Multi-level information fusion nuclear power device health state prediction method and equipment

Abstract

According to the multi-level information fusion nuclear power device health state prediction method and equipment, data driving is taken as a core method, a valve in a nuclear power production device is taken as a main object, and multi-level and multi-dimensional health related data are integrated, so that fusion of comprehensive information of the valve is realized; and a more comprehensive view angle is provided for health state evaluation. According to the method, graph deep learning is introduced in the aspect of data modeling, and the model has excellent graph structure adaptability and sensitivity to multi-level association, so that the system can more flexibly adapt to a complex nonlinear relationship, the integration capability to multi-source data is enhanced, and the valve health assessment has higher depth and precision. By adopting a multi-level information fusion method, the system can comprehensively consider a plurality of factors in valve health assessment, including multi-dimensional data such as the current state and the operation history, so that an individualized health state assessment model is established, and assessment is more comprehensive and deep.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and particularly to a multi-level information fusion method and device for predicting the health state of nuclear power devices. Background Art

[0002] A nuclear power plant refers to a device that uses the fission reaction of nuclear fuel in a nuclear reactor to generate heat energy and convert it into power, and has a complex structure and diverse components. These key components need to withstand complex operating conditions during operation, and their safety and stability directly affect the operation safety of nuclear power equipment, and further affect the reliability of the entire system.

[0003] Currently, most nuclear power plants are subject to regular inspections. If damage or abnormalities are found in the structure, targeted maintenance will be carried out. However, this method has strong passivity, resulting in potential safety hazards in the equipment. Traditional health assessment methods usually adopt linear models, which are difficult to capture complex linear relationships, resulting in insufficient accuracy in health state assessment, and are limited in information fusion, making it difficult to comprehensively consider multi-level information. Therefore, the assessment results may not be comprehensive enough. In addition, the assessment of the health state of valves requires the ability of real-time monitoring and dynamic tracking, while existing methods usually lack the corresponding real-time monitoring ability, resulting in insufficient real-time feedback on the health changes of valves. And for the health state assessment of key components of nuclear power plants, most of them adopt the process of regular inspection, manual analysis and management of inspection data, and subjective judgment of the health state based on expert knowledge. Existing methods have the disadvantages of high labor cost, long time consumption, low efficiency, chaotic data management, and subjective assumption.

[0004] Therefore, it is of great significance to adopt certain data management, data analysis, machine learning and other methods to form an autonomous and objectively based health feature measurement method. Summary of the Invention

[0005] A multi-level information fusion method, device and storage medium for predicting the health state of nuclear power devices proposed by the present invention can at least solve one of the technical problems in the background art.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A multi-level information fusion method for predicting the health state of nuclear power devices, which is executed by a computer device through the following steps:

[0008] Step 1: Obtain the fault data of key components of the nuclear power plant;

[0009] According to the long-term monitoring and regular test records of key components of the nuclear power plant, obtain the fault data including fault types, frequencies and severities;

[0010] Step 2: Classify the obtained fault data according to the orthogonal defect classification method;

[0011] Using the orthogonal defect classification method, through feature extraction and feature selection of the fault data, map the multi-dimensional fault data to an orthogonal feature space, so that different types of faults have obvious separation in this space, and classify different faults according to this separation and assign classification labels;

[0012] Step 3: Assign weights to different fault types according to the subjective and objective fusion weighting method;

[0013] Use the subjective and objective fusion weighting method to assign weights related to the health state characteristics to different fault types classified in Step 2. This can be achieved by combining expert experience and data analysis methods to quantitatively evaluate the importance of different fault types to the health state characteristics of the nuclear power plant and assign weight labels;

[0014] Step 4: Perform data preprocessing on the data before neural network training based on normalization and sliding window data preprocessing operations;

[0015] Normalize the data labeled in the early stage, map all training set and test set data to the same scale to eliminate the influence brought by the differences in dimension and value range between different features. At the same time, set the width and step size of the sliding window to ensure the reliability and consistency of the data;

[0016] Step 5: Construct and train a health assessment model based on graph deep learning;

[0017] Statistically analyze the data preprocessed in Step 4, establish a fault knowledge graph, and build a health assessment model of graph deep learning based on this to realize the health assessment of key equipment of the nuclear power plant.

[0018] Furthermore, the fault data acquisition method in Step 1 of the present invention specifically includes the following steps:

[0019] Step 1.1: Collect monitoring data during the long-term operation of key components of the nuclear power plant, test data from regular tests, historical data stored in the background, and simulation data simulated by experts;

[0020] Step 1.2: Combine expert experience and operation and maintenance records to subjectively classify and label the collected data, where the data types include: fault type, fault frequency, and fault severity.

[0021] Furthermore, the orthogonal defect classification method in Step 2 of the present invention classifies the obtained fault data, and the specific steps are as follows:

[0022] Step 2.1: Classify the fault types of the key components of the nuclear power plant obtained in Step 1 into m categories. There are n samples for each fault type, and the feature vector of the fault data is represented as x i , where i is the sample index;

[0023] Step 2.2: Through the principal component analysis (PCA) method, perform an orthogonal transformation on the classified fault data to map the feature vector x i into the orthogonal feature space to obtain the orthogonal feature vector y i ;

[0024] Step 2.3: For the j-th feature, calculate its variance ratio y jk in the k-th fault type. The variance ratio is calculated by the following formula:

[0025]

[0026] where, represents the orthogonal eigenvalue of the i-th sample under the j-th feature in the k-th fault type, and Var represents the variance. By calculating the variance ratio of each feature in different fault types, evaluate the contribution degree of this feature to distinguishing different fault types. According to the size of the variance ratio, select the feature subspace with significant influence, and then use it for the classification and identification of fault data.

[0027] Furthermore, in Step 3 of the present invention, the subjective and objective fusion weighting method assigns weights to different fault types. The specific steps are as follows:

[0028] Step 3.1: Use the subjective weighting method to assign weights α j to different fault types, where j represents the fault type. Through long-term operation and maintenance data and historical experience summary, summarize the influence degree of different fault types on the health state, and combine expert knowledge to agree on the weights α j of the subjective weighting method for different fault types. In special cases, the analytic hierarchy process (AHP) may also be used to derive α j ;

[0029] Step 3.2: Use the objective weighting method to assign weights β j to different fault types, where j represents the fault type. The entropy weight method is used for calculation. For nuclear power plants in different health states in historical data, collect the statistical data of their different fault types, calculate the entropy value of each fault type. When the health state is different, the statistical data of various fault types will also be different. For different health states, if the statistical data of the fault type varies greatly, it means its entropy value is large and it has a great influence on the health state. Therefore, a larger weight β j is assigned. At the same time, the objective weighting method can also use the TOPSIS method, or combine the two methods to calculate the objective fault weight β j;

[0030] Step 3.3: Combine the two weighting methods. Here, this weighting method is called the subjective and objective fusion weighting method, or the combined weighting method for short. The weight W of this method j is expressed as:

[0031]

[0032] where j represents the fault type, and α j is the weight of the subjective weighting method, and β j is the weight of the objective weighting method (entropy weight method).

[0033] Furthermore, the data preprocessing operation based on normalization and sliding window in step 4 of the present invention is specifically as follows:

[0034] Step 4.1: Adopt linear function normalization. Using the maximum and minimum values of each parameter variable in the dataset, scale the original data proportionally to the interval [0, 1]. The formula is as follows:

[0035]

[0036] where x i is the original data of the valve component of the nuclear power production device, x' i is the normalized data, is the maximum value in the original data of this parameter, is the minimum value in the original data of this parameter;

[0037] Step 4.2: On the basis of normalization, in order to make full use of all the data, perform a sliding window operation on the data. Combine the sliding window and slicing to divide the data x1, x2, x3,..., x n of the training set and test set into samples. Set the width and sliding step of the sliding window, and use the historical sliced data to test the data of the valve parameter to be detected at the current moment. The expression is as follows:

[0038] X(t) = {x(t - 1), x(t - 2), …, x(t - D)}

[0039] Y(t) = y(t)

[0040] where X(t) is the data input into the model at time t, Y(t) is the value of the parameter to be detected at time t, x(t) is the data of this parameter at time t, D is the number of sliding window slices, that is, the sampling dimension, and x(t - 1) is the data of this parameter at the previous 1 timestamp after the sliding window operation.

[0041] Furthermore, the method for constructing and training the health assessment model based on graph deep learning in step 5 of the present invention is specifically as follows:

[0042] Step 5.1: By statistically analyzing a large amount of data, find the linear and non-linear correlations between different parameters of valve components, establish a fault knowledge graph, and record the behavior patterns and characteristics of different valves under different fault conditions.

[0043] Step 5.2: Identify the threat factors and transmission relationships of valves in the nuclear power plant through correlation analysis and the fault knowledge graph.

[0044] Step 5.3: Based on the basic theory of graph theory, construct the adjacency matrix of the graph neural network to accurately reflect the association relationship between valves, providing a basis for subsequent information fusion and health status assessment.

[0045] Step 5.4: Through the extension of the trusted adjacency matrix, combined with multi-source boundary conditions, incorporate additional data sources and boundary conditions such as variable working conditions and dynamic ambient temperature into the adjacency matrix to improve the accuracy and reliability of the assessment. This includes data related to the operating status of the valve device obtained from flow sensors, pressure sensors, etc., to help detect the health status of the valve, and design a multi-level information fusion health status assessment model using graph deep learning theory.

[0046] Step 5.5: Conduct multiple iterative trainings on the graph convolution model.

[0047] In the k-th layer of graph convolution, the update formula for node features is:

[0048]

[0049] where, H (k) represents the node features of the k-th layer, is the normalized adjacency matrix, is the angular measurement matrix, and W (k) is the weight matrix of the k-th layer. Each element in the weight matrix represents the connection strength between nodes, that is, it reflects the association degree between different valves, and σ(·) represents the activation function.

[0050] Step 5.6: Through multiple levels of graph convolution layers, obtain feature representations at different levels. By fusing the features of each layer, a health feature representation of multi-level information fusion can be obtained:

[0051] H = Concat(H (1) , H (2) ,..., H (K) )

[0052] where, H represents the health feature matrix after multi-level fusion, and K represents the number of graph convolution layers.

[0053] Step 5.7: Use the fused health feature matrix to perform health status assessment by methods such as classification and regression, and implement it through a fully connected layer and a softmax function:

[0054] Y = softmax(HW)

[0055] where Y represents the predicted result of the health status, and W is the weight matrix of the fully connected layer;

[0056] Step 5.8: According to the usage scenarios and characteristics of the valve components, divide the health status into 4 levels: "healthy", "sub-healthy", "near-failure", and "failed".

[0057] On the other hand, the present invention also discloses a computer device, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0058] As can be seen from the above technical solutions, a multi-level information fusion nuclear power device health status prediction method of the present invention takes data-driven as the core method, takes the valves in the nuclear power production device as the main object, and realizes the fusion of comprehensive multi-level and multi-dimensional health-related data, providing a more comprehensive perspective for health status assessment. In data modeling, graph deep learning is introduced. This model has excellent adaptability to graphic structures and sensitivity to multi-level associations, so that the system can more flexibly adapt to complex non-linear relationships, enhance the integration ability of multi-source data, and make the valve health assessment more in-depth and accurate. By adopting the method of multi-level information fusion, the system can comprehensively consider multiple factors in valve health assessment, including current status, operation history and other multi-dimensional data, so as to establish an individualized health status assessment model, making the assessment more comprehensive and in-depth.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0060] 1. By using methods such as subjective and objective fusion weighting, with the assistance of experts' historical experience, scientific and objective weight indicators are formed to achieve the goal of unifying the health feature performance indicators of components, further reducing subjective factors, thereby improving the accuracy of health feature measurement and making the measurement results consistent.

[0061] 2. By comprehensively integrating multi-level and multi-dimensional health-related data, the fusion of comprehensive information of nuclear power devices is realized, a set of model parameters with strong generality is formed, and then the health features of key components of nuclear power plants are effectively identified and measured. This process is an automated identification and measurement process, reducing the participation of personnel and greatly reducing labor costs. Brief Description of the Drawings

[0062] Figure 1 Schematic diagram of the data acquisition process of the present invention;

[0063] Figure 2 It is a block diagram of the model in an embodiment of the present invention. Specific implementation manners

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention.

[0065] As Figure 1 shown, the multi-level information fusion nuclear power device health state prediction method described in this embodiment, that is, a health feature measurement method for key components of a nuclear power plant, first obtains the fault data of key components of the nuclear power plant including information such as type, frequency, and severity. On this basis, the orthogonal defect classification method is used to classify the fault data, and then the subjective and objective fusion weighting method is used to assign weights to different fault types for the health state features. Thus, through data preprocessing such as normalization and sliding window, the fusion of health state monitoring information in each state monitoring data is realized, and the association mapping relationship between the state monitoring data and the health state features is extracted to accurately measure the health state features of the key components of the nuclear power plant.

[0066] The specific steps are as follows:

[0067] Step 1: Obtain the fault data of the key components of the nuclear power plant;

[0068] Step 2: Classify the obtained fault data according to the orthogonal defect classification method;

[0069] Step 3: Assign weights to different fault types according to the subjective and objective fusion weighting method;

[0070] Step 4: Based on the data preprocessing operations of normalization and sliding window, preprocess the data before neural network training;

[0071] Step 5: Construct and train a health assessment model based on graph deep learning.

[0072] Furthermore, the specific steps for obtaining the fault data are as follows:

[0073] Step 1.1: Collect the monitoring data during the long-term operation of the key components of the nuclear power plant, the test data from regular tests, the historical data stored in the background, the simulation data simulated by experts, etc.;

[0074] Step 1.2: Combine expert experience and operation and maintenance records to subjectively classify and label the collected data. The data types include: fault types, fault frequencies, and fault severities, etc. The specific types still need to be classified according to the evaluation model.

[0075] The orthogonal defect classification method mentioned above is a commonly used fault data classification method. Its basic principle is to classify fault data according to different characteristics to identify and distinguish different types of faults. This method extracts and selects features from fault data, maps multi-dimensional fault data to an orthogonal feature space, so that different types of faults have obvious separation degrees in this space. Specifically, the orthogonal defect classification method is based on orthogonal transformation, transforms the original fault data into orthogonal features, and determines the feature subspace that has a significant impact on the health state features by calculating the variance ratio of each feature in each fault type. The basic principle of the orthogonal defect classification method is to transform the original multi-dimensional fault data into an orthogonal feature space, determine the feature subspace that has a significant impact on the health state features by calculating the variance ratio of the features, so as to classify and distinguish different types of faults, and provide support for fault diagnosis and health state assessment. The specific steps are as follows:

[0076] Step 2.1: Classify the fault types of the key components of the nuclear power plant obtained in Step 1 into m categories. Each fault type has n samples, and the feature vector of the fault data is represented as x i , where i is the sample index;

[0077] Step 2.2: Through the principal component analysis (PCA) method, perform orthogonal transformation on the classified fault data, and map the feature vector x i to the orthogonal feature space to obtain the orthogonal feature vector y i ;

[0078] Step 2.3: For the jth feature, calculate its variance ratio y jk in the kth fault type. The variance ratio is calculated by the following formula:

[0079]

[0080] where, represents the orthogonal eigenvalue of the ith sample under the jth feature in the kth fault type, and Var represents the variance. By calculating the variance ratio of each feature in different fault types, evaluate the contribution degree of this feature to distinguishing different fault types. According to the size of the variance ratio, select the feature subspace with significant influence, and then use it for the classification and identification of fault data.

[0081] The following takes the two fault types of jamming and stalling as examples for illustration:

[0082] There are two types of faults, namely jamming and stalling, in the key components of the current nuclear power plant, and the data includes current, torque, temperature, etc. By using the orthogonal defect classification method, the variance ratios of the three types of data under these two types of faults can be obtained, that is, by calculating the variance ratios of different data characteristics of multiple samples through formulas and then performing average fusion.

[0083] Calculate the variance ratio during jamming. First, assume that jamming and stalling are fault types 1 and 2 respectively, and current, torque, and temperature are characteristics 1, 2, and 3 respectively. Then, obtain the orthogonal eigenvalue of current during jamming for the i-th sample through the PCA method Orthogonal eigenvalue of torque And orthogonal eigenvalue of temperature Similarly, the orthogonal eigenvalues during stalling are respectively Next, calculate six variances from the six groups of obtained orthogonal eigenvalues. Finally, calculate the ratios of the six variances respectively, which are the variance ratios of fault type - characteristic. It is easy to see that the variance ratio of torque is large during jamming, and the variance ratio of current is large during stalling. Then, determine the contribution degree of each data characteristic to the fault through the variance ratios of different data. Finally, further classify and identify the faults using the contribution degree of the data characteristics. At the same time, this method also conforms to expert knowledge and historical experience. Because during jamming, the torque will change mainly and show periodic oscillation; during stalling, the current changes mainly. Therefore, the contribution degree of the torque characteristic is large during jamming, and the contribution degree of the current characteristic is large during stalling.

[0084] The described subjective and objective fusion weighting method is a common method for assigning weights related to the health state characteristics to different fault types. It can quantitatively evaluate the importance of different fault types to the health state characteristics of the nuclear power plant by combining expert experience and data analysis methods.

[0085] The subjective weighting method has greater advantages than the objective weighting method in determining weights according to the intentions of decision-makers, but its objectivity is relatively poor and its subjectivity is relatively strong; while the objective weighting method has objective advantages, but it cannot reflect the degree of importance that decision-makers attach to different indicators, and there will be a certain degree where the sum of weights is contrary to the actual indicators. Therefore, when assigning weights to indicators, the internal statistical laws and authoritative values between indicator data should be considered. A reasonable decision-making indicator weighting method is given, that is, a combined weighting method that combines the subjective weighting method and the objective weighting method to make up for the deficiencies brought by single weighting. The specific steps of the subjective and objective weighting method are as follows:

[0086] Step 3.1: Use the subjective weighting method to assign weights α to different fault types j, where j represents the type of fault. Through long-term operation and maintenance data and historical experience, the influence degree of different fault types on the health status is summarized. Combining with expert knowledge, the weight α of the subjective weighting method for different fault types is agreed upon. j In special cases, the analytic hierarchy process (AHP) is also used to derive α. j ;

[0087] Step 3.2: Use the objective weighting method to assign the weight β to different fault types. j , where j represents the type of fault. The entropy weight method is used for calculation. For the nuclear power plant with different health statuses in the historical data, the statistical data of different fault types are collected, and the entropy value of each fault type is calculated. When the health status is different, the statistical data of various fault types will also be different. If, for different health statuses, the statistical data of a certain fault type vary greatly, it indicates that its entropy value is large and it has a great influence on the health status. Therefore, a larger weight β is assigned. j At the same time, the TOPSIS method can also be used for the objective weighting method, or the two methods can be combined to calculate the objective fault weight β. j ;

[0088] Step 3.3: Combine the two weighting methods. Here, this weighting method is called the subjective and objective fusion weighting method, or the combined weighting method for short. The weight W of this method j is expressed as:

[0089]

[0090] where j represents the type of fault, and α j is the weight of the subjective weighting method, and β j is the weight of the objective weighting method (entropy weight method).

[0091] The subjective and objective fusion weighting method mainly assigns weights related to the health status characteristics to various fault types of the key components of the nuclear power plant through different experiences, knowledge, and information, converting the subjective language of the health status into an objective mathematical data representation.

[0092] Before formally training the neural network, data preprocessing is required. Data preprocessing mainly includes data normalization and sliding window processing, as shown in Figure 2 the data set analysis and processing section. To map all the data in the training set and test set to the same scale to eliminate the influence caused by the differences in dimension and value range between different features, the specific method is as follows;

[0093] Step 4.1: Adopt linear function normalization. Using the maximum and minimum values of each parameter variable in the data set, the original data is scaled proportionally to the interval [0, 1]. The formula is as follows:

[0094]

[0095] Among them, x i is the original data of the valve component of the nuclear power production device, and x' i is the normalized data, is the maximum value in the original data of this parameter, is the minimum value in the original data of this parameter;

[0096] Step 4.2: On the basis of normalization, in order to make full use of all the data, perform a sliding window operation on the data, and divide the data x1, x2, x3,..., x of the training set and the test set by combining the sliding window and slicing. n Set the width of the sliding window and the sliding step size, and use the historical sliced data to test the data of the valve parameter to be detected at the current moment. The expression is as follows:

[0097] X(t) = {x(t - 1), x(t - 2), …, x(t - D)}

[0098] Y(t) = y(t)

[0099] Among them, X(t) is the data input into the model at time t, Y(t) is the value of the parameter to be detected at time t, x(t) is the data of this parameter at time t, D is the number of sliding window slices, that is, the sampling dimension, x(t - 1) is the data of this parameter at the previous 1 timestamp after the sliding window operation, and so on. According to the basic principle of the Fibonacci sequence, this patent samples the data points on the data set to obtain 18 timestamps on the data set - 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610, 987, 1597, 2584, 4181, and uses this as the sampling dimension, and selects 100 as the sliding step size to perform the sliding window processing operation on the data set, so as to ensure the reliability and consistency of the data.

[0100] The specific construction method of the health assessment model based on graph deep learning in step 5 is as follows:

[0101] Step 5.1: Through statistical analysis of a large amount of data, find out the linear and non - linear correlations between different parameters of the valve component, and establish a fault knowledge graph to record the behavior patterns and characteristics of different valves under different fault conditions;

[0102] Step 5.2: Identify the threat factors and transmission relationships of the valves in the nuclear power plant through correlation analysis and the fault knowledge graph;

[0103] Step 5.3: Based on the basic theory of graph theory, construct the adjacency matrix of the graph neural network to accurately reflect the association relationship between the valves, providing a basis for subsequent information fusion and health status assessment;

[0104] Step 5.4: Through the extension of the trusted adjacency matrix and the combination of multi-source boundary conditions, additional data sources and boundary conditions such as the change work compliance and dynamic ambient temperature are incorporated into the adjacency matrix to improve the accuracy and reliability of the assessment. This includes data related to the operating status of the valve device obtained by flow sensors, pressure sensors, etc., to help detect the health status of the valve, and a multi-level information fusion health status assessment model is designed using graph deep learning theory.

[0105] Step 5.5: Perform multiple iterative trainings on the graph convolution model.

[0106] In the k-th layer of graph convolution, the update formula for node features is:

[0107]

[0108] where, H (k) represents the node features of the k-th layer, is the normalized adjacency matrix, is the angular measurement matrix, W (k) is the weight matrix of the k-th layer. Each element in the weight matrix represents the connection strength between nodes, that is, it reflects the correlation degree between different valves, and σ(·) represents the activation function.

[0109] Step 5.6: Through multi-level graph convolution layers, obtain feature representations at different levels. By fusing the features of each layer, a health feature representation with multi-level information fusion can be obtained:

[0110] H = Concat(H (1) , H (2) ,..., H (K) )

[0111] where, H represents the health feature matrix after multi-level fusion, and K represents the number of graph convolution layers.

[0112] Step 5.7: Use the fused health feature matrix to perform health status assessment using methods such as classification and regression, and implement it through a fully connected layer and a softmax function:

[0113] Y = softmax(HW)

[0114] where, Y represents the prediction result of the health status, and W is the weight matrix of the fully connected layer.

[0115] Step 5.8: According to the usage scenario and characteristics of the valve components, divide their health status into 4 levels: "healthy", "sub-healthy", "near failure", and "failure". The corresponding relationship between the system health status classification levels and the system health status prediction results is shown in Table 1 below:

[0116] Table 1 Corresponding Relationship between Health Status Classification Levels and Health Indexes

[0117]

[0118] A multi-level information fusion nuclear power device health status prediction method according to an embodiment of the present invention is executed by a computer device through the following steps: obtaining fault data of a nuclear power plant through a long-term operating monitoring device, stored historical records, regularly conducted test records, or other data sources. Based on orthogonal transformation through the orthogonal defect classification method, the original fault data is transformed into orthogonal features. By calculating the variance ratio of each feature in each fault type, a feature subspace that has a significant impact on the health status features is determined. By calculating the variance ratio of the features, a feature subspace that has a significant impact on the health status features is determined to classify and distinguish different types of faults. Further, the subjective and objective fusion weighting method is used to assign weights related to the health status features to different fault types.

[0119] As can be seen from the above technical solutions, for the multi-level information fusion nuclear power device health status prediction method and system of the present invention, this technology takes data-driven as the core method, takes the valves in the nuclear power production device as the main object, and realizes the fusion of comprehensive information of the valves by integrating multi-level and multi-dimensional health-related data, providing a more comprehensive perspective for health status assessment. This innovative method aims to thoroughly improve the limitations of traditional health assessments, thereby more accurately reflecting the operating conditions of the valves. Graph deep learning is introduced in data modeling. This model has excellent adaptability to graph structures and sensitivity to multi-level associations, so that the system can more flexibly adapt to complex non-linear relationships, enhance the integration ability of multi-source data, and make the valve health assessment more in-depth and accurate. By adopting the multi-level information fusion method, the system can comprehensively consider multiple factors in valve health assessment, including multi-dimensional data such as the current state and operating history, so as to establish an individualized health status assessment model, making the assessment more comprehensive and in-depth.

[0120] On the other hand, the present invention also discloses a computer device, including a memory and a processor. When the computer program stored in the memory is executed by the processor, the processor executes the steps of the above method.

[0121] In another embodiment provided by the present application, a computer program product containing instructions is also provided. When it runs on a computer, the computer is made to execute any of the above-mentioned mobile source emission prediction methods based on temporal feature migration.

[0122] It is understandable that the system provided by the embodiments of the present invention corresponds to the method provided by the embodiments of the present invention. For the explanations, examples and beneficial effects of related content, reference can be made to the corresponding parts in the above method.

[0123] An embodiment of the present application further provides an electronic device, including a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete communication with each other through the communication bus.

[0124] The memory is used to store a computer program.

[0125] When the processor is used to execute the program stored in the memory, it implements the above-mentioned mobile source emission prediction method based on temporal feature migration.

[0126] The communication bus mentioned in the above electronic device may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0127] The communication interface is used for communication between the above electronic device and other devices.

[0128] The memory may include a Random Access Memory (RAM), and may also include a Non-Volatile Memory (NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located far from the aforementioned processor.

[0129] The above-mentioned processor may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0130] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer, or a data storage device such as a server or data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)).

[0131] It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise", or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device that includes a series of elements includes not only those elements, but also other elements that are not explicitly listed, or elements that are inherent to such process, method, article, or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article, or device that includes the element.

[0132] Each embodiment in this specification is described in a related manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiment.

[0133] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-level information fusion nuclear power device health state prediction method, characterized in that, It includes the following steps: Step 1: Obtain fault data based on the long-term monitoring and regular test records of the key components of the nuclear power plant. Step 2: Using the orthogonal defect classification method, through feature extraction and feature selection of the fault data, map the multi-dimensional fault data to an orthogonal feature space, so that different types of faults have obvious separation degrees in this space, and classify different faults according to this separation degree and assign classification labels. Step 3: Use the subjective and objective fusion weighting method to assign weights related to the health status characteristics to different fault types classified in Step 2. By combining expert experience and data analysis methods, quantitatively evaluate the importance of different fault types to the health status characteristics of the nuclear power plant and assign weight labels. Step 4: Normalize the data labeled in the early stage, map all training set and test set data to the same scale to eliminate the influence brought by the differences in dimension and value range between different features. At the same time, set the width and step size of the sliding window to ensure the reliability and consistency of the data. Step 5: Statistically analyze the data preprocessed in Step 4, establish a fault knowledge graph, and build a health assessment model of graph deep learning based on this to realize the health assessment of the key equipment of the nuclear power plant.

2. The multi-level information fusion nuclear power device health state prediction method according to claim 1, characterized in that, The fault data obtained in Step 1 includes fault type, frequency and severity. The specific steps are as follows: Step 1.1: Collect the monitoring data during the long-term operation of the key components of the nuclear power plant, the test data of regular tests, the historical data stored in the background, and the simulation data simulated by experts. Step 1.2: Combine expert experience and operation and maintenance records to subjectively classify and label the collected data, and the data types include: fault type, fault frequency and fault severity.

3. The multi-level information fusion nuclear power device health status prediction method according to claim 1, characterized in that The orthogonal defect classification method in Step 2 classifies the obtained fault data. The specific steps are as follows: Step 2.1: Classify the fault types of the key components of the nuclear power plant obtained in Step 1 into m categories. There are n samples for each fault type, and the feature vector of the fault data is represented as x i , where i is the sample index; Step 2.2: By using the principal component analysis (PCA) method, perform an orthogonal transformation on the classified fault data to map the feature vector x i into the orthogonal feature space, obtaining the orthogonal feature vector y i ; Step 2.

3. For the j-th feature, calculate its variance ratio y in the k-th fault type jk , and the variance ratio is calculated by the following formula: Among them, represents the orthogonal eigenvalue of the i-th sample under the j-th feature for the k-th fault type. Var represents variance. By calculating the variance ratio of each feature among different fault types, the contribution degree of this feature to differentiating different fault types is evaluated. According to the magnitude of the variance ratio, a feature subspace with significant influence is selected, which is then used for the classification and identification of fault data.

4. The method for predicting the health state of a multi-level information fusion nuclear power device according to claim 1, characterized in that The subjective and objective fusion weighting method in Step 3 assigns weights to different fault types. The specific steps are as follows: Step 3.1: Use the subjective weighting method to assign weights α to different fault types j , where j represents the fault type. Summarize the influence degrees of different fault types on the health state through long-term operation and maintenance data and historical experience, and combine with expert knowledge to agree on the weights α of the subjective weighting method for different fault types j , or use the Analytic Hierarchy Process (AHP) to derive α j ; Step 3.2: Use the objective weighting method to assign weights β to different fault types j , where j represents the fault type. The entropy weight method is used for calculation. For nuclear power plants in different health states in historical data, collect the statistical data of their different fault types, and calculate the entropy value of each fault type. When the health state is different, the statistical data of various fault types will also be different. For different health states, if the difference in the statistical data of fault types is large, it means that its entropy value is large and it has a great impact on the health state. Therefore, a larger weight β is assigned j . At the same time, the TOPSIS method can also be used in the objective weighting method, or the two methods can be combined to calculate the objective fault weight β j ; Step 3.3: Combine the two weighting methods. Here, this weighting method is called the subjective and objective fusion weighting method, or the combined weighting method for short. The weight W of this method j is expressed as: Among them, j represents the fault type, α j is the weight of the subjective weighting method, and β j is the weight of the objective weighting method, i.e., the entropy weight method.

5. The method for predicting the health state of a multi-level information fusion nuclear power device according to claim 1, characterized in that, The data preprocessing operation based on normalization and sliding window in Step 4 is as follows: Step 4.1: Adopt linear function normalization, use the maximum and minimum values of each parameter variable in the data set to scale the original data proportionally to the interval [0,1]. The formula is as follows: Among them, x i is the original data of the valve component of the nuclear power production device, and x' i is the normalized data, is the maximum value in the original data of this parameter, is the minimum value in the original data of this parameter; Step 4.

2. On the basis of normalization, in order to make full use of all data, perform a sliding window operation on the data, and divide the data of the training set and the test set x1, x2, x3,..., x n into samples. Set the width and sliding step of the sliding window, and use the historical slice data to test the data of the valve parameters to be detected at the current moment. The expression is as follows: X(t) = {x(t - 1), x(t - 2), …, x(t - D)} Y(t) = y(t) Where X(t) is the data input into the model at time t, Y(t) is the value of the parameter to be detected at time t, x(t) is the data of this parameter at time t, D is the number of sliding window slices, that is, the sampling dimension, and x(t - 1) is the data of this parameter at the previous 1 time stamp after the sliding window operation.

6. The multi-level information fusion nuclear power device health status prediction method according to claim 1, wherein, Step 5: Build and train a health assessment model based on graph deep learning. The specific method is as follows: Step 5.1: Through statistical analysis of a large amount of data, find out the linear and non-linear correlations between different parameters of the valve components, and establish a fault knowledge graph to record the behavior patterns and characteristics of different valves under different fault conditions. Step 5.2: Identify the threat factors and transmission relationships of the valves in the nuclear power plant through correlation analysis and the fault knowledge graph. Step 5.3: Construct the adjacency matrix of the graph neural network based on the basic theory of graph theory to accurately reflect the correlation between valves, providing a basis for subsequent information fusion and health status assessment; Step 5.4: Through the extension of the credible adjacency matrix and in combination with multi-source boundary conditions, incorporate additional data sources and boundary conditions such as variable working loads and dynamic ambient temperatures into the adjacency matrix to improve the accuracy and reliability of the assessment. This includes data related to the operating status of the valve device obtained from flow sensors, pressure sensors, etc., to help detect the health status of the valves, and apply the theory of graph deep learning to design a multi-level information fusion health status assessment model; Step 5.5: Conduct multiple iterative trainings on the graph convolution model; In the k-th layer of graph convolution, the update formula for node features is: Among them, H (k) represents the node features of the k-th layer, is the normalized adjacency matrix, is the angular metric matrix, and W (k) is the weight matrix of the k-th layer. Each element in the weight matrix represents the connection strength between nodes, that is, it reflects the correlation degree between different valves. σ(·) represents the activation function; Step 5.6: Through multi-level graph convolution layers, obtain feature representations at different levels. By fusing the features of each layer, a health feature representation of multi-level information fusion can be obtained: H = Concat(H (1) , H (2) ,..., H (K) ) where H represents the health feature matrix after multi-level fusion, and K represents the number of graph convolution layers; Step 5.7: Use the fused health feature matrix to conduct health status assessment using methods such as classification and regression, and implement it through a fully connected layer and a softmax function: Y = softmax(HW) where Y represents the predicted result of the health status, and W is the weight matrix of the fully connected layer; Step 5.8: According to the usage scenarios and characteristics of valve components, divide the health status into 4 levels: "healthy", "sub-healthy", "near failure", and "failure".

7. A computer device, comprising a memory and a processor. When the computer program stored in the memory is executed by the processor, the processor is caused to execute the steps of the method according to any one of claims 1 to 6.

Images