A method for early warning of faults in intermediate range detectors of an off-core nuclear measurement system

Abstract

This invention relates to the field of fault early warning technology for intermediate-range detectors in off-core nuclear testing systems. It discloses a fault early warning method for intermediate-range detectors in off-core nuclear testing systems. The method involves acquiring channel current data from the intermediate-range detector, filling the data using linear interpolation, downsampling the original training and test data, identifying outliers in the intermediate-range channel current data and resetting them to null values, interpolating the null values ​​using linear interpolation, applying sliding window processing to the outlier-processed data, calculating the mutual information of each channel current in the intermediate range pairwise, identifying the fault time by judging changes in the magnitude of the mutual information, predicting the current signal of the intermediate-range detector from the fault initiation time, and constructing a confidence interval for fault early warning. This invention utilizes the temporal characteristics of intermediate-range channel current data based on mutual information for fault time identification.

Description

Technical Field

[0001] This invention relates to the field of fault early warning technology for intermediate range detectors in off-core nuclear testing systems, and particularly to a fault early warning method for intermediate range detectors in off-core nuclear testing systems. Background Technology

[0002] To ensure the safe and stable operation of nuclear power units and avoid potential risks, monitoring the operating status of nuclear power units is essential. The external nuclear monitoring system (EPS) is the foundation for reactor core monitoring and plays a crucial role in ensuring nuclear power unit safety. EPS detectors measure reactor core power, and their importance to both safe operation and the economic benefits of nuclear power plants is self-evident. Fault early warning can detect potential abnormal states before equipment fails completely, preventing equipment failures from developing into serious accidents. For fault early warning of intermediate-range detectors in EPS, it is necessary to identify the fault initiation time of the channel current data and interpret the channel current trend to provide relevant operational and maintenance decision-making information. The data-driven intermediate-range detector fault early warning method is a method that meets both the needs of safe nuclear power system operation and the economic benefits of nuclear power plants, and has significant engineering implications.

[0003] Currently, some scholars both domestically and internationally have implemented fault diagnosis for neutron detectors using various methods such as generalized likelihood ratio, principal component analysis, and the k-means algorithm. While some methods exist for detecting anomalies and faults in neutron detectors within off-core nuclear testing systems, they fail to provide early warnings for faults in intermediate-range detectors. Fault early warning is a predictive maintenance method aimed at taking measures before and during fault acquisition to prevent equipment damage and production interruptions. It involves analyzing and modeling historical equipment data to establish predictive models, forecasting future data, and providing early warnings for faults. Currently, research on fault early warning for intermediate-range detectors in off-core nuclear testing systems is limited both domestically and internationally. Summary of the Invention

[0004] The purpose of this invention is to provide a method for early warning of faults in intermediate range detectors of off-core nuclear measurement systems, providing a solution for early warning of latent faults in intermediate range detectors.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for early warning of faults in intermediate range detectors of an off-core nuclear measurement system includes the following steps:

[0007] Step 1: Acquire channel current data using the intermediate range detector of the external nuclear measurement system;

[0008] Step 2: Use linear interpolation to fill in the data to ensure consistent data sampling frequency;

[0009] Step 3: Data downsampling, downsampling the original training data and test data;

[0010] Step 4: Identify outliers in the intermediate range channel current data and reset them to null values;

[0011] Step 5: Use linear interpolation to interpolate the missing values ​​to ensure that the processed curve is highly similar to the original curve in terms of features;

[0012] Step 6: Perform sliding window processing on the outlier-handled data;

[0013] Step 7: Calculate the mutual information of each channel current in the intermediate range pairwise. The calculation formula is as follows:

[0014]

[0015] In the formula, I(X;Y) represents the mutual information between variables X and Y. When X and Y are completely independent, I(X;Y) is minimized to zero. p(x), p(y), and p(x,y) represent the marginal probability distribution functions of variables X, Y, and X and Y, respectively.

[0016] Step 8: Identify the fault time by judging the change in mutual information;

[0017] Step 9: Using the NerualProphet model, predict the current signal of the intermediate range detector starting from the fault initiation time;

[0018] Step 10: Construct confidence intervals and perform fault early warning.

[0019] In step 3, a sliding window is used to downsample the data.

[0020] Furthermore, a sliding window with a length of 3600 and a step size of 3600 is used to sample the original training data and test data.

[0021] In step 4, the constant fluctuation detector differential algorithm is used to identify outliers.

[0022] In step 6, the left pointer of the window is always fixed at the leftmost position of the array, and the right pointer of the window moves 24 bits to the right after each sliding window operation to obtain the time characteristics of the intermediate range channel current.

[0023] In step 7, using the formula Calculate information entropy.

[0024] In the formula, H(X) is the information entropy value, p(x) is the probability of taking the value x, and b is usually taken as the natural constant e as the base of the logarithm.

[0025] In step 9, the Neural Prophet model consists of the Prophet model and the AR-Net model, a p-order autoregressive model.

[0026] The Prophet model is:

[0027] P(t)=g(t)+s(t)+h(t)+ε t ,

[0028]

[0029] In the formula, g(t) represents the trend term; C(t) represents the carrying capacity, which is a function that changes with time and limits the maximum value of the curve fitting; s(t) represents the seasonal term; h(t) represents the holiday term; κ ~ Normal(0,v) 2 ) follows a normal distribution, κ i The influence of the event on the model is represented by i, which represents a holiday or other event, and D. i ε represents the variable of the i-th independent model; t The remaining terms represent the Gaussian distribution and indicate the fluctuations encountered during model fitting;

[0030] The p-order autoregressive model of the AR-Net model is:

[0031]

[0032] In the formula, y t-1 ,…,y t-p w represents the p-th order lag term. i The p-th order weights, also known as autoregressive coefficients, are e t represents noise, and c represents a constant term;

[0033] The MSE calculation formula is as follows, which is an evaluation metric used to calculate the loss function:

[0034]

[0035] To introduce sparsity into the autoregressive coefficients, a regularization term R is added to the loss function, and the formula for calculating the loss function is:

[0036]

[0037] In the formula, s represents the sparsity of the autoregressive coefficients, and c λ c1 and c2 represent the degree of regularization, and c1 and c2 represent the regularization parameters.

[0038] In step 10, a confidence interval is constructed using six sigma (similar to the 3 sigma function, but with the internal parameter 3 set to 6 to construct a larger confidence interval and avoid false alarms). The confidence interval is constructed by using the predicted values ​​of the maximum channel current and the predicted values ​​of the fault channel current to infer the time when the intermediate range channel current ratio is less than 1.9 times.

[0039] Compared with existing technologies, the fault early warning method for intermediate range detectors in off-core nuclear measurement systems provided by this invention has the following advantages:

[0040] The data cleaning process provided by this invention can effectively solve the problem of intermediate range channel current data and improve data quality.

[0041] This invention identifies fault time based on mutual information, effectively utilizing the time characteristics of intermediate range channel current data.

[0042] Furthermore, the NerualProphet model provided by this invention introduces functions such as AR-Net to construct time series autocorrelation, gradient descent optimization, and feedforward neural network on the basis of the Prophet model, which effectively improves the prediction accuracy of the model. Attached Figure Description

[0043] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the technical description will be briefly introduced below.

[0044] Figure 1 The flowchart shows the fault early warning method for the intermediate range detector of the off-core nuclear measurement system provided by the present invention. Detailed Implementation

[0045] The following detailed description provides further details on specific implementation methods.

[0046] like Figure 1 As shown, the present invention provides a method for early warning of faults in intermediate range detectors of an off-core nuclear measurement system, comprising the following steps:

[0047] Step 1: Obtain channel current data through the intermediate range detector of the off-core nuclear measurement system.

[0048] Step 2: To address the issue of inconsistent data sampling frequency, linear interpolation is used to fill in the data, ensuring a consistent data sampling frequency.

[0049] Step 3: Data downsampling. This involves shifting the original frequency of the data to another frequency, such as from a high frequency to a low frequency (e.g., from one sampling point per minute to one sampling point every ten minutes). To address data redundancy, a sliding window is used for downsampling. A sliding window with a length of 3600 and a step size of 3600 is used to sample both the original training and test data.

[0050] Step 4: To address the issues of large data fluctuations and data drift, a constant fluctuation detector differential algorithm is used to identify outliers in the intermediate range channel current data and reset them to null values.

[0051] Step 5: Use linear interpolation to interpolate the null values ​​to ensure that the processed curve is highly similar to the original curve in terms of features.

[0052] Step 6: Perform sliding window processing on the outlier-handled data. The left pointer of the window is always fixed at the leftmost position of the array, and the right pointer of the window moves 24 bits to the right after each sliding window operation to obtain the time characteristics of the intermediate range channel current.

[0053] Step 7: The detectors in each channel of the intermediate range are uniformly distributed and have spatial correlation. Calculate the mutual information of the current in each channel of the intermediate range pairwise.

[0054] In information theory, information entropy is a basic unit of information, used to measure the degree of disorder among variables. The higher the information entropy, the greater the disorder of the variables, and the more information they contain; conversely, the lower the information entropy, the less disorder the variables, and the more limited the information they contain. For any variable, the formula for calculating information entropy is:

[0055]

[0056] In the formula: H(X) is the information entropy value, p(x) is the probability of taking the value x, and b is usually taken as the natural constant e as the base of the logarithm.

[0057] Mutual information (MI) in information theory represents the correlation between two random variables X and Y, that is, it measures the degree of dependence between the two variables. It can be viewed as the amount of information that one random variable X contains about another random variable Y; that is, the stronger the correlation between variables X and Y, the greater the mutual information value, and vice versa. The formula for calculating the mutual information of two random variables X and Y is:

[0058]

[0059] In the formula: I(X;Y) represents the mutual information between variables X and Y. When X and Y are completely independent, I(X;Y) is minimized to zero. p(x), p(y), and p(x,y) represent the marginal probability distribution function of variables X, the marginal probability distribution function of Y, and the joint probability distribution function of X and Y, respectively.

[0060] Step 8: Identify the fault time by judging the change in mutual information. If the mutual information value is increasing and then continues to decrease after a certain point in time, the detector has failed; otherwise, the detector is not faulty.

[0061] Step 9: Predict from the fault initiation time using the NerualProphet model.

[0062] It should be noted that the Neural Prophet model is a decomposable time series model developed based on PyTorch. It retains the main functionalities of the Prophet open-sourced by the Facebook team in 2017, and introduces features such as AR-Net for constructing time series autocorrelation, gradient descent optimization, and feedforward neural networks. The Neural Prophet model consists of Prophet and AR-Net. Prophet is an additive combination, composed of a trend term, a seasonal term, and a holiday term, with the basic formula as follows:

[0063] P(t)=g(t)+s(t)+h(t)+ε t

[0064] g(t) represents the trend term, which indicates the non-periodic trend of the time series. It can be fitted to a non-linear curve by adjusting the smoothness parameter, and the calculation formula is as follows:

[0065]

[0066] In the formula: C(t) represents the carrying capacity, which is a function that varies with time and limits the maximum value of the curve fit. k(t) represents the growth rate, which can be changed by the algorithm or manually specified to make the model change when facing certain specific moments or potential cycles. m(t) represents the offset, which will be adjusted when the growth rate k(t) changes.

[0067] s(t) represents the seasonal term, or periodic term, usually expressed in weeks or years. Its periodicity is approximated by a Fourier series, and the formula is as follows:

[0068]

[0069] In the formula: N represents the number of periodic types used in the model. A larger N value can fit a more complex seasonal curve, but if it is too large, it is easy to overfit. T S It represents a fixed period.

[0070] h(t) represents the holiday term, indicating the impact of non-fixed-period holidays or major events on the curve in the time series, such as substation maintenance and repair, typhoon weather, etc. Since different events have varying degrees of impact on the curve, the holiday model can treat the impact of each event at different points in time as multiple independent models, and can set the range of time periods before and after the impact of different events. The calculation formula is as follows:

[0071]

[0072] In the formula: κ~Normal(0,v) 2 ) follows a normal distribution, κ i The influence of the event on the model is represented by i, which represents a holiday or other event, and D. i Let D represent the variable of the i-th independent model. If the time of event i belongs to this variable, then D i The value is 1 if it is 1, otherwise it is 0.

[0073] ε t The remainder term represents the Gaussian distribution and indicates the fluctuations encountered during model fitting.

[0074] In addition, Neural Prophet introduces the AR-Net model. The first layer of the neural network is a p-order autoregressive structure. To improve curve fitting accuracy, a certain number of neurons are added to the second layer, and the mean squared error (MSE) is defined as the loss function. The basic formula for the p-order autoregressive model is:

[0075]

[0076] In the formula: y t-1 ,…,y t-p w represents the p-th order lag term. i The p-th order weights, also known as autoregressive coefficients, are e t represents noise, and c represents a constant term.

[0077] The formula for calculating MSE is:

[0078]

[0079] To introduce sparsity into the autoregressive coefficients, a regularization term R is added to the loss function, and the loss function calculation formula is changed to:

[0080]

[0081] In the formula: s is the sparsity of the autoregressive coefficients, c λ The degree of regularization depends on the noise of the data. c1 and c2 are regularization parameters that depend on the range of the autoregressive coefficients. When the autoregressive coefficients are normalized to [0,1], then c1≈c2≈3.

[0082] Step 10: Construct a confidence interval using Six Sigma (similar to the 3sigema function, but with the internal parameter 3 set to 6 to create a larger confidence interval and avoid false alarms). By constructing the confidence interval based on the predicted maximum channel current and the predicted fault channel current, infer the time when the intermediate range channel current ratio is less than 1.9 times.

[0083] This invention utilizes machine learning algorithms to mine the spatiotemporal characteristics of the intermediate-range detector channel current data in the off-core nuclear testing system, enabling fault early warning for the intermediate-range detectors. Due to issues such as inconsistent sampling frequency, high data redundancy, data fluctuations, and data drift in the intermediate-range channel current data of the off-core nuclear testing system, it is difficult to distinguish between faulty and normal data. This invention proposes an intermediate-range channel current data cleaning process to improve data quality.

[0084] In fault early warning systems, determining the fault initiation time allows for timely intervention to prevent further escalation and facilitates better analysis of the fault's causes and mechanisms. However, incorrect fault initiation time identification can lead to misjudging normal operation as abnormal, resulting in false alarms and wasted resources. Therefore, this invention proposes a precise and rapid method for identifying the fault time of intermediate-range channel current in an off-core nuclear testing system. Based on this method, the invention uses a Neuralprophet model to predict the fault current and infer the time when the intermediate-range channel current ratio of the off-core nuclear testing system is less than 1.9 times, thus providing fault early warning for the intermediate-range detector.

[0085] This invention addresses the problems of inconsistent sampling frequency, large data redundancy, data fluctuation, and data drift in intermediate range channel current data by proposing a data cleaning process that can effectively improve the quality of intermediate range channel current data.

[0086] This invention utilizes mutual information-based fault time identification, leveraging the spatiotemporal characteristics of intermediate-range channel current data to accurately and rapidly identify fault times. Based on this, the NeuralProphet model is used to predict channel currents, construct confidence intervals, and infer the time when the intermediate-range channel current ratio is less than 1.9 times, thus providing early warning of intermediate-range detector faults.

[0087] The above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. An ex-core nuclear instrumentation system intermediate range detector failure warning method characterized by, The method comprises the following steps: Step 1: Obtain channel current data through the out-of-core nuclear measurement system intermediate range detector; Step 2: Fill the data using linear interpolation to ensure consistent data sampling frequency; Step 3: Downsample the data, and downsample the original training data and test data; Step 4: Identify outliers in the intermediate range channel current data and reset the outliers to null values; Step 5: Interpolate the null values using linear interpolation to ensure that the processed curve is highly similar to the original curve characteristics; Step 6: Perform sliding window processing on the data after outlier processing; Step 7: Calculate mutual information between each pair of intermediate range channel currents, and the calculation formula is as follows: In the formula, I(X;Y) represents the mutual information of variables X and Y, and when X and Y are completely independent, I(X;Y) is minimum, which is zero, p(x), p(y) and p(x,y) represent the edge probability distribution function of variable X, the edge probability distribution function of variable Y and the joint probability distribution function of X and Y, respectively; Step 8: Identify the fault time by judging the size change of mutual information; Step 9: Predict the current signal of the intermediate range detector from the start time of the fault through the NeuralProphet model; Step 10: Construct a confidence interval to perform fault warning.

2. The method of claim 1, wherein the method further comprises: In step 3, the data is downsampled using a sliding window.

3. The method of claim 2, wherein the method further comprises: The original training data and test data are sampled using a sliding window with a length of 3600 and a step of 3600.

4. The method of claim 1, wherein the method further comprises: In step 4, the constant fluctuation detector difference algorithm is used to identify outliers.

5. The method of claim 1, wherein the method further comprises: In step 6, the left pointer of the window is always fixed at the leftmost side of the array, and the right pointer of the window is moved 24 bits to the right after each execution of the sliding window to obtain the characteristics of the intermediate range channel current in time.

6. The method of claim 1, wherein the method further comprises: In step 7, the information entropy is calculated by the formula H = -∑p(x)log2p(x) In the formula, H(X) is the information entropy value, p(x) is the probability of taking value x, and b is usually taken as the base of the logarithm, which is the natural constant e.

7. The method of claim 1, wherein the method further comprises: In step 9, the Neural Prophet model is composed of a Prophet model and a p-order autoregressive model of an AR-Net model, The Prophet model is: P(t) = g(t) + s(t) + h(t) + ε t , In the formula, g(t) represents the trend term; C(t) represents the carrying capacity, which is a function that changes with time and limits the maximum value of the curve fitting; s(t) represents the seasonal term; h(t) represents the holiday term; κ ~ Normal(0,v) 2 ) follows a normal distribution, κ i The influence of the event on the model is represented by i, which represents a holiday or other event, and D. i ε represents the variable of the i-th independent model; t The remaining terms represent the Gaussian distribution and indicate the fluctuations encountered during model fitting; The p-order autoregressive model of the AR-Net model is: where y t-1 ,…,y t-p represent p-order lag terms, w i is a p-order weight or autoregressive coefficient, e t represents noise, and c represents a constant term. The MSE calculation formula is: In order to introduce sparsity to the autoregressive coefficients, a regularization term R is added to the loss function, and the loss function calculation formula is: where s represents the sparsity of the auto-regressive coefficients, c λ represents the regularization degree, and c1, c2 represent the regularization parameters.

8. The method of claim 1, wherein the method further comprises: In step 10, a six-sigma confidence interval is constructed, a confidence interval is constructed based on the maximum channel current prediction value and the fault channel current prediction value, and the time when the intermediate range channel current ratio is less than 1.9 times is inferred.

Images