A nuclear reactor core detector multi-model anomaly detection method and system

Abstract

This invention discloses a multi-model anomaly detection method and system for nuclear reactor core detectors, relating to the field of nuclear reactor monitoring and fault diagnosis technology. This invention determines the horizontal distance between adjacent sample detector nodes by acquiring power distribution observation data of all sample detector nodes within the nuclear reactor core. r At each sample detection node, select r 2 r , r The neighboring node data and distance data are input into a machine learning regressor to construct four spatial neighborhood observation models, obtaining predicted power distribution observations and calculating the relative errors between these predictions and actual power distribution observations. A statistical fusion technique is used to fuse the four relative errors of each target detection node into an integrated error. A normal upper limit is determined based on the statistical value of the integrated error and a global threshold coefficient; values ​​exceeding this limit are considered abnormal, and the anomaly detection result is displayed. This method improves the accuracy of anomaly detection in nuclear reactor core detectors.

Description

Technical Field

[0001] This application relates to the field of nuclear reactor monitoring and fault diagnosis technology, and in particular to a method and system for multi-model anomaly detection of nuclear reactor core detectors. Background Technology

[0002] With the continuous development of digital transformation and predictive maintenance technologies in nuclear power plants, online monitoring and anomaly detection of in-core detectors have become one of the core tasks in the field of safe nuclear reactor operation. Accurately identifying detector malfunctions and abnormal core physical conditions can provide important data support for reactor operation and maintenance management, maintenance decisions during shutdown and refueling, radiation protection optimization, and nuclear safety assessment. Its core purpose is to solve the problem of anomaly detection at multiple detector nodes in complex core layouts, overcome the limitations of traditional detection methods such as low efficiency and insufficient diagnostic information, and meet the needs of refined operation and maintenance management of nuclear power plants.

[0003] In existing technologies, anomaly detection in nuclear reactor core detectors typically employs pure signal processing or a single machine learning model to analyze detector data, presenting the anomaly determination results to the operator in numerical form, which is then manually verified by the operator, and maintenance is scheduled during reactor shutdowns and refueling. However, existing anomaly detection methods for nuclear reactor core detectors still face many technical bottlenecks, such as high false alarm and false negative rates, insufficient robustness, high maintenance costs, poor interpretability, and limited adaptability, which restrict the accuracy of anomaly detection. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, system, medium, and equipment for multi-model anomaly detection of nuclear reactor core detectors to address the above-mentioned technical problems.

[0005] The following technical solution is adopted in this specification: This specification provides a multi-model anomaly detection method for nuclear reactor core detectors, including: Acquire power distribution observation data of all sample detection nodes within the nuclear reactor core and determine the horizontal distance r between adjacent sample detection nodes; At each sample detection node, select multiple adjacent nodes that are r units away from the sample detection node, and nodes that are r units away from the sample detection node. Multiple neighboring nodes of a unit, multiple neighboring nodes 2r units away from the sample detection node, and nodes r units away from the sample detection node and The power distribution observation data of multiple neighboring nodes of a unit are used as 4 sets of input data, which are then input into the corresponding machine learning regressors to construct four spatial neighborhood observation models. Four types of observation data were collected from the target detection node at multiple consecutive time points; For any target detection node at any given time, the four types of observation data of the target detection node are input into the four pre-constructed spatial neighborhood observation models, and the four predicted power distribution observations of the target detection node are output. For any spatial neighborhood observation model, the relative error of the spatial neighborhood observation model on the target detection node is determined based on the actual observed power distribution and the predicted power distribution of the target detection node at all times. A statistical fusion technique is used to fuse the relative errors of four spatial neighborhood observation models at the target detection node into an integrated error of the target detection node; The normal upper limit is determined based on the statistical value of the integration error and the global threshold coefficient. When the integration error of the target detection node is greater than the normal upper limit, it is judged as abnormal and the abnormal detection result is displayed.

[0006] Optionally, the machine learning regressor is selected from... K Nearest neighbor regressor, random forest regressor, gradient boosting decision tree regressor, extreme gradient boosting tree regressor, or fully connected neural network regressor.

[0007] Optionally, based on the actual observed power distribution and the predicted power distribution of the target detection node at all times, the relative error of the spatial neighborhood observation model at the target detection node is determined, including: The ratio of the absolute value of the difference between the actual observed power distribution value and the predicted power distribution value at each time moment to the actual observed power distribution value is taken as the relative error at that time moment. The arithmetic mean of the relative errors at all times is then calculated as the relative error of the spatial neighborhood observation model at the target detection node.

[0008] Optionally, a statistical fusion technique is employed to fuse the relative errors of the four spatial neighborhood observation models at the target detection node into an integrated error for the target detection node, including: The confidence interval filter technique is used to calculate the mean and standard deviation of the four relative errors of each target detection node, and retain the relative errors within the range of mean plus or minus a preset multiple of standard deviation. The arithmetic mean of the retained relative errors is taken as the integrated error of the target detection node.

[0009] Optionally, a statistical fusion technique is employed to fuse the relative errors of the four spatial neighborhood observation models at the target detection node into an integrated error for the target detection node, including: Obtain the standard deviation of the errors for the four spatial neighborhood observation models; The odd-even space averaging technique is used to calculate the square root of the sum of the squares of the standard deviations of any two models in the four spatial neighborhood observation models, and obtain the root mean square accuracy. For any two models, if the absolute value of the difference between the relative errors of the two models is less than or equal to their root mean square accuracy, then a preset weight is assigned to the corresponding model. The weighted average of the relative errors of each model is determined as the integrated error of the target detection node.

[0010] Optionally, the statistics of the integration error include the mean and standard deviation of the integration error. A normal upper limit is determined based on the statistics of the integration error and the global threshold coefficient, including: The sum of the products of the average value of the integration error and the global threshold coefficient with the standard deviation of the integration error is determined as the normal upper limit.

[0011] Optionally, the global threshold coefficient can be determined in the following ways: Traverse the threshold coefficient range, for any sample threshold coefficient in the threshold coefficient range, detect abnormal samples of the sample detection node, and determine the accuracy of the sample threshold coefficient based on the detected abnormal samples. The sample threshold coefficient corresponding to the highest accuracy is determined as the global threshold coefficient.

[0012] Optionally, the method further includes: A heatmap is generated based on the integrated error of all target detection nodes, and then displayed.

[0013] This specification provides a multi-model anomaly detection system for nuclear reactor core detectors, including: The data acquisition unit consists of multiple self-powered neutron detector nodes. The detector nodes are divided into multiple layers along the axial direction, with multiple nodes arranged in each layer for real-time acquisition of power distribution data. The data transmission unit, connected to the data acquisition unit, is used to transmit the probe data to the processing unit in real time; The processing unit, connected to the data transmission unit, is used to acquire power distribution observation data of all sample detection nodes within the nuclear reactor core and determine the horizontal distance between adjacent sample detection nodes. r At each sample detection node, a distance of [distance] from that sample detection node is selected. r Multiple adjacent nodes of the unit, and the distance from the sample detection node is Multiple adjacent nodes of the unit, with a distance of 2 from the sample detection node r Multiple adjacent nodes of the unit, and the distance from the sample detection node is r Units and The power distribution observation data of multiple neighboring nodes of a unit are used as four sets of input data, which are respectively input into the corresponding machine learning regressors to construct four spatial neighborhood observation models. Four types of observation data of the target detection node are collected at multiple consecutive time points. For any target detection node at any time point, the four types of observation data of the target detection node are respectively input into the four constructed spatial neighborhood observation models, outputting four predicted power distribution observation values ​​for that target detection node. For any spatial neighborhood observation model, the relative error of the spatial neighborhood observation model on the target detection node is determined based on the actual observed power distribution and predicted power distribution observation values ​​of the target detection node at all time points. A statistical fusion technique is used to fuse the relative errors of the four spatial neighborhood observation models on the target detection node into an integrated error for the target detection node. A normal upper limit is determined based on the statistical value of the integrated error and the global threshold coefficient. When the integrated error of the target detection node exceeds the normal upper limit, it is judged as an anomaly, and the anomaly detection result is displayed.

[0014] Optionally, the system also includes a visual interaction unit; The visualization and interaction unit, connected to the processing unit, is used to display heatmaps, hyperparameter fine-tuning, and annotation of abnormal results.

[0015] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described multi-model anomaly detection method for nuclear reactor core detectors.

[0016] This specification provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described multi-model anomaly detection method for nuclear reactor core detectors.

[0017] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: In the multi-model anomaly detection method for nuclear reactor core detectors provided in this specification, power distribution observation data of all sample detector nodes within the nuclear reactor core are acquired, and the horizontal distance between adjacent sample detector nodes is determined. r At each sample detection node, select respectively r , , 2r , r and The distance data of neighboring nodes are input into four machine learning regressors to construct four spatial neighborhood observation models. The target detection node data are input into the models to obtain four predicted power distribution observations, and the relative errors between these predictions and the actual power distribution observations are calculated. A statistical fusion technique is used to fuse the four relative errors of each target detection node into an integrated error. The normal upper limit is determined based on the statistical value of the integrated error and the global threshold coefficient. If the error exceeds the normal upper limit, it is judged as an anomaly, and the anomaly detection result is displayed.

[0018] This invention addresses the complex spatial distribution of detectors in nuclear reactor cores by selecting [a specific location] at each sample detection node. r , , 2r , r and By constructing four spatial neighborhood observation models based on the distance data of neighboring nodes, and through multi-model parallel prediction and statistical fusion technology, the spatial correlation information of neighboring nodes in different directions and at different distances is comprehensively utilized, which effectively avoids misjudgment of local anomalies or global distortions by a single spatial neighborhood model. At the same time, the upper limit of normal is determined by combining the statistical value of the integrated error with the global threshold coefficient, which improves the anomaly judgment from the traditional fixed threshold judgment to adaptive statistical inference. It can dynamically adjust the judgment benchmark according to the overall power distribution of the reactor core, thereby significantly improving the accuracy of anomaly detection and reducing the false alarm rate and false negative rate. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0020] Figure 1 This document provides a schematic flowchart of a multi-model anomaly detection method for a nuclear reactor core detector. Figure 2 This document provides schematic diagrams of the spatial layout of four neighborhood-based models. Figure 3 This is an example of a layered heatmap corresponding to the first detector layer provided in this specification; Figure 4 This is the flowchart for the anomaly detection provided in this manual; Figure 5 This specification provides an overall architecture diagram of a multi-model anomaly detection system for a nuclear reactor core detector. Figure 6 This is a schematic diagram of the visual interactive unit interface provided in this manual; Figure 7This is a schematic diagram of a computer device for implementing a multi-model anomaly detection method for nuclear reactor core detectors, as provided in this specification. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0022] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0023] Figure 1 This is a schematic diagram of a multi-model anomaly detection method for a nuclear reactor core detector as described in this specification, which specifically includes the following steps: S101: Acquire power distribution observation data of all sample detection nodes within the nuclear reactor core and determine the horizontal distance between adjacent sample detection nodes. r .

[0024] In one embodiment of this specification, the power distribution observation data comes from the actual operation records of a Hualong One (HPR1000) unit of China General Nuclear Power Group Co., Ltd., and was continuously collected for 166 hours at a sampling frequency of 1 Hz, generating a total of 597,600 time steps. A total of 294 detector nodes (7 layers * 42 nodes per layer) are arranged within the reactor core, forming a total of 175,694,400 valid data points. No additional normalization or standardization processing is required, and these data points can be directly used as model input.

[0025] The detection nodes are arranged in a grid-like, regularly spaced pattern on a two-dimensional plane (radial) within the nuclear reactor core, and uniformly layered in the vertical direction (axial direction), forming a three-dimensional monitoring network to achieve comprehensive and accurate sensing of the core's condition. The horizontal or vertical distance between adjacent sample detection nodes is [missing information]. r This distance is also the minimum distance between adjacent detection nodes of the sample.

[0026] S102: At each sample detection node, select a distance of [distance from the sample detection node]. r Multiple adjacent nodes of the unit, and the distance from the sample detection node is Multiple adjacent nodes of the unit, with a distance of 2 from the sample detection node r Multiple adjacent nodes of the unit, and the distance from the sample detection node is r Units and The power distribution observation data of multiple neighboring nodes of a unit are used as four sets of input data, which are then input into the corresponding machine learning regressors to construct four spatial neighborhood observation models.

[0027] Among them, the machine learning regressor is selected from K Nearest neighbor regressor, random forest regressor, gradient boosting decision tree regressor, extreme gradient boosting tree regressor, or fully connected neural network regressor. In one embodiment of this specification, a regressor is selected. K The nearest neighbor regressor uses GridSearchCV for adaptive parameter tuning and five-fold cross-validation to determine the optimal parameters. Its training speed is 2 seconds per model, which is much faster than the random forest regressor (19 minutes per model) and the fully connected neural network regressor (25 minutes per model). Moreover, its prediction error differs from other regressors by ≤0.0014, and it supports rapid deployment and iteration.

[0028] In one embodiment of this specification, the four spatial neighborhood observation models (M1-M4) are specifically configured such that, at each sample detection node: M1: Select a distance from the detection node of this sample. r The power distribution observation data of the unit's four neighboring nodes are used as input data, and the corresponding machine learning regressor is used to train the M1 spatial neighborhood observation model.

[0029] M2: Select a distance from the detection node of this sample. The power distribution observation data of the unit's four neighboring nodes are used as input data, and the corresponding machine learning regressor is used to train the M2 spatial neighborhood observation model.

[0030] M3: Select a distance of 2 from the detection node of this sample. r The power distribution observation data of the unit's four neighboring nodes are used as input data, and the corresponding machine learning regressor is used to train the M3 spatial neighborhood observation model.

[0031] M4: Select a distance from the detection node of this sample. r Units and The power distribution observation data of the unit's eight neighboring nodes are used as input data, and the corresponding machine learning regressor is used to train the M4 spatial neighborhood observation model.

[0032] like Figure 2 As shown, Figure 2 The following are schematic diagrams of the spatial layout of the four neighborhood-based models provided in this specification: (a) is the M1 spatial neighborhood observation model, (b) is the M2 spatial neighborhood observation model, (c) is the M3 spatial neighborhood observation model, and (d) is the M4 spatial neighborhood observation model. Red dots represent sample detection nodes, and black dots represent selected adjacent detection nodes.

[0033] In one embodiment of this specification, to improve the robustness of the machine learning regressor to industrial noise, a noise augmentation training technique is used to train the model. Specifically, this is achieved through the formula:

[0034] Gaussian noise is injected into the original training data, where Mean 0, standard deviation Gaussian noise, The value range is [0.01, 0.10]. Specifically, To inject noise levels into the training set during the training phase: each original training sample is independently sampled for noise 500 times, thus expanding the training set size by 500 times. The noise level of the validation set is injected during the testing phase to simulate measurement disturbances in a real industrial environment. After training, a [system / structure] is established. and optimal The mapping relationship is used to achieve adaptive noise matching, that is, to select a certain... The trained model, when subjected to When the minimum prediction error is achieved on the perturbed test set, this That is, the corresponding The optimal training noise. In noise-enhanced training, and optimal The mapping relationship is: when The optimal values ​​are 0.00, 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, and 0.10. These correspond to values ​​of 0.00, 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.08, and 0.09, respectively. Noise-enhanced training reduced the difference between the prediction error and the actual noise error by 53%. When the value is 0.10, the prediction error is only 70%-80% of that of the noiseless training model, which improves the robustness of anomaly detection.

[0035] S103: Collect four types of observation data from the target detection node at multiple consecutive time points.

[0036] In one embodiment of this specification, continuous data acquisition is performed. Group power distribution observation data, as a continuous Four types of observation data at each time point, among which It is an integer greater than 1.

[0037] S104: For any target detection node at any given time, input the four types of observation data of the target detection node into the four existing spatial neighborhood observation models, and output the four predicted power distribution observations of the target detection node.

[0038] S105: For any spatial neighborhood observation model, determine the relative error of the spatial neighborhood observation model on the target detection node based on the actual observed value and the predicted observed value of the power distribution of the target detection node at all times.

[0039] In one embodiment of this specification, the relative error of the spatial neighborhood observation model on the target detection node is determined based on the actual observed values ​​and predicted observed values ​​of the power distribution of the target detection node at all times, including: The ratio of the absolute value of the difference between the actual observed power distribution value and the predicted power distribution value at each time step to the actual observed power distribution value is taken as the relative error at that time step. The arithmetic mean of the relative errors at all times is then calculated as the relative error of the spatial neighborhood observation model at the target detection node. The calculation formula is as follows: in, In the first i At the target detection node, the actual observed power distribution value is the same as that at the target detection node. m The relative error of the power distribution prediction observations obtained from the spatial neighborhood observation model is, in one embodiment of this specification, 20 sets of continuous observation data are collected, namely =20, j Integer and j The value range is [1, 20]. For the first i The target detection node at the ... j Actual observed power distribution values ​​at each time point. For the first i The target detection node at the ... j The moment of the first Power distribution prediction observations obtained from a spatial neighborhood observation model.

[0040] S106: Using a statistical fusion technique, the relative errors of the four spatial neighborhood observation models at the target detection node are fused into the integrated error of the target detection node.

[0041] In one embodiment of this specification, a statistical fusion technique is employed to fuse the relative errors of four spatial neighborhood observation models at the target detection node into an integrated error for the target detection node. This includes: using a confidence interval filter technique to calculate the mean and standard deviation of the four relative errors for each target detection node, retaining the relative errors within the range of the mean plus or minus a preset multiple of the standard deviation, and taking the arithmetic mean of the retained relative errors as the integrated error of the target detection node. The integrated error is calculated as follows: Calculate the mean of the four relative errors for each target detection node. with standard deviation Integral coefficients of confidence interval filters =1.4, retain error in [ , For data within the range, take the average as the first... i Integrated error of individual target detection nodes .

[0042] In another embodiment of this specification, a statistical fusion technique is used to fuse the relative errors of four spatial neighborhood observation models at the target detection node into an integrated error of the target detection node, including the following steps: S201, Obtain the standard deviation of the errors of the four spatial neighborhood observation models. , , , .

[0043] S202 uses the odd-even space averaging technique to calculate the square root of the sum of the squares of the error standard deviations of any two models in the four spatial neighborhood observation models, and obtains the root mean square accuracy; for any two models, if the absolute value of the difference between the relative errors of the two models is less than or equal to their root mean square accuracy, then a preset weight is assigned to the corresponding model.

[0044] The root mean square accuracy is: in, , Let the standard deviation of the error of any two models be taken. , , , For any two models, if the absolute value of the difference between the relative errors of the two models is less than or equal to their root mean square accuracy, then... ,but , Increment the weights of both models by 1, and iterate through all models. After pairing the models, the cumulative vote weights of each model are obtained. .

[0045] S203, the weighted average of the relative errors of each model is determined as the integrated error of the target detection node.

[0046] in, Let be the integration error of the i-th target detection node. The weights of model m are obtained after average voting in the even and odd spaces. Let be the relative error between the actual observed power distribution at the i-th target detection node and the predicted power distribution obtained from the m-th spatial neighborhood observation model. If a model is not included in the consistency set in all pairwise comparisons, its weight is 0 and it does not participate in the final integration.

[0047] In one embodiment of this specification, generating and displaying a heatmap based on the integration error of all target detection nodes includes: generating seven layered heatmaps corresponding to seven detector layers based on the integration error of all target detection nodes, such as... Figure 3 As shown, Figure 3 This is an example of a layered heatmap corresponding to the first detector layer provided in this specification. In addition, 28 layered heatmaps for 7 detector layers × 4 neighborhood models were generated.

[0048] In the heatmap, the intensity of color is positively correlated with the magnitude of the error, and color bars on the right are used for quantitative reference. The layered heatmap visually presents the error distribution, and the darker blocks accurately locate the detection nodes of abnormal targets, improving the efficiency of manual verification by more than 60%, effectively solving the technical problem of poor interpretability of anomaly detection in existing nuclear reactor core detectors.

[0049] S107: Determine the normal upper limit based on the statistical value of the integration error and the global threshold coefficient. When the integration error of the target detection node is greater than the normal upper limit, it is judged as abnormal and the abnormal detection result is displayed.

[0050] like Figure 4 As shown, Figure 4 This is a flowchart for the exception determination provided in this manual.

[0051] In one embodiment of this specification, the statistics of the integration error include the mean and standard deviation of the integration error. Determining the normal upper limit based on the statistics of the integration error and the global threshold coefficient includes: determining the normal upper limit by summing the products of the mean of the integration error, the global threshold coefficient, and the standard deviation of the integration error.

[0052] Specifically, the normal upper limit is , The average value of the integrated error of all target detection nodes. The standard deviation of the integrated error for all target detection nodes. For the global threshold coefficient, when At that time, the target detection node was determined to be abnormal.

[0053] The global threshold coefficient is dynamically adjusted based on the number of abnormal sample detection nodes. In one embodiment, the method for determining the global threshold coefficient includes: traversing the threshold coefficient range, detecting abnormal samples of the sample detection nodes for any sample threshold coefficient in the threshold coefficient range, determining the accuracy of the sample threshold coefficient based on the detected abnormal samples, and determining the sample threshold coefficient corresponding to the highest accuracy as the global threshold coefficient.

[0054] Specifically, before deployment, based on real data, a search is performed within the threshold coefficient range [2.8, 8.3]. For any sample threshold coefficient, the four types of observation data of the sample detection node are input into the four constructed spatial neighborhood observation models, and the four predicted power distribution observations of the sample detection node are output. The relative error between the actual power distribution observations and the predicted power distribution observations of the sample detection node is determined. A statistical fusion technique is used to fuse the four relative errors of the sample detection node into an integrated error. The normal upper limit is determined based on the statistical value of the integrated error and the sample threshold coefficient. When the integrated error of the sample detection node is greater than the normal upper limit, the sample detection node is determined to be an abnormal sample.

[0055] After comparing the simulated anomaly sample set with the actual anomaly sample set set in advance, it is determined whether the simulated anomaly detection was successful. Success is considered only when the detected anomaly sample set is equal to the actual anomaly sample set; otherwise, it is considered a failure. After multiple simulation experiments, the average success rate under the sample threshold coefficient is calculated, and the average success rate is determined as the accuracy rate under the sample threshold coefficient.

[0056] When the accuracy is highest It is determined to be the global threshold coefficient.

[0057] During deployment, based on the current noise and the expected number of anomalies, the corresponding values ​​are obtained through table lookup or fitting. When the preset number of anomalies is 1, The value is usually [8.1, 8.3]; when the preset number of exceptions is 10, it is usually [2.8, 3.1].

[0058] The methods provided in this specification can be implemented by a server, which can be a server set up on a business platform, or a device such as a desktop computer or laptop computer capable of executing the solutions in this specification. In one embodiment of this specification, an embedded industrial control computer is used, with a CPU frequency ≥ 2.0 GHz and memory ≥ 8 GB.

[0059] based on Figure 1The method shown is a multi-model anomaly detection method for nuclear reactor core detectors. Considering the complex spatial distribution of nuclear reactor core detectors, it selects at each sample detection node... r , , 2r , r and By constructing four spatial neighborhood observation models based on the distance data of neighboring nodes, and through multi-model parallel prediction and statistical fusion technology, the spatial correlation information of neighboring nodes in different directions and at different distances is comprehensively utilized, which effectively avoids misjudgment of local anomalies or global distortions by a single spatial neighborhood model. At the same time, the upper limit of normal is determined by combining the statistical value of the integrated error with the global threshold coefficient, which improves the anomaly judgment from the traditional fixed threshold judgment to adaptive statistical inference. It can dynamically adjust the judgment benchmark according to the overall power distribution of the reactor core, thereby significantly improving the accuracy of anomaly detection and reducing the false alarm rate and false negative rate.

[0060] When applying the multi-model anomaly detection method for nuclear reactor core detectors provided in this specification, it is not necessary to consider... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this manual does not impose any restrictions on it.

[0061] The above describes a multi-model anomaly detection method for a nuclear reactor core detector, provided by one or more embodiments of this specification. Based on the same idea, this specification also provides a corresponding multi-model anomaly detection system for a nuclear reactor core detector, such as... Figure 5 As shown.

[0062] Figure 5 This specification provides an overall architecture diagram of a multi-model anomaly detection system for a nuclear reactor core detector, including: The data acquisition unit consists of multiple self-powered neutron detector nodes. The detector nodes are divided into multiple layers along the axial direction, with multiple nodes arranged in each layer for real-time acquisition of power distribution data. The data transmission unit, connected to the data acquisition unit, is used to transmit the probe data to the processing unit in real time; The processing unit, connected to the data transmission unit, is used to acquire power distribution observation data of all sample detection nodes within the nuclear reactor core and determine the horizontal distance r between adjacent sample detection nodes. At each sample detection node, multiple adjacent nodes with a distance of r units from that node and nodes with a distance of r units from that node are selected. Multiple neighboring nodes of a unit, multiple neighboring nodes 2r units away from the sample detection node, and nodes r units away from the sample detection node and The power distribution observation data of multiple neighboring nodes of a unit are used as four sets of input data, which are respectively input into the corresponding machine learning regressors to construct four spatial neighborhood observation models. Four types of observation data of the target detection node are collected at multiple consecutive time points. For any target detection node at any time point, the four types of observation data of the target detection node are respectively input into the four constructed spatial neighborhood observation models, outputting four predicted power distribution observation values ​​for that target detection node. For any spatial neighborhood observation model, the relative error of the spatial neighborhood observation model on the target detection node is determined based on the actual observed power distribution and predicted power distribution observation values ​​of the target detection node at all time points. A statistical fusion technique is used to fuse the relative errors of the four spatial neighborhood observation models on the target detection node into an integrated error for the target detection node. A normal upper limit is determined based on the statistical value of the integrated error and the global threshold coefficient. When the integrated error of the target detection node exceeds the normal upper limit, it is judged as an anomaly, and the anomaly detection result is displayed.

[0063] In one embodiment of this specification, the data transmission unit adopts an industrial Ethernet interface, which supports real-time transmission of probe data to the processing unit with a transmission delay of ≤100ms; the processing unit adopts an embedded industrial computer with a CPU frequency of ≥2.0GHz and memory of ≥8GB.

[0064] In one embodiment of this specification, the system further includes a visualization interaction unit: connected to the processing unit, used to display heatmaps, hyperparameter fine-tuning, and annotation of abnormal results. Figure 6 As shown, Figure 6 This is a schematic diagram of the visual interactive unit interface provided in this manual. The visual interactive unit is equipped with a touch screen of 10 inches or larger, allowing operators to view layered heatmaps and manually verify abnormal results; it also provides a hyperparameter fine-tuning interface for adjusting... , The parameters enable the collaboration between automatic detection and manual decision-making, avoiding unnecessary downtime maintenance, reducing ineffective maintenance costs, adapting to predictive maintenance needs, and extending the lifespan of the detector.

[0065] Specific limitations regarding the multi-model anomaly detection system for a nuclear reactor core detector can be found in the limitations of the multi-model anomaly detection method for a nuclear reactor core detector described above, and will not be repeated here. Each module in the aforementioned multi-model anomaly detection system for a nuclear reactor core detector can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0066] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 A multi-model anomaly detection method for nuclear reactor core detectors is provided.

[0067] This instruction manual also provides Figure 7 The schematic diagram of the computer device shown is as follows: Figure 7 At the hardware level, the computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to achieve the above-mentioned functions. Figure 1 A multi-model anomaly detection method for nuclear reactor core detectors is provided.

[0068] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0069] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for detecting anomalies in a nuclear reactor core detector using multiple models, characterized in that, include: Acquire power distribution observation data for all sample detection nodes within the nuclear reactor core and determine the horizontal distance between adjacent sample detection nodes. r ; At each sample detection node, a distance of [distance] from that sample detection node is selected. r Multiple adjacent nodes of the unit, and the distance from the sample detection node is Multiple adjacent nodes of the unit, with a distance of 2 from the sample detection node r Multiple adjacent nodes of the unit, and the distance from the sample detection node is r Units and The power distribution observation data of multiple neighboring nodes of a unit are used as 4 sets of input data, which are then input into the corresponding machine learning regressors to construct four spatial neighborhood observation models. Four types of observation data were collected from the target detection node at multiple consecutive time points; For any target detection node at any given time, the four types of observation data of the target detection node are input into the four pre-constructed spatial neighborhood observation models, and the four predicted power distribution observations of the target detection node are output. For any spatial neighborhood observation model, the relative error of the spatial neighborhood observation model on the target detection node is determined based on the actual observed power distribution and the predicted power distribution of the target detection node at all times. A statistical fusion technique is used to fuse the relative errors of four spatial neighborhood observation models at the target detection node into an integrated error of the target detection node; The normal upper limit is determined based on the statistical value of the integration error and the global threshold coefficient. When the integration error of the target detection node is greater than the normal upper limit, it is judged as abnormal and the abnormal detection result is displayed.

2. The method as described in claim 1, characterized in that, The machine learning regressor is selected from K Nearest neighbor regressor, random forest regressor, gradient boosting decision tree regressor, extreme gradient boosting tree regressor, or fully connected neural network regressor.

3. The method as described in claim 1, characterized in that, The step of determining the relative error of the spatial neighborhood observation model at the target detection node based on the actual observed values ​​and predicted observed values ​​of the power distribution at all times includes: The ratio of the absolute value of the difference between the actual observed power distribution value and the predicted power distribution value at each time moment to the actual observed power distribution value is taken as the relative error at that time moment. The arithmetic mean of the relative errors at all times is then calculated as the relative error of the spatial neighborhood observation model at the target detection node.

4. The method as described in claim 1, characterized in that, A statistical fusion technique is employed to fuse the relative errors of four spatial neighborhood observation models at the target detection node into an integrated error for the target detection node, including: The confidence interval filter technique is used to calculate the mean and standard deviation of the four relative errors of each target detection node, and retain the relative errors within the range of mean plus or minus a preset multiple of standard deviation. The arithmetic mean of the retained relative errors is taken as the integrated error of the target detection node.

5. The method as described in claim 1, characterized in that, A statistical fusion technique is employed to fuse the relative errors of four spatial neighborhood observation models at the target detection node into an integrated error for the target detection node, including: Obtain the standard deviation of the errors for the four spatial neighborhood observation models; The odd-even space averaging technique is used to calculate the square root of the sum of the squares of the standard deviations of any two models in the four spatial neighborhood observation models, and obtain the root mean square accuracy. For any two models, if the absolute value of the difference between the relative errors of the two models is less than or equal to their root mean square accuracy, then a preset weight is assigned to the corresponding model. The weighted average of the relative errors of each model is determined as the integrated error of the target detection node.

6. The method as described in claim 1, characterized in that, The statistical values ​​of the integration error include the mean and standard deviation of the integration error. The determination of the normal upper limit based on the statistical values ​​of the integration error and the global threshold coefficient includes: The sum of the products of the average value of the integration error and the global threshold coefficient with the standard deviation of the integration error is determined as the normal upper limit.

7. The method as described in claim 6, characterized in that, The method for determining the global threshold coefficient includes: Traverse the threshold coefficient range, for any sample threshold coefficient in the threshold coefficient range, detect abnormal samples of the sample detection node, and determine the accuracy of the sample threshold coefficient based on the detected abnormal samples. The sample threshold coefficient corresponding to the highest accuracy is determined as the global threshold coefficient.

8. The method as described in claim 1, characterized in that, The method further includes: A heatmap is generated based on the integrated error of all target detection nodes, and the heatmap is displayed.

9. A multi-model anomaly detection system for a nuclear reactor core detector, characterized in that, The system includes a data acquisition unit, a data transmission unit, and a processing unit; The data acquisition unit consists of multiple self-powered neutron detector nodes. The detector nodes are divided into multiple layers along the axial direction, with multiple nodes arranged in each layer for real-time acquisition of power distribution data. A data transmission unit, connected to the data acquisition unit, is used to transmit the detection data to the processing unit in real time; The processing unit, connected to the data transmission unit, is used to acquire power distribution observation data of all sample detection nodes in the nuclear reactor core and determine the horizontal distance r between adjacent sample detection nodes. At each sample detection node, a distance of [distance] from that sample detection node is selected. r Multiple adjacent nodes of the unit, and the distance from the sample detection node is Multiple adjacent nodes of the unit, with a distance of 2 from the sample detection node r Multiple adjacent nodes of the unit, and the distance from the sample detection node is r Units and The power distribution observation data of multiple neighboring nodes of a unit are used as four sets of input data, which are respectively input into the corresponding machine learning regressors to construct four spatial neighborhood observation models. Four types of observation data of the target detection node are collected at multiple consecutive time points. For any target detection node at any time point, the four types of observation data of the target detection node are respectively input into the four constructed spatial neighborhood observation models, outputting four predicted power distribution observation values ​​for that target detection node. For any spatial neighborhood observation model, the relative error of the spatial neighborhood observation model on the target detection node is determined based on the actual observed power distribution and predicted power distribution observation values ​​of the target detection node at all time points. A statistical fusion technique is used to fuse the relative errors of the four spatial neighborhood observation models on the target detection node into an integrated error for the target detection node. A normal upper limit is determined based on the statistical value of the integrated error and the global threshold coefficient. When the integrated error of the target detection node exceeds the normal upper limit, it is judged as an anomaly, and the anomaly detection result is displayed.

10. The system as described in claim 9, characterized in that, The system also includes a visual interaction unit; A visualization interaction unit, connected to the processing unit, is used to display heatmaps, hyperparameter fine-tuning, and annotation of abnormal results.

Images