probability of the sample belonging to the kth category at the threshold te, and Cla is the set of all categories. S503, in each splitting process of the decision tree, the fuzzy membership degree is calculated, and the splitting feature and point are selected according to the fuzzy membership degree, so as to optimize the structure of the tree, and the calculation manner is represented as: where K u is a steepness parameter that controls the membership degree; is a kernel function at the threshold te; η is a kernel function at the threshold te; η u is a smoothness parameter that controls the membership degree; The kernel function is calculated as In the formula, is a bandwidth parameter of the adaptive kernel function; The calculation manner of the bandwidth parameter of the adaptive kernel function is represented as: where a u is a tuning factor; median() is a median calculation function, ) denotes the median of the Euclidean distances between the sample and its neighborhood samples and, S504, in the growth process of the decision tree, each node is split until the stop condition is met; S505, after all the decision trees stop classification, that is, the model training is completed, and the output of the trained model is output through voting to obtain the final classification result.

5. The real-time monitoring method of the steam generator state based on artificial intelligence according to claim 1, wherein, In step S202, a quantum information gain loss function c The calculation is performed using the quantum state-based Wasserstein loss, and the calculation manner is represented as: where represents the expectation, |ψ real > and |ψ gen > respectively represent the real and generated quantum states; |ψ interp > is the interpolation between the real and generated quantum states; D c () denotes the discriminant function of the quantum discriminator; λ c is the balancing coefficient; ||| | | is the L2 norm.

6. The real-time monitoring method of the steam generator state based on artificial intelligence according to claim 2, wherein, In step S304, the calculation manners of the weight parameter adjustment amount and the bias parameter adjustment amount obtained based on the fruit fly algorithm are represented as: where p p is the search radius of the fruit fly algorithm; randn(0,1) generates a random number of standard normal distribution in the interval (0,1); sgn() is a sign function; denotes the loss function p the gradient with respect to the weight; denotes the loss function p the gradient with respect to the bias.

7. The artificial intelligence based real time monitoring method of the steam generator status as claimed in claim 4 wherein, In step S502, the blur probability P k is calculated as follows: wherein is the sample fuzzy membership at threshold te belonging to the kth class; is the sample fuzzy membership at threshold te.

Images