A reliability analysis system and method for in-pile components based on data mining
Through the reliability analysis system of internal components based on data mining, supervised learning and deep learning technology, the problem of fault prediction caused by the dependency theory of internal components in the existing technology is solved, and the accuracy of internal components is achieved. Reliability assessment and forecasting improves safety and economicality.
Patent Information
- Application Number
- CN202011211049.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-11-03
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2040-11-03
AI Technical Summary
The prior art relies on deterministic methods when designing components in the stack, resulting in the inability to effectively estimate the failure and its risk probability under complex operating conditions, unable to meet actual engineering needs, and the diversity of parts makes it difficult to classify reliability evaluations.
Adopt a reliability analysis system for internal components based on data mining, including data preprocessing, reliability evaluation index establishment, feature construction, dimension analysis and predictive analysis modules, and use supervised learning, automatic encoder and long-term memory recurrent neural network to Cross-entropy loss function and Adam optimizer build models to evaluate and predict the reliability of components within the heap.
It realizes accurate monitoring and safety assessment of the health status of components in the stack, has high recognition accuracy and generalization capabilities, low prediction errors, and design results not only meet the overall requirements but also improve economicality.
Smart Images

Figure CN112380763B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to state recognition and fault diagnosis of in-core components, and in particular to a system and method for analyzing the reliability of in-core components based on data mining. Background Art
[0002] The safety of a pressurized water reactor nuclear power plant mainly depends on the safety of the primary circuit. According to nuclear safety regulations, all mechanical equipment and pipelines of the primary circuit belong to nuclear safety level 1 and earthquake resistance category 1. One of the main equipment of the primary circuit is the reactor pressure vessel. The components in the container are called internal components. The internal components mainly include the lower core support components, the upper core support components, and the internal measurement devices. When the reactor is in operation, the components in the reactor are in a harsh environment such as high temperature, high pressure, corrosion, strong radiation, high heat release rate, and multiple load impacts. There are many reasons for the damage of the main equipment and piping system of the primary circuit, such as unreasonable thermal hydraulic design, mechanical design, manufacturing, material selection, welding, installation, etc., as well as reactive fluctuations and earthquakes. Once the internal components of the reactor vibrate, the degree of rupture will be further expanded. Therefore, preventing the vibration of the internal components of the reactor is one of the important links to ensure the safe and reliable operation of nuclear power plants.
[0003] At present, the design of reactor internals mostly relies on deterministic methods, which often leads to overly conservative designs in some aspects, while existing theories do not fully consider other aspects, and cannot predict the possible failures and risk probabilities under complex operating conditions, resulting in the fact that reactor internals still do not meet the needs of actual projects. At the same time, due to the large number of parts and failure forms of reactor internals, targeted classification and analysis are required to determine the corresponding reliability evaluation indicators and methods. Summary of the invention
[0004] In order to better identify and diagnose the health status of in-core components, the present invention provides a system and method for analyzing the reliability of in-core components based on data mining. The specific technical solution is as follows:
[0005] A reliability analysis system for in-pile components based on data mining, the system comprises: a data preprocessing module, an in-pile component reliability evaluation index establishment module, a reliability evaluation feature construction module, a reliability evaluation dimension analysis module, and a reliability prediction analysis module, wherein:
[0006] The data preprocessing module performs outlier processing, null value processing, discretization processing and normalization processing on the data collected from the in-pile components;
[0007] The module for establishing the reliability evaluation index of the reactor internal components uses supervised learning to establish evaluation indexes in the modeling process to measure the quality of the model, and obtains two modeling indexes for evaluating the reliability of the reactor internal components;
[0008] The reliability evaluation feature construction module performs dimensionality reduction processing on the collected data based on the determined reliability evaluation indicators of the in-core components, and constructs the business features of the reliability evaluation;
[0009] The reliability evaluation dimension analysis module reconstructs the current sample into a sample set with time series characteristics based on the time series characteristics of the operating status data of the in-core components;
[0010] The reliability prediction and analysis module establishes a model framework through the long short-term memory recurrent neural network prediction method; the model training process is based on the prediction purpose, selects the cross entropy loss function and Adam optimizer, finds the algorithm model parameters, and predicts the reliability of the components inside the pile.
[0011] Furthermore, in the data preprocessing module, data outlier processing and null value processing are achieved through data screening and data cleaning. Discretization processing and normalization processing are achieved through feature simplification and data transformation.
[0012] Furthermore, in the module for establishing the reliability evaluation index of the in-pile components, the two modeling indexes include the ROC curve area index and the F-value series index.
[0013] Furthermore, the ROC curve uses area as an indicator to evaluate the overall performance of the model under all classification thresholds, and describes the relationship curve between the true positive rate (TPR) and the false positive rate (FPR), wherein TPR and FPR are obtained based on the calculation of the binary classification confusion matrix.
[0014] Furthermore, the F-value series of indicators evaluates the prediction results of a specific classification threshold model, including three indicators: accuracy, precision, and recall. Accuracy represents the ratio of the number of samples correctly predicted by the classifier to the total number of samples in a given test data set. Its calculation formula is shown in (1):
[0015]
[0016] Among them, TP represents the number of samples whose actual value and predicted value are both positive examples; FP represents the number of samples whose actual value is negative but predicted as positive examples; FN represents the number of samples whose actual value is positive but predicted as negative examples; TN represents the number of samples whose actual value and predicted value are both negative examples.
[0017] Precision refers to the proportion of instances predicted to be positive that are actually positive. Its calculation formula is shown in (2):
[0018]
[0019] Recall focuses on the prediction accuracy of a single category and measures how many positive examples are predicted as positive examples. Its calculation formula is shown in (3):
[0020]
[0021] Furthermore, the dimensionality reduction in the reliability evaluation feature construction module adopts the method of autoencoder, which is a commonly used unsupervised learning model in deep learning. The error comes from the comparison between the reconstructed data and the original input data. The input signal is reproduced as much as possible through the unsupervised learning method.
[0022] Furthermore, the reliability evaluation dimension analysis module is a sample reconstruction method based on multi-dimensional time series prediction. From the perspectives of the data set and the label set, the collected time point data set is reconstructed into a time series sample, and the current state label set is reconstructed into a predicted state label set. The specific implementation is shown in Formulas (4) to (5).
[0023]
[0024] Among them, the matrix A represents the original labeled sample set, a i represents the i-th sample, c i represents the label corresponding to the i-th sample, m represents the number of dimensions, and n represents the number of sample points.
[0025]
[0026] Among them, the matrix B represents the labeled sample set reconstructed based on the time series characteristics. The parameters i and j are the input data time series length and the predicted state time point.
[0027] Furthermore, the long short-term memory recurrent neural network (LSTM) prediction method in the reliability prediction and analysis module introduces a memory storage module - cell state (Cell State) in the hidden layer, and adds an input control gate (input), a forget control gate (forget), and an output control gate (output) to determine the information to be memorized in the cell state, thereby realizing a long-term memory function.
[0028] A reliability analysis method for in-pile components based on data mining, the steps are as follows:
[0029] 1) Initially set the sample ratio and cost-sensitive learning parameters, batch size, and batches based on experience;
[0030] 2) Given the value range and value granularity of the time series length parameter i and the prediction state time interval parameter j, determine the values of i and j, and generate m groups of data using a cross-combination method;
[0031] 3) Use the combined values of m groups i and j to reconstruct samples based on multidimensional time series, and obtain m groups of sample sets with different input data time series lengths and prediction state time intervals. At the same time, in order to ensure the comparability of the modeling results of the m groups of data, the consistency of different sample sets at the prediction time points must be ensured;
[0032] 4) Model and predict the m groups of sample sets in turn, and compare and evaluate the application effects of different models on the test set. The combination of i and j with the best application effect is the parameter value of i and j;
[0033] 5) Based on the optimal combination of i and j, batch p and batch n are initially set, the value range of sample proportion and cost-sensitive learning parameters are given, grid search is used for optimization, and the optimal combination of sample proportion and cost-sensitive learning parameters is determined by the application effect of the model on the test set;
[0034] 6) Based on the optimal combination of i and j and the optimal combination of given sample proportion and cost-sensitive learning parameters, a grid search is used to find the optimal value of the given batch p and batch n, and the optimal combination of p and n is determined by the application effect of the model on the test set;
[0035] 7) Determine the optimal weighting method and number of data subsets based on the weighted results of the application effects of different numbers of data subset models on the test set.
[0036] The beneficial effects of the present invention are mainly manifested in that the healthy operating status of the components in the pile can be monitored based on an accurate model, the safety level and economic rationality of the components in the pile can be comprehensively evaluated based on the design method of data mining, the probability of failure is calculated by probabilistic design, and the reliability probability that must be possessed is reasonably allocated to each component (part) during the probabilistic design process. The present invention has high recognition accuracy and generalization ability, low prediction error, good prediction effect, and the design result not only meets the overall requirements but also improves the economy. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 Schematic diagram of the structure of the in-pile component reliability analysis system based on data mining.
[0038] Figure 2 This is a diagram of the preprocessing process of the in-pile component data.
[0039] Figure 3 This is a statistical chart of the accuracy results of the vibration index prediction model for pile internal components applied to the test set.
[0040] Figure 4 Statistical chart of abnormal recognition rate and abnormal misjudgment rate when the vibration index prediction model of pile internal components is applied to the same test set.
[0041] Figure 5 The anomaly recognition rate and misjudgment rate of the (p, n) test set application of the in-pile component reliability prediction model. DETAILED DESCRIPTION
[0042] The present invention can be more fully described with reference to the accompanying drawings, which show certain embodiments of the present invention, but not all embodiments. In fact, the present invention can be embodied in many different forms and should not be considered to be limited to the embodiments described here, but should be considered to be provided for the purpose of satisfying the applicable legal requirements of the disclosure of the present invention. The present invention is described in detail below in conjunction with the drawings and specific implementations of the specification.
[0043] Figure 1 The functions of each module of the in-pile component reliability analysis system based on data mining and the logical relationship between the modules are listed.
[0044] The data preprocessing module processes the data collected from the in-pile components by performing outlier processing, null value processing, discretization processing, and normalization processing, thus preparing for data mining and analysis.
[0045] The module for establishing reliability evaluation indexes of in-core components uses a supervised learning method to establish evaluation indexes in the modeling process to measure the quality of the model, and obtains two modeling indexes for evaluating the reliability of in-core components.
[0046] The reliability evaluation feature construction module performs dimensionality reduction processing on the collected data based on the determined reliability evaluation indicators of the in-core components, and constructs the business features of the reliability evaluation;
[0047] The reliability evaluation dimension analysis module reconstructs the current sample into a sample set with time series characteristics based on the time series characteristics of the operating status data of the in-core components;
[0048] The reliability prediction and analysis module establishes a model framework through the long short-term memory recurrent neural network (LSTM) prediction method; the model training process is based on the prediction purpose, selects the cross entropy loss function and the Adam optimizer, finds the algorithm model parameters, and predicts the reliability of the components in the pile. The cross entropy loss function completes the error calculation by measuring the similarity between the distribution of the predicted value and the distribution of the true value; the Adam optimizer is a gradient-based optimization algorithm with the characteristics of fast convergence speed, efficient calculation, and low memory usage.
[0049] like Figure 2The figure shows the preprocessing process of the in-pile component data. Data quality has a great impact on the analysis results of data mining. The acquired raw data of the in-pile components contains a large number of attributes, with erroneous values and outliers. It is necessary to conduct preliminary screening of the data, remove erroneous values and outliers, ensure the accuracy of the data, remove attributes irrelevant to mining, and ensure the diversity of sample data and the completeness of feature information. In addition, the data needs to be processed according to the algorithm requirements so that the data meets the input requirements of the algorithm.
[0050] like Figure 3 The figure shows the statistical chart of the accuracy results of the prediction model for the vibration index of the internal components of the pile applied to the test set. Through the statistics of the accuracy of the test sample set, we can roughly analyze the impact of different i and j values on the results. From the model accuracy comparison chart with different j values on the right, it is found that the larger the j value, the lower the model accuracy under different i values. From the model accuracy comparison chart with different i values at the bottom, it is found that when j is 4s, 6s, and 8s, the model accuracy is high when the i value is larger. The values of i and j need to be further determined based on the abnormal recognition rate and abnormal misjudgment rate.
[0051] like Figure 4 The figure shows the abnormal recognition rate and abnormal misjudgment rate of the in-pile component vibration index prediction model applied to the same test set. The left figure shows the abnormal recognition rate, and the right figure shows the abnormal misjudgment rate. Based on business needs, the model works well when the abnormal recognition rate is high and the abnormal misjudgment rate is low. Figure 4 , according to the legend, the darker i and j values are the desired values.
[0052] Figure 5 The abnormal recognition rate and misjudgment rate of the test set of the reliability prediction model (p, n) for the in-pile components. The prediction algorithm model parameters refer to batch p and batch n. Based on the principle of batch training, the batch p and batch n of different sample sets are different. Therefore, it is necessary to find the optimal p and n combination for each training set of group a.
[0053] Set the value of a to 5, and select 5 sets of training samples that meet the parameter values of i, j, k' (sample proportion, the ratio of abnormal class to normal class), c (cost-sensitive learning parameter). Set the value of p to 50, 100, 200, 500, 1000, 2000, and set n to an arithmetic progression with a minimum value of 200, a maximum value of 3000, and a difference of 200, with a total of 90 combinations. Complete 90 sets of model training for each of the 5 sets of samples, and find the parameter values of batch p and batch n that have the best test set application effect in the 90 sets of models.
[0054] Based on the application results of 90 different prediction models in the test set, the relationship between the abnormal recognition rate and the abnormal misjudgment rate of the sample set is plotted as follows: Figure 5As shown in the figure, the p and n values of some data points are marked in the form of "p_n", and a four-quadrant graph is drawn based on a recognition rate of 75% and a misjudgment rate of 30%. In the lower right quadrant, the model with a smaller batch size has better overall performance.
[0055] exist Figure 5 In the two marked models, the model parameters are "200_2800" and "100_600", that is, the model performance is better when the batch size is 200 and the training is 2800 times, and the model performance is better when the batch size is 100 and the training is 600 times. A set of results is randomly selected as the parameter values of batch p and batch n in the prediction model of group 1 sample set. The same method is used to determine the parameter values of batch p and batch n for the other 4 groups of sample sets, and the results are shown in Table 1.
[0056] Table 1 Batch p and batch n parameter values of the shell vibration prediction model for other data subsets and the application results of the test set
[0057] Group p-value n value Abnormal recognition rate Abnormal misjudgment rate Group 2 500 1000 79.2% 22.5% Group 3 200 600 79.2% 24.3% Group 4 100 1000 80.0% 23.0% Group 5 200 1600 80.8% 25.0%
[0058] from Figure 5 It can be seen from Table 1 that the prediction error is low, the predicted value can basically reflect the changing trend of the numerical value, and the prediction effect of the model is good.
[0059] The above is only a preferred embodiment of the present invention. Although the present invention has been disclosed as a preferred embodiment, it is not intended to limit the present invention. Any technician familiar with the art can make many possible changes and modifications to the technical solution of the present invention by using the above disclosed methods and technical contents without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment of equivalent changes. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.
Claims
1. A reliability analysis system for in-pile components based on data mining, characterized in that: The system includes: data preprocessing module, in-core component reliability evaluation index establishment module, reliability evaluation feature construction module, reliability evaluation dimension analysis module, reliability prediction and analysis module, among which: The data preprocessing module performs outlier processing, null value processing, discretization processing and normalization processing on the data collected from the in-pile components; among them, data outlier processing and null value processing are realized through data screening and data cleaning; discretization processing and normalization processing are realized by feature simplification and data transformation; The module for establishing the reliability evaluation index of the in-core components uses supervised learning to establish evaluation indexes in the modeling process to measure the quality of the model, and obtains two modeling indexes for evaluating the reliability of the in-core components, including the ROC curve area as an index and the F-value series of indicators; the ROC curve uses the area as an indicator to evaluate the overall performance of the model under all classification thresholds, and describes the relationship curve between the positive rate TPR and the negative rate FPR, where TPR and FPR are obtained based on the calculation of the binary classification confusion matrix; the F-value series of indicators evaluates the prediction results of a specific classification threshold model, including three indicators: accuracy, precision, and recall; The accuracy rate refers to the ratio of the number of samples correctly predicted by the classifier to the total number of samples in a given test data set. The calculation formula is shown in (1): Among them, TP represents the number of samples whose actual value and predicted value are both positive examples; FP represents the number of samples whose actual value is negative but predicted as positive examples; FN represents the number of samples whose actual value is positive but predicted as negative examples; TN represents the number of samples whose actual value and predicted value are both negative examples; The precision rate indicates the proportion of instances predicted to be positive that are actually positive, and its calculation formula is shown in (2): The recall rate focuses on the prediction accuracy of a single category, measuring how many positive examples are predicted as positive examples. Its calculation formula is shown in (3): The reliability evaluation feature construction module performs dimensionality reduction processing on the collected data based on the determined reliability evaluation indicators of the in-core components, and constructs the business features of the reliability evaluation. The dimensionality reduction adopts the method of autoencoder, which is an unsupervised learning model in deep learning. The error comes from the comparison between the reconstructed data and the original input data. The input signal is reproduced as much as possible through the unsupervised learning method. The reliability evaluation dimension analysis module reconstructs the current sample into a sample set with time series characteristics based on the time series characteristics of the operating status data of the in-core components; specifically, based on the sample reconstruction method of multi-dimensional time series prediction, the collected time point data set is reconstructed into a time series sample from the perspective of the data set and the label set, and the current state label set is reconstructed into a predicted state label set. The specific implementation is shown in equations (4) to (5); Among them, the matrix A represents the original labeled sample set, a i represents the i-th sample, c i represents the label corresponding to the i-th sample, m represents the number of dimensions, and n represents the number of sample points; Among them, the matrix B represents the labeled sample set reconstructed based on the time series characteristics; the parameters i and j are the input data time series length and the predicted state time point; The reliability prediction and analysis module establishes a model framework through the long short-term memory recurrent neural network prediction method; the model training process is based on the prediction purpose, selects the cross entropy loss function and the Adam optimizer, finds the algorithm model parameters, and predicts the reliability of the components in the pile; the long short-term memory recurrent neural network prediction method introduces a memory storage module - unit state in the hidden layer, and adds an input control gate, a forget control gate, and an output control gate to determine the information to be remembered in the unit state, thereby realizing a long-term memory function; The method for analyzing the reliability of a pile internal component implemented based on the pile internal component reliability analysis system comprises the following steps: 1) Initially set the sample ratio and cost-sensitive learning parameters, batch size, and batches based on experience; 2) Given the value range and value granularity of the time series length parameter i and the prediction state time interval parameter j, determine the values of i and j, and generate m groups of data using a cross-combination method; 3) Use the combined values of m groups i and j in turn to reconstruct samples based on multidimensional time series, and obtain m groups of sample sets with different input data time series lengths and prediction state time intervals; at the same time, in order to ensure the comparability of the modeling results of the m groups of data, it is necessary to ensure the consistency of different sample sets at the prediction time points; 4) Model and predict the m groups of sample sets in turn, and compare and evaluate the application effects of different models on the test set. The combination of i and j with the best application effect is the parameter value of i and j; 5) Based on the optimal combination of i and j, batch p and batch n are initially set, the value range of sample proportion and cost-sensitive learning parameters are given, grid search is used for optimization, and the optimal combination of sample proportion and cost-sensitive learning parameters is determined by the application effect of the model on the test set; 6) Based on the optimal combination of i and j and the optimal combination of given sample proportion and cost-sensitive learning parameters, a grid search is used to find the optimal value of the given batch p and batch n, and the optimal combination of p and n is determined by the application effect of the model on the test set; 7) Determine the optimal weighting method and number of data subsets based on the weighted results of the application effects of different numbers of data subset models on the test set.
