Nuclear power plant sensor abnormal data processing method based on hybrid model

Abstract

The invention discloses a nuclear power plant sensor abnormal data processing method based on a hybrid model, belongs to the technical field of data processing, and solves the problems that the sensor abnormal data processing of an existing detection system is missing, and the compensation and correction method for the sensor abnormal data is limited. The method comprises the steps of reading sensor data and sensor information of a nuclear power plant reactor, and marking a failure sensor; normalizing the sensor data, and identifying a data abnormal mode based on a deep neural network; performing abnormal data processing on the identified data abnormal mode, outputting the processed sensor abnormal data and the time and position when the sensor is abnormal, outputting a real-time monitoring result of the nuclear power plant operation process parameters and a sensor response value calculation result, and comparing; according to the method, the data reliability of the sensor is improved by screening and removing the over-limit value and the outlier and correcting the data deviation, and the fault of the sensor can be found and processed in time.

Description

Technical Field

[0001] The invention belongs to the technical field of data processing, and in particular relates to a method for processing abnormal data of a nuclear power plant sensor based on a hybrid model. Background Art

[0002] With the continuous growth of global energy demand and the enhancement of environmental protection awareness, the safe and efficient use of nuclear energy has become an important part of the energy strategy of various countries. As the main carrier of nuclear energy utilization, the safe and stable operation of nuclear power plants is particularly important. Nuclear power plant data detection, real-time monitoring of various parameters of nuclear reactors and their surrounding environment, and ensuring that nuclear reactors operate within the set parameter range are of great significance to the safe operation of nuclear power plants.

[0003] In the process of data detection in nuclear power plants, sensor equipment monitoring is the main source of data detection, but it is easily affected by various factors, resulting in abnormal detection data. This abnormality not only reduces the reliability of the data, but may also mislead the judgment of the operating status of the nuclear power plant in the event of an accident, affecting the fault diagnosis results, leading to misoperation, and thus worsening the accident; therefore, strengthening the management of abnormal data from nuclear power plant sensors is of great significance to ensuring the safe operation of nuclear power plants, especially data mining and management under accident conditions, which can effectively improve the accuracy and reliability of data and provide strong support for the safe operation of nuclear power plants.

[0004] Due to the particularity of the operating environment of nuclear power plants, most of the sensors in nuclear power plants work in harsh environments, such as high radiation, high temperature and high pressure, complex electromagnetic fields, corrosion, etc., which leads to abnormal output data of nuclear power plant sensors. From the results, the sensor data abnormalities are caused by:

[0005] Complete failure, unable to output any signal or outputting out-of-range values;

[0006] Drift, a continuous or incremental change in sensor output over a general period of time, making the measurement inaccurate;

[0007] Bias, there is a fixed error between the sensor output and the actual measured physical quantity;

[0008] The accuracy decreases and the error of the sensor during the measurement process increases, resulting in a larger difference between the measurement result and the actual value.

[0009] Due to the harsh working environment of nuclear sensors and the low efficiency of manual inspection and maintenance, the causes and patterns of abnormal sensor data in nuclear power plants are diverse and difficult to distinguish, which increases the difficulty of abnormal data processing.

[0010] Traditional methods for processing sensor data anomalies, such as harmonic expansion method and wavelet transform, are mostly suitable for analyzing harmonic distortion in power systems, but are not effective in processing complex non-sinusoidal signals in nuclear power plants. Neural networks have strong self-learning and adaptive capabilities, and can optimize their performance through training to adapt to different types of signals and distortion conditions, and can handle complex nonlinear relationships. For example, long short-term memory networks (LSTMs) can effectively process long time series data generated by nuclear power plants and accurately predict the trend of core parameter changes in abnormal conditions, thus providing a basis and method for the management of sensor abnormal data. At present, nuclear detection systems are lacking in sensor abnormal data processing, and there are limited methods for compensating and correcting sensor abnormal data. Therefore, it is urgent to invent a fast and reliable sensor abnormal data processing method for real-time monitoring and decision-making by nuclear power plant staff, improve the credibility of sensor data, and provide strong support for the safe operation of nuclear power plants. Summary of the invention

[0011] The purpose of the present invention is to address the deficiencies of the prior art and provide a method for processing abnormal data of nuclear power plant sensors based on a hybrid model, which solves the problems that the existing detection system is deficient in processing abnormal data of sensors and has limited compensation and correction methods for abnormal data of sensors.

[0012] The existing detection system has deficiencies in sensor abnormal data processing, and the compensation and correction methods for sensor abnormal data are limited. To address the above problems, we propose a nuclear power plant sensor abnormal data processing method based on a hybrid model. When the method is implemented, the nuclear power plant sensor data is first read, the current measurement value and its position information are read from each sensor, and the obtained sensor data is normalized according to the total number of sensors. The sensor data abnormal pattern is identified based on the recognition model of the deep neural network, and the identified data abnormal pattern is processed for abnormal data. The processed sensor abnormal data and the time and location of the sensor abnormality are output, and the real-time monitoring results of the nuclear power plant operation process parameters and the sensor response value calculation results are output and compared. The present invention can timely discover and process sensor faults, effectively reduce the maintenance cost and safety risks of nuclear power plants, and through accurate data monitoring, can truly grasp the operating status of nuclear power plants, provide strong support for the safe operation of nuclear power plants, and realize accurate diagnosis of faults and timely response to accidents, effectively avoid personnel misoperation, prevent accidents from worsening, and significantly improve the overall safety of nuclear power plants.

[0013] The present invention is implemented as follows: a method for processing abnormal data of a nuclear power plant sensor based on a hybrid model, the method for processing abnormal data of a nuclear power plant sensor based on a hybrid model comprising:

[0014] S10, reading sensor data of the nuclear power plant, where the sensor data is the operating process parameters of the nuclear power plant reactor, and the operating process parameters include nuclear reactor power, reactivity, coolant temperature, control rod position, coolant flow rate, and the corresponding total number of sensors, sensor channel number and name of each sensor;

[0015] S20, reading the current measurement value and its position information from each sensor, and checking whether the number of the read measurement values ​​is consistent according to the total number of sensors. If they are consistent, continue; if they are inconsistent, determine and mark the name of the failed sensor according to the position of the missing measurement value;

[0016] S30, normalizing the obtained sensor data, and identifying abnormal patterns of the sensor data based on a recognition model of a deep neural network, where the abnormal patterns of the data are used to identify sensor data that exceeds a limit, outlier data, drift, bias, and reduced accuracy;

[0017] S40, performing abnormal data processing on the identified data abnormality pattern, wherein the abnormal data processing includes removing null values, negative values, and over-limit data that exceed a reasonable range, and supplementing them with interpolation of data points in the previous and next time series that are within the interval; removing outlier data, and supplementing them with interpolation of data points in the previous and next time series that are within the interval; reconstructing and replacing data with strong nonlinearity such as drift, bias, and precision degradation using a prediction model based on a deep neural network;

[0018] S50, outputting the processed sensor abnormality data and the time and location of the sensor abnormality, and at the same time outputting the real-time monitoring results of the nuclear power plant operation process parameters and the sensor response value calculation results, and making comparisons.

[0019] Preferably, when the obtained sensor data is normalized, the normalization calculation formula is:

[0020]

[0021] In the formula, x min is the lower limit value of the sensor instrument, x max is the upper limit value of the sensor instrument, x * The data is normalized by the sensor.

[0022] Preferably, the training method of the recognition model based on a deep neural network comprises:

[0023] S301, input data, import historical real data or nuclear island operation data generated by a simulator into the deep neural network model;

[0024] S302, data division, dividing the sensor abnormal data into a training set and a validation set, wherein the ratio of the training set to the validation set is 8:2;

[0025] S303, training the deep neural network model, using the training set as input to the built deep neural network model for training, and selecting the cross entropy loss function, mean square error MSE, and root mean square error RMSE as the loss function of the deep neural network model;

[0026] S304, model evaluation, input the validation set into the trained deep neural network model for testing, and evaluate the model output and the true value. The evaluation criteria include root mean square error RMSE, determination coefficient R 2 .

[0027] Preferably, the method of reconstructing and replacing data with strong nonlinearity such as drift, bias and precision degradation using a prediction model based on a deep neural network specifically includes:

[0028] S401, using a prediction model based on a deep neural network to test sensor abnormal data to predict operating process parameters of the nuclear power plant and reconstruct sensor response values, wherein a training method of the prediction model based on the deep neural network is the same as a recognition model based on the deep neural network;

[0029] S402, calculating the relative error of each sensor according to the prediction result of the prediction model based on the deep neural network and the actual sensor measurement value;

[0030] Among them, the relative error calculation formula is expressed as:

[0031]

[0032] S403, judging whether the sensor is failed according to the magnitude of the relative error and the safety standard of nuclear power plant operation. If the relative error is greater than the set threshold, the sensor is regarded as a failed sensor, and its measured value is discarded and supplemented with the reconstructed value to obtain sensor abnormality processing data;

[0033] S404, obtaining the maximum relative error from the relative errors, setting the relative error limit according to the selection of the prediction model of the deep neural network and the sensor accuracy, if the maximum relative error is greater than the relative error limit, repeating steps S401 to S403, and using the sensor data after the sensor abnormality processing data obtained in step S403 to test the prediction model of the deep neural network again, until the maximum relative error is less than the relative error limit.

[0034] Preferably, when removing outlier data, thresholds and nodes are set according to the experience and rules of nuclear power plant operation, and each normalized data is judged. If it exceeds n or 1 / n times the average value of several node data before and after the node, it is regarded as an outlier and removed, and supplemented by interpolation of the previous and next time series data points.

[0035] Preferably, the deep neural network includes: a convolutional neural network CNN, a recurrent neural network RNN, a long short-term memory network LSTM or a generative adversarial network GAN.

[0036] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0037] The present invention ensures accurate monitoring of the operating process parameters of the nuclear power plant by diagnosing sensor failures, including complete failures and deviations from normal values, and can respond quickly, thereby significantly improving the stability of the nuclear power plant monitoring system, reducing the risks of false alarms and missed alarms due to sensor failures, and enabling operators to make decisions based on accurate sensor data, optimize reactor operating efficiency, and improve the economic benefits of the nuclear power plant. At the same time, the present invention can promptly detect and handle sensor failures, effectively reducing the maintenance costs and safety risks of the nuclear power plant. Through accurate data monitoring, the nuclear power plant can truly grasp the operating status of the nuclear power plant, provide strong support for the safe operation of the nuclear power plant, and achieve accurate diagnosis of faults and timely response to accidents, effectively avoiding human misoperation and preventing accidents from worsening, significantly improving the overall safety of the nuclear power plant. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the implementation flow of the hybrid model-based nuclear power plant sensor abnormal data processing method provided by the present invention.

[0039] Figure 2 It is a flow chart of drift, bias and precision degradation data processing of the present invention.

[0040] Figure 3 It is a training flow chart of the recognition model based on deep neural network in the present invention.

[0041] Figure 4 It is a comparison diagram of the sensor abnormal data processing effect provided by the present invention. DETAILED DESCRIPTION

[0042] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of this application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0043] The existing detection system has deficiencies in sensor abnormal data processing, and the compensation and correction methods for sensor abnormal data are limited. To address the above problems, we propose a nuclear power plant sensor abnormal data processing method based on a hybrid model. When the method is implemented, the sensor data of the nuclear power plant reactor is first read, the current measurement value and its position information are read from each sensor, and the obtained sensor data is normalized according to the total number of sensors. The sensor data abnormal pattern is identified based on the recognition model of the deep neural network, and the identified data abnormal pattern is processed for abnormal data. The processed sensor abnormal data and the time and location of the sensor abnormality are output, and the real-time monitoring results of the nuclear power plant operation process parameters and the sensor response value calculation results are output and compared. The present invention can timely discover and process sensor faults, effectively reduce the maintenance cost and safety risks of nuclear power plants, and through accurate data monitoring, can truly grasp the operating status of nuclear power plants, provide strong support for the safe operation of nuclear power plants, and realize accurate diagnosis of faults and timely response to accidents, effectively avoid personnel misoperation, prevent accidents from worsening, and significantly improve the overall safety of nuclear power plants.

[0044] The embodiment of the present invention provides a method for processing abnormal data of nuclear power plant sensors based on a hybrid model, such as Figure 1 As shown, the hybrid model-based nuclear power plant sensor abnormal data processing method includes:

[0045] S10, reading sensor data of the nuclear power plant, where the sensor data is the operating process parameters of the nuclear power plant reactor, and the operating process parameters include but are not limited to nuclear reactor power, reactivity, coolant temperature, control rod position, coolant flow rate, and the corresponding total number of sensors, sensor channel number and name of each sensor;

[0046] For example, the embodiment of the present invention selects abnormal data of the water temperature sensor of the replacement water storage tank (IRWST) compartment in the containment when the main steam pipe in the containment is broken (MSLB), such as Figure 4 As shown, the solid grey line represents abnormal data, and abnormal patterns include exceeding the limit, outliers, and decreased accuracy.

[0047] S20, reading the current measurement value and its position information from each sensor, and checking whether the number of the read measurement values ​​is consistent according to the total number of sensors. If they are consistent, continue; if they are inconsistent, determine and mark the name of the failed sensor according to the position of the missing measurement value;

[0048] S30, normalizing the obtained sensor data, taking the upper limit value of the sensor instrument as 1 and the lower limit value as 0, linearizing the valid data to the range of [0, 1], and identifying the abnormal pattern of the sensor data based on the recognition model of the deep neural network. The data abnormal pattern is used to identify the sensor data with over-limit data, outlier data, drift, bias and reduced accuracy;

[0049] In this embodiment, when the obtained sensor data is normalized, the normalization calculation formula is:

[0050]

[0051] In the formula, x min is the lower limit value of the sensor instrument, x max is the upper limit value of the sensor instrument, x * In the embodiment of the present invention, the real data fluctuates between 313°C and 362°C, the lower limit of the sensor instrument is 300°C, and the upper limit of the sensor instrument is 370°C.

[0052] S40, performing abnormal data processing on the identified data abnormality pattern, wherein the abnormal data processing includes removing null values, negative values, and over-limit data that exceed a reasonable range, and supplementing them with interpolation of data points in the previous and next time series that are within the interval; removing outlier data, and supplementing them with interpolation of data points in the previous and next time series that are within the interval; reconstructing and replacing data with strong nonlinearity such as drift, bias, and precision degradation using a prediction model based on a deep neural network;

[0053] S50, output the processed sensor abnormality data and the time and location of the sensor abnormality, and at the same time output the real-time monitoring results of the nuclear power plant operation process parameters and the sensor response value calculation results, and compare them for real-time monitoring and decision-making by nuclear power plant staff.

[0054] In this embodiment, if Figure 3 As shown, the training method of the recognition model based on the deep neural network includes:

[0055] S301, input data, import historical real data or nuclear island operation data generated by a simulator into the deep neural network model;

[0056] S302, data division, dividing the sensor abnormal data into a training set and a validation set, wherein the ratio of the training set to the validation set is 8:2;

[0057] S303, training the deep neural network model, using the training set as input to the built deep neural network model for training, and selecting the cross entropy loss function as the loss function of the deep neural network model;

[0058] S304, model evaluation, input the validation set into the trained deep neural network model for testing, and evaluate the model output and the true value. The evaluation criteria include root mean square error RMSE, determination coefficient R 2 .

[0059] In this embodiment, the evaluation criteria expressions are:

[0060]

[0061] Where n is the number of test points, y i represents the true value at the i-th sample, is the model prediction value at this point, is the mean of the true values. R 2 RMSE is used to evaluate the prediction accuracy of the model. 2 The closer the value is to 1 or the smaller the RMSE value is, the higher the accuracy of the model prediction is.

[0062] It should be noted that the deep neural network includes but is not limited to convolutional neural network CNN, recurrent neural network RNN, long short-term memory network LSTM, generative adversarial network GAN and other commonly used deep neural network structures and their combination algorithms.

[0063] like Figure 2 As shown, the method of reconstructing and replacing data with strong nonlinearity such as drift, bias and precision degradation using a prediction model based on a deep neural network specifically includes:

[0064] S401, using a prediction model based on a deep neural network to test sensor abnormal data to predict operating process parameters of the nuclear power plant and reconstruct sensor response values, wherein a training method of the prediction model based on the deep neural network is the same as a recognition model based on the deep neural network;

[0065] S402, calculating the relative error of each sensor according to the prediction result (i.e., the reconstructed value) of the prediction model based on the deep neural network and the actual sensor measurement value;

[0066] Among them, the relative error calculation formula is expressed as:

[0067]

[0068] S403, judging whether the sensor is failed according to the magnitude of the relative error and the safety standard of nuclear power plant operation. If the relative error is greater than a set threshold (such as N%), the sensor is regarded as a failed sensor, and its measured value is discarded and supplemented with a reconstructed value to obtain sensor abnormality processing data;

[0069] S404, obtaining the maximum relative error from the relative errors, setting the relative error limit according to the selection of the prediction model of the deep neural network and the sensor accuracy, if the maximum relative error is greater than the relative error limit, repeating steps S401 to S403, and using the sensor data after the sensor abnormality processing data obtained in step S403 to test the prediction model of the deep neural network again, until the maximum relative error is less than the relative error limit.

[0070] In the embodiment of the present invention, when the outlier data is removed, the outlier data, as a nuclear island process parameter, will not change several times instantly to the point where the sensor cannot capture it. Therefore, even the change of the transient process sensor should be within a certain range. Severely exceeding the average value by too much or too little can be determined as sensor abnormality. According to the experience and rules of nuclear power plant operation, thresholds and nodes are set, and each normalized data is judged. If it exceeds n or 1 / n times the average value of several node data before and after the node, it is regarded as an outlier and removed, and supplemented by interpolation of the previous and next time series data points, where the number of nodes and the n value are given by the type analysis of different sensor abnormalities.

[0071] For the over-limit data, in the implementation, negative values ​​appear at 122-145s, 675s, and 1034-1038s, and data greater than 1 appears at 424s. Obviously, these data are over-limit abnormal, so these data are removed and supplemented by interpolation of the data points before and after the time series and located in the interval. In the embodiment, the number of nodes is 5, and the value of n is 2. In the embodiment, the value at 328s is 2.06 times, the value at 354s is 2.32 times, and the value at 515s is 2.90 times, which exceeds the set value and is judged as outlier data. These data are regarded as outliers and removed, and supplemented by interpolation of the data points before and after the time series.

[0072] In order to verify the ability of the present invention to manage sensor data anomalies, the following design is made: Figure 4 In the example shown, the dotted line is the true value of the water temperature in the IRWST compartment of the containment when the main steam pipe in the containment is breached (MSLB), the gray solid line is the abnormal data output by the sensor, and the black solid line is the data after abnormal processing. The abnormal mode includes complete failure, data drift, and sensor accuracy degradation, which is manifested as Figure 4 Some of the data are zero, out of limit, or oscillating. The results are as follows Figure 4 As shown, after applying the present invention to process abnormal sensor data, although the output data after treatment shows a small fluctuation, it is basically consistent with the real data, and compared with the abnormal data, the problems of some data being zero and exceeding the limit are completely treated, and the data oscillation problem is greatly improved, indicating that the treatment result of the present invention for sensor anomalies is accurate and reliable.

[0073] In summary, the present invention provides a method for processing abnormal data of nuclear power plant sensors based on a hybrid model. The particularity of this application lies in the integration of a framework of multiple algorithms. After identifying multiple abnormal modes of nuclear power plant sensors, corresponding processing methods are adopted respectively to effectively process abnormal data. The model training part is based on historical real data or nuclear island operation data generated by a simulator. There is no specific requirement for the network for model training. The selection of the model includes but is not limited to the use of deep neural network structures such as convolutional neural network CNN, recurrent neural network RNN, long short-term memory network LSTM, gated recurrent unit GRU, generative adversarial network GAN, and their combined algorithms. Different deep neural networks can be selected for optimization for different sensors and different abnormal modes.

[0074] The present invention ensures accurate monitoring of the operating process parameters of the nuclear power plant by diagnosing sensor failures, including complete failures and deviations from normal values, and can respond quickly, thereby significantly improving the stability of the nuclear power plant monitoring system, reducing the risks of false alarms and missed alarms due to sensor failures, and enabling operators to make decisions based on accurate sensor data, optimize reactor operating efficiency, and improve the economic benefits of the nuclear power plant. At the same time, the present invention can promptly detect and handle sensor failures, effectively reducing the maintenance costs and safety risks of the nuclear power plant. Through accurate data monitoring, the nuclear power plant can truly grasp the operating status of the nuclear power plant, provide strong support for the safe operation of the nuclear power plant, and achieve accurate diagnosis of faults and timely response to accidents, effectively avoiding human misoperation and preventing accidents from worsening, significantly improving the overall safety of the nuclear power plant.

[0075] It should be noted that, for the above-mentioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described order of actions, because according to the present invention, some steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0076] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A method for processing abnormal data of nuclear power plant sensors based on a hybrid model, characterized in that: The hybrid model-based nuclear power plant sensor abnormal data processing method includes: S10, reading sensor data of the nuclear power plant, where the sensor data is the operating process parameters of the nuclear power plant reactor, and the operating process parameters include nuclear reactor power, reactivity, coolant temperature, control rod position, coolant flow rate, and the corresponding total number of sensors, sensor channel number and name of each sensor; S20, reading the current measurement value and its position information from each sensor, and checking whether the number of the read measurement values ​​is consistent according to the total number of sensors. If they are consistent, continue; if they are inconsistent, determine and mark the name of the failed sensor according to the position of the missing measurement value; S30, normalizing the obtained sensor data, and identifying abnormal patterns of the sensor data based on a recognition model of a deep neural network, where the abnormal patterns of the data are used to identify sensor data that exceeds a limit, outlier data, drift, bias, and reduced accuracy; S40, performing abnormal data processing on the identified data abnormal pattern, wherein the abnormal data processing includes removing null values, negative values, and over-limit data that exceed a reasonable range, and supplementing them with interpolation of data points in the previous and next time series that are within the interval; removing outlier data, and supplementing them with interpolation of data points in the previous and next time series that are within the interval; reconstructing and replacing data with strong nonlinearity of drift, bias, and precision drop using a prediction model based on a deep neural network; S50, outputting the processed sensor abnormality data and the time and location of the sensor abnormality, and at the same time outputting the real-time monitoring results of the nuclear power plant operation process parameters and the sensor response value calculation results, and making comparisons.

2. The method for processing abnormal data of nuclear power plant sensors based on a hybrid model according to claim 1, characterized in that: When the obtained sensor data is normalized, the normalization calculation formula is: In the formula, x min is the lower limit value of the sensor instrument, x max is the upper limit value of the sensor instrument, x * The data is normalized by the sensor.

3. The method for processing abnormal data of nuclear power plant sensors based on a hybrid model according to claim 1, characterized in that: The training method of the recognition model based on the deep neural network includes: S301, input data, import historical real data or nuclear island operation data generated by a simulator into a deep neural network model; S302, data division, dividing the sensor abnormal data into a training set and a validation set, wherein the ratio of the training set to the validation set is 8:2; S303, training the deep neural network model, using the training set as input to the built deep neural network model for training, and selecting the cross entropy loss function, mean square error MSE, and root mean square error RMSE as the loss function of the deep neural network model; S304, model evaluation, input the validation set into the trained deep neural network model for testing, and evaluate the model output and the true value. The evaluation criteria include root mean square error RMSE, determination coefficient R 2 .

4. The method for processing abnormal data of nuclear power plant sensors based on a hybrid model as claimed in claim 3, characterized in that: The method of reconstructing and replacing data with strong nonlinear drift, bias and precision drop using a prediction model based on a deep neural network specifically includes: S401, using a prediction model based on a deep neural network to test sensor abnormal data to predict operating process parameters of the nuclear power plant and reconstruct sensor response values, wherein a training method of the prediction model based on the deep neural network is the same as a recognition model based on the deep neural network; S402, calculating the relative error of each sensor according to the prediction result of the prediction model based on the deep neural network and the actual sensor measurement value; The relative error calculation formula is expressed as: S403, judging whether the sensor is failed according to the magnitude of the relative error and the safety standard of nuclear power plant operation. If the relative error is greater than the set threshold, the sensor is regarded as a failed sensor, and its measured value is discarded and supplemented with the reconstructed value to obtain sensor abnormality processing data; S404, obtaining the maximum relative error from the relative errors, setting the relative error limit according to the selection of the prediction model of the deep neural network and the sensor accuracy, if the maximum relative error is greater than the relative error limit, repeating steps S401 to S403, and using the sensor data after the sensor abnormality processing data obtained in step S403 to test the prediction model of the deep neural network again, until the maximum relative error is less than the relative error limit.

5. The method for processing abnormal data of nuclear power plant sensors based on a hybrid model as claimed in claim 4, characterized in that: When removing outlier data, thresholds and nodes are set according to the experience and rules of nuclear power plant operation, and each normalized data is judged. If it exceeds n or 1 / n times the average value of the data of several nodes before and after the node, it is regarded as an outlier and removed, and supplemented by interpolation of the previous and next time series data points.

6. The method for processing abnormal data of nuclear power plant sensors based on a hybrid model as claimed in claim 3, characterized in that: The deep neural network includes: convolutional neural network CNN, recurrent neural network RNN, long short-term memory network LSTM or generative adversarial network GAN.

Images