A method for multi-sensor failure detection and signal reconstruction in nuclear power plants

Abstract

A method for failure detection and signal reconstruction of multiple sensors in a nuclear power plant relates to the technical field of sensor measurement and signal reconstruction. The method comprises the following steps: obtaining measurement data of sensors in a nuclear power plant during normal operation in real time; constructing a training set after preprocessing the measurement data; extracting part of the training set data, simulating sensor failures by using a sensor failure mechanism model, obtaining simulated sensor failure data, and constructing a test set; detecting the working state of the sensors in the nuclear power plant based on the training set and the test set data, locating the specific position of the faulty sensors, and outputting normal sensor data or faulty sensor position data; reconstructing the failure data of the faulty sensors by using a local fusion method, and outputting the reconstructed sensor data; the method can accurately determine the working state of the sensors in the nuclear power plant.

Description

Technical Field

[0001] The present invention relates to the technical field of sensor measurement and signal reconstruction, and in particular to a multi-sensor failure detection and signal reconstruction method for a nuclear power plant. Background Art

[0002] Under harsh conditions such as high temperature, high pressure, high corrosion and high radioactivity, sensors in nuclear power plants may age or fail to varying degrees as they work longer or are accidentally damaged, causing sensor data drift, reduced accuracy or even complete failure, which can lead to misjudgment by operators or systems, and even induce erroneous control measures, thus seriously threatening the safety of nuclear facilities.

[0003] However, since it is difficult to directly obtain the fault data of nuclear power plant sensors, relevant algorithms are generally used to carry out failure detection of sensors and reconstruction of failed sensor signals to accurately detect, isolate and reconstruct the signals of faulty sensors, ensuring that nuclear facilities can still operate safely and stably when sensor failures occur. This plays an important role in improving the nuclear facility's ability to prevent sensor failures and accidents, as well as its ability to adapt to harsh environments.

[0004] Traditionally, a single neural network algorithm is used to determine whether a sensor is faulty. However, it is difficult to accurately locate the faulty sensor, which increases the difficulty of troubleshooting for staff and affects the operation of nuclear power plants.

[0005] The Chinese patent (CN114004148A) is a method for filtering failure data of mining catalytic sensors based on neural networks. The neural network is applied to the sensor system to determine the failure data and filter it, thereby improving the detection accuracy of the sensor. However, for the operating system of a nuclear power plant, since the operating environment of the sensor is more severe and the types of threats it encounters are more numerous, it is impossible to accurately determine whether the sensor has failed only through the neural network.

[0006] Therefore, we propose a method that can accurately determine the working status of sensors in nuclear power plants. Summary of the invention

[0007] In order to overcome the deficiencies in the background technology, the present invention discloses a method for multi-sensor failure detection and signal reconstruction in a nuclear power plant, which can reconstruct signals in the event of sensor failure, improve the reliability of nuclear facility sensors, and enhance the nuclear facility's ability to prevent sensor failures and accidents, as well as its adaptability in harsh environments.

[0008] In order to achieve the above-mentioned invention object, the present invention adopts the following technical scheme:

[0009] A method for multi-sensor failure detection and signal reconstruction in a nuclear power plant, specifically:

[0010] Real-time acquisition of measurement data from nuclear power plant sensors during normal operation;

[0011] After preprocessing the measured data, it is constructed as a training set;

[0012] Extract part of the training set data, use the sensor failure mechanism model to simulate the sensor failure, obtain the simulated sensor failure data, and construct it as a test set;

[0013] Based on the training set and test set data, detect the working status of the nuclear power plant sensors, locate the specific location of the faulty sensors, and output the normal sensor data or the faulty sensor location data;

[0014] The failure data of the faulty sensor is reconstructed by local fusion, and the reconstructed sensor data is output.

[0015] Furthermore, the step of preprocessing the measurement data is specifically as follows:

[0016] Use FMEA method to determine the data set of key sensors;

[0017] Usage Improvements The criterion is to remove outlier data from the data set;

[0018] The measurement data after removing the outlier is reconstructed by taking the average of the two measured values ​​before and after the outlier.

[0019] The wavelet packet method is used to reduce the noise of the reconstructed data.

[0020] Furthermore, the AANN method and the KPCA method are used to detect the working status of the nuclear power plant sensor.

[0021] Furthermore, the steps of using the AANN method to detect the working status of the nuclear power plant sensor are specifically as follows:

[0022] Construct AANN sensor state detection model;

[0023] The training set is used as input data of the AANN sensor state detection model, and the output data of the sensor state detection model is obtained;

[0024] Calculate the residual between input data and output data;

[0025] Determine whether the residual exceeds the limit value;

[0026] If the residual exceeds the limit value, it is judged as a sensor failure, and the specific location of the faulty sensor is searched by searching the fault combination;

[0027] If the residual does not exceed the limit value, the sensor is judged to be normal and the normal sensor data is output.

[0028] Furthermore, the steps of using the KPCA method to detect the working status of the nuclear power plant sensor are specifically as follows:

[0029] Based on the training set data, the KNN algorithm is used to determine the current operating conditions of the nuclear power system;

[0030] The KPCA model trained with the test set data closest to the current operating conditions is selected;

[0031] Calculate the statistical residuals of the KPCA model;

[0032] Determine whether the residual exceeds the limit value;

[0033] If the residual exceeds the limit value, it is judged as a sensor failure, and the specific position of the corresponding faulty sensor is located;

[0034] If the residual does not exceed the limit value, the sensor is judged to be normal and the normal sensor data is output.

[0035] Furthermore, the failure data reconstruction method of the faulty sensor is specifically as follows:

[0036] The reconstruction of failure data is achieved by iterative calculation of KPCA model;

[0037] The IPSO-RBF method is used to achieve the reconstruction of failure data;

[0038] The local fusion processing method is used to fuse the calculation results of the two reconstruction algorithms to obtain an accurate reconstructed signal.

[0039] Furthermore, the method of implementing failure data reconstruction using the IPSO-RBF method comprises the following specific steps:

[0040] Initialize RBF model parameters and set IPSO algorithm parameters;

[0041] Initialize the individual extreme value positions and corresponding individual extreme values ​​of particles, and the global extreme value positions and corresponding global extreme values;

[0042] Calculate the fitness values ​​of all particles;

[0043] Compare the obtained individual extreme value positions, individual extreme values, global extreme value positions and global extreme values;

[0044] Update individual extreme values ​​and global extreme values ​​according to the comparison results;

[0045] Update the speed and position of the particle and make them within the specified range;

[0046] If the number of iterations or the cutoff accuracy is reached, the iteration ends; otherwise, return to the step of initializing the particle extremum.

[0047] Furthermore, the specific steps of the local fusion method are:

[0048] Use the KNN algorithm to find the historical signal that is closest to the reconstructed signal;

[0049] Calculate the distance between the reconstructed signal and the historical signal;

[0050] This distance is used as a basis for measuring the weight of the reconstructed signal in the local fusion process;

[0051] Perform local fusion and calculate the fusion result.

[0052] Furthermore, the calculation formula of the fusion result is:

[0053] ;

[0054] in, represents the reconstructed signal value after local fusion, represents the deviation correction of the reconstructed signal, Represents the weight correction of the reconstructed signal, represents the number of fusion models, The data value reconstructed for the algorithm.

[0055] Furthermore, the deviation correction amount of the reconstructed signal and weight correction The calculation formulas are:

[0056] ;

[0057] ;

[0058] in, It represents the actual sensor signal set that is close to the model reconstruction value calculated by KNN. Represents the neighboring point data in the KNN historical normal sensor data set, It represents the maximum value of the distance between the reconstructed signal and the actual signal in the nearest neighbor set of each model; k represents the number of neighbor points selected by KNN, and i represents the i-th reconstructed model.

[0059] Due to the adoption of the above-mentioned technical solution, the present invention has the following beneficial effects:

[0060] The present invention discloses a method for multi-sensor failure detection and signal reconstruction in a nuclear power plant. By constructing a training set and a test set, the fault detection algorithm is trained and verified using simulated fault data, which helps to improve the accuracy of fault detection. By using multiple means to reconstruct the signal and combining the local fusion technology, the accuracy of signal reconstruction can be significantly improved. Therefore, the method can be operated efficiently in the complex multi-operating environment of the nuclear power plant, improve the nuclear facility's ability to prevent sensor failures and accidents, and enhance the system's adaptive ability in harsh environments.

[0061] In addition, data preprocessing can improve the accuracy of fault detection, reduce the impact of data anomalies on detection results, maintain data integrity, and improve signal clarity and accuracy;

[0062] In addition, the AANN method is used to detect the working status of nuclear power plant sensors, which can further improve the accuracy and reliability of sensor working status detection;

[0063] In addition, the key parameters of the RBF neural network are optimized using the improved particle swarm optimization algorithm IPSO, which can improve the generalization ability and accuracy of the model; and the KNN method is used to find the historical signal closest to the reconstructed signal and calculate the weight of the reconstructed signal based on the distance, which helps to improve the accuracy of the reconstructed signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0065] Figure 2 This is a schematic diagram of the failure detection and signal reconstruction technology route;

[0066] Figure 3 For improved Schematic diagram of the outlier reconstruction calculation process of the criterion;

[0067] Figure 4 It is a schematic diagram of the AANN structure;

[0068] Figure 5 This is a schematic diagram of the sensor condition monitoring technology route based on KPCA;

[0069] Figure 6 Figure 2 is a flow chart of the IPSO-RBF algorithm. DETAILED DESCRIPTION

[0070] The technical solution of the present invention will be described below in conjunction with the drawings in the embodiments of the present invention. In the description, it should be understood that if there are terms such as "upper", "lower", "front", "back", "left", "right" and the like indicating directions or positional relationships, they are only corresponding to the drawings of the present invention, for the convenience of describing the present invention, and do not indicate or imply that the devices or components referred to must have a specific direction:

[0071] Embodiment 1 is:

[0072] Combined with Figure 1-2 The method for multi-sensor failure detection and signal reconstruction in a nuclear power plant is specifically as follows:

[0073] Real-time acquisition of measurement data from nuclear power plant sensors during normal operation;

[0074] After preprocessing the measured data, it is constructed as a training set;

[0075] Extract part of the training set data, use the sensor failure mechanism model to simulate the sensor failure, obtain the simulated sensor failure data, and construct it as a test set;

[0076] Based on the training set and test set data, detect the working status of the nuclear power plant sensors, locate the specific location of the faulty sensors, and output the normal sensor data or the faulty sensor location data;

[0077] The failure data of the faulty sensor is reconstructed by local fusion, and the reconstructed sensor data is output.

[0078] Embodiment 2 is:

[0079] Based on Example 1, Figure 3 , the preprocessing step is further defined, that is, the step of preprocessing the measurement data is:

[0080] The FMEA method (Failure Modes and Effects Analysis) used is a systematic risk assessment tool that determines the data set of key sensors. Since there are many sensors in nuclear power plants and the usage frequencies of each sensor are different, the failure probability of each sensor is also different. Therefore, key sensors are determined in advance to reduce the workload of subsequent sensor working status detection, thereby effectively improving the efficiency of subsequent sensor fault detection;

[0081] Key sensor data Under normal operating conditions, the sensor x exist i The measured value before x (1) to x ( i ), two adjacent measurement pointsx ( j )and x ( j -1) can be expressed as:

[0082] (1)

[0083] exist i Before the moment, the sensor x The average rate of change between all adjacent measurement points, that is:

[0084] (2)

[0085] according to i The average rate of change of all adjacent points before the moment can be calculated i Sensor before time x The standard deviation of the change in the measured values ​​of two adjacent points ,Right now:

[0086] (3)

[0087] After obtaining the normal operation data of the nuclear power plant sensors, the improved Criteria and wavelet packet method for preprocessing;

[0088] According to the classic Criterion: If the change of sensor x at the current moment x(i) and the previous moment x(i-1) Greater than ,Right now or less than ,Right now

[0089] (4)

[0090] The current measurement value x(i) of sensor x is considered to be an outlier and needs to be removed from the measurement data. After removing the outliers from the original data, in order to ensure the integrity of the training sample matrix, the removed outliers need to be reconstructed. The average of the two measurements before and after the outlier is used. Approximately replace the removed outlier point x(i), that is:

[0091] (5)

[0092] Aiming at the random noise abnormal data in the normal operation of nuclear power plant sensors, the wavelet packet algorithm is used to remove the environmental noise of sensor measurement signals.

[0093] By determining the wavelet packet basis and the number of wavelet packet decomposition layers, the sensor measurement signal is decomposed by wavelet packet, and the wavelet packet coefficients of each decomposition layer are extracted.

[0094] A threshold is selected for each wavelet packet coefficient extracted to perform threshold quantization processing. Points whose absolute values ​​of wavelet packet coefficients are greater than the selected threshold are set to zero, and points whose absolute values ​​are less than the threshold are retained and their differences are taken.

[0095] The low-frequency wavelet packet coefficients of the first layer and the high-frequency wavelet packet coefficients of the remaining layers are reconstructed to obtain the wavelet packet denoised signal of the sensor.

[0096] The specific process of preprocessing by improving the criterion and wavelet packet method is as follows:

[0097] 1) Get i All measurements before time x (1) x (2) … , x ( i ), use formula (3) to calculate the standard deviation of the variation between adjacent points;

[0098] 2) Compare the change in the sensor measurement value at the current moment and The size of Greater than or less than , then the sensor x Current measurement value x ( i ) is an outlier, remove it and go to step 3); otherwise return to step 1) and proceed to the next moment x ( i+ 1) Check whether it is an outlier;

[0099] 3) Reconstructing outliers in sensor measurements x ( i ): x ( i )=( x ( i- 1)+ x ( i+ 1)) / 2 ;

[0100] 4) After outlier extraction, the wavelet packet algorithm is used to remove random noise abnormal data from the operating data.

[0101] In addition, the construction process of the test set is as follows: intercept some normal operating data after preprocessing, combine it with the sensor fault model, and simulate typical faults such as constant gain, constant deviation, drift, etc. of the sensor through fault injection to obtain sensor status monitoring test data. The injection method is as follows:

[0102] For any sensor, its measured value x at a certain moment can be expressed as follows:

[0103] ;

[0104] In the formula, Indicates the true value of the measured parameter at the time of measurement; Indicates the systematic error of the measured parameter at the time of measurement; Represents the random error or noise in the measurement process.

[0105] generally Mainly caused by sensor failure, It is caused by the uncertainty of sensor measurement accuracy or measurement process. is an unpredictable quantity, but a large number of experiments have shown that noise It obeys the normal distribution, that is, it is white noise. To simplify the simulation analysis of sensor failure, it is assumed that the random error of the sensor measurement process is white noise. Therefore, different types of sensor failures can be expressed as different form.

[0106] a. Sensor drift failure:

[0107] Sensor drift fault refers to the deviation between the sensor output value and the actual value of the measured parameter. It increases over time. Drift failure may be caused by environmental factors, electromagnetic interference, temperature changes, working years, etc. It will gradually change the accuracy and robustness of the sensor. The mathematical expression of sensor drift failure is as follows:

[0108] ;

[0109] In the formula, Indicates the actual value of the measured parameter; It indicates the systematic error of the measured parameter at the time of measurement, which is mainly caused by sensor failure; represents a constant, which characterizes the drift degree of the sensor output value x; t and They represent the current time and the time when the drift fault occurs respectively. It represents the random error or noise in the measurement process, which is mainly caused by the measurement accuracy of the sensor or the uncertainty of the measurement process. Generally, it conforms to the normal distribution.

[0110] b. Sensor fixed deviation failure:

[0111] The sensor fixed deviation fault means that the output value of the sensor The actual value of the measured parameter The deviation between Does not change with time; fixed offset fault may be caused by bias current or bias voltage; the mathematical expression of sensor fixed offset fault is as follows:

[0112] ;

[0113] In the formula, Represents a constant that represents the output value of the sensor The measured value The degree of deviation between them.

[0114] c. Accuracy reduction failure

[0115] Accuracy degradation fault refers to sensor system error Beyond its allowable range, the sensor's measurement capability deteriorates and its accuracy decreases. When the accuracy level decreases, the average value of the measurement does not change, but the variance of the measurement changes, so it can also be called a measurement variance failure. Assume that the noise of sensor x in the measurement process follows a mean of 0 and a variance of The standard normal distribution of . That is:

[0116] ;

[0117] When x has a precision degradation failure, the increase in sensor measurement variance is , then the system error satisfy:

[0118] ;

[0119] Since the sum of multiple normal distributions is still a normal distribution, the output value of the sensor can be expressed as:

[0120] ;

[0121] According to this formula, the sensor accuracy degradation failure can be approximated as an increase in random errors in the measurement process.

[0122] Embodiment 3 is:

[0123] On the basis of Example 1, the method for detecting the working state of the nuclear power plant sensor is further limited, that is, the AANN method, i.e., Auto Associative Neural Networks (AANN) and the KPCA method, i.e., Kernel Principal Component Analysis (KPCA), are used to detect the working state of the nuclear power plant sensor;

[0124] In addition, combined with Figure 4,The steps of using AANN method to detect the working status of nuclear power plant sensors are as follows:

[0125] 1) Construct an AANN sensor state detection model;

[0126] 2) Using the training set as the input data of the AANN sensor state detection model and obtaining the output data of the sensor state detection model;

[0127] 3) Calculate the residual between input data and output data;

[0128] And the specific calculation formula of residual SSE is:

[0129] ;

[0130] in, X represents the AANN input data matrix, Y Represents the AANN output data matrix

[0131] 4) Determine whether the residual exceeds the limit value;

[0132] 5) If the residual does not exceed the limit value, the sensor is judged to be normal and the normal sensor data is output;

[0133] 6) When there is a fault in the sensor set, locate the faulty sensor. The specific location process is as follows:

[0134] In order to find the location of the faulty sensor, the fault combination search method is used to search for the faulty sensor. Assume there are n sensors , where m sensors fail, the operation of locating the sensors by “searching by fault combination” is as follows:

[0135] a. Determine the sensor failure combination

[0136] According to the permutations and combinations, there are n sensors, among which m sensors fail. The possible combinations are For example, there are 1, 2, 3, 4, and 5 sensors, and the possible combinations of two sensors failing are (1, 2), (1, 3)…, (4, 5).

[0137] b. Increase or decrease the sensor value for each fault combination

[0138] Take the case of two sensors failing as an example. Suppose there are five sensors in total, and their component vectors are , assuming that sensors 1 and 4 fail, then and Compared with the normal state, there will be an increase or decrease and , then the vector under fault condition can be expressed as , due to the increase or decrease and This will cause the residual SSE of the AANN output to increase. If we give each faulty sensor combination a small increment Or reduce the amount , then each sensor combination must perform the following 9 steps:

[0139] Step 1: Add a small increment to the faulty sensor combination under consideration ( , )

[0140] Step 2: Reduce the faulty sensor combination under consideration by a small increment ( , )

[0141] Step 3: The first sensor adds an increment and the second sensor subtracts an increment ( , )

[0142] Step 4: The first sensor decreases by one increment and the second sensor adds one increment ( , )

[0143] Step 5: The first sensor decreases by one increment and the second sensor remains unchanged ( ,0)

[0144] Step 6: Add an increment to the first sensor and keep the second sensor constant ( ,0)

[0145] Step 7: The first sensor remains unchanged and the second sensor decreases by one increment (0, )

[0146] Step 8: The first sensor remains unchanged and the second sensor adds an increment (0, )

[0147] Step 9: Both sensors keep their current values ​​unchanged (0, 0)

[0148] in Indicates a small step size.

[0149] If the formula is used, it can be expressed as (taking the combination of sensors 1 and 4 as an example):

[0150] Step 1:

[0151] Step 2:

[0152] Step 3:

[0153] Step 4:

[0154] Step 5:

[0155] Step 6:

[0156] Step 7:

[0157] Step 8:

[0158] Step 9:

[0159] From the above steps, we can see that if , and When the numbers are of different signs, SSE will decrease, and when they are of the same sign, SSE will increase. Therefore, the above steps gradually increase If the SSE of one step is reduced to the corresponding training accuracy (threshold), it means that the corresponding faulty sensor combination has been found.

[0160] c. Training accuracy (threshold)

[0161] When training AANN, the mean square error is used as the training indicator of AANN, and the corresponding mean square error (training accuracy) is generally set to be less than 1e-4. When normal data is input into AANN as test data, its training accuracy will be less than 1e-4.

[0162] d. Repeat step b until all sensors are traversed and a set of faulty sensors is obtained.

[0163] According to the needs, the process of using the KPCA method to detect the working status of the nuclear power plant sensor is as shown in the attached Figure 5 As shown, the present invention adopts Statistics and Q Statistics monitor the status of key sensor sets. Since KPCA learns the distribution information of sensor signals, and the distribution information of sensor signals under different working conditions or even variable working conditions is different, the KNN algorithm (K-Nearest Neighbor) is selected. It is a supervised machine learning algorithm to determine the current operating conditions of the nuclear power system and select the KPCA model trained by the sensor data closest to the current operating conditions to achieve adaptive selection of the KPCA model and improve the adaptability of the KPCA model to the current working conditions.

[0164] After selecting the KPCA model, we further determine the Control Threshold and Q Control Threshold ,in;

[0165] (1) The statistic represents the change of the measurement vector x in the principal component space. Its mathematical meaning is the distance from the projection point of x in the principal component space to the center point of the sample mean. It describes the degree to which x deviates from the KPCA model in terms of change trend and change amplitude, and is a measure of the internal changes in the principal component space. The statistic is defined as:

[0166] ;

[0167] In the formula , λ is the eigenvalue; express Statistics threshold, express When the sensor is in a normal state, its measurement data obeys the multivariate normal distribution. The statistic falls on within the range. It can be approximately calculated as follows:

[0168] ;

[0169] In the formula, represents a probability with k and nk degrees of freedom and a confidence level of The critical value of the F distribution.

[0170] (2) The Q statistic describes the square value of the projection of the measurement vector x in the residual space, that is, the Euclidean distance from x to the principal component space defined by the model, and quantifies the degree of mismatch between x and the KPCA model. Therefore, the Q statistic is a measure of the internal changes in the residual space, converting multidimensional data into one-dimensional data for analysis. The definition is as follows:

[0171] ;

[0172] In the formula express After simplifying the mk eigenvectors, we can get:

[0173] ;

[0174] In the normal state The projection in the residual space is very small. The statistical value is When a system failure occurs, the destruction of the correlation between variables leads to In the residual space, the projection increases, and the correlation between variables is destroyed, resulting in The projection in the residual space is increased so that The Euclidean distance to the model principal component space exceeds the threshold . It can be calculated according to the following formula. Definition:

[0175] ;

[0176] ;

[0177] In the formula, for ; k is the number of principal components selected by the PCA model.

[0178] Under certain assumptions, Approximately obeys the normal distribution, that is:

[0179] ;

[0180] When the confidence level is selected as When , the upper threshold of Q statistic can be obtained The calculation expression is:

[0181] ;

[0182] In the formula, Represents the standard normal distribution (1- ) confidence limits.

[0183] according to n The test vector of the sensor set composed of sensors , get the test vector and Q Statistics, by comparing and Q The statistics and the respective thresholds are used to obtain the anomaly detection results, and the detection results of AANN are used as verification comparison;

[0184] After obtaining the abnormal detection results, when the detection results are "abnormal", different fault mode recognition methods are used to locate the faulty components of the system abnormality; when the test vector of or Q When the statistic exceeds its upper threshold, it indicates that x is abnormal and fault identification is required to locate the test vector Sensors that have abnormalities in the

[0185] Based on the specific Or the contribution rate analysis method of Q statistics to locate the faulty sensor, as shown below;

[0186] 1) Contribution rate analysis based on Q statistics

[0187] Test Vector The total variation of the n sensors in the residual space constitutes Therefore, the contribution rate of each sensor variable to the Q statistic is calculated separately, and the operating status of the sensor is evaluated according to the size of the variable contribution rate. Definition The contribution rate of the i-th sensor to the Q statistic is;

[0188] ;

[0189] in, Before the Representative k The principal element space is composed of principal elements. represents the projection matrix of the principal element space, The projection matrix representing the residual space;

[0190] Under normal conditions, the contribution of each sensor variable to the Q statistic should be the same. A sensor in When the contribution rate of the sensor increases, it proves that An exception occurred;

[0191] 2) Based on Contribution Rate Analysis Method of Statistics

[0192] Test Vector The total change of all sensor variables in the principal element space is of Statistics. The change of each sensor variable in the principal element space is The contribution rate of the statistic indicates the operating status of this sensor. Sensor variables in right The contribution rate of the statistic can be calculated as follows;

[0193] 2.1) Calculating sensor variables For the j-th score vector Contribution rate:

[0194] (6)

[0195] In the formula, t express The score vector of ; Pj,i express The eigenvector of Pj No. i elements; , which will be i elements remain unchanged, and the rest are set to zero;

[0196] 2.2) Calculating sensor variables right Contribution rate of statistics:

[0197] (7)

[0198] Under normal conditions, each sensor variable The contribution rate of the statistic should also be the same. x A sensor in 1 When the contribution rate of An exception occurred.

[0199] In order to ensure that the detection results are accurate and reliable, the statistical analysis method is used to eliminate false alarms. The specific steps are as follows:

[0200] (1) Establish the KPCA condition monitoring model: According to the following formula, and Q statistic false alarm probability and ;

[0201] (8)

[0202] (2) Determine the false alarm probability of the statistic and the sliding window length n: According to the inequality of formula (9), we can obtain and the second control limit of the Q statistic and , that is, the maximum number of alarms allowed by the two types of statistics in the sliding window;

[0203] (9)

[0204] (3) Confirmation or elimination of the test sample alarm signal: Calculate the value within the sliding window n according to formula (10) T and Q statistics alarm times and .

[0205] (10)

[0206] If the current test sample The Q statistic and If the statistic exceeds its first control limit, continue to calculate the Q statistic and the Q statistic of all test samples in the current sliding window n. The total number of times the statistic exceeds its first control limit. This occurs if and only if the total number of test sample alarms within the sliding window n is greater than the Q statistic or Second control limit for the statistic and When the current test sample The status is defined as a fault status, that is, alarm confirmation.

[0207] Embodiment 4 is:

[0208] On the basis of Example 1, the sensor signal reconstruction method is further limited, that is, the failure data reconstruction method of the faulty sensor is specifically:

[0209] The reconstruction of failure data is achieved by iterative calculation of KPCA model;

[0210] The reconstruction process is as follows: based on fault identification, the fault data is reconstructed through iterative calculation in the principal element space:

[0211] (11)

[0212] When the sensor When a fault occurs, use formula (11) Projection matrix in principal space The projection value As a sensor Compared to the original measurement value of the sensor , estimated value Closer to the true value of the sensor ;

[0213] The estimated value As a sensor An iterative estimate of ,Will Continue to the KPCA model Projection, get the second iteration estimate Compared with the one-iteration estimate , Obviously closer to the sensor The true value of In this way, when the number of iterations is large enough, the sensor can be used of n Iteration estimate Approximately as a sensor The true value of Sensor The best reconstruction value.

[0214] In addition, KNN's operating condition identification and KPCA model adaptive selection technology determines the operating conditions of the nuclear power system and selects the KPCA model that best matches the current operation. This solves the problem of the KPCA model not matching the current operating conditions due to nonlinear changes in the system's operating characteristics when the nuclear power system is operating frequently, thereby improving the accuracy of KPCA model sensor fault detection and the accuracy of data reconstruction.

[0215] The IPSO-RBF method is used to achieve the reconstruction of failure data;

[0216] Combined with Figure 6 As shown in the figure, the IPSO algorithm calculates the distance between a particle and other particles in the population during iteration. If the ratio of the distance between them and the maximum distance is less than the threshold, it means that the particle belongs to the neighborhood range, and the improved algorithm is used to update the particle's speed and position; if the distance ratio between the particle and other particles does not meet the above situation, the standard PSO algorithm is used. The calculation steps of the IPSO-RBF algorithm are:

[0217] 1) Initialize the RBF model parameters, including the optimization range of the hidden layer center and basis function width parameters of the RBF neural network; set the IPSO algorithm parameters, including learning factor, number of iterations, and population size;

[0218] 2) Initialize the individual extreme position of the particle p b and the corresponding individual extreme values p bf, and the global extreme value position g b and the corresponding global extreme value g bf;

[0219] 3) Calculate the fitness values ​​of all particles p f;

[0220] 4) Comparison p b. p bf g b. g bf, updates individual extreme values ​​and global extreme values;

[0221] like p f< p bf, p bf = p f, p b= ;otherwise p b. p bf remains unchanged;

[0222] like p f< g bf,g bf = p f, g b= ;otherwise g b. g bf remains unchanged;

[0223] 5) Update the speed and position of the particle and make the speed and position within the specified range;

[0224] 6) If the number of iterations or the cutoff accuracy is reached, the iteration ends; otherwise, return to 2).

[0225] In addition, the sensor signal reconstruction technology of IPSO-RBF neural network, based on the influence relationship between the system operation parameters measured by the sensor, uses the good local function approximation effect of RBF neural network to establish the mapping function of normal sensor signals and abnormal sensor signals, and realizes accurate reconstruction of sensor signals. At the same time, the improved particle swarm optimization algorithm IPSO is used to optimize the key parameters of RBF to improve the accuracy and generalization ability of RBF model.

[0226] The local fusion processing method is used to fuse the calculation results of the two reconstruction algorithms to obtain an accurate reconstructed signal;

[0227] The local fusion is to compare the reconstructed signal results of the two methods with the historical sensor data, find the historical signal closest to the reconstructed signal through the KNN method, calculate the distance between the reconstructed signal and the historical signal, and use the distance as the basis for measuring the weight of the reconstructed signal in the local fusion process, perform local fusion, and calculate the fusion result.

[0228] The calculation formula of the fusion result is as follows:

[0229] (12)

[0230] in, represents the reconstructed signal value after local fusion, represents the deviation correction of the reconstructed signal, Represents the weight correction of the reconstructed signal, represents the number of fusion models, Data values ​​reconstructed for the algorithm;

[0231] The deviation correction of the reconstructed signal is and weight correction The calculation formulas are:

[0232] (13)

[0233] (14)

[0234] in, It represents the actual sensor signal set that is close to the model reconstruction value calculated by KNN. Represents the neighboring point data in the KNN historical normal sensor data set, It represents the maximum value of the distance between the reconstructed signal and the actual signal in the nearest neighbor set of each model; k represents the number of neighbor points selected by KNN, and i represents the i-th reconstructed model.

[0235] The main steps are as follows:

[0236] (1) Obtain the data value reconstructed by each algorithm ;

[0237] (2) According to each reconstructed value Use KNN to calculate the distance The nearest k points;

[0238] (3) According to each reconstructed value The corresponding k points calculate n and ;

[0239] (4) Finally, the fused reconstructed signal is calculated according to formula (12).

[0240] In addition, the KNN method is used to find the historical signal that is closest to the reconstructed signal, and the distance between the reconstructed signal and the historical signal is calculated. This distance is used as the basis for measuring the weight of the reconstructed signal in the local fusion process. The signals reconstructed by the IPSO-RBF and KPCA methods are locally fused, and the fusion result is calculated to improve the accuracy of the reconstructed signal.

[0241] The parts of the present invention that are not described in detail are prior art. It is obvious to those skilled in the art that the present invention is not limited to the details of the above-mentioned exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention; therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is limited by the appended claims rather than the above description. Therefore, it is intended to include all changes within the meaning and scope of the equivalent elements of the claims in the present invention, and any figure marks in the claims should not be regarded as limiting the claims involved.

Claims

1. A method for multi-sensor failure detection and signal reconstruction in a nuclear power plant, characterized by: Specifically: Real-time acquisition of measurement data from nuclear power plant sensors during normal operation; After preprocessing the measured data, it is constructed as a training set; Extract part of the training set data, use the sensor failure mechanism model to simulate the sensor failure, obtain the simulated sensor failure data, and construct it as a test set; Based on the training set and test set data, detect the working status of the nuclear power plant sensors, locate the specific location of the faulty sensors, and output the normal sensor data or the faulty sensor location data; The AANN method and the KPCA method are used to detect the working status of the nuclear power plant sensor; The steps of using AANN method to detect the working status of nuclear power plant sensors are as follows: Construct AANN sensor state detection model; The training set is used as input data of the AANN sensor state detection model, and the output data of the sensor state detection model is obtained; Calculate the residual between input data and output data; Determine whether the residual exceeds the limit value; If the residual exceeds the limit value, it is judged as a sensor failure, and the specific location of the faulty sensor is searched by searching the fault combination; If the residual does not exceed the limit value, the sensor is judged to be normal and the normal sensor data is output; The steps of using the KPCA method to detect the working status of the nuclear power plant sensor are specifically as follows: Based on the training set data, the KNN algorithm is used to determine the current operating conditions of the nuclear power system; The KPCA model trained with the test set data closest to the current operating conditions is selected; Calculate the statistical residuals of the KPCA model; Determine whether the residual exceeds the limit value; If the residual exceeds the limit value, it is judged as a sensor failure, and the specific position of the corresponding faulty sensor is located; If the residual does not exceed the limit value, the sensor is judged to be normal and the normal sensor data is output; Reconstruct the failure data of the faulty sensor by using local fusion method, and output the reconstructed sensor data; The failure data reconstruction method of the faulty sensor is specifically as follows: The reconstruction of failure data is achieved by iterative calculation of KPCA model; The IPSO-RBF method is used to achieve the reconstruction of failure data; The local fusion processing method is used to fuse the calculation results of the two reconstruction algorithms to obtain an accurate reconstructed signal.

2. The method for multi-sensor failure detection and signal reconstruction in a nuclear power plant according to claim 1, characterized in that: The steps of preprocessing the measurement data are specifically as follows: Use FMEA method to determine the data set of key sensors; Usage Improvements The criterion is to remove outlier data from the data set; The measurement data after removing the outlier is reconstructed by taking the average of the two measured values ​​before and after the outlier. The wavelet packet method is used to reduce the noise of the reconstructed data.

3. The method for multi-sensor failure detection and signal reconstruction in a nuclear power plant according to claim 1, characterized in that: The method for implementing failure data reconstruction using the IPSO-RBF method comprises the following specific steps: Initialize RBF model parameters and set IPSO algorithm parameters; Initialize the individual extreme value positions and corresponding individual extreme values ​​of particles, and the global extreme value positions and corresponding global extreme values; Calculate the fitness values ​​of all particles; Compare the obtained individual extreme value positions, individual extreme values, global extreme value positions and global extreme values; Update individual extreme values ​​and global extreme values ​​according to the comparison results; Update the speed and position of the particle and make them within the specified range; If the number of iterations or the cutoff accuracy is reached, the iteration ends; otherwise, return to the step of initializing the particle extremum.

4. The method for multi-sensor failure detection and signal reconstruction in a nuclear power plant according to claim 3, characterized in that: The specific steps of the local fusion method are: Use the KNN algorithm to find the historical signal that is closest to the reconstructed signal; Calculate the distance between the reconstructed signal and the historical signal; This distance is used as a basis for measuring the weight of the reconstructed signal in the local fusion process; Perform local fusion and calculate the fusion result.

5. The method for multi-sensor failure detection and signal reconstruction in a nuclear power plant according to claim 4, characterized in that: The calculation formula of the fusion result is: ； in, represents the reconstructed signal value after local fusion, represents the deviation correction of the reconstructed signal, Represents the weight correction of the reconstructed signal, represents the number of fusion models, The data value reconstructed for the algorithm.

6. The method for multi-sensor failure detection and signal reconstruction in a nuclear power plant according to claim 5, characterized in that: The deviation correction amount of the reconstructed signal and weight correction The calculation formulas are: ； ； in, It represents the actual sensor signal set that is close to the model reconstruction value calculated by KNN. Represents the neighboring point data in the KNN historical normal sensor data set, It represents the maximum value of the distance between the reconstructed signal and the actual signal in the nearest neighbor set of each model; k represents the number of neighbor points selected by KNN, and i represents the i-th reconstructed model.

Images