KR20220124990A - A Sensor Fault-Tolerant Accident Diagnosis System - Google Patents

Abstract

Emergency situations in nuclear power plants involve automatic shutdown of a reactor, placing a heavy workload on plant operators under stressful conditions. Diagnosis of an incident that has occurred is an essential sequence for optimal incident outcome mitigation. However, incident identification results are also an important source of error, as the incident identification results determine the workflow linked to all subsequent operations. To support nuclear power plant accident identification, recent recurrent neural network (RNN)-based approaches have shown excellent performance. However, despite these achievements, the robustness of RNN models is not promising as it has been shown that incorrect inputs degrade the performance of RNNs more than other methods in some applications. In this embodiment, an accident diagnosis system that is tolerant of sensor failures was developed based on the existing RNN model and tested with expected sensor failures. After receiving and processing sensor defect information, an accident is diagnosed using a GRU algorithm.

Description

A Sensor Fault-Tolerant Accident Diagnosis System {A Sensor Fault-Tolerant Accident Diagnosis System}

The present technology relates to an accident diagnosis system that is robust against sensor faults.

In safety-critical systems, prompt response to anomalies is an important factor in minimizing the associated consequences. Nuclear fuels that generate fission energy in nuclear power plants (NPPs) can pose a threat to the public if large amounts of radioactive material are released after a nuclear accident. Various systems must be prepared to maintain plant safety and stability in terms of component failure and external threats. In preparation for anticipated accidents, the NPP has several safety systems driven by process parameters exceeding thresholds or manual actions by plant operators. Responses to emergencies follow specific procedures that include sequential actions. In order to know how to alleviate the accident according to the symptoms of the accident, the accident diagnosis is carried out by specifying the exact accident type. Incident diagnosis is therefore important because it determines the specific optimal recovery procedure (ORP) with the necessary mitigation actions. Although the accident diagnosis procedure provides intuitive logic for identifying an accident based on a series of symptom identifications, it can be a daunting task for plant operators because the early stages of an emergency can affect the outcome of an accident. In this regard, an incorrect diagnosis may lead to the selection of an incorrect optimal recovery procedure (ORP), which may lead to human error.

Plant health is monitored through numerous sensors connected to the components, and the sensor values act as the basis for the plant operators' status awareness and the cause of automatic safety system operation. In this regard, misinformation from the sensor can confuse the operator and lead to erroneous judgment. In fact, misdiagnosis due to sensor failure has been discussed as a kind of deterioration factor in serious accidents such as the Three Mile Island (TMI-2) and Fukushima accidents. In particular, the TMI-2 accident represents a case where a serious misdiagnosis error occurred due to a sensor defect as follows. Electromagnetic relief valves, also known as pilot-operated relief valves (PORVs), were accidentally locked open. However, the relevant indicator indicated that the valve was closed. The error caused the operator to misunderstand the situation and turn off an automatically activated safety injection system to cool the reactor core, causing catastrophe.

In the existing diagnostic procedures or developed diagnostic models, countermeasures for sensor failure during an accident sequence were not discussed. In this context, the current diagnostic model, which lacks consideration of diagnostic failure due to sensor failure, seems vulnerable to sensor failure. Most of the related prior studies on sensor faults in the nuclear field are aimed at normal operation, which has not yet been studied in transient situations, including accidents.

A nuclear reactor accident diagnosis method according to the present embodiment includes the steps of detecting occurrence of a reactor trip; checking whether a sensor defect has occurred; After receiving the sensor defect information, processing it, and diagnosing the accident using the GRU algorithm, and outputting the diagnosis result.

The data-based methods including the short-term memory and GRU of this embodiment perform learning and testing using design-based accident data produced using a compact nuclear simulator (CNS) developed by the Korea Atomic Energy Research Institute.

In the reactor accident diagnosis method of this embodiment, as an example, the step of determining whether a sensor defect occurs is a sensor value when a consistency index value output by a long short-term memory is greater than 0.7 is considered normal, and when the consistency index value is less than 0.7, the sensor is determined to be defective.

In the nuclear reactor accident diagnosis method of this embodiment, as an example, the step of diagnosing the accident using the GRU algorithm is performed by using the GRU algorithm and an algorithm for substituting the data of the sensor in which the defect has occurred.

In the reactor accident diagnosis method of the present embodiment, as an example, an algorithm for replacing data of a sensor having a defect includes any one of a K-nearest neighbors (KNN) algorithm, a MICE algorithm, and a Missforess algorithm.

In the reactor accident diagnosis method of this embodiment, as an example, the KNN algorithm replaces the average of K parameters close to the data of the defective sensor with the data of the defective sensor.

In the reactor accident diagnosis method of the present embodiment, as an example, the MICE algorithm includes the steps of simply replacing the missing data with a place holder, and returning the missing data as the variable with the largest missing data. And, the process of regressing the variable from another variable and the regression are repeatedly performed until the result converges.

In the reactor accident diagnosis method of this embodiment, as an example, the Missforest algorithm includes constructing a plurality of decision trees, integrating the decision trees to generate a mean regression, and performing iterative random forest regression.

In the nuclear reactor accident diagnosis method of this embodiment, as an example, the step of diagnosing the accident using the GRU algorithm is performed without replacing the data of the sensor in which the defect occurs with the GRU algorithm.

In the nuclear reactor accident diagnosis method of this embodiment, for example, the step of diagnosing the accident without replacing the data of the sensor having the defect with the GRU algorithm is performed using the GRUD (GRU-Decay) algorithm.

In the reactor accident diagnosis method of this embodiment, for example, the GRUD algorithm is

is expressed as

According to the present embodiment, an advantage is provided that an accident diagnosis system robust to sensor failure is provided.

1 is a diagram illustrating a symptom-based emergency operating procedure (EOP) package.
2 is a diagram illustrating the overall framework of the accident diagnosis system robust to sensor defects according to the present embodiment.
3 shows a gated recurrent unit (GRU) based thought analysis algorithm of a time series multivariate input and a SoftMax output.
4 is a diagram illustrating a result of identifying a sensor state from a normal data set and an error-injected data set.
5 is a diagram illustrating a drift error and a sticking error.
6 is a diagram illustrating an example of a GRUD method when there is no current input.
7 is a diagram illustrating an accident diagnosis success tendency from a normal data set and a diagnosis failure tendency from a result of injecting a defect.
8 is a diagram illustrating a comparison result of substitution results.
9 is a diagram illustrating the accuracy of a diagnosis result.

Hereinafter, this embodiment will be described with reference to the accompanying drawings. In the NPP, several operating procedures are in place for anomalies of various scales. Alarm response procedures or abnormal operating procedures are implemented in response to individual component malfunctions or potential threats to core integrity. Frequent accidents of this kind are called anticipated operational occurrences. A more severe deviation beyond the expected operational occurrence is defined as an “accident”. In the event of an accident, the reactor will trip automatically when the reactor protection system detects a deviation from the predefined setpoint. Design-based incidents have no or no radiation effects outside the exclusion zone. This is because the installed safety features can completely mitigate accidents. Certain accident situations are called emergency situations and require emergency operating procedures (EOP) to be implemented. EOP provides procedural guidance focused on preventing core damage. Depending on the type of incident, which depends on the location of the failure or loss of a particular water supply, the optimal response is completely different. Diagnosis of the incident is therefore essential to prepare for appropriate mitigation actions.

Diagnostic tasks include a variety of other tasks for plant operators. Errors can be classified into three types: global false positives, local false positives, and slips. Global misdiagnosis refers to errors resulting from the selection of inappropriate response procedures. Local misdiagnosis refers to errors due to intentional or improper measures, and may result from misunderstanding of relevant procedures and information. A slip is an error that occurs unintentionally and due to improper performance when performing a task. According to this classification, misdiagnosis of plant operators includes misdiagnosis such as incorrect ORP selection as well as misunderstanding of procedural work or related information. Both types are possible candidates for human error, but global misdiagnosis can have more serious consequences as it involves numerous actions to mitigate the incident.

Emergencies are unfamiliar to plant operators because they rarely occur in practice. Therefore, performing EOP is a complex and stressful task. Operators must specify the type of incident to know the best response while surrounded by numerous alarms. For proper accident diagnosis, proper situational awareness to identify various accident symptoms is essential. To guide the operator in a standardized manner, the diagnostic procedure provides a combinational logic that specifies the type of accident from the symptoms. 1 shows the diagnostic flow of the EOP package, the EOP package includes an early response to an incident, a diagnostic procedure, and an appropriate ORP according to a specific incident. These structures are susceptible to incorrect process parameters due to sensor faults. One erroneous transition within a procedure can lead to diagnostic failures, preventing an optimal response from it, as well as unnecessary work, or even harmful actions. Ensuring robust and accurate accident diagnosis performance in nuclear power plants is essential for safe operation.

Fault-tolerant accident diagnosis system framework

2 is a diagram illustrating the overall framework of the accident diagnosis system robust to sensor defects according to the present embodiment. Referring to FIG. 2 , the system according to this embodiment includes two subsystems, as illustrated in FIG. 2 , a sensor fault detection system and a sensor fault-tolerant mitigation system. Sensor fault monitoring is performed at the head of the system initiated upon a reactor trip (3.2). In the absence of sensor faults, the incident diagnosis system diagnoses the incident (3.1) and generates an output to identify the incident. When the sensor fault detection system detects a sensor fault, the fault information is sent to the sensor fault-tolerant mitigation system 3.3, where it replaces or attenuates the impact of faulty sensor data through a data replacement method. Finally, fault-tolerant accident diagnosis results are generated.

gru Accident diagnosis algorithm used

)와 업데이트 게이트(

)를 포함한다. 리셋 게이트는 시그모이드 함수를 사용하여 이전 상태(

)를 어느 정도 반영해야 하는지 결정하고 업데이트 게이트는 이전 상태와 입력 데이터에서 현재 상태를 업데이트하는 방법을 결정한다. 후보 함수(

)와 이전 상태는 현재 은닉 상태(

)를 결정하기 위해 업데이트 게이트에 의해 변조된다. 결정된 숨김 상태는 다음 GRU 셀로 전송되거나 출력으로 종료된다. 게이트 함수, 은닉 상태 및 후보 함수는 아래의 수학식 1과 같다.This embodiment proposes an emergency accident diagnosis model using an RNN-based algorithm. RNNs use advanced models such as long and short-term memory networks or gated recurrent units (GRUs) to classify accidental or anomalous data. The GRU has two main gate functions: a reset gate (

) and update gates (

) is included. The reset gate uses a sigmoid function to set the previous state (

) should be reflected, and the update gate determines how to update the current state from the previous state and the input data. candidate function (

) and the previous state are the current hidden state (

) is modulated by the update gate to determine The determined hidden state is transmitted to the next GRU cell or terminated with an output. The gate function, the hidden state, and the candidate function are expressed in Equation 1 below.

[Equation 1]

,

,

는 각각 업데이트 게이트, 후보 게이트 및 리셋 게이트이고 W, U는 가중치가 적용된 벡터이고 x, h는 각각 입력 및 숨김 상태이다. ⊙는 요소 별 곱셈(element-wise multiplication)이다. here

,

,

are update gates, candidate gates and reset gates, respectively, W and U are weighted vectors, and x and h are input and hidden states, respectively. ⊙ is element-wise multiplication.

3 shows a GRU-based accident diagnosis algorithm. GRU receives time series data of nuclear accident situation as input and learns about the accident label, which is the correct answer at every step of the time. Learning proceeds by adjusting the weight and bias values by back propagation to process the input value according to Equation 1 above and output the correct thinking label. The calculated hidden state directly determines the output value or is transmitted along the time axis.

In the case of the system of this embodiment, a GRU-based accident diagnosis algorithm was built and the performance degradation due to an incorrect input was tested to confirm the diagnosis performance. This algorithm has one hidden layer with 64 nodes and Adam optimizer is applied for training. Hyperparameters including the number of hidden layers and nodes were determined to achieve sufficient diagnostic performance. The output result is generated by a SoftMax function that produces an output normalized to a probability distribution. The SoftMax regularization expression is as Equation 2 below.

[Equation 2]

The function of the GRU is suitable for accident diagnosis due to the forward propagation of the connection between cells and the forward propagation of situation information. However, although RNN has high performance, there may be a problem of robustness along with omission of data values of sensors. Therefore, it is necessary to prepare possible measures to secure the robustness of the RNN model against sensor defects.

Sensor Fault Detection System

는 측정값,

는 실제값,

는 상대 오차,

는 계산된 일관성 지수를 의미한다. 학습된 장단기 메모리는 발전소 변수를 입력으로 받아 특정 센서의 일관성 지수를 초마다 출력한다. 도 4는 센서 상태 식별의 예를 도시하며, 도 4(a)는 정상 데이타 테스트 결과, (b)는 오류가 주입된 테스트 결과이다. 감지 시스템의 출력은 센서 상태에 대한 수치적으로 정규화된 인덱스인 일관성 지수의 형태로 생성된다. A sensor fault detection system is presented that utilizes the remarkable performance of RNNs with multiple time-series multivariate data to account for sensor faults in emergency situations. The system using long-term memory (LSTM), a type of RNN, quantifies the consistency index of the target sensor with respect to normal accident data and sensor fault injection accident data, and learns it in the long-term memory. The coherence index is calculated by squaring the relative distance between the value currently displayed by the sensor and the actual value. The formula for calculating the consistency index is as Equation 3,

is the measured value,

is the actual value,

is the relative error,

is the calculated consistency index. The learned long-term memory receives power plant variables as input and outputs the consistency index of a specific sensor every second. Fig. 4 shows an example of sensor state identification, Fig. 4 (a) is a normal data test result, (b) is a test result in which an error is injected. The output of the sensing system is generated in the form of a consistency index, which is a numerically normalized index of the sensor state.

The consistency index is maintained at about 1 in a steady state as shown in FIG. 4(a) and decreases according to the degree of sensor signal deviation. In the empirical test results, the standard of the consistency index, which is the error threshold that considers both the detection speed and the uncertainty, has a value of 0.7. A masking input can be obtained by calculating Equation (6) of Equation 4 below from the consistency index, and sensor defect information can be derived together with the masking input. Here, '1' means that the sensor value is normally observed, and '0' means that data is missing.

[Equation 3]

[Equation 4]

The sensor failure mode was chosen considering its connectivity and typicality with human error. Drift and stuck errors were injected into the incident data and consistency was assessed. In this embodiment, a sensor defect is implemented to check the performance degradation of the diagnostic algorithm. Drift errors of 2x and 10x rates of change in both directions (up and down) and a sticking error at zero point are injected into the accident data.

Fig. 5(a) is a diagram illustrating a drift error, and Fig. 5(b) is a diagram illustrating a zero stuck error. 5 shows the case of drift and sticking errors occurring at 100 sec. In the example of Figure 5, all error injections were added to 1 second of data.

Error-tolerant accident diagnosis system

From the result of the sensor fault detection model, it is possible to monitor the sensor status in the event of an accident as information that can be used continuously. In the event of a sensor failure, the inherently unreliable incoming data must be removed from the defective sensor. However, an accident diagnosis algorithm based on GRU cannot tolerate empty inputs because the GRU structure specifies that each input affects all functions and outputs through interconnection. Therefore, to mitigate the effect of missing data, it is necessary to estimate the missing values or apply a modified RNN structure. A fault-tolerant accident diagnosis system was constructed by applying the imputation method, a process of replacing missing data with estimated values, and transforming the GRU model.

First, we considered simple imputation methods for time series data, including moving window imputation or last observed carried forward imputation. This is inappropriate in the case of this embodiment because deviation data is generated by In other words, a simple approach cannot reflect the characteristics of multivariate time series data. At least one statistical model should be included in the imputation model to reflect the various accident symptoms according to the accident type. Considering typical usage and performance, the three imputation methods are compared to replace the sensor fault data.

K-Nearest Neighbors (KNN)

KNN (K-nearest neighbors) is a single imputation method in which only a single computation is performed. The basic principle of KNN imputation is estimation by calculating the average of several neighbors. Based on the K parameter settings, the closest data are grouped into specific distance calculations including Manhattan distance, Euclidean distance, and correlation distance. Missing data is replaced with a weighted average of the nearest neighbors. In this embodiment, data is substituted using the Euclidean distance-based KNN method.

Multivariate Imputation with Chained Equations (MICE)

Several imputation methods have emerged to overcome the limitations of single imputation methods. The multiple imputation process is as follows. (1) A simple imputation (eg averaging) is performed with placeholders for all missing data. (2) The variable with the largest missing part is returned as missing data. (3) The variable is regressed from another variable. (4) The regression is repeated until the results converge. Regression models generally include linear, logistic, and Poisson regression. Specific regression models and convergence criteria are different for each MICE software package. In the model according to this example, linear regression with a convergence criterion of Δ < 0.1 was applied (ie, the relative change of the new imputation value from the old imputation value was less than 0.1).

Miss Forest

The Missforest imputation method proposed to deal with big data with missing sections is a multiple imputation method applicable to both categorical and numeric data. It has a similar imputation process to MICE, but the regression method is different. Missforest performs iterative random forest regression, which constructs numerous decision trees from training and integrates multiple decision trees to produce a mean regression. The computation time of Missforest can be controlled by the number of inputs, the number of trees, and the number of iterations.

GRUD (GRU-Decay)

In addition to replacing the missing data itself, it is also possible to utilize the missing data by modifying the structure of the RNN. Structurally, RNNs contain the same parameters for all time series data, so the input data must match the number of data dimensions of the trained model. Existing RNN models cannot produce an output if the input is missing, as accidental data loss can occur for all variables and scales. A GRUD model was proposed to utilize multivariate time series data with missing sections.

The GRUD model adds imputation and weight decay mechanisms to the underlying GRU structure to reduce the impact of missing data. The attenuation term (γ) represents the reduction in missing data and is determined in training. The attenuation mechanism in which the attenuation term is γ is expressed in Equation 5 below.

[Equation 5]

where γ _t is the attenuation term and m is the masking showing the missing state of the data. The attenuation term is determined by an exponentialized negative rectifier with trained weights and variable bias as in Equation (7) of Equation 5. The input and hidden states are modified input and hidden states as in Equations (8) and (9) of Equation 5, and decrease by the decay term over time.

로 대체된다. 이 값은 마지막으로 관찰된 값에서 감쇄항 γ_xt가 있는 매개 변수의 평균값으로 감쇠된다. 은닉 상태

은 감쇠율 γ_ht로 감쇠하며, 이는 출력 생성시 입력의 영향이 축소됨을 의미한다. 이것은 사고 진단 분류를 위한 시스템과 같이 실제 작업을 수행하는 RNN 모델에서 GRUD와 대치 로직을 결합하는 것과 같은 통합 모델 설계에 유리하다. 센서 결함 완화 시스템에서 누락된 데이터를 나타내는 필수 마스킹 입력 (수학식 4의 (6)식 참조)은 상술한 바와 같이 결함 모니터링 시스템에서 제공될 수 있다. 6 shows the structure of a GRUD in which an input is omitted. Since the current input x _t is not available

is replaced by This value is attenuated from the last observed value to the average of the parameters with the decay term γ _xt . stash

is attenuated with the attenuation factor γ _ht , which means that the influence of the input on generating the output is reduced. This is advantageous for integrated model design, such as combining GRUD and substitution logic in RNN models that perform real tasks, such as systems for accident diagnosis classification. The required masking input (refer to Equation (6) of Equation 4) representing missing data in the sensor fault mitigation system may be provided by the fault monitoring system as described above.

Using the above four approaches, two fault-tolerant diagnostic constructs were developed. The first structure is a method of replacing missing data with a regression-based imputation method (KNN, MICE or Missforest) and diagnosing it with GRU. The second structure is to input data with missing parts to GRUD. The three methods of the first structure are compared below.

4. Experimental results of a system resistant to sensor faults

Nuclear accident data were generated on a small scale on a compact nuclear simulator (CNS) at the Westinghouse 940 MWe pressurized waterway. The adopted CNS was developed by the Korea Atomic Energy Research Institute (KAERI). This simulator has been used as a source for several data-driven machine learning applications in the nuclear sector. The CNS can generate emergency or accident data with detailed malfunction options. It is a simplified one-dimensional model with theoretical assumptions that cannot simulate all accident phenomena, but can generate large amounts of data in a short time. Among the 2217 process parameters generated by the CNS, 41 parameters were selected here based on the existing diagnostic procedure of EOP and parameters representing specific accident symptoms were included. The data collection period was set to 900 seconds, referring to the recommended incident diagnosis time limit in the IAEA safety report.

All selected parameters show non-linear and unstable changes in emergency situations with reactor trips and operation of various safety systems. The data also includes detailed malfunction options, such as unexpected phenomena such as vibrations caused by vaporization, as well as various other accident symptoms. Table 1 lists the simulated incidents along with the number of relevant data sets.

By extracting 9 detailed accident sequence data from 5 broad nuclear accident categories, 1850 data were extracted, divided by severity or break location, generating various accident characteristics. The coolant loss accident (LOCA) data was divided into small / medium LOCA and large LOCA according to the break size in the standard, and a pilot operated relief valve (PORV) LOCA was added as a symptom different from other LOCA types. Excess Steam Demand Events (ESDEs), also known as main steam line breaks, have been separated into internal and external isolation as each exhibits very different symptoms. Reactor coolant pump (RCP) failure and reactor protection system (RPS) failure were selected as common spurious reactor trip accidents. Of the data, 453 test sets were randomly selected and the remaining 1397 were used for training and validation.

Before training the diagnostic algorithm, min-maximal normalization was performed on all training and test data based on the maximum and minimum variable data collected among all datasets for efficient training of the neural network model.

Accident Diagnosis Algorithm Test Results

The SoftMax function at the end of the GRU model produces a numerically normalized output, but there is no exact criterion for identifying the state. Although there are different symptoms depending on the type and scale of the accident, many studies have shown that RNNs can successfully identify accidents. To evaluate the stability and robustness of a diagnostic algorithm, it is necessary to generate a consistently high output of real accident labels. For this reason, a simple criterion for the thorough evaluation of the diagnostic algorithm was established. Sufficient time is required for symptoms to appear after an accident. In this embodiment, it is assumed that the success criterion of the accident diagnosis is when the actual SoftMax output maintains the maximum value from 200 seconds to the end of the simulation (900 seconds).

Before testing the injected sensor fault, the diagnostic algorithm composed of the GRU should be initially evaluated. The test results of the GRU model are described in Table 2 to classify the accident types among the 453 test sets with defect-free data. An instability trend was observed in the S/MLOCA and LLOCA test data with fracture sizes near the threshold. Nevertheless, the output maintained a correct diagnosis in all time sequences. Next, the degradation of the diagnostic algorithm was analyzed with five selected sensor failure modes, providing a total of 2265 test data for each sensor. The sensor fault data were generated for seven process parameters, which have different trends depending on the type of accident and have a significant impact on the diagnostic procedure sequence. In this embodiment, only a single sensor failure is assumed. As can be seen in Table 2, sensor faults significantly degrade diagnostic performance because the degree of degradation varies depending on the specific fault-injected sensor. Because this parameter is an important factor in differentiating steam generator tube rupture (SGTR) from other accidents, the greatest degradation occurred in secondary radiation sensor faults. Figure 7 shows the successful identification of SGTR in normal data and diagnostic failure due to secondary radiation sensor failure.

Evaluating the performance of imputation models

Accuracy was measured for the imputation methods described in section 3.3 to select the most precise imputation model. Mean imputation was also included in the comparative study because it is the default method for the three imputation methods of KNN, MICE, and Missforest. Errors in the data reconstructed from the original values were collected in a time-averaged format. Calculation time was also collected for performance comparison because it is an important factor in an accident situation that requires a quick response. Adjustment of model parameters affects both imputation accuracy and computation time, but these adjustments are not a major determinant of model performance here. The parameters for the imputation model (e.g., the number of nearest neighbors in KNN or the number of decision trees in Missforest) were fixed based on pilot tests that achieved the best performance in less than 20 seconds of computation time. Since only a single sensor fault was assumed in this study, MICE and Missforest deal with a single fit of linear and random forest regression in their tests.

Some process parameters in the NPP contain multiple zero values (eg certain radiation alarms that remain zero without radioactive material leakage). However, the actual value of the sensor should be inserted as the denominator of the error percentage metric, such as mean absolute percentage error. To solve this problem, Armstrong proposed a symmetric mean absolute percentage error (sMAPE) metric, which is defined as Equation 6 below.

[Equation 6]

where A _t represents the actual measured value at time t and F _t represents the imputed value. Although sMAPE has an asymmetry problem, it is not a problem for non-negative real data. Here, an evaluation using sMAPE was performed to select the optimal imputation method among four models (mean, KNN, MICE, and Missforest). Each method has important parameters that determine computation time and imputation accuracy. Comparison of imputation models was performed based on regression of missing sensor variables in the same sample data as the training data set. The percentage error and calculation time are listed in Tables 3 and 4, respectively. Five variables were randomly selected among 41 plant parameters to compare the characteristics of the variables and the reconstruction performance of the independent method.

Overall, Missforest provided the most stable results with the lowest errors. MICE seems to be a characteristic of linear regression, showing good performance on a simple pattern of variables (Var #4) regardless of the accident. However, in the case of Var #2 and #3 including constant zero data, the performance was rapidly decreased as shown in FIG. 8(d). KNN showed good performance at Min error, but showed an abnormal tendency due to unstable performance. As mentioned earlier, in terms of calculation time, which is an important factor determining applicability to actual nuclear emergency situations, the average calculation time for MICE was less than 5 seconds, Missforest about 9 seconds, and KNN about 17 seconds. In this structure, since the calculation time is determined by the number of branches, Missforest presented various calculation times between variables. Therefore, the regression of sensor values with various trends, such as Var #4, required a longer time to calculate.

Results showed that the Missforest-based imputation method performed best among the methods tested in terms of mean error, maximum peak error, and low computation time. Therefore, Missforest was chosen as the replacement tool in this work for signal reconstruction to replace missing data in sensor defects.

Defect Mitigation Results

GRUD and Missforest imputations were applied to unreliable sensor data to confirm the sensor fault mitigation strategy. Each imputation method was tested with a test set containing seven sensor defects. GRUD showed significant performance recovery under error conditions as shown in Table 5. The recovered diagnostic accuracy was directly affected by the decrease in accuracy. For example, the lowest accuracy among fault data for 'Secondary RAD' is associated with the lowest recovery accuracy in the mitigated results. For Missforest, all sensor faults were repaired, resulting in full diagnostic accuracy. The performance degradation and recovered diagnostic accuracy associated with 7 sensor faults are shown in FIG. 9 .

In this embodiment, first, it can be confirmed that the basic diagnostic algorithm using the GRU can successfully diagnose the prepared 453 test data with the assumed threshold value. After implanting the sensor fault, the diagnostic accuracy dropped to an average of 82.71%. That is, out of a total of 13,113 test data, 2742 failure cases occurred. By applying the GRUD-based fault tolerance strategy as a mitigation strategy for the first fault, a remarkable recovery of 96.75% was achieved, and 103 failure cases were observed out of a total of 3171 test data. As a second strategy, Missforest achieved full diagnostic accuracy recovery, ie, 100% with zero failures of 3171 test data. Table 5 lists the final test results.

Although it has been described with reference to the embodiment shown in the drawings in order to help the understanding of the present invention, this is an embodiment for implementation, it is merely an example, and various modifications and equivalents from those of ordinary skill in the art It will be appreciated that other embodiments are possible. Accordingly, the true technical protection scope of the present invention should be defined by the appended claims.

3.1, 3.2, 3.3: Each stage of the accident diagnosis system

Claims (10)

detecting a reactor trip occurrence;
checking whether a sensor defect has occurred;
Diagnosing an accident using a GRU algorithm for the sensor defect; and
and outputting the diagnosis result. According to claim 1,
The step of checking whether the sensor defect occurs,
If the consistency index value is greater than 0.7, it is determined that the sensor value is normal,
A reactor accident diagnosis method for determining that the sensor is defective when the consistency index value is less than 0.7. According to claim 1,
Diagnosing an accident using the GRU algorithm comprises:
A reactor accident diagnosis method performed by using the GRU algorithm and an algorithm for substituting data from a sensor in which a defect has occurred. 4. The method of claim 3,
The algorithm that replaces the data of the defective sensor is,
A reactor accident diagnosis method comprising any one of a K-nearest neighbors (KNN) algorithm, a MICE algorithm, and a Missforess algorithm. 5. The method of claim 4,
The KNN algorithm is
A reactor accident diagnosis method, which is an algorithm for replacing the average of K parameters close to the data of the defective sensor with the data of the defective sensor. 5. The method of claim 4,
The MICE algorithm is
The steps of simply replacing the missing data with a place holder, the step of returning the variable having the largest missing data as the missing data, and the process and result of regressing the variable from the other variable are converged A method of diagnosing a nuclear accident, an algorithm that iterates through regression until According to claim 1,
The Missforest algorithm is
constructing a plurality of decision trees; and
The method of diagnosing a nuclear accident, which is an algorithm comprising the step of integrating the decision tree to generate a mean regression and performing iterative random forest regression. According to claim 1,
Diagnosing an accident using the GRU algorithm comprises:
A method of diagnosing a nuclear accident that is performed without replacing the data of a defective sensor with the GRU algorithm. 9. The method of claim 8,
The step of diagnosing the accident without replacing the data of the sensor in which the defect has occurred with the GRU algorithm,
A nuclear accident diagnosis method performed using the GRUD (GRU-Decay) algorithm. 10. The method of claim 9,
The GRUD algorithm is
formula

A method of diagnosing a nuclear accident represented by .
(γt: attenuation term, m: masking,

where is a candidate gate, W: a weighted vector, x and h are input and hidden states, respectively, and ⊙ is element-wise multiplication)

Images