A method for detecting shaft displacement failure of a nuclear reactor main circulation pump

Abstract

The present application belongs to the technical field of fault detection, and particularly relates to a nuclear reactor main circulating pump shaft displacement fault detection method. The method comprises the following steps: first, obtaining main pump shaft displacement data through a monitoring device; second, performing data sliding window sampling; third, performing data enhancement, and adding padding data with a length of m on both sides of the sliding window; fourth, inputting the data set into a spectral residual network, performing saliency research, and extracting data spectral residual values; fifth, inputting the spectral residual data into a DAGMM network, and performing anomaly detection; and sixth, outputting the anomaly detection result through data analysis and comparison. The beneficial effects are as follows: (1) the spectral residual network technology means realizes saliency research on the main pump shaft displacement data, successfully distinguishes normal data from abnormal data, improves the accuracy of anomaly detection, and avoids misjudgment in the anomaly detection process.

Description

Technical Field

[0001] This invention belongs to the field of fault detection technology, specifically relating to a method for detecting displacement faults in the main circulation pump shaft of a nuclear reactor. Background Technology

[0002] The main circulating pump for nuclear reactor coolant, also known as the main pump, is located between the reactor and the steam generator in the primary coolant loop. It serves as the pressure boundary and a key component of the reactor coolant system, and is the only high-speed rotating device in the primary loop. Situated at the heart of the nuclear island, the main pump pumps hot water into the evaporator for heat conversion. It is crucial for controlling the water circulation in nuclear power plant operations, and its stable performance directly impacts the plant's profitability. Measuring the axial displacement of the main pump is an important method for evaluating its vibration and assessing its long-term safe operation.

[0003] Current techniques for detecting anomalies in main pump shaft displacement rely solely on historical spectrum analysis. These methods primarily analyze online shaft displacement measurements by comparing the spectra of local temporary auxiliary measurements with those of online data. This approach is too simplistic for detecting main pump displacement anomalies, resulting in low detection efficiency and a high risk of misdiagnosis.

[0004] Furthermore, the main pump shaft displacement data is affected by various factors, such as pressure changes in the fluid inside the nuclear reactor and the instability of liquid flow, resulting in significant fluctuations in the dataset. These fluctuations make it difficult to distinguish between anomalous and normal data, and traditional anomaly detection algorithms cannot effectively handle such data, leading to misjudgments of anomaly detection results. In addition, the main pump shaft displacement dataset is usually very large, and traditional anomaly detection algorithms may be limited by computational and storage constraints, making it impossible to effectively handle large-scale datasets. Summary of the Invention

[0005] The purpose of this invention is to provide a method for detecting shaft displacement faults in the main circulation pump of a nuclear reactor. By applying artificial intelligence algorithms to deeply mine historical shaft displacement data of the main pump, the method can achieve timely and accurate identification of abnormal fluctuations in the shaft displacement of the main pump, thereby improving the operating efficiency and safety of the nuclear reactor.

[0006] The technical solution of the present invention is as follows: A method for detecting shaft displacement faults in a nuclear reactor main circulation pump, comprising the following steps:

[0007] Step 1: Obtain main pump shaft displacement data through monitoring equipment;

[0008] Step 2: Perform data sliding window sampling;

[0009] Step 3: Perform data augmentation by adding fill data of length m to both sides of the sliding window;

[0010] Step 4: Input the dataset into the spectral residual network, conduct a significance study, and extract the spectral residual values;

[0011] Step 5: Input the spectral residual data into the DAGMM network for anomaly detection;

[0012] Step 6: Output anomaly detection results through data analysis and comparison.

[0013] The second step involves using a sliding window to identify the synchronization signal, sampling the original training data and test data using a sliding window of length k and step size j.

[0014] The third step involves adding several estimators to both sides of each sliding window. The formula for calculating the estimators is as follows:

[0015]

[0016]

[0017] Wherein, g(x) i ,x j-i ) is x j With x j-i The gradient between two points, x j For the current data, x j-i Let i be the first i data points, and m be the number of data points to add. x is an estimated value. j-m+1 For the first j-m+1 Data number, x j+1 To enhance the data.

[0018] The fourth step includes the following steps:

[0019] Step 1: For a certain main pump shaft displacement historical dataset x n We take e time windows of length w, that is, the number of sampling points is w;

[0020] Step 2: Determine the number of data points b needed for data augmentation at both ends of each sliding window, and calculate x. b With x' b ;

[0021] Step 3: Use Fourier transform to obtain the amplitude spectrum A(f) and phase spectrum P(f);

[0022] Step 4: Calculate the spectral residual of the data based on the amplitude spectrum A(f). The spectral residual is the difference between the image log spectrum and the average spectrum.

[0023] Step 5: Finally, perform an inverse Fourier transform on the residual spectrum R(f) and the phase spectrum P(f) from the prior information to obtain the saliency spectrum S(f) of the input image. The specific formula is shown below:

[0024]

[0025] In the formula: ξ(x) is the Fourier transform, Amplitude and Phase are the calculated amplitude and phase angles, respectively, L(f) is the logarithm of A(x), AL(f) is the average of L(f), iP(f) is the value of P(f) as an imaginary number, S(x) is the spectral residual data, h q (f) is to find the mean matrix, where q is the length of ξ(x), and the specific expression is:

[0026]

[0027] The fifth step involves a deep autoencoder Gaussian mixture model, which includes a dimensionality reduction network and an anomaly detection network.

[0028] The dimensionality reduction network described above compresses the data dimensionality through a deep autoencoder, and the specific process is as follows:

[0029] Step 1: Compress the input data X using the encoding part of the autoencoder to obtain z. c ;

[0030] z c =h(x; θ) e )

[0031] In the formula: h(·) represents the encoding part of the depth autoencoder, θ e These are the parameters of the depth autoencoder, z c Encoded data;

[0032] Step 2: Reconstruct x' from the dimensionality-reduced data using the decoding part of the autoencoder, and calculate the reconstruction error z. r ;

[0033] x'=g(z c ;θ d )

[0034] z r =f(x,x')

[0035] In the formula: g(·) represents the decoding part of the depth autoencoder, θ d These are the parameters of the depth self-decoder, z r To reconstruct the data;

[0036] Step 3: Low-dimensional data z after decoding by the deep autoencoder c With reconstruction error z r Combined, it serves as the input to the anomaly detection network;

[0037] z = [z c ,z r ].

[0038] The anomaly detection network described above is jointly trained using a multilayer perceptron and a data dimensionality reduction network. The low-level data z obtained from the data dimensionality reduction network is input into the multilayer perceptron, and the activation function of the final output layer of the multilayer perceptron is softmax(), as shown in the following formula:

[0039] p = MLP(z; θ) m )

[0040]

[0041] In the formula: P is derived from θ m The output of the multilayer perceptron with parameters, It is the output after passing through P and the softmax activation function;

[0042] When the number of Gaussian components in the Gaussian mixture model is K, using The parameters of the Gaussian mixture model are updated in conjunction with the sample size N, as shown in the following formula:

[0043]

[0044]

[0045]

[0046] Where: φ k μ k ∑k represents the mixture probability, mean, and covariance in the Gaussian mixture model, respectively; N is the sample size; k is the number of Gaussian models; and y ik This represents the classification result of the i-th data point in k Gaussian models;

[0047] The formula for calculating the likelihood function evaluation parameters is as follows:

[0048]

[0049] An end-to-end joint training structure is adopted, and the objective equation for training is as follows:

[0050]

[0051] In the formula: L(x) i ,x' i ) is the loss function, representing the reconstruction error of the deep autoencoder, E(z) is the likelihood function loss, and λ1 and λ2 are fixed parameters of the model.

[0052] A Gaussian probability distribution of shaft displacement data is constructed using a DAGMM network fault detection model. The data points are then judged to be faulty waveforms by determining whether the sample points exceed the distribution.

[0053] The beneficial effects of this invention are as follows: (1) The spectral residual network technique enables the saliency study of the main pump shaft displacement data, successfully distinguishing normal data from abnormal data. Based on this technique, the accuracy of anomaly detection can be improved, and misjudgment during the anomaly detection process can be avoided; (2) The Gaussian hybrid autoencoder (DAGMM) achieves efficient anomaly detection for a large amount of main pump shaft displacement data. Compared with traditional methods, this model can process large-scale datasets more efficiently, and the anomaly detection results are significantly higher than other models. Attached Figure Description

[0054] Figure 1 The flowchart of a method for detecting shaft displacement faults in a nuclear reactor main circulation pump provided by the present invention is shown. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] This invention proposes a method for detecting faults in the main circulation pump shaft displacement of a nuclear reactor. The method utilizes a combination of a spectral residual network and a Gaussian mixture autoencoder (GJE) for anomaly detection. First, the collected main pump displacement data is input into the spectral residual network, and significance analysis is performed to improve the differentiation between abnormal and normal data distributions. Then, the spectral residual data is input into the GJE hybrid model, leveraging the efficient data processing capabilities of deep learning autoencoders to improve the model's data processing efficiency. Finally, the anomaly detection capability of the GJE model is used to detect anomalies in the main pump displacement data.

[0057] like Figure 1 As shown, a method for detecting shaft displacement faults in a nuclear reactor main circulation pump includes the following steps:

[0058] Step 1: Obtain the main pump shaft displacement data through monitoring equipment.

[0059] The main pump shaft displacement data is first acquired through sensors, and then obtained through the plant's digital instrumentation and control system (DCS).

[0060] Step 2: Perform data sliding window sampling.

[0061] Spectral residual networks primarily utilize Fourier transform and are generally used for data analysis where frequency and amplitude remain constant. However, the operating state of the main pump shaft changes with factors such as temperature and maintenance. Therefore, in this embodiment of the invention, a sliding window is used to identify the synchronization signal, ensuring the accuracy of the significance study. A sliding window of length k and step size j is used to sample the original training and test data.

[0062] Step 3: Perform data augmentation by adding fill data of length m to both sides of the sliding window.

[0063] Since a sliding window is equivalent to analyzing the product of an infinitely long signal and a rectangular window in the time domain, the main lobe of the rectangular window function is wide, and the side lobes decay slowly, causing spectral energy leakage. Because the signal is truncated by the time window, the spectrum obtained after the Fourier transform shows some spurious spectral lines with small amplitudes at other frequency points on both sides of the true spectral lines, resulting in spectral leakage. To avoid side-lobe spectral leakage, in this embodiment of the invention, several estimators are added on both sides of each sliding window. The formula for calculating the estimators is as follows:

[0064]

[0065]

[0066] In the formula: g(x) i ,x j-i ) is x j With x j-i The gradient between two points, x j For the current data, x j-i Let i be the first i data points, and m be the number of data points to add. x is an estimated value. j-m+1 For the first j-m+1 Data number, x j+1 To enhance the data.

[0067] Step 4: Input the dataset into the spectral residual network, conduct a significance study, and extract the spectral residual values.

[0068] The spectral residual algorithm includes the following steps:

[0069] Step 1: For a certain main pump shaft displacement historical dataset x n We take e time windows of length w, that is, the number of sampling points is w.

[0070] Step 2: Determine the number of data points b needed for data augmentation at both ends of each sliding window, and calculate x. b With x' b .

[0071] Step 3: Use Fourier transform to obtain the amplitude spectrum A(f) and phase spectrum P(f).

[0072] Step 4: Calculate the spectral residual of the data based on the amplitude spectrum A(f). The spectral residual is the difference between the image log spectrum and the average spectrum.

[0073] Step 5: Finally, perform an inverse Fourier transform on the residual spectrum R(f) and the phase spectrum P(f) from the prior information to obtain the saliency spectrum S(f) of the input image. The specific formula is shown below:

[0074]

[0075] In the formula: ξ(x) is the Fourier transform, Amplitude and Phase are the calculated amplitude and phase angles, respectively. L(f) is the logarithm of A(x), AL(f) is the average of L(f), iP(f) is the value of P(f) as an imaginary number, S(x) is the spectral residual data, h q (f) is to find the mean matrix, where q is the length of ξ(x), and the specific expression is:

[0076]

[0077] Step 5: Input the spectral residual data into a deep autoencoder Gaussian mixture model (DAGMM) network for anomaly detection.

[0078] The deep autoencoder Gaussian mixture model consists of two main components: a dimensionality reduction network and an anomaly detection network. The dimensionality reduction network primarily compresses the data dimensions using a deep autoencoder, as detailed below:

[0079] Step 1: Compress the input data X using the encoding part of the autoencoder to obtain z. c ;

[0080] z c =h(x; θ) e )

[0081] In the formula: h(·) represents the encoding part of the depth autoencoder, θ e These are the parameters of the depth autoencoder, z c For encoded data.

[0082] Step 2: Reconstruct x' from the dimensionality-reduced data using the decoding part of the autoencoder, and calculate the reconstruction error z. r ;

[0083] x'=g(z c ;θ d )

[0084] z r =f(x,x')

[0085] In the formula: g(·) represents the decoding part of the depth autoencoder, θ d These are the parameters of the depth self-decoder, z r To reconstruct the data.

[0086] Step 3: Low-dimensional data z after decoding by the deep autoencoder c With reconstruction error z r Combined, it serves as the input to the anomaly detection network;

[0087] z = [z c ,z r ]

[0088] Gaussian Mixture Model (GMM) is a probabilistic clustering method that assumes input samples follow a Gaussian distribution with k unknown parameters; samples following the same distribution are clustered into one class. Unlike traditional methods that use expectation-maximization (EM) to fit multiple Gaussian distributions, this invention employs a Multilayer Perceptron (MLP) to jointly train an anomaly detection network and a data dimensionality reduction network. The low-level data z obtained from the data dimensionality reduction network is input into the MLP, and the final output layer activation function of the MLP is softmax(), as shown in the following formula:

[0089] p = MLP(z; θ) m )

[0090]

[0091] In the formula: P is derived from θ m The output of the multilayer perceptron (MLP) parameters. It is the output after passing through P and the softmax activation function.

[0092] When the number of Gaussian components in the Gaussian mixture model is K, using The parameters of the Gaussian mixture model are updated in conjunction with the sample size N, as shown in the following formula:

[0093]

[0094]

[0095]

[0096] Where: φ k μ k ∑k represents the mixture probability, mean, and covariance in the Gaussian mixture model, respectively; N is the sample size; k is the number of Gaussian models; and y ik This represents the classification result of the i-th data in k Gaussian models.

[0097] The formula for calculating the likelihood function evaluation parameters is as follows:

[0098]

[0099] To avoid data loss and misjudgments of reactor faults caused by training the dimensionality reduction network and anomaly detection network separately, resulting in local optima, this invention proposes a DAGMM model that employs an end-to-end joint training structure. The objective equation for training is shown below:

[0100]

[0101] In the formula: L(x) i ,x' i ) is the loss function, representing the reconstruction error of the deep autoencoder, E(z) is the likelihood function loss, and λ1 and λ2 are fixed parameters of the model.

[0102] Step 6: Output anomaly detection results through data analysis and comparison.

[0103] A Gaussian probability distribution of shaft displacement data is constructed using a DAGMM network fault detection model. The data points are then judged to be faulty waveforms by determining whether the sample points exceed the distribution.

[0104] The advantages of this invention are as follows:

[0105] 1. Considering that the main pump shaft displacement data is affected by external factors, resulting in large fluctuations and unclear abnormal data, a spectral residual network is used to conduct a significance study to amplify the abnormal information in the data.

[0106] 2. To address the issue that traditional anomaly detection models cannot effectively detect anomalies in large-scale datasets due to the large volume of main pump shaft displacement data, a DAGMM model is adopted. This model leverages the efficient data processing capabilities of a deep autoencoder (DA) and the efficient anomaly detection capabilities of a Gaussian mixture model (GMM) to perform anomaly detection on the main pump shaft displacement data.

Claims

1. A method for detecting displacement faults in the main circulation pump shaft of a nuclear reactor, characterized in that, Includes the following steps: Step 1: Obtain main pump shaft displacement data through monitoring equipment; Step 2: Perform data sliding window sampling; Step 3: Perform data augmentation by adding fill data of length m to both sides of the sliding window; Step 4: Input the dataset into the spectral residual network, conduct a significance study, and extract the spectral residual values; The fourth step includes the following steps: Step 1: For a certain main pump shaft displacement historical dataset x n We take e time windows of length w, that is, the number of sampling points is w; Step 2: Determine the number of data points b needed for data augmentation at both ends of each sliding window, and calculate x. b With x' b ; Step 3: Use Fourier transform to obtain the amplitude spectrum A(f) and phase spectrum P(f); Step 4: Calculate the spectral residual of the data based on the amplitude spectrum A(f). The spectral residual is the difference between the image log spectrum and the average spectrum. Step 5: Finally, perform an inverse Fourier transform on the residual spectrum R(f) and the phase spectrum P(f) from the prior information to obtain the saliency spectrum S(f) of the input image. The specific formula is shown below: In the formula: ξ(x) is the Fourier transform, Amplitude and Phase are the calculated amplitude and phase angles, respectively, L(f) is the logarithm of A(x), AL(f) is the average of L(f), iP(f) is the value of P(f) as the imaginary number, S(x) is the spectral residual data, h q (f) is to find the mean matrix, where q is the length of ξ(x), and the specific expression is: Step 5: Input the spectral residual data into a deep autoencoder Gaussian mixture model network for anomaly detection; Step 6: Output anomaly detection results through data analysis and comparison.

2. The method for detecting displacement faults in the main circulation pump shaft of a nuclear reactor as described in claim 1, characterized in that: The second step involves using a sliding window to identify the synchronization signal, sampling the original training data and test data using a sliding window of length k and step size j.

3. The method for detecting displacement faults in the main circulation pump shaft of a nuclear reactor as described in claim 1, characterized in that: The third step involves adding several estimators to both sides of each sliding window. The formula for calculating the estimators is as follows: Wherein, g(x) i ,x j-i ) is x j With x j-i The gradient between two points, x j For the current data, x j-i Let m be the number of data items to be added, representing the first i data items. x is an estimated value. j-m+1 For the first j-m+1 Data number, x j+1 To enhance the data.

4. The method for detecting displacement faults in the main circulation pump shaft of a nuclear reactor as described in claim 1, characterized in that: The fifth step involves a deep autoencoder Gaussian mixture model, which includes a dimensionality reduction network and an anomaly detection network.

5. A method for detecting displacement faults in the main circulation pump shaft of a nuclear reactor as described in claim 4, characterized in that, The dimensionality reduction network described above compresses the data dimensionality through a deep autoencoder, and the specific process is as follows: Step 1: Compress the input data X using the encoding part of the autoencoder to obtain z. c ; z c =h(x;θ e ) In the formula: h(·) represents the encoding part of the depth autoencoder, θ e These are the parameters of the depth autoencoder, z c Encoded data; Step 2: Reconstruct x' from the dimensionality-reduced data using the decoding part of the autoencoder, and calculate the reconstruction error z. r ; x'=g(z c ;θ d ) z r =f(x,x') In the formula: g(·) represents the decoding part of the depth autoencoder, θ d These are the parameters of the depth self-decoder, z r To reconstruct the data; Step 3: Low-dimensional data z after decoding by the deep autoencoder c With reconstruction error z r Combined, it serves as the input to the anomaly detection network; z=[z c ,With r ]。 6. The method for detecting displacement faults in the main circulation pump shaft of a nuclear reactor as described in claim 4, characterized in that, The anomaly detection network described above is jointly trained using a multilayer perceptron and a data dimensionality reduction network. The low-level data z obtained from the data dimensionality reduction network is input into the multilayer perceptron, and the activation function of the final output layer of the multilayer perceptron is softmax(), as shown in the following formula: p=MLP(z;θ m ) In the formula: P is derived from θ m The output of the multilayer perceptron with parameters, It is the output after passing through P and the softmax activation function; When the number of Gaussian components in the Gaussian mixture model is K, using The parameters of the Gaussian mixture model are updated in conjunction with the sample size N, as shown in the following formula: Where: φ k μ k ∑k represents the mixture probability, mean, and covariance in the Gaussian mixture model, respectively; N is the sample size; k is the number of Gaussian models; and y ik This represents the classification result of the i-th data point in k Gaussian models; The formula for calculating the likelihood function evaluation parameters is as follows: An end-to-end joint training structure is adopted, and the objective equation for training is as follows: In the formula: L(x) i ,x i ') is the loss function, representing the reconstruction error of the deep autoencoder, E(z) is the likelihood function loss, and λ1 and λ2 are fixed parameters of the model.

7. The method for detecting displacement faults in the main circulation pump shaft of a nuclear reactor as described in claim 6, characterized in that: A Gaussian probability distribution for shaft displacement data is constructed using a deep autoencoder Gaussian mixture model network for fault detection. The data points are then judged to be faulty waveforms by determining whether the sample points exceed the distribution.

Images