Devices and methods for using neural network models to track the basis for judging abnormal states

Abstract

This invention relates to an apparatus for tracking the criteria for judging abnormal states using a neural network model, comprising: an abnormality type classification unit, used to classify abnormal states into multiple faults in an abnormal operation scenario storing multiple scenarios related to the abnormal state; an operation variable derivation unit, used to derive operation variables that affect the abnormal state judgment result for each of the classified multiple faults; a power plant operation variable weighting unit, used to weight the variables related to the abnormal state among the operation variables; and an abnormal state judgment criteria generation unit, used to track the criteria for judging abnormal states from the abnormal state judgment results generated by power plant operation variables with assigned weight values.

Description

Technical Field

[0001] The research related to this patent was conducted under the supervision of the Ministry of Trade, Industry and Energy and with the support of the Nuclear Power Core Technology Development Project (Project Name: Development of Resource Technology for Start-up and Shutdown Operation of Nuclear Power Plants Based on Artificial Intelligence, Project No.: 1415159084).

[0002] This invention relates to an apparatus and method for tracking the criteria for judging abnormal states, and more specifically, to an apparatus and method for using a neural network model to judge abnormal states and track the criteria for judgment. Background Technology

[0003] Nuclear power plants are susceptible to various abnormal operating conditions. When an abnormal operating condition occurs, an alarm is generated in the main control room, and the relevant status of the power plant changes. These changes include temperature, pressure, and flow rate.

[0004] Based on the power plant's alarms, operators determine the type of abnormal operating condition that has occurred and take appropriate measures according to the procedures for that abnormal operating condition.

[0005] However, since there are hundreds of abnormal operating conditions in nuclear power plants, and inexperienced operators may not be able to accurately identify abnormal conditions, they may not be able to take the correct measures to deal with them.

[0006] Therefore, a method is currently being researched that uses neural network models to learn from abnormal operating data to determine the abnormal state of nuclear power plants. This would allow operators to promptly identify abnormal states caused by equipment or facility malfunctions and then take appropriate measures. However, neural network models struggle to track which changes in the plant's operating data are the basis for determining abnormal states.

[0007] [Related Technical Documents]

[0008] [Patent Documents]

[0009] (Patent Document 1) Korean Patent 2095653 (Device and method for diagnosing abnormal operation status using neural network model) of Korea Nuclear Power Co., Ltd.

[0010] (Patent Document 2) US Patent 10452845 (Generic framework to detect cyber threats in electric power grid).

[0011] (Patent Document 2) (US Patent 20190164057 (Mapping and quantification of influence of neural network features for explainable artificial intelligence) by Intel Corporation). Summary of the Invention

[0012] Technical issues

[0013] Therefore, the present invention was made with regard to the aforementioned problems that have occurred in the related art, and one object of the present invention is to provide a method for estimating operating variables as a basis for judging the abnormal state of a nuclear power plant using a neural network model, and an apparatus for tracking the basis for judging the abnormal state using a neural network model.

[0014] Technical solution

[0015] In one aspect, an apparatus for tracking abnormal state judgment criteria using a neural network model includes: an abnormality classification unit for classifying abnormal states into multiple faults in an abnormal operation scenario storing multiple scenarios related to the abnormal state; an operation variable derivation unit for deriving operation variables that affect the abnormal state judgment result for each of the classified multiple faults; a power plant operation variable weighting unit for assigning weight values ​​to variables related to the abnormal state among the operation variables; and an abnormal state judgment criteria generation unit for tracking the abnormal state judgment criteria from the abnormal state judgment result generated by the power plant operation variables with assigned weight values.

[0016] In addition, the power plant operation variable weighting unit, which assigns weight values ​​to variables related to abnormal states in the operating variables, classifies and assigns weight values ​​to physical variables related to abnormal states, taking into account the physical correlation of the system with respect to abnormal states.

[0017] In addition, the anomaly classification unit categorizes abnormal operating scenarios into at least one of valve leakage, pump failure, heat exchanger failure, and coolant leakage.

[0018] Furthermore, the operational variable derivation unit for deriving operational variables that affect the abnormal state judgment result for each of the multiple classified faults includes: when the fault is classified as valve leakage, deriving the flow rate of the system associated with the corresponding valve; when the fault is classified as pump fault, deriving the flow rate and pressure of the system associated with the corresponding pump; when the fault is classified as heat exchanger fault, deriving the temperature of the system associated with the corresponding heat exchanger; and when the fault is classified as coolant leakage, deriving the radiation level at the leakage site.

[0019] In addition, the physical variables related to abnormal states are established with reference to abnormal procedures or actual power plant operating history.

[0020] In addition, the basis for abnormal state judgment is the running variable that can be distinguished from different abnormal states, and it is used to verify the abnormal state judgment logic used in the abnormal state judgment system.

[0021] In addition, the criteria for judging abnormal states are used to verify the judgment logic recorded in the abnormal program, and the abnormal program records the runtime variables that change when an abnormal state occurs.

[0022] In addition, a method for generating abnormal state judgment criteria using a neural network model includes: learning from power plant operation data and a neural network model to generate abnormal state judgment results in the final stage of the neural network model; and extracting variable values ​​that affect the abnormal state judgment results by performing an impact analysis on the abnormal state judgment results in a fully connected layer before generating the abnormal state judgment results.

[0023] In addition, the variable values ​​that affect the abnormal state judgment results are extracted as follows: virtual input change data is generated by applying visualization algorithms; the impact of input change data on the abnormal state judgment results is analyzed by using virtual input change data as input for neural network model calculation; and the input change data that contributes the most to the derivation of the abnormal state judgment results is extracted.

[0024] Invention Effects

[0025] The apparatus and method for tracking abnormal state judgment criteria using a neural network model according to embodiments of the present invention can accurately determine the type of abnormal state in a short time when various types of abnormal states occur and provide the type to the operators, thereby enabling rapid and accurate response to abnormal states in nuclear power plants and improving the safety of nuclear power plants. Attached Figure Description

[0026] Figure 1 This is a schematic diagram illustrating an existing abnormal state detection device that uses a neural network model.

[0027] Figure 2 This diagram schematically illustrates the operation of a device that uses a neural network model to track the basis for judging abnormal states according to an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram illustrating the process of using a neural network model to extract operational variables that affect the judgment result of an abnormal state according to an embodiment of the present invention.

[0029] Figure 4 This is a diagram illustrating the operation process of generating the criteria for judging abnormal states according to an embodiment of the present invention.

[0030] Figure 5 This is a schematic block diagram illustrating the configuration of an apparatus for extracting criteria for judging abnormal states according to an embodiment of the present invention.

[0031] Figure 6 This is a diagram illustrating the use of an abnormal state judgment criterion derived by applying a neural network model according to an embodiment of the present invention.

[0032] [Explanation of reference numerals in the attached figures]

[0033] 100: Abnormal operating status judgment device

[0034] 200, 500: Devices used to determine abnormal states.

[0035] 210: Power Plant Operation Data

[0036] 230: Neural Networks

[0037] 510, 600: Abnormal operating scenarios

[0038] 520: Anomaly Classification Unit

[0039] 530: Runtime Variable Derivation Unit

[0040] 540: Weighted Unit for Power Plant Operation Variables

[0041] 550: Abnormal state judgment basis generation unit. Detailed Implementation

[0042] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Some components unrelated to the key points of the present invention will be omitted or reduced, but the omitted components are not unnecessary in the present invention and can be combined and used by those skilled in the art.

[0043] Figure 1 This is a schematic diagram illustrating an existing abnormal state detection device that uses a neural network model.

[0044] like Figure 1As shown, the abnormal operation state judgment device 100 using a neural network model may include an abnormal operation state data generation unit 110, an abnormal operation state data learning unit 130, an abnormal operation state judgment unit 150, and an abnormal operation state monitoring unit 170.

[0045] The abnormal operation status data generation unit 110 is a component that virtually generates abnormal operation status data based on information about abnormal operation status, and may include a scenario database 111 and a simulator 112.

[0046] Scenario database 111 is configured to include multiple scenarios related to abnormal operating states. These scenarios include scenarios based on operating variables related to temperature changes in the nuclear power plant unit, scenarios based on operating variables related to turbine bearing vibration, etc., and scenarios related to various abnormal operating states are stored in scenario database 111. Here, operating variables are operating factors of the operating state of the nuclear power plant unit, and for each unit, approximately 1000 to 2000 operating variables may be included. These operating variables may include pressure, temperature, flow rate, etc.

[0047] Simulator 112 is configured to simulate abnormal operating states in relation to scenes selected from scenes related to abnormal operating states stored in scene database 111. Therefore, data regarding abnormal operating states can be virtually generated.

[0048] The abnormal operation state data learning unit 130 is configured to visualize and learn abnormal operation states by applying visualization algorithms based on the abnormal operation state data generated in the abnormal operation state data generation unit 110. The abnormal operation state data learning unit 130 may include a first visualization arrangement unit 131 and a second visualization arrangement unit 132.

[0049] The first visualization layout unit 131 is configured to arrange the devices provided in the nuclear power plant based on the physical location of the operating variables. That is, the operating variables can be arranged according to the same layout as the actual nuclear power plant.

[0050] The second visualization arrangement unit 132 is configured to prioritize the arrangement of physically identical operating variables. For example, temperature-related operating variables are arranged in the same area to show the characteristics of each event when a temperature change occurs.

[0051] The abnormal operating state determination unit 150 is configured to learn abnormal operating states based on the operating variables indicated by the abnormal operating state data learning unit 130 using a visualization algorithm, and to determine whether an abnormal operating state has occurred based on the operating variables of the device acquired by the process monitoring and warning system of the nuclear power plant. This abnormal operating state determination unit 150 may include a neural network model 151 and a signal matching unit 152.

[0052] The neural network model 151 is configured to learn abnormal operating state data visualized by the first visualization arrangement unit 131 and the second visualization arrangement unit 132 based on the visualization algorithm.

[0053] The signal matching unit 152 is configured to transmit information about the monitoring signal, including information about abnormal operating conditions, to the corresponding device.

[0054] The abnormal operating status monitoring unit 170 is configured to monitor whether the operating status of each unit provided in the nuclear power plant is within the normal range. This type of abnormal operating status monitoring unit 170 can periodically acquire monitoring signals including information about the operating variables of each unit and transmit the monitoring signals to the abnormal operating status judgment unit 150.

[0055] Figure 2 This diagram schematically illustrates the operation of a device that uses a neural network model to track the basis for judging abnormal states according to an embodiment of the present invention.

[0056] like Figure 2 As shown, the abnormal state judgment tracking device 200 uses power plant operation data 210 and a neural network model 230 to perform learning to generate an abnormal state judgment result 250. The neural network model 230 performs calculations through a multi-layer neural network (deep learning) to effectively learn each abnormal state. In the final stage of the neural network model 230, the abnormal state judgment result 250 is generated. According to the present invention, on the fully connected layer before generating the abnormal state judgment result 250, an impact analysis 270 is performed on the abnormal state judgment result 250 to extract the variable values ​​that affect the abnormal state judgment result 250.

[0057] Figure 3 This is a schematic diagram illustrating the process of using a neural network model to extract operational variables that affect the judgment result of an abnormal state according to an embodiment of the present invention.

[0058] like Figure 3As shown, according to the present invention, firstly, visualized input change data 310 is virtually generated by applying a visualization algorithm. The impact of the virtual input change data 310 on the change in result 350 of each item is analyzed by using the virtual input change data 310 as input for computation by the neural network model 330, and the input change data 310 that contributes the most to the change in the derivation result 350 is extracted.

[0059] Therefore, visualization algorithms are beneficial for extracting corresponding features during the preprocessing convolution and pooling of neural network models 330 when collecting and arranging running variables at the location where actual events (such as pipe breaks) occur.

[0060] Figure 4 This is a diagram illustrating the operation process of generating the criteria for judging abnormal states according to an embodiment of the present invention.

[0061] Figure 5 This is a schematic block diagram illustrating the configuration of an apparatus for extracting criteria for judging abnormal states according to an embodiment of the present invention.

[0062] like Figure 4 and Figure 5 As shown, the abnormal state judgment basis tracking device 500 may include an abnormal operation scenario 510, an abnormal classification unit 520, an operation variable derivation unit 530, a power plant operation variable weighting unit 540, and an abnormal state judgment basis generation unit 550.

[0063] Abnormal operating scenario 500 provides multiple scenarios related to abnormal operating states. These scenarios include those based on operating variables related to temperature changes in nuclear power plant equipment and those based on operating variables related to turbine bearing vibration. The anomaly classification unit 520 classifies abnormal operating scenario 500 into categories such as valve leakage, pump failure, heat exchanger failure, and coolant leakage. For each classified failure, the operating variable derivation unit 530 derives the operating variables that affect the anomaly judgment result. According to an embodiment, in the case of valve leakage, the flow rate of the system related to the corresponding valve is derived; in the case of pump failure, the flow rate and pressure of the system related to the corresponding valve are derived; in the case of heat exchanger failure, the temperature of the system related to the corresponding heat exchanger is derived; and in the case of coolant leakage, the radiation level at the leak location is derived. The input range affecting the abnormal state result includes multiple uncertainties. To extract the basis for the abnormal state judgment result, it is important to change the inputs physically related to the corresponding abnormal state.

[0064] Therefore, considering the physical correlation of the relevant systems, the operating variables used to determine the extracted abnormal states are used as the basis for judging the neural network. Based on the information about the abnormal states, physical classification is performed, and weight values ​​are assigned to the physical variables associated with each corresponding abnormal state.

[0065] In this embodiment, abnormal conditions are classified as valve leaks, and a high-weighted value is assigned to the system flow rate, which is correlated with the corresponding system thermo-hydraulic pressure and used to explain the basis for judging the abnormal condition. Operating variables physically related to the abnormal condition are selected with reference to abnormal procedures or actual power plant operating history.

[0066] The power plant operation variable weighting unit 540 assigns weight values ​​to the physical variables related to the abnormal state in the operation variables of each fault, and finally, the abnormal state judgment basis generation unit 550 can more accurately track the basis by reflecting the results of the weight values ​​in it.

[0067] Figure 6 This is a diagram illustrating the use of an abnormal state judgment criterion derived by applying a neural network model according to an embodiment of the present invention.

[0068] like Figure 6 As shown, the abnormal state judgment criterion 620 derived from abnormal operation simulation data 610, which simulates an abnormal operation scenario selected from multiple scenarios related to abnormal states, is a power plant operation variable that can distinguish the corresponding abnormal state from other abnormal states. This can be used to verify 630 the abnormal state judgment logic used in the abnormal state judgment system. That is, it can be confirmed whether the abnormal state judgment logic is developed by utilizing operation variables that change due to abnormal states.

[0069] Furthermore, the exception procedure records the runtime variables that change when the corresponding exception state occurs. The exception procedure's judgment logic can be verified using exception procedure judgment logic verification 640 and exception state judgment criteria 620. The program can also be verified and modified, allowing operators to more effectively judge exception states.

Claims

1. An apparatus for tracking abnormal state judgment criteria using a neural network model, the apparatus comprising: An anomaly classification unit is used to classify the anomaly state into multiple faults in an anomaly operation scenario that stores multiple scenarios related to the anomaly state. The variable derivation unit is used to derive the operational variables that affect the abnormal state judgment result for each of the multiple classified faults. A power plant operation variable weighting unit is used to assign weight values ​​to variables among the operation variables that are related to the abnormal state. as well as An abnormal state judgment basis generation unit is used to track the abnormal state judgment basis from the abnormal state judgment results generated by power plant operation variables assigned the weight values. The abnormal state judgment generation unit generates the abnormal state judgment result by utilizing power plant operation data and a trained neural network model. In the fully connected layer of the neural network model immediately preceding the output of the abnormal state judgment result, the influence of the input change data that contributes to the derivation of the abnormal state judgment result is analyzed, and the variable values ​​that affect the abnormal state judgment result are extracted.

2. The apparatus for tracking abnormal state judgment criteria using a neural network model according to claim 1, wherein, The power plant operating variable weighting unit, used to assign the weight values ​​to the variables in the operating variables that are related to the abnormal state, classifies the variables by taking into account the physical correlation of the system related to the abnormal state and assigns the weight values ​​to the physical variables related to the abnormal state.

3. The apparatus for tracking abnormal state judgment criteria using a neural network model according to claim 1, wherein, The anomaly classification unit classifies the abnormal operating scenarios into at least one of valve leakage, pump failure, heat exchanger failure, and coolant leakage.

4. The apparatus for tracking abnormal state judgment criteria using a neural network model according to claim 3, wherein, In the operating variable derivation unit for each of the multiple faults classified as such, which derives the operating variable that affects the abnormal state judgment result: when the fault is classified as valve leakage, the flow rate of the system associated with the corresponding valve is derived; When the fault is classified as a pump fault, derive the flow rate and pressure of the system associated with the corresponding valve; when the fault is classified as a heat exchanger fault, derive the temperature of the system associated with the corresponding heat exchanger; and when the fault is classified as a coolant leak, derive the radiation level at the leak location.

5. The apparatus for tracking abnormal state judgment criteria using a neural network model according to claim 2, wherein, The physical variables associated with the abnormal state are established with reference to the abnormal procedure or the actual operating history of the power plant.

6. The apparatus for tracking abnormal state judgment criteria using a neural network model according to claim 1, wherein, The abnormal state judgment criteria are the running variables that can be distinguished from different abnormal states, and are used to verify the abnormal state judgment logic used in the abnormal state judgment system.

7. The apparatus for tracking abnormal state judgment criteria using a neural network model according to claim 5, wherein, The abnormal state judgment criteria are used to verify the judgment logic recorded in the abnormal program, and the abnormal program records the running variables that change when the abnormal state occurs.

8. The apparatus for tracking abnormal state judgment criteria using a neural network model according to claim 1, wherein, The variable derivation unit is configured to: virtually generate visualized input change data by applying a visualization algorithm; and analyze the impact of the input change data on the abnormal state judgment result by using the virtual input change data as input for the calculation of the neural network model. And extract the input change data that contributes the most to the derivation of the abnormal state judgment result.

Images