Sensor fault monitoring and failure information reconstruction method and system

Abstract

This invention proposes a sensor fault monitoring and failure information reconstruction method and system. First, it uses a robust improved principal component analysis method for sensor fault detection and identification in nuclear power plants. Then, it reconstructs the failure information of nuclear power plant sensors based on maximum mutual information coefficients, convolutional autoencoders, long short-term memory networks, and self-attention mechanisms. Finally, it uses an improved particle swarm optimization algorithm for automatic hyperparameter optimization. This method achieves more robust sensor fault detection and faster, more accurate sensor failure information reconstruction. It can accurately detect abnormal readings from steam generator water level sensors, accurately determine whether the abnormality is caused by a "false water level" or a sensor malfunction, and accurately and quickly reconstruct the abnormal water level information of the steam generator, ensuring the normal operation of the control system and improving the safety and economy of nuclear power plants.

Description

Technical Field

[0001] This invention belongs to the field of sensor information reconstruction technology, specifically, it relates to a sensor fault monitoring and failure information reconstruction method and system. Background Technology

[0002] Due to the complexity and safety concerns of nuclear power plants, monitoring the operational status of instruments and equipment is of paramount importance. Ensuring the accuracy and validity of sensor measurements is fundamental to achieving condition monitoring. Because of the unique operating environment of sensors in nuclear power plants, the probability of sensor failure is relatively high. As a key piece of equipment connecting the primary and secondary loops in a nuclear power plant, the U-tube steam generator is a complex system with nonlinearity and minimum phase, exhibiting a phenomenon known as "false water level." This means that sudden changes in power output leading to sudden changes in feedwater and steam flow can cause the steam generator's water level to reverse in a short period. This reverse water level change manifests similarly to a sensor malfunction and can lead to incorrect judgments by the control system. Therefore, in the actual operation of a nuclear power plant, it is necessary to monitor the steam generator's water level sensors and be able to distinguish between malfunctions caused by "false water levels" and those caused by sensor failures. Furthermore, for data anomalies caused by "false water levels" or sensor malfunctions, it is necessary to reconstruct the steam generator's water level information to ensure the normal operation of the control system.

[0003] Therefore, research is needed on sensor fault detection and failure information reconstruction; the reconstructed value can be temporarily replaced with the failure value to provide a buffer period for manual replacement of the failed sensor. When an abnormal reading of the steam generator water level sensor is detected, it is possible to accurately determine whether the abnormality is caused by a "false water level" or a sensor failure, so as to ensure the normal operation of the nuclear power plant. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a sensor fault monitoring and failure information reconstruction method and system. It employs principal component analysis (PCA) from data-driven state monitoring methods to detect faults in nuclear power plant sensors, and deep learning to reconstruct sensor failure information. Furthermore, it utilizes a Savitzky-Golay filter from the field of signal processing to process the statistical curves of the test data.

[0005] This invention is achieved through the following technical solution:

[0006] A method for sensor fault monitoring and failure information reconstruction:

[0007] The method specifically includes the following steps:

[0008] Step 1: Establish a sensor model for the nuclear power plant and use a data-driven principal component analysis method for sensor fault detection and identification.

[0009] Step 2: Based on the sensor model established in Step 1, obtain the operating data of each sensor under different working conditions, perform correlation analysis of different working condition features based on the maximum mutual information coefficient, select reconstruction reference features, and construct a sensor failure information reconstruction model.

[0010] Step 3: Reconstruct the model based on the failure information from Step 2, use the improved particle swarm optimization algorithm to automatically optimize hyperparameters, and use the historical optimal position of the population obtained from multiple rounds of iterative calculation as the value of the hyperparameters of the reconstructed model.

[0011] Step 4: After completing Step 3, perform model verification and testing; manually insert sensor fault data, call the reconstruction model to reconstruct the fault data, and analyze the error between the reconstructed value and the true value after normalization.

[0012] Further, in step 1,

[0013] Step 1.1: Use a full-range nuclear power plant simulator to acquire raw sensor measurement data X under various operating conditions during normal operation. m×n Where n is the total number of sensors to be monitored and m is the total number of measurement samples, a sensor model for a nuclear power plant is established;

[0014]

[0015] Step 1.2: Based on equations (1) and (2), process the original measurement data X. m×n Standardize using standard deviation;

[0016]

[0017] In the formula, X j This represents the raw operating data of the j-th sensor; E(X) j D(X) represents the mean of the raw operating data of the j-th sensor; j X is the standard deviation of the raw operating data of the j-th sensor; j * X is the standardized data from the j-th sensor. j * The mean is 0 and the standard deviation is 1; X m×n * The matrix is ​​the standardized result of the original measurement data;

[0018] Step 1.3: Based on equation (3), approximately solve for the standardized matrix X. * The covariance matrix ∑;

[0019]

[0020] Step 1.4: According to the matrix decomposition theorem, perform eigenvalue decomposition on ∑ to obtain the eigenvalues ​​λ1, λ2, ..., λ of ∑. n (λ1≥λ2≥...≥λ n ) and the corresponding standard eigenvectors P1, P2, ..., P n ;

[0021] Σ=P×diag(λ1,λ2,…,λ n )×P T (4)

[0022] In the formula, the orthogonal matrix P = [P1, P2, ..., P n ];diag(λ1,λ2,...,λ n ) represents the expression with λ1,λ2,...,λ n It is an n-order diagonal matrix with diagonal elements;

[0023] Step 1.5: Using the cumulative variance contribution rate method, select k principal components based on the criterion that CPV ≥ 90%;

[0024]

[0025] In the formula, the eigenvalue λ reflects the amount of information contained in each principal element; PV represents the percentage of information contained in the i-th principal element relative to the total amount of information in the original data; CPV represents the percentage of the total original information contained in the first k principal elements.

[0026] Step 1.6: Use the eigenvectors of the first k principal components to form the principal component space. The residual space is formed by the remaining (nk) principal elements. The test vector at a certain moment can be represented by equation (6):

[0027]

[0028] In the formula, x n×1 A test vector at a certain moment contains the readings of n sensors at that moment; and They are x n×1 Projection in principal space and residual space; The projection matrix of the principal element space. The residual space projection matrix;

[0029] Step 1.7: Based on equations (7) and (8), calculate Hotelling's T according to the partitioned principal space and residual space. 2 That is, T 2 Threshold T of the statistic α 2And the squared prediction error SPE, i.e., the threshold of the Q statistic Q. α ;

[0030]

[0031]

[0032] In the formula, F α (k,nk) represents the critical value of the F-distribution with k and nk degrees of freedom and a confidence level of α; λ i Let k be the eigenvalues ​​of Σ; k be the number of principal components; c be the eigenvalues ​​of Σ. α This is the (1-α) confidence limit for the standard normal distribution.

[0033] Furthermore, step 1 also includes:

[0034] Step 1.8: Begin fault detection by inputting the real-time sensor data test vector x. n×1 And according to equation (2), it is standardized to obtain x * ;

[0035] Step 1.9: Calculate the test vector x at this moment according to equations (9) and (10). n×1 T 2 Statistic and Q statistic;

[0036]

[0037]

[0038] In the formula, Λ k =diag(λ1,λ2,…,λ) k );

[0039] Step 1.10: Based on equation (11), process T using the SG filtering method. 2 Curves showing the changes of the statistic and the Q statistic over time;

[0040]

[0041] In the formula, Y represents the fitted value of the corresponding statistic. j+1 C represents the original input value for the corresponding statistic; i is the coefficient for filtering the i-th statistic; m is the width of half a filtering window; N is the filter length, which is equal to the width of the sliding array (2m+1);

[0042] Step 1.11: If T after processing by the SG filtering method 2 If neither the statistic nor the Q statistic exceeds the corresponding threshold, it indicates that the system is operating normally.

[0043] If T is processed by the SG filtering method 2 If the statistic or Q statistic exceeds the corresponding threshold, it indicates an anomaly, but the cause of the anomaly still needs further analysis and judgment.

[0044] Step 1.12: Perform fault identification and calculate x according to equations (12) and (13). n×1 The contribution rate of the i-th sensor to the Q-statistic Q i and against T 2 Contribution rate T of the statistic i 2 Locate the test vector x n×1 An abnormal sensor was detected in the middle;

[0045] If the contribution rates of each sensor are evenly distributed, it indicates that the cause of the anomaly is a system-level fault.

[0046] If the contribution rate of a certain sensor is significantly higher than normal, it indicates that the abnormality is caused by the failure of that sensor.

[0047]

[0048]

[0049] In the formula, t represents x n×1 The score vector, Represents sensor variable x j For the j-th score vector t j Contribution rate; x i =[0,…0,x i ,0…,0] T .

[0050] Furthermore, in step 2,

[0051] Step 2.1: If the fault source is the target sensor, then call the reconstruction model to reconstruct the data of the failed sensor;

[0052] Step 2.2: Use a full-range nuclear power plant simulator to obtain the operating data of each sensor under the target operating conditions;

[0053] Step 2.3: Analyze the statistical correlation between sensors; calculate the maximum mutual information coefficient between the reference sensor and the target sensor under different power steady-state conditions and various accident conditions according to Equation (14);

[0054]

[0055] In the formula, a(X;Y) represents the maximum mutual information coefficient between X and Y; This represents the maximum mutual information value between X and Y for a given set of values ​​for M and N.

[0056] Step 2.4: Select the reconstruction reference sensor; select the top ten sensors with the largest mutual information coefficient under steady-state conditions, and select the top five sensors with the largest mutual information coefficient under other accident conditions. Then, take the union of the selected sensors to obtain the reconstruction model reference sensor combination.

[0057] Step 2.5: Construct the reconstructed model dataset; process the full-range simulator running data, remove sensors with low correlation to the target sensor, and retain the sensors selected in Step 2.4;

[0058] Step 2.6: Data standardization processing; According to equation (15), the sensor operation data selected in step 2.5 is subjected to min-max standardization processing to eliminate the influence of different sensor dimensions and accelerate the convergence speed of the reconstruction model.

[0059]

[0060] In the formula, x * represents the normalized data of a certain sensor; x represents the data of a certain sensor before normalization; min represents the minimum value in the sample data of a certain sensor; max represents the maximum value in the sample data of a certain sensor.

[0061] Step 2.7: Construct the input data for the reconstruction model using the sliding window method. The data execution time interval is 0.25s, and the sliding window length is set to 2.5s. Convert the two-dimensional tensor (n×m) obtained in Step 2.6 into a three-dimensional tensor ((n-10+1)×(10×m)).

[0062] Furthermore, step 2 also includes,

[0063] Step 2.8: Build a deep learning network structure in the TensorFlow framework, consisting of a convolutional autoencoder, a long short-term memory network, and a self-attention mechanism. The model is composed of an input layer, multiple convolutional layers, pooling layers, upsampling layers, multiple long short-term memory network layers, a self-attention layer, a fully connected layer, and an output layer connected layer by layer.

[0064] Step 2.9: Construct the encoding layer of the convolutional autoencoder by combining convolutional and pooling layers; use one-dimensional convolution operations to extract features from the data;

[0065] A decoding layer of a convolutional autoencoder is constructed by combining convolutional layers and upsampling layers to restore the data features to the original data; the data is then output to a long short-term memory network layer after passing through multiple layers of convolutional autoencoders.

[0066] Step 2.10: Construct a Long Short-Term Memory (LSTM) network; the LSTM network uses a gate mechanism to control the flow and loss of information; represented as follows:

[0067] Input gate (threshold): determines the proportion of information that will be stored in the current cell state;

[0068] i t =σ(W i· [h t-1 ,x t ]+b i (16)

[0069] Forgetting gate: Selectively forgets information in the cellular state;

[0070] f t =σ(W f ·[h t-1 ,x t ]+b f (17)

[0071] Output gate (threshold): Selectively outputs information from the cell state;

[0072] o t =σ(W o ·[h t-1 ,x t ]+b o (18)

[0073] In the above three formulas, W i W f W o b is the parameter matrix to be trained; i b f b o σ is the bias term to be trained; σ is the sigmoid activation function.

[0074] memory h t The short-term memory is represented by the output gate obtained from the current cell state:

[0075] h t =o t *tanh(C t (19)

[0076] Candidate state The new knowledge to be stored in the cell state, represented and inductively derived, is the input feature x at the current moment. t and the short-term memory of the previous moment h t-1 Functions:

[0077]

[0078] Cellular state C t This represents long-term memory, which is equal to the long-term memory C from the previous moment. t-1 The value of the forget gate and the new knowledge summarized at the current moment. Sum of values ​​from the input gate:

[0079]

[0080] The overall calculation process is as follows: the forget gate, input gate, and output gate use the sigmoid function to determine the content f to be discarded from the information in the previous time step. t Content that needs to be retained t and the content to be output o t The content output at this moment. t With cellular state C t The memory h is determined t The memory h from the previous moment t-1 and the input x at this moment t The candidate state at that moment was determined. Input gate i at this moment t Output gate o t The previous cell state C t and the candidate states at this moment Together, they determine the cell state C at this moment. t Then, the calculation is repeated for each time step in this manner.

[0081] Step 2.11: Add a Dropout operation between the Long Short-Term Memory (LSTM) network layer and the self-attention layer;

[0082] Step 2.12: Introduce a self-attention mechanism; first, perform a linear transformation on the output X of the Long Short-Term Memory Network according to equations (22), (23) and (24) to obtain Q, K and V;

[0083]

[0084]

[0085]

[0086] In the formula, W Q W K and W V It is the parameter matrix to be trained;

[0087] The output vector of the self-attention layer can be calculated using equation (25):

[0088]

[0089] In the formula, i,j∈[1,N] represent the positions of the output and input vectors, and the connection weight α ij Dynamically generated by attention mechanisms;

[0090] Step 2.13: According to equation (26), the ELU function is used as the activation function of the fully connected layer of the model, and the value of α is 0.2;

[0091]

[0092] Step 2.14: Based on equation (27), the mean square error (MSE) is used as the loss function;

[0093]

[0094] Furthermore, in step 3,

[0095] Step 3.1: Train the reconstructed model; Randomly shuffle the dataset and divide it into training and test sets according to a certain ratio, and use the Adam optimizer to train the model;

[0096] Step 3.2: Review all hyperparameters of the reconstructed model and determine the feasible solution domains of the corresponding hyperparameters: number of convolutional layers, kernel size, stride, pooling layer size, upsampling layer size, number of layers in the long short-term memory network and the size of the hidden layers within the unit, number of fully connected layers and the number of neurons per layer, and the proportion of Dropout.

[0097] Step 3.3: Initialize the position information of each particle in the particle swarm. The position of each particle can be represented by equation (28);

[0098] X i =(x i1 ,x i2 ,…,x iD (28)

[0099] In the formula, X i It is the position of the i-th particle; x i1 ,x i2 ,…,x iD The values ​​of the hyperparameters described in step 27;

[0100] Step 3.4: Calculate the fitness value f for each particle. i ; based on the particle's position X i As the hyperparameter values ​​for the reconstruction model, the loss function and optimization method described in steps 2.13 and 2.14 are used to train the model, and the reconstruction error of the trained model is used as the fitness value of the particles.

[0101] Step 3.5: Update the individual historical best fitness value f for each particle. i,pand the individual's historical best position P i,pbest If the particle's current fitness value is less than its historical best fitness value, then update the particle's historical best fitness value and best position.

[0102] Step 3.6: Update the historical best fitness value f of the population. g and optimal position P gbest If the individual best fitness value of a particle is smaller than the best fitness value of the population, then update the best fitness value and best position of the population.

[0103] Step 3.7: Update the velocity of each particle according to equation (29);

[0104]

[0105] In the formula, V i k+1 Let be the velocity of the i-th particle in the (k+1)-th iteration. Let be the velocity of the i-th particle in the k-th iteration; Let be the optimal position of the i-th particle in the k-th iteration. This represents the optimal position of the population during the k-th iteration. c1 represents the position of the i-th particle in the k-th iteration; c2 is the individual learning factor with a value of 2; c2 is the group learning factor with a value of 2; r1 and r2 are random numbers in the interval [0,1]; ω k For adaptive inertia weights, ω k Update according to equation (30), where ω start Take an empirical value of 0.9, ω end Take the empirical value of 0.4;

[0106]

[0107] Step 3.8: Update the position of each particle according to equation (31);

[0108]

[0109] Step 3.9: Repeat steps 3.3 to 3.7 until the maximum number of iterations is reached, and calculate the group's historical optimal position P obtained from multiple iterations. gbest The values ​​of the hyperparameters used to reconstruct the model.

[0110] A sensor fault monitoring and failure information reconstruction system:

[0111] The system includes a model building and fault detection module, a failure information reconstruction module, a hyperparameter automatic optimization module, and a verification and testing module.

[0112] The model building and fault detection module is used to build a sensor model of a nuclear power plant and uses a data-driven principal component analysis method for sensor fault detection and fault identification.

[0113] The failure information reconstruction module, based on the sensor model established by the model building and fault detection module, acquires the operating data of each sensor under different working conditions, performs correlation analysis of different working condition features based on the maximum mutual information coefficient, selects reconstruction reference features, and constructs a sensor failure information reconstruction model.

[0114] The hyperparameter automatic optimization module reconstructs the model based on the failure information of the failure information reconstruction module, and uses an improved particle swarm optimization algorithm to perform hyperparameter automatic optimization, taking the historical optimal position of the population obtained from multiple rounds of iterative calculation as the value of the hyperparameter of the reconstruction model.

[0115] The verification and testing module is used to perform model verification and testing after completing the automatic optimization of hyperparameters; manually insert sensor fault data, call the reconstruction model to reconstruct the fault data, and analyze the error between the reconstructed value and the true value after normalization.

[0116] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.

[0117] A computer-readable storage medium for storing computer instructions that, when executed by a processor, implement the steps of the above-described method.

[0118] Beneficial effects of the invention

[0119] This invention employs a data-driven principal component analysis method with significant noise reduction effect to detect sensor faults and uses deep learning methods to reconstruct sensor failure information. It does not require the establishment of an accurate mathematical model of the system, nor does it require the analysis of causal relationships between sensor variables within the system or a detailed expert knowledge base, which has great advantages for the large number of sensors in nuclear power plants.

[0120] Two key statistics in principal component analysis are highly sensitive to data noise and prone to false alarms. This invention employs a Savitzky-Golay filter from the field of signal processing to process the statistical curves of the test data. This filter removes noise while ensuring the shape and width of the signal remain unchanged, suppressing sudden increases in statistics caused by random noise. The T values ​​processed by the Savitzky-Golay filter are then improved. 2 The curves of the statistical measure and the Q-statistic become smoother, reducing false alarms in the model and making the fault detection results more reliable.

[0121] This invention improves the accuracy and generalization ability of the model by simultaneously considering the characteristics of time and space dimensions. It also employs a self-attention mechanism, which assigns different weights to different locations of data to achieve a reasonable allocation of computational resources. This focuses on the most important parts of the current task, reducing attention to other information, effectively reducing the computational burden of the model and improving its operating efficiency. The particle swarm optimization algorithm is used to find the optimal combination of hyperparameters for reconstructing the model. It can reconstruct the water level sensor of a steam generator that exhibits "false water level" phenomena, ensuring the normal operation of the control system and demonstrating significant application value.

[0122] This invention constructs a deep learning model, using the readings of other sensors at different times as input to the model to reconstruct the failure information of the steam generator's water level sensor. Utilizing the characteristic that mutual information can capture nonlinear correlations between variables, the maximum mutual information coefficient between each sensor and the target sensor is calculated before establishing the reconstruction model. Based on the maximum mutual information coefficient, a set of sensors highly correlated with the steam generator's water level sensor are selected as input to the reconstruction model, avoiding problems such as the "curse of dimensionality" and "overfitting," effectively improving the model's accuracy and generalization ability.

[0123] In the process of building a deep learning model, the method of this invention utilizes a convolutional autoencoder and a long short-term memory network to simultaneously extract the spatial and temporal features of the original running data, which effectively improves the model's fitting ability. Furthermore, a self-attention mechanism is introduced between the long short-term memory network and the fully connected output layer, which achieves reasonable resource allocation by assigning weights to different parts of the data, thereby improving the model's running efficiency.

[0124] This invention replaces the constant inertia weight in the particle swarm optimization algorithm with a nonlinear adaptive inertia weight, so that the inertia weight is automatically adjusted with the iteration process, which enhances the global search capability of the algorithm and effectively avoids the hyperparameter optimization process from getting stuck in local optima.

[0125] In summary, compared with the prior art, the method of the present invention can achieve more robust sensor fault detection and faster, more accurate sensor failure information reconstruction; it can accurately detect abnormal readings of steam generator water level sensors, accurately determine whether the abnormality is caused by "false water level" or sensor failure, and accurately and quickly reconstruct the abnormal water level information of the steam generator, ensuring the normal operation of the control system and improving the safety and economy of nuclear power plants. Attached Figure Description

[0126] Figure 1 This invention relates to a robust improved principal component analysis method for sensor fault detection and identification processes.

[0127] Figure 2 This is a sensor failure information reconstruction model based on maximum mutual information coefficient, convolutional autoencoder, long short-term memory network and self-attention mechanism;

[0128] Figure 3 This invention presents an automatic hyperparameter optimization process based on an improved particle swarm optimization algorithm. Detailed Implementation

[0129] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0130] A method for sensor fault monitoring and failure information reconstruction:

[0131] The method specifically includes the following steps:

[0132] Step 1: Establish a sensor model for the nuclear power plant and use a data-driven principal component analysis method for sensor fault detection and identification.

[0133] Step 2: Based on the sensor model established in Step 1, obtain the operating data of each sensor under different working conditions, perform correlation analysis of different working condition features based on the maximum mutual information coefficient, select reconstruction reference features, and construct a sensor failure information reconstruction model.

[0134] Step 3: Reconstruct the model based on the failure information from Step 2, use the improved particle swarm optimization algorithm to automatically optimize hyperparameters, and use the historical optimal position of the population obtained from multiple rounds of iterative calculation as the value of the hyperparameters of the reconstructed model.

[0135] Step 4: After completing Step 3, perform model verification and testing; manually insert sensor fault data, call the reconstruction model to reconstruct the fault data, and analyze the error between the reconstructed value and the true value after normalization.

[0136] In step 1, step 1.1: Use a nuclear power plant full-range simulator to acquire raw sensor measurement data X under various operating conditions. m×n Where n is the total number of sensors to be monitored and m is the total number of measurement samples, a sensor model for a nuclear power plant is established;

[0137]

[0138] Step 1.2: Based on equations (1) and (2), process the original measurement data X. m×nStandard deviation standardization is performed to eliminate differences in the order of magnitude of readings from different sensors during actual operation;

[0139]

[0140] In the formula, X j This represents the raw operating data of the j-th sensor; E(X) j D(X) represents the mean of the raw operating data of the j-th sensor; j X is the standard deviation of the raw operating data of the j-th sensor; j * X is the standardized data from the j-th sensor. j * The mean is 0 and the standard deviation is 1; X m×n * The matrix is ​​the standardized result of the original measurement data;

[0141] Step 1.3: Based on equation (3), approximately solve for the standardized matrix X. * The covariance matrix ∑;

[0142]

[0143] Step 1.4: According to the matrix decomposition theorem, perform eigenvalue decomposition on ∑ to obtain the eigenvalues ​​λ1, λ2, ..., λ of ∑. n (λ1≥λ2≥...≥λ n ) and the corresponding standard eigenvectors P1, P2, ..., P n ;

[0144] Σ=P×diag(λ1,λ2,…,λ n )×P T (4)

[0145] In the formula, the orthogonal matrix P = [P1, P2, ..., P n ];diag(λ1,λ2,...,λ n ) represents the expression with λ1,λ2,...,λ n It is an n-order diagonal matrix with diagonal elements;

[0146] Step 1.5: Using the Cumulative Percentage Variance (CPV) method, k principal components are selected based on CPV ≥ 90% to achieve the goal of characterizing the operating status of all sensors with a small number of principal components;

[0147]

[0148] In the formula, the eigenvalue λ reflects the amount of information contained in each principal element; PV represents the percentage of information contained in the i-th principal element relative to the total amount of information in the original data; CPV represents the percentage of the total original information contained in the first k principal elements.

[0149] Step 1.6: Use the eigenvectors of the first k principal components to form the principal component space. The residual space is formed by the remaining (nk) principal elements. The test vector at a certain moment can be represented by equation (6):

[0150]

[0151] In the formula, x n×1 A test vector at a certain moment contains the readings of n sensors at that moment; and They are x n×1 Projection in principal space and residual space; The projection matrix of the principal element space. The residual space projection matrix;

[0152] Step 1.7: Based on equations (7) and (8), calculate Hotelling's T according to the partitioned principal space and residual space. 2 That is, T 2 Threshold T of the statistic α 2 And the squared prediction error (SPE), which is the threshold of the Q statistic Q. α ;

[0153]

[0154]

[0155] In the formula, F α (k,nk) represents the critical value of the F-distribution with k and nk degrees of freedom and a confidence level of α; λ i Let k be the eigenvalues ​​of Σ; k be the number of principal components; c be the eigenvalues ​​of Σ. α This is the (1-α) confidence limit for the standard normal distribution.

[0156] Step 1.8: Begin fault detection by inputting the real-time sensor data test vector x. n×1 And according to equation (2), it is standardized to obtain x * ;

[0157] Step 1.9: Calculate the test vector x at this moment according to equations (9) and (10). n×1 T 2Statistic and Q statistic;

[0158]

[0159]

[0160] In the formula, Λ k =diag(λ1,λ2,…,λ) k );

[0161] Step 1.10: Based on equation (11), process T using the SG filtering method. 2 The curves showing the changes of the statistic and Q statistic over time eliminate fluctuations caused by random noise and reduce false alarms.

[0162]

[0163] In the formula, Y represents the fitted value of the corresponding statistic. j+1 C represents the original input value for the corresponding statistic; i is the coefficient for filtering the i-th statistic; m is the width of half a filtering window; N is the filter length, which is equal to the width of the sliding array (2m+1);

[0164] Step 1.11: If T after processing by the SG filtering method 2 If neither the statistic nor the Q statistic exceeds the corresponding threshold, it indicates that the system is operating normally.

[0165] If T is processed by the SG filtering method 2 If the statistic or Q statistic exceeds the corresponding threshold, it indicates an anomaly, but the cause of the anomaly still needs further analysis and judgment.

[0166] Step 1.12: Perform fault identification and calculate x according to equations (12) and (13). n×1 The contribution rate of the i-th sensor to the Q-statistic Q i and against T 2 Contribution rate T of the statistic i 2 Locate the test vector x n×1 An abnormal sensor was detected in the middle;

[0167] If the contribution rates of each sensor are evenly distributed, it indicates that the cause of the anomaly is a system-level fault.

[0168] If the contribution rate of a certain sensor is significantly higher than normal, it indicates that the abnormality is caused by the failure of that sensor.

[0169]

[0170]

[0171] In the formula, t represents x n×1 The score vector, Represents sensor variable x j For the j-th score vector t j Contribution rate; x i =[0,…0,x i ,0…,0] T .

[0172] In step 2, step 2.1: If the source of the fault is the target sensor, then the reconstruction model is invoked to reconstruct the data of the failed sensor;

[0173] Step 2.2: Use a full-range nuclear power plant simulator to obtain the operating data of each sensor under the target operating conditions;

[0174] Step 2.3: Analyze the statistical correlation between sensors; calculate the maximum mutual information coefficient (MIC) between the reference sensor and the target sensor under different power steady-state conditions and various accident conditions according to Equation (14);

[0175]

[0176] In the formula, a(X;Y) represents the maximum mutual information coefficient between X and Y, which is essentially a normalized mutual information that masks the order-of-magnitude difference in the absolute value of the mutual information. This represents the maximum mutual information value between X and Y for a given set of values ​​for M and N.

[0177] Step 2.4: Select the reconstruction reference sensor; Since nuclear power plants are in steady-state conditions most of the time and only in accident conditions a very small number of times, this invention selects the top ten sensors with the largest mutual information coefficient under steady-state conditions, and the top five sensors with the largest mutual information coefficient under other accident conditions. Then, the union of the selected sensors is taken to obtain the reconstruction model reference sensor combination.

[0178] Step 2.5: Construct the reconstructed model dataset; process the full-range simulator running data, remove sensors with low correlation to the target sensor, and retain the sensors selected in Step 2.4;

[0179] Step 2.6: Data standardization processing; According to equation (15), the sensor operation data selected in step 2.5 is subjected to min-max standardization processing to eliminate the influence of different sensor dimensions and accelerate the convergence speed of the reconstruction model.

[0180]

[0181] In the formula, x * represents the normalized data of a certain sensor; x represents the data of a certain sensor before normalization; min represents the minimum value in the sample data of a certain sensor; max represents the maximum value in the sample data of a certain sensor.

[0182] Step 2.7: Construct the input data for the reconstruction model using the sliding window method. The data execution time interval is 0.25s, and the sliding window length is set to 2.5s. Convert the two-dimensional tensor (n×m) obtained in Step 2.6 into a three-dimensional tensor ((n-10+1)×(10×m)).

[0183] Step 2.8: Construct a deep learning network structure within the TensorFlow framework, consisting of a convolutional autoencoder, a long short-term memory network, and a self-attention mechanism; for example... Figure 2 As shown, the model consists of an input layer, multiple convolutional layers, pooling layers, upsampling layers, multiple long short-term memory network layers, self-attention layers, fully connected layers, and an output layer connected layer by layer.

[0184] Step 2.9: Construct the encoding layer of the convolutional autoencoder by combining convolutional and pooling layers; use one-dimensional convolution operations to extract features from the data;

[0185] A decoding layer of a convolutional autoencoder is constructed by combining convolutional layers and upsampling layers to restore the data features to the original data; the data is then output to a long short-term memory network layer after passing through multiple layers of convolutional autoencoders.

[0186] Step 2.10: Construct a Long Short-Term Memory (LSTM) network; the LSTM network uses a gate mechanism to control the flow and loss of information; represented as follows:

[0187] Input gate (threshold): determines the proportion of information that will be stored in the current cell state;

[0188] i t =σ(W i· [h t-1 ,x t ]+b i (16)

[0189] Forgetting gate: Selectively forgets information in the cellular state;

[0190] f t =σ(W f ·[h t-1 ,x t ]+b f (17)

[0191] Output gate (threshold): Selectively outputs information from the cell state;

[0192] o t =σ(W o ·[h t-1 ,x t ]+b o (18)

[0193] In the above three formulas, W i W f W o b is the parameter matrix to be trained; i b f b o σ is the bias term to be trained; σ is the sigmoid activation function.

[0194] memory h t The short-term memory is represented by the output gate obtained from the current cell state:

[0195] h t =o t *tanh(C t (19)

[0196] Candidate state The new knowledge to be stored in the cell state, represented and inductively derived, is the input feature x at the current moment. t and the short-term memory of the previous moment h t-1 Functions:

[0197]

[0198] Cellular state C t This represents long-term memory, which is equal to the long-term memory C from the previous moment. t-1 The value of the forget gate and the new knowledge summarized at the current moment. Sum of values ​​from the input gate:

[0199]

[0200] The overall calculation process is as follows: the forget gate, input gate, and output gate use the sigmoid function to determine the content f to be discarded from the information in the previous time step. t Content that needs to be retained t and the content to be output o t The content output at this moment. t With cellular state C t The memory h is determined t The memory h from the previous moment t-1 and the input x at this moment t The candidate state at that moment was determined. Input gate i at this moment t Output gate o t The previous cell state Ct and the candidate states at this moment Together, they determine the cell state C at this moment. t Then, the calculation is repeated for each time step in this manner.

[0201] Step 2.11: Add a Dropout operation between the Long Short-Term Memory (LSTM) network layer and the self-attention layer;

[0202] Step 2.12: Introduce a self-attention mechanism; first, perform a linear transformation on the output X of the Long Short-Term Memory Network according to equations (22), (23) and (24) to obtain Q, K and V;

[0203]

[0204]

[0205]

[0206] In the formula, W Q W K and W V It is the parameter matrix to be trained;

[0207] The output vector of the self-attention layer can be calculated using equation (25):

[0208]

[0209] In the formula, i,j∈[1,N] represent the positions of the output and input vectors, and the connection weight α ij Dynamically generated by attention mechanisms;

[0210] Step 2.13: According to equation (26), the ELU function is used as the activation function of the fully connected layer of the model. The value of α in this invention is 0.2;

[0211]

[0212] Step 2.14: Based on equation (27), the mean square error (MSE) is used as the loss function;

[0213]

[0214] In step 3,

[0215] Step 3.1: Train the reconstructed model; Randomly shuffle the dataset and divide it into training and test sets according to a certain ratio, and use the Adam optimizer to train the model;

[0216] Step 3.2: Review all hyperparameters of the reconstructed model and determine the feasible solution domains of the corresponding hyperparameters: number of convolutional layers, kernel size, stride, pooling layer size, upsampling layer size, number of layers in the long short-term memory network and the size of the hidden layers within the unit, number of fully connected layers and the number of neurons per layer, and the proportion of Dropout.

[0217] Step 3.3: Initialize the position information of each particle in the particle swarm. The position of each particle can be represented by equation (28);

[0218] X i =(x i1 ,x i2 ,…,x iD (28)

[0219] In the formula, X i It is the position of the i-th particle; x i1 ,x i2 ,…,x iD The values ​​of the hyperparameters described in step 27;

[0220] Step 3.4: Calculate the fitness value f for each particle. i ; based on the particle's position X i As the hyperparameter values ​​for the reconstruction model, the loss function and optimization method described in steps 2.13 and 2.14 are used to train the model, and the reconstruction error of the trained model is used as the fitness value of the particles.

[0221] Step 3.5: Update the individual historical best fitness value f for each particle. i,p and the individual's historical best position P i,pbest If the particle's current fitness value is less than its historical best fitness value, then update the particle's historical best fitness value and best position.

[0222] Step 3.6: Update the historical best fitness value f of the population. g and optimal position P gbest If the individual best fitness value of a particle is smaller than the best fitness value of the population, then update the best fitness value and best position of the population.

[0223] Step 3.7: Update the velocity of each particle according to equation (29);

[0224]

[0225] In the formula, V i k+1 Let be the velocity of the i-th particle in the (k+1)-th iteration. Let be the velocity of the i-th particle in the k-th iteration; Let be the optimal position of the i-th particle in the k-th iteration. This represents the optimal position of the population during the k-th iteration. c1 represents the position of the i-th particle in the k-th iteration; c2 is the individual learning factor with a value of 2; c2 is the group learning factor with a value of 2; r1 and r2 are random numbers in the interval [0,1]; ω k For adaptive inertia weights, ω k Update according to equation (30), where ω start Take an empirical value of 0.9, ω end Take the empirical value of 0.4;

[0226]

[0227] Step 3.8: Update the position of each particle according to equation (31);

[0228]

[0229] Step 3.9: Repeat steps 3.3 to 3.7 until the maximum number of iterations is reached, and calculate the group's historical optimal position P obtained from multiple iterations. gbest The values ​​of the hyperparameters used to reconstruct the model.

[0230] A sensor fault monitoring and failure information reconstruction system:

[0231] The system includes a model building and fault detection module, a failure information reconstruction module, a hyperparameter automatic optimization module, and a verification and testing module.

[0232] The model building and fault detection module is used to build a sensor model of a nuclear power plant and uses a data-driven principal component analysis method for sensor fault detection and fault identification.

[0233] The failure information reconstruction module, based on the sensor model established by the model building and fault detection module, acquires the operating data of each sensor under different working conditions, performs correlation analysis of different working condition features based on the maximum mutual information coefficient, selects reconstruction reference features, and constructs a sensor failure information reconstruction model.

[0234] The hyperparameter automatic optimization module reconstructs the model based on the failure information of the failure information reconstruction module, and uses an improved particle swarm optimization algorithm to perform hyperparameter automatic optimization, taking the historical optimal position of the population obtained from multiple rounds of iterative calculation as the value of the hyperparameter of the reconstruction model.

[0235] The verification and testing module is used to perform model verification and testing after completing the automatic optimization of hyperparameters; manually insert sensor fault data, call the reconstruction model to reconstruct the fault data, and analyze the error between the reconstructed value and the true value after normalization.

[0236] The reason why this invention can achieve more robust and accurate identification of sensor faults lies in steps 1.6, 1.7, 1.9, and 1.12.

[0237] Step 1.7 calculates the statistical threshold based on the principal space and residual space partitioned in Step 1.6. and Q α As a basis for fault identification, this invention can determine the fault identification basis without fault data, thus reducing the cost of data acquisition while ensuring the accuracy of the model.

[0238] Step 1.9: Based on the principal space and residual space partitioned in Step 1.6, the statistic T of the test vector is calculated. 2 and Q:T 2 This describes the degree to which the test vector deviates from the principal component analysis model in terms of its trend and magnitude of change, while Q quantifies the degree of mismatch between the test vector and the principal component analysis model. Therefore, this invention calculates T... 2 The Q statistic enables effective identification of sensor anomalies.

[0239] Step 1.12 Calculate the Q statistic and T statistic for the test vector pair. 2 The contribution rate of the statistics enables this invention to accurately determine whether the cause of the fault lies in the system or the sensor.

[0240] The reason why this invention can reduce false alarms lies in step 10.

[0241] Step 10: Process T using an SG filter 2 The T statistic and Q statistic curve reduce the T statistic caused by random noise. 2 The fluctuations in the statistical and Q-statistics reduce the frequency of false alarms caused by random noise and enhance the robustness of the fault detection model to random noise.

[0242] T, a test data processed using the Savitzky-Golay filter from the field of signal processing. 2 The T statistic and Q statistic curves are used to reduce the impact of noise on the model's detection results, thereby reducing false alarms and improving the model's robustness. From equations (7) and (8), it can be seen that T... 2 The T-statistic and Q-statistic are highly sensitive to data noise; noise in the test sample can easily distort the T-statistic. 2The failure detection model may issue false alarms due to the statistic and Q-statistic exceeding the threshold. The Savitzky-Golay filter is a smoothing filtering algorithm based on local polynomial least squares fitting in the time domain. Compared to other smoothing filtering algorithms such as moving average filtering, the biggest advantage of the Savitzky-Golay filter is that it can remove noise while ensuring that the shape and width of the signal remain unchanged. The T signal after processing by the Savitzky-Golay filter... 2 The curves of the statistical quantity and the Q-statistic become smoother, which alleviates the false alarm phenomenon caused by a large increase in the statistical quantity due to noise, improves the robustness of the fault detection model based on principal component analysis, and makes the results of the fault detection model more reliable.

[0243] The reason why this invention can achieve faster and more accurate reconstruction of sensor failure information with small reconstruction error is due to the overall implementation of steps 2.3, 2.4, 2.9, 2.10, 2.11, 2.12, 2.13, 3.2, 3.3, 3.4, 3.5, 3.6, 3.7, 3.8, and 3.9.

[0244] Steps 2.3 and 2.4 perform correlation analysis on a large number of sensors in the nuclear power plant by calculating the maximum mutual information coefficient between each sensor and the target sensor. Based on the correlation analysis results, a set of reference sensors for the reconstruction model is selected. Given the large number of sensors in a nuclear power plant, selecting all sensors as reference sensors for the reconstruction model could easily lead to problems such as the "curse of dimensionality" and "overfitting," thereby reducing the performance of the reconstruction model.

[0245] The greatest advantage of the maximum mutual information number is that it can characterize the degree of nonlinear correlation between variables, making it suitable for complex and nonlinear dynamic systems such as nuclear power plants. Therefore, this invention can accurately select a set of reference sensors that are highly correlated with the target sensor, reducing the input dimension of the reconstruction model, alleviating the overfitting problem while reducing the computational burden, and improving the accuracy of the reconstruction model.

[0246] Based on the above issues, it is necessary to select reference sensors for reconstruction, choosing those with high correlation to the target sensor and eliminating those with low correlation. Mutual information is a criterion for measuring the correlation between two random variables based on information entropy theory. Unlike ordinary similarity measurement methods that can only describe linear relationships between variables, mutual information can capture nonlinear statistical correlations between variables, and therefore can be considered to measure and reflect the true dependence between variables to a greater extent. The maximum mutual information coefficient is an excellent variation of mutual information, utilizing normalized mutual information and possessing universality and balance, making it suitable for quantitatively analyzing the degree of correlation between variables. The method of this invention uses the maximum mutual information coefficient to measure the correlation between each sensor and the target sensor, which can largely uncover the statistical correlation between different variables in a complex, nonlinear dynamic system like a nuclear power plant. By selecting a set of sensors with high correlation to the target sensor as reconstruction reference sensors and eliminating sensors irrelevant to the target sensor, the accuracy and generalization ability of the reconstruction model are improved.

[0247] Step 2.9 Establish a convolutional autoencoder model to fully leverage its advantages, extract spatial features from the original data, reduce data noise interference, and improve the model's computational speed and accuracy.

[0248] Step 2.10 establishes a long short-term memory network model, which effectively extracts the time series features of the data and improves the accuracy of the reconstructed values.

[0249] Step 2.11 introduces the Dropout operation to avoid overfitting and reduce reconstruction error.

[0250] Step 2.12 introduces a self-attention mechanism, focusing on the most important information for the current task and reducing attention to other information. This not only improves the model's running speed but also enhances its prediction accuracy to some extent. A self-attention mechanism is introduced between the Long Short-Term Memory network and the fully connected output layer. Self-attention is a resource allocation scheme. It can focus on important tasks within limited computational resources and also address the problem of information overload. The data fitting ability of the reconstructed model stems from its large parameter system. More parameters result in better network fitting ability but also increase the space required to store the model and potentially lead to information overload. By introducing a self-attention mechanism, focusing on the most important parts of the current task and reducing attention to other information reduces the model's computational burden, improving its efficiency while maintaining accuracy.

[0251] Step 2.13 uses the ELU function as the activation function of the fully connected layer. Compared with the traditional ReLU function, it does not have the problems of gradient explosion or vanishing and neuron death, thus improving the convergence speed.

[0252] Steps 3.2 to 35 utilize the particle swarm optimization algorithm to automatically obtain the optimal hyperparameters of the reconstructed model, avoiding the uncertainty of manually selecting hyperparameters and ensuring the accuracy of the model.

[0253] Step 3.7 replaces the constant weights with nonlinear adaptive inertia weights, allowing the inertia weights to automatically adjust following the iteration process, effectively preventing the particle swarm algorithm from getting trapped in local optima. To avoid the time and effort required for manually selecting hyperparameters of the reconstruction model, the method of this invention uses the particle swarm algorithm to automatically find the optimal hyperparameter combination. In the velocity evolution equation (28) of the particle swarm algorithm, the first part V i k To ensure the global convergence performance of the algorithm, the second and third parts enable the algorithm to achieve local convergence. Introducing an inertial weight ω balances global and local convergence capabilities, improving the algorithm's optimization performance. In global search algorithms, it is desirable to have a larger inertial weight ω in the early stages to achieve higher search capability, and a smaller inertial weight ω in the later stages to accelerate convergence. The method of this invention employs adaptive inertial weights, allowing the inertial weights to automatically adjust following the iteration process. This results in a higher search capability in the early stages and a faster convergence speed in the later stages, avoiding the problem of particle swarm optimization getting trapped in local optima.

[0254] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.

[0255] A computer-readable storage medium for storing computer instructions that, when executed by a processor, implement the steps of the above-described method.

[0256] The memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory of the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0257] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. 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 via wired means such as coaxial cable, optical fiber, digital subscriber line, DSL, or wireless means such as infrared, wireless, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more 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 high-density digital video disc, DVD; or a semiconductor medium such as a solid-state disk, SSD, etc.

[0258] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0259] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, 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, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as execution by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0260] The present invention has provided a detailed description of a sensor fault monitoring and failure information reconstruction method and system, and has explained the principles and implementation methods of the present invention. The above description of the embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for sensor fault monitoring and failure information reconstruction, characterized in that: The method specifically includes the following steps: Step 1: Establish a sensor model for the nuclear power plant and use a data-driven principal component analysis method for sensor fault detection and identification. Step 2: Based on the sensor model established in Step 1, obtain the operating data of each sensor under different working conditions, perform correlation analysis of different working condition features based on the maximum mutual information coefficient, select reconstruction reference features, and construct a sensor failure information reconstruction model. Step 2.1: If the fault source is the target sensor, then call the reconstruction model to reconstruct the data of the failed sensor; Step 2.2: Use a full-range nuclear power plant simulator to obtain the operating data of each sensor under the target operating conditions; Step 2.3: Analyze the statistical correlation between sensors; calculate the maximum mutual information coefficient between the reference sensor and the target sensor under different power steady-state conditions and various accident conditions according to Equation (14); (14) In the formula, express X and Y The maximum mutual information coefficient; Indicates a certain M, N Under the given value, X and Y The maximum mutual information value; Step 2.4: Select the reconstruction reference sensor; select the top ten sensors with the largest mutual information coefficient under steady-state conditions, and select the top five sensors with the largest mutual information coefficient under other accident conditions. Then, take the union of the selected sensors to obtain the reconstruction model reference sensor combination. Step 2.5: Construct the reconstructed model dataset; process the full-range simulator running data, remove sensors with low correlation to the target sensor, and retain the sensors selected in Step 2.4; Step 2.6: Data standardization processing; According to equation (15), the sensor operation data selected in step 2.5 is subjected to min-max standardization processing to eliminate the influence of different sensor dimensions and accelerate the convergence speed of the reconstruction model. (15) In the formula, This represents the normalized data from a specific sensor. x This represents the data of a sensor before normalization; min represents the minimum value in the sample data of a sensor; max represents the maximum value in the sample data of a sensor. Step 2.7: Construct the input data for the reconstructed model using the sliding window method. The data processing time interval is 0.25s, and the sliding window length is set to 2.5s. The two-dimensional tensor obtained in Step 2.6 ( Convert to a three-dimensional tensor ; Step 3: Reconstruct the model based on the failure information from Step 2, use the improved particle swarm optimization algorithm to automatically optimize hyperparameters, and use the historical optimal position of the population obtained from multiple rounds of iterative calculation as the value of the hyperparameters of the reconstructed model. Step 4: After completing Step 3, perform model verification and testing; manually insert sensor fault data, call the reconstruction model to reconstruct the fault data, and analyze the error between the reconstructed value and the true value after normalization.

2. The method according to claim 1, characterized in that: In step 1, Step 1.1: Use a nuclear power plant full-range simulator to acquire raw sensor measurement data under various operating conditions during normal operation. ,in n The total number of sensors to be monitored. m To measure the total number of samples, a sensor model of the nuclear power plant was established; (1) Step 1.2: Based on equations (1) and (2), process the original measurement data. Standardize using standard deviation; (2) In the formula, For the first j Raw operating data from each sensor; For the first j The average of the raw operating data of each sensor; It is the first j The standard deviation of the raw operating data of each sensor; It is the first j Standardized data from each sensor The mean is 0 and the standard deviation is 1. The matrix is ​​the standardized result of the original measurement data; Step 1.3: Based on equation (3), approximately solve for the normalized matrix. covariance matrix ; (3) Step 1.4: According to the matrix decomposition theorem, for Perform eigenvalue decomposition to obtain eigenvalues and the corresponding standard feature vector ; (4) In the formula, the orthogonal matrix ; Indicated by For diagonal elements n A diagonal matrix of order 1; Step 1.5: Use the cumulative variance contribution rate method to... Selection of criteria for judgment k One principal element; (5) In the formula, eigenvalues This reflects the amount of information contained in each principal element; PV Indicates the first i The percentage of information contained in each principal element relative to the total amount of information in the original data; CPV Indicates the preceding k The percentage of original information contained in each principal element; Step 1.6: Before use k The eigenvectors of each principal element constitute the principal element space. and the remainder The residual space is composed of principal elements. Then the test vector at a certain moment can be represented by equation (6): (6) In the formula, For a test vector at a certain time, it contains the values ​​at that time. n The readings of each sensor; They are Projection in principal space and residual space; The projection matrix of the principal element space. The residual space projection matrix; Step 1.7: Based on equations (7) and (8), calculate Hotelling's... T 2 ,Right now T 2 Threshold of statistics And the squared prediction error SPE, i.e. Threshold of statistics ; (7) (8) In the formula, Indicates having and With degrees of freedom and a confidence level of [value missing], the following values ​​are given: of The critical value of the distribution; for eigenvalues; Number of principal elements; For standard normal distribution Confidence limits.

3. The method according to claim 2, characterized in that: Step 1 also includes: Step 1.8: Begin fault detection by inputting real-time sensor data test vectors. And according to equation (2), it is standardized to obtain ; Step 1.9: Calculate the test vector at this moment according to equations (9) and (10). of T 2 Statistics and Q Statistic; (9) (10) In the formula, ; Step 1.10: Process using the SG filtering method according to equation (11). T 2 Statistics and Q The curve of how the statistic changes over time; (11) In the formula, These are the fitted values ​​of the corresponding statistics; These are the original input values ​​for the corresponding statistics; For the first i The coefficients for filtering individual statistics; m The width is half the width of the filter window; N The filter length is equal to the width of the sliding array (2). m +1); Step 1.11: If the result after processing by the SG filtering method T 2 Statistics and Q If none of the statistical values ​​exceed the corresponding threshold, it indicates that the system is operating normally; If processed by the SG filtering method T 2 Statistic or Q If the statistic exceeds the corresponding threshold, it indicates an anomaly. Step 1.12: Perform fault identification and calculate according to equations (12) and (13). The Middle i Each sensor pair Q Contribution rate of statistics and against T 2 Contribution rate of statistics Locating test vectors An abnormal sensor was detected in the middle; If the contribution rates of each sensor are evenly distributed, it indicates that the cause of the anomaly is a system-level fault. If the contribution rate of a certain sensor is significantly higher than normal, it indicates that the abnormality is caused by the failure of that sensor. (12) (13) In the formula, t express The score vector, ; Representing sensor variables For the score vectors The contribution rate; .

4. The method according to claim 3, characterized in that: Step 2 also includes, Step 2.8: Build a deep learning network structure in the TensorFlow framework, consisting of a convolutional autoencoder, a long short-term memory network, and a self-attention mechanism. The model is composed of an input layer, multiple convolutional layers, pooling layers, upsampling layers, multiple long short-term memory network layers, a self-attention layer, a fully connected layer, and an output layer connected layer by layer. Step 2.9: Construct the encoding layer of the convolutional autoencoder by combining convolutional and pooling layers; use one-dimensional convolution operations to extract features from the data; A decoding layer of a convolutional autoencoder is constructed by combining convolutional layers and upsampling layers to restore the data features to the original data; the data is then output to a long short-term memory network layer after passing through multiple layers of convolutional autoencoders. Step 2.10: Construct a Long Short-Term Memory (LSTM) network; the LSTM network uses a gate mechanism to control the flow and loss of information; represented as follows: Input gate (threshold): determines the proportion of information that will be stored in the current cell state; (16) Forgetting gate: Selectively forgets information in the cellular state; (17) Output gate (threshold): Selectively outputs information from the cell state; (18) Of the three formulas above, It is the parameter matrix to be trained; It is the bias term to be trained; It is the sigmoid activation function; memory The short-term memory is represented by the output gate obtained from the current cell state: (19) Candidate state The new knowledge to be stored in the cell state, derived from characterization and induction, is the input feature at the current moment. Short-term memory of the previous moment Functions: (20) Cellular state This refers to long-term memory, which is equivalent to the long-term memory from the previous moment. The value of the forget gate and the new knowledge summarized at the current moment. Sum of values ​​from the input gate: (21) The overall calculation process is as follows: the forget gate, input gate, and output gate use the sigmoid function to determine the content to be discarded from the information in the previous time step. f t Content that needs to be retained i t and the content to be output o t The content output at this moment o t With cellular state C t Determines the memory h t The memory of the previous moment h t-1 and the input at this moment x t The candidate state at that moment was determined. Input gate at this moment i t Output gate o t The cellular state at the previous moment C t and the candidate states at this moment Together they determine the cell state at this moment. C t Then, the calculation is repeated for each time step in this manner. Step 2.11: Add a Dropout operation between the Long Short-Term Memory (LSTM) network layer and the self-attention layer; Step 2.12: Introduce a self-attention mechanism; first, analyze the output of the Long Short-Term Memory network based on equations (22), (23), and (24). X Perform a linear transformation to obtain Q, K, and V; (22) (23) (24) In the formula, W Q W K and W V It is the parameter matrix to be trained; The output vector of the self-attention layer can be calculated using equation (25): (25) In the formula, Connect the weights to the positions of the output and input vectors. Dynamically generated by attention mechanisms; Step 2.13: Based on equation (26), the ELU function is used as the activation function for the fully connected layers of the model. The value is 0.2; (26) Step 2.14: Based on equation (27), the mean square error (MSE) is used as the loss function; (27)。 5. The method according to claim 4, characterized in that: In step 3, Step 3.1: Train the reconstructed model; Randomly shuffle the dataset and divide it into training and test sets according to a certain ratio, and use the Adam optimizer to train the model; Step 3.2: Review all hyperparameters of the reconstructed model and determine the feasible solution domains of the corresponding hyperparameters: number of convolutional layers, kernel size, stride, pooling layer size, upsampling layer size, number of layers in the long short-term memory network and the size of the hidden layers within the unit, number of fully connected layers and the number of neurons per layer, and the proportion of Dropout. Step 3.3: Initialize the position information of each particle in the particle swarm. The position of each particle can be represented by equation (28); (28) In the formula, It is the first i The position of each particle; These are the values ​​of the hyperparameters described in step 3.2; Step 3.4: Calculate the fitness value of each particle. Based on the position of the particle As the hyperparameter values ​​for the reconstruction model, the loss function and optimization method described in steps 2.13 and 2.14 are used to train the model, and the reconstruction error of the trained model is used as the fitness value of the particles. Step 3.5: Update the individual historical best fitness value of each particle. and individual historical best position If the particle's current fitness value is less than its historical best fitness value, then update the particle's historical best fitness value and best position. Step 3.6: Update the historical best fitness value of the population and optimal position ; If the individual best fitness value of a particle is smaller than the best fitness value of the population, then update the best fitness value and best position of the population. Step 3.7: Update the velocity of each particle according to equation (29); (29) In the formula, For the first i The particle in the first Speed ​​during round iteration; For the first i The particle in the first k Speed ​​during round iteration; For the first i The particle in the first k The optimal position of an individual during round iteration; For the first k The optimal position of the population during round iteration; Indicates the first i Individual particles k Position during round iteration; The individual learning factor has a value of 2. The group learning factor has a value of 2. For interval Random numbers within; For adaptive inertia weights, Update according to formula (30), where Take the empirical value of 0.

9. Take the empirical value of 0.4; (30) Step 3.8: Update the position of each particle according to equation (31); (31) Step 3.9: Repeat steps 3.3 to 3.7 until the maximum number of iterations is reached, and calculate the group's historical optimal position from multiple iterations. The values ​​of the hyperparameters used to reconstruct the model.

6. A sensor fault monitoring and failure information reconstruction system, characterized in that: The system is based on the sensor fault monitoring and failure information reconstruction method according to any one of claims 1 to 5; The system includes a model building and fault detection module, a failure information reconstruction module, a hyperparameter automatic optimization module, and a verification and testing module. The model building and fault detection module is used to build a sensor model of a nuclear power plant and uses a data-driven principal component analysis method for sensor fault detection and identification. The failure information reconstruction module, based on the sensor model established by the model building and fault detection module, acquires the operating data of each sensor under different working conditions, performs correlation analysis of different working condition features based on the maximum mutual information coefficient, selects reconstruction reference features, and constructs a sensor failure information reconstruction model. Step 2.1: If the fault source is the target sensor, then call the reconstruction model to reconstruct the data of the failed sensor; Step 2.2: Use a full-range nuclear power plant simulator to obtain the operating data of each sensor under the target operating conditions; Step 2.3: Analyze the statistical correlation between sensors; The maximum mutual information coefficients between the reference sensor and the target sensor under different power steady-state conditions and various accident conditions are calculated according to Equation (14); (14) In the formula, express X and Y The maximum mutual information coefficient; Indicates a certain M, N Under the given value, X and Y The maximum mutual information value; Step 2.4: Select the reconstruction reference sensor; The top ten sensors with the largest mutual information coefficient under steady-state conditions are selected, and the top five sensors with the largest mutual information coefficient under other accident conditions are selected. Then, the union of the selected sensors is taken to obtain the reference sensor combination for the reconstruction model. Step 2.5: Construct the reconstructed model dataset; process the full-range simulator running data, remove sensors with low correlation to the target sensor, and retain the sensors selected in Step 2.4; Step 2.6: Data standardization processing; According to equation (15), the sensor operation data selected in step 2.5 is subjected to min-max standardization processing to eliminate the influence of different sensor dimensions and accelerate the convergence speed of the reconstruction model. (15) In the formula, This represents the normalized data from a specific sensor. x This represents the data of a sensor before normalization; min represents the minimum value in the sample data of a sensor; max represents the maximum value in the sample data of a sensor. Step 2.7: Construct the input data for the reconstructed model using the sliding window method. The data processing time interval is 0.25s, and the sliding window length is set to 2.5s. The two-dimensional tensor obtained in Step 2.6 ( Convert to a three-dimensional tensor ; The hyperparameter automatic optimization module reconstructs the model based on the failure information of the failure information reconstruction module, and uses an improved particle swarm optimization algorithm to perform hyperparameter automatic optimization, taking the historical optimal position of the population obtained from multiple rounds of iterative calculation as the value of the hyperparameter of the reconstruction model. The verification and testing module is used to perform model verification and testing after completing the automatic optimization of hyperparameters; manually insert sensor fault data, call the reconstruction model to reconstruct the fault data, and analyze the error between the reconstructed value and the true value after normalization.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 5.

Images