Nuclear power device fault diagnosis method based on feature vectors and Bayesian reasoning

Abstract

The invention discloses a nuclear power device fault diagnosis method based on feature vectors and Bayesian reasoning, belongs to the technical field of Bayesian reasoning, and particularly relates to nuclear power device fault diagnosis. The method solves the problem that a traditional Bayesian model only uses parameter values as a fault reasoning basis and ignores a large amount of parameter time-frequency feature information, so that the model is difficult to distinguish feature similar faults and realize accurate positioning of the faults, and the fault diagnosis accuracy and the model robustness are affected. The method comprises the following steps: a model learning step: performing structure learning and parameter learning on a Bayesian model according to fault historical operation data to obtain a Bayesian vector reasoning model; the nuclear power device fault diagnosis method based on the feature vector and Bayesian reasoning is suitable for fault diagnosis of a nuclear power device.

Description

Technical Field

[0001] The invention relates to the technical field of Bayesian reasoning, and in particular to fault diagnosis of nuclear power plants. Background Art

[0002] The system equipment and operating environment of nuclear power plants are relatively complex. If a failure occurs and there is no timely response, it is easy to cause a series of serious consequences, such as core damage, which in turn causes significant property losses. Therefore, real-time monitoring and fault diagnosis of the equipment status of nuclear power plants are particularly important.

[0003] At present, the threshold monitoring method is mainly used to monitor the equipment status of nuclear power plants. However, the threshold monitoring method can only provide prompts that the parameters are too high or too low, but cannot deeply reveal the specific reasons affecting the parameters and the linkage effects of other parameters that may be triggered. When the on-site operators receive these alarms, it is often difficult to effectively solve the problem by local adjustments alone. Therefore, how to grasp the operating status of the nuclear power system and establish a new and effective fault diagnosis method to quickly predict and discover potential faults, reduce the negative impact of accidents, and improve the operating stability of the system is of vital importance to ensure the safe operation of nuclear power plants.

[0004] At present, in order to solve the above problems, those skilled in the art have developed a series of fault diagnosis methods and models, such as diagnostic methods based on expert knowledge, data-driven methods, and methods based on Bayesian reasoning models. However, these methods still have the following problems:

[0005] For diagnostic methods based on expert knowledge: they have high requirements on knowledge quality and are difficult to manually construct complex models;

[0006] For data-driven methods: they have high requirements on the volume and quality of training data and the model has insufficient interpretability;

[0007] For methods based on Bayesian reasoning models: Bayesian reasoning models are essentially a fault diagnosis method that combines graph theory with probability theory. They have interpretability and good fault reasoning capabilities for nonlinear complex systems. However, traditional Bayesian models only use parameter values ​​as the basis for fault reasoning, ignoring a large amount of parameter time-frequency feature information, making it difficult for the model to distinguish faults with similar features and accurately locate faults, thus affecting the accuracy of fault diagnosis and the robustness of the model.

[0008] At the same time, although fault diagnosis technology has been applied in nuclear power plant related systems, there are still some urgent problems to be solved in the research of reactor coolant system fault diagnosis. As the core component of nuclear power plants, the stability and reliability of the reactor system are of great importance to ensure the safe and stable operation of nuclear power plants. At present, the monitoring of reactor coolant systems is still mainly based on condition monitoring, and the on-site abnormal conditions mainly rely on the judgment of staff. The function of the fault diagnosis system is relatively single, and the guiding role for subsequent fault handling is limited. Therefore, the research on the fault diagnosis method of the reactor coolant system has important theoretical value and practical significance. Summary of the invention

[0009] The present invention proposes a nuclear power plant fault diagnosis method based on feature vectors and Bayesian reasoning, which solves the problem that the traditional Bayesian model only uses parameter values ​​as the basis for fault reasoning, ignoring a large amount of parameter time-frequency feature information, making it difficult for the model to distinguish faults with similar features and to accurately locate the faults, thus affecting the accuracy of fault diagnosis and the robustness of the model.

[0010] The nuclear power plant fault diagnosis method based on feature vector and Bayesian reasoning of the present invention comprises the following steps:

[0011] Model learning and real-time monitoring steps: used to obtain the Bayesian model and determine whether the Bayesian model has executed the model learning step:

[0012] If the model learning step has not been executed, execute the model learning step;

[0013] If the model learning step has been performed, the fault reasoning and adaptive update steps are performed;

[0014] Data acquisition step: used to obtain the fault history operation data of the nuclear power plant and the collected real-time operation data;

[0015] Data screening step: for the historical fault operation data or the collected real-time operation data of the nuclear power plant, the data of k sensor nodes of the fault location equipment, associated equipment, subsystems and related systems are selected as needed according to the specific fault, and each sensor node is regarded as a vector node;

[0016] Data preprocessing step: used to preprocess the data of k vector nodes to obtain the data feature vectors of k vector nodes; the data feature vector of the i-th vector node is x i dimensional vector data;

[0017] Model learning step: used to perform structure learning and parameter learning on the Bayesian model according to the fault history operation data to obtain the Bayesian vector reasoning model; the model learning step includes the following steps:

[0018] Historical data preprocessing step: used to preprocess the fault historical operation data by using the data screening step and the data preprocessing step, and obtain the data feature vectors of k vector nodes corresponding to the fault historical operation data as a training set;

[0019] Prior probability setting step: used to set the prior probability of the faulty node;

[0020] Structural learning step: It is used to perform structural learning on the Bayesian model by using the maximum and minimum hill climbing algorithm for the data feature vector to obtain the Bayesian model after structural learning;

[0021] Parameter learning step: Based on the Bayesian model after structure learning and the training set, the maximum likelihood estimation algorithm is used to establish the conditional probability information of the Bayesian model structure, find the parameter value that maximizes the probability of data appearing in the training set, and learn the parameters of the Bayesian model;

[0022] Learning completion step: the Bayesian model that has undergone the structure learning step and the parameter learning step is used as the Bayesian vector inference model;

[0023] Fault reasoning and adaptive updating steps: used to perform fault reasoning diagnosis using a Bayesian vector reasoning model based on the collected real-time operating data to obtain a fault diagnosis result; and also used to adaptively update the Bayesian vector reasoning model.

[0024] Further, a preferred embodiment is provided, wherein the data preprocessing step comprises the following steps:

[0025] Feature selection step: used to perform feature selection on the data of k vector nodes based on the sliding time window algorithm to obtain multi-dimensional discretized feature data of the k vector nodes: the feature selection step includes the following steps:

[0026] Normalization step: used to normalize the data of each vector node to obtain one-dimensional time series data of each vector node;

[0027] Sliding window step: used to select data of n time lengths from the one-dimensional time series data of each vector node using a sliding time window algorithm;

[0028] Feature extraction step: for extracting features from data of n time lengths using p feature extraction algorithms, each feature extraction algorithm extracts a data feature, and combines the data of n time lengths to obtain p+1-dimensional feature data of each vector node;

[0029] Discretization step: for dividing the p+1-dimensional feature data of each vector node into integer values ​​of 1-10 by using the max-min discretization method, and obtaining the p+1-dimensional discretized feature data of each vector node;

[0030] Feature screening step: for performing feature screening on the p+1-dimensional discretized feature data of each vector node by using principal component analysis to obtain the data feature vector of each vector node; the feature screening step includes the following steps:

[0031] Statistic calculation steps: used to calculate the contribution rates of T2 statistics and Q statistics of p+1-dimensional discretized feature data respectively;

[0032] Dimensionality reduction step: It is used to screen the parameters used to characterize the fault characteristics according to the contribution rate of T2 statistics and Q statistics, and obtain the discretized characteristic data of each vector node; wherein, the characteristic dimension of the i-th vector node is reduced from p+1 dimension to

[0033] x i Dimension, get the x of the i-th vector node i dimensional discretized feature data, x i <=p+1;

[0034] The step of obtaining vector data is used to use the vector form data of the discretized feature data of each vector node as a data feature vector, and the data feature vector is used as the input of the Bayesian model or the Bayesian vector reasoning model.

[0035] Furthermore, a preferred implementation manner is provided, in the prior probability setting step, the prior probability of the faulty node is obtained in the following manner:

[0036] Obtain the prior probability of faulty nodes through expert experience knowledge;

[0037] Alternatively, the training set is used to train the Bayesian model to obtain the prior probability of the faulty node.

[0038] Further, a preferred embodiment is provided, wherein the structure learning step comprises the following steps:

[0039] The data feature vectors of k vector nodes corresponding to the fault history operation data are divided into One-dimensional feature data;

[0040] Learning using the max-min hill climbing algorithm The connection relationship between the 1-dimensional feature data is determined, and the feature data belonging to the same vector node are prohibited from being connected to each other;

[0041] Design a node relationship storage matrix to store the association relationship of vector nodes in a two-dimensional matrix, and use an index algorithm to query the association relationship of each feature data under the vector node;

[0042] Complete the structural learning step and obtain the Bayesian model after structural learning.

[0043] Further, a preferred implementation is provided, wherein the fault reasoning and adaptive updating steps include the following steps:

[0044] Real-time data preprocessing step: used to preprocess the collected real-time operation data by using the data screening step and the data preprocessing step, and obtain the data feature vectors of k vector nodes corresponding to the real-time operation data as the data to be diagnosed;

[0045] Fault reasoning diagnosis step: used to input the data to be diagnosed into the Bayesian vector reasoning model for fault reasoning to obtain a fault diagnosis result; the fault diagnosis result includes the fault occurrence result and probability corresponding to the data to be diagnosed at each moment;

[0046] Update judgment step: used to obtain verification data of the fault occurrence result and probability, and compare it with the fault occurrence result and probability obtained in the fault reasoning diagnosis step:

[0047] If the result accuracy of the fault reasoning diagnosis step is greater than a given threshold, then continue to execute the fault reasoning diagnosis step;

[0048] Otherwise, perform the adaptive update step;

[0049] Adaptive update step: used to use the verification data of the fault occurrence results and probabilities as a training set, execute the parameter learning step in the model learning step, and establish new conditional probability information of the Bayesian model structure; also used to fuse the new conditional probability information of the Bayesian model structure and the original conditional probability information of the Bayesian model structure to obtain updated conditional probability information and the Bayesian vector inference model:

[0050] a2=(1-y)*a1+y*b;

[0051] Among them, a1 is the original conditional probability information; b is the new conditional probability information; a2 is the updated conditional probability information; y is the learning rate, and its value range is 0 to 1;

[0052] The updated Bayesian vector reasoning model is used to perform the fault reasoning diagnosis step.

[0053] Furthermore, a preferred implementation is provided, in which, in the normalization step, a z-score regularization method is used to normalize the data of each vector node.

[0054] Further, a preferred embodiment is provided, wherein the p feature extraction algorithms are the following 15 feature extraction algorithms:

[0055] Mean value extraction algorithm, variance extraction algorithm, standard deviation extraction algorithm, skewness extraction algorithm, kurtosis extraction algorithm, maximum value extraction algorithm, minimum value extraction algorithm, peak-to-peak value extraction algorithm, root mean square extraction algorithm, peak factor extraction algorithm, waveform factor extraction algorithm, pulse factor extraction algorithm, margin factor extraction algorithm, kurtosis factor extraction algorithm and energy extraction algorithm.

[0056] The present invention also proposes a nuclear power plant fault diagnosis device based on feature vectors and Bayesian reasoning, the device comprising the following modules:

[0057] Model learning and real-time monitoring module: used to obtain the Bayesian model and determine whether the Bayesian model has executed the model learning module:

[0058] If the model learning module has not been executed, execute the model learning module;

[0059] If the model learning module has been executed, the fault reasoning and adaptive update module will be executed;

[0060] Data acquisition module: used to obtain the fault history operation data of the nuclear power plant and the collected real-time operation data;

[0061] Data screening module: for the historical fault operation data of the nuclear power plant or the collected real-time operation data, select the data of k sensor nodes of the fault location equipment, associated equipment, subsystems and related systems according to the specific fault, and each sensor node is regarded as a vector node;

[0062] Data preprocessing module: used to preprocess the data of k vector nodes to obtain the data feature vectors of k vector nodes; the data feature vector of the i-th vector node is x i dimensional vector data;

[0063] Model learning module: used to perform structure learning and parameter learning on the Bayesian model according to the fault history operation data to obtain the Bayesian vector reasoning model; the model learning module includes the following submodules:

[0064] The historical data preprocessing submodule is used to preprocess the fault historical operation data using the data screening module and the data preprocessing module, and obtain the data feature vectors of k vector nodes corresponding to the fault historical operation data as a training set;

[0065] Prior probability setting submodule: used to set the prior probability of the faulty node;

[0066] Structural learning submodule: used to perform structural learning on the Bayesian model using the maximum and minimum hill climbing algorithm for the data feature vector, and obtain the Bayesian model after structural learning;

[0067] Parameter learning submodule: Based on the Bayesian model after structure learning and the training set, the maximum likelihood estimation algorithm is used to establish the conditional probability information of the Bayesian model structure, find the parameter value that maximizes the probability of data appearing in the training set, and learn the parameters of the Bayesian model;

[0068] Learning completion submodule: the Bayesian model after the structure learning submodule and the parameter learning submodule is used as the Bayesian vector inference model;

[0069] Fault reasoning and adaptive update module: used to perform fault reasoning diagnosis based on the collected real-time operation data using the Bayesian vector reasoning model to obtain fault diagnosis results; also used to adaptively update the Bayesian vector reasoning model.

[0070] The present invention also proposes a computer storage medium, in which a computer program is stored. When the computer program is run, any one of the above-mentioned nuclear power plant fault diagnosis methods based on feature vectors and Bayesian reasoning is executed.

[0071] The present invention also proposes a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of any one of the above-mentioned nuclear power plant fault diagnosis methods based on feature vectors and Bayesian reasoning.

[0072] The present invention has the following beneficial effects:

[0073] 1. The nuclear power plant fault diagnosis method based on feature vector and Bayesian reasoning described in the present invention aims at the deficiencies of existing fault diagnosis technology and combines signal processing, feature screening, intelligent learning, probability decision-making and other methods to form a nuclear power plant fault diagnosis method based on data feature vector and Bayesian reasoning to improve the accuracy and robustness of fault diagnosis results.

[0074] 2. The nuclear power plant fault diagnosis method based on feature vector and Bayesian reasoning of the present invention proposes an improvement on the Bayesian model, replacing the model threshold node with a vector node with multi-dimensional features, thereby providing more data information, so that the model can more comprehensively reflect the relevance of fault features. And a complete fault diagnosis process integrating data collection, feature extraction, model learning, reasoning diagnosis, and model updating is designed to realize the engineering application of nuclear power plant fault diagnosis technology.

[0075] 3. The nuclear power plant fault diagnosis method based on feature vectors and Bayesian reasoning described in the present invention first uses a sliding time window algorithm to perform real-time feature extraction on sensor detection data to obtain more comprehensive data feature information; and uses principal component analysis to perform feature screening on the data to eliminate features with lower contributions to form a data feature vector; then the vector data is imported into an improved structural learning and parameter model to obtain a Bayesian vector reasoning model that best fits the historical fault data; finally, the Bayesian vector reasoning model is used to achieve effective diagnosis of fault results, and an adaptive update method is proposed to automatically update the model.

[0076] 4. The nuclear power plant fault diagnosis method based on feature vector and Bayesian reasoning described in the present invention is higher in accuracy and rationality than other existing nuclear power plant fault diagnosis methods based on Bayesian models. Specifically: by adopting the model learning step and the fault reasoning and adaptive updating step, the accuracy of fault diagnosis detection is improved; by adopting the model learning step, the rationality of fault diagnosis detection is improved.

[0077] 5. The method for diagnosing a nuclear power plant fault based on feature vector and Bayesian reasoning described in the present invention, the model learning step learns the prior probability of the fault cause node by combining expert experience knowledge and historical operation data. This step not only provides a solid foundation for subsequent model learning, but also innovatively introduces vector data, so that the model can capture more fault feature information. Then, a maximum and minimum hill climbing algorithm for vectors is proposed to conduct in-depth structural learning of the Bayesian model. The algorithm cleverly divides the fault vector data into feature nodes, avoiding the interconnection between features under the same node, thereby ensuring the rationality and accuracy of the model structure. Compared with threshold data, vector data provides more dimensional information, so that the model can more comprehensively reflect the relevance of fault features. At the same time, through the maximum and minimum hill climbing algorithm, the model can find the network structure with the largest Bayesian information criterion score, further improving the accuracy and generalization ability of fault diagnosis. Finally, the maximum likelihood estimation algorithm is used to learn the parameters of the model. After inputting the model file (Bayesian model after structure learning) and vector data obtained by structure learning, the algorithm can establish the conditional probability information of the model structure and find the parameter value that maximizes the probability of occurrence of data samples. Compared with threshold data, vector data provides more sample information and feature dimensions for parameter learning, enabling the model to estimate the Bayesian model parameter values ​​more accurately.

[0078] 6. The nuclear power plant fault diagnosis method based on feature vectors and Bayesian reasoning described in the present invention, the model learning step ensures the high accuracy and practicality of the Bayesian model in fault cause diagnosis, and by subdividing the parameter nodes (sensor nodes) into secondary nodes, it avoids the defect that cyclic graphs cannot appear in the Bayesian model, allowing loops to appear at the parameter node level in the structure, thereby greatly increasing the rationality of the Bayesian model.

[0079] 7. The fault diagnosis method for nuclear power plant based on feature vector and Bayesian reasoning described in the present invention, the fault reasoning and adaptive update step imports the data to be diagnosed into the Bayesian network reasoning model to realize real-time monitoring and fault diagnosis. This step is the key to the operation of the system, which ensures that the system can receive and process the monitoring data from the nuclear power plant in real time. Through the Bayesian network reasoning model, the system can calculate the fault result and occurrence probability at each moment based on real-time data, which is crucial for timely detection and handling of faults. When the accuracy of the reasoning result is higher than the given threshold, the system will continue to execute the real-time monitoring process to ensure continuous monitoring of the nuclear power plant. This mechanism ensures the stability and reliability of the system under normal operating conditions. The adaptive update is triggered when the accuracy of the reasoning result is lower than the given threshold. Its main function is to import this set of inaccurate data as a fault training set into the Bayesian network structure learning and parameter learning steps to further optimize and improve the performance of the model. By executing the Bayesian network parameter learning step, the system can learn new conditional probability information based on new fault data and update the original conditional probability information according to the learning rate y. The introduction of this step enables the system to continuously learn from new data and optimize itself, thereby improving the accuracy and adaptability of fault diagnosis. The setting of the learning rate y plays a key role in adaptive learning. It determines the degree of influence of the newly learned conditional probability information on the original conditional probability. When the y value is large, the newly learned conditional probability information has a greater impact on the original conditional probability, which helps the system to quickly adapt to new fault conditions; when the y value is small, the system pays more attention to maintaining the original stability and reliability. By setting the y value reasonably, the system can strike a balance between maintaining stability and adapting to new situations. In summary, the fault reasoning and adaptive update steps constitute the real-time monitoring and adaptive learning mechanism of the nuclear power plant fault diagnosis system, ensuring that the system can diagnose faults in real time and accurately, and continuously optimize and improve performance based on new fault data.

[0080] The nuclear power plant fault diagnosis method based on characteristic vectors and Bayesian reasoning described in the present invention is suitable for nuclear power plant fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0082] Figure 1 A schematic flow chart of a method for diagnosing a nuclear power plant fault based on feature vectors and Bayesian reasoning in one embodiment of the present invention;

[0083] Figure 2 A schematic diagram of a flow chart of a feature selection step in one embodiment of the present invention;

[0084] Figure 3 The figure is a flowchart of the structure learning step in one embodiment of the present invention. DETAILED DESCRIPTION

[0085] In order to make the technical solutions and advantages of the present invention more clearly described, the specific implementation methods of the present invention will be further described in detail and completely in conjunction with the accompanying drawings. The various implementation methods described below are only part of the preferred solutions of the present invention, rather than all implementation methods; the various implementation methods described below are intended to explain the present invention and cannot be understood as limitations on the present invention; the reasonable combination of technical features defined in the various implementation methods of the present invention, and all other implementation methods obtained by ordinary technicians in the field without creative work based on the implementation methods of the present invention, are within the scope of protection of the present invention.

[0086] In one embodiment, a method for diagnosing a nuclear power plant fault based on a feature vector and Bayesian reasoning is provided, the method comprising the following steps:

[0087] Model learning and real-time monitoring steps: used to obtain the Bayesian model and determine whether the Bayesian model has executed the model learning step:

[0088] If the model learning step has not been executed, execute the model learning step;

[0089] If the model learning step has been performed, the fault reasoning and adaptive update steps are performed;

[0090] Data acquisition step: used to obtain the fault history operation data of the nuclear power plant and the collected real-time operation data;

[0091] Data screening step: for the historical fault operation data or the collected real-time operation data of the nuclear power plant, the data of k sensor nodes of the fault location equipment, associated equipment, subsystems and related systems are selected as needed according to the specific fault, and each sensor node is regarded as a vector node;

[0092] Data preprocessing step: used to preprocess the data of k vector nodes to obtain the data feature vectors of k vector nodes; the data feature vector of the i-th vector node is x i dimensional vector data;

[0093] Model learning step: used to perform structure learning and parameter learning on the Bayesian model according to the fault history operation data to obtain the Bayesian vector reasoning model; the model learning step includes the following steps:

[0094] Historical data preprocessing step: used to preprocess the fault historical operation data by using the data screening step and the data preprocessing step, and obtain the data feature vectors of k vector nodes corresponding to the fault historical operation data as a training set;

[0095] Prior probability setting step: used to set the prior probability of the faulty node;

[0096] Structural learning step: It is used to perform structural learning on the Bayesian model by using the maximum and minimum hill climbing algorithm for the data feature vector to obtain the Bayesian model after structural learning;

[0097] Parameter learning step: Based on the Bayesian model after structure learning and the training set, the maximum likelihood estimation algorithm is used to establish the conditional probability information of the Bayesian model structure, find the parameter value that maximizes the probability of data appearing in the training set, and learn the parameters of the Bayesian model;

[0098] Learning completion step: the Bayesian model that has undergone the structure learning step and the parameter learning step is used as the Bayesian vector inference model;

[0099] Fault reasoning and adaptive updating steps: used to perform fault reasoning diagnosis using a Bayesian vector reasoning model based on the collected real-time operating data to obtain a fault diagnosis result; and also used to adaptively update the Bayesian vector reasoning model.

[0100] In this embodiment, the Bayesian model diagnoses the fault based on the fault node and the sensor node. The fault node represents the fault type and is the root node; the sensor node represents the sensor parameter, that is, the data feature vector, and is the leaf node. The data feature vector is x-dimensional vector data from the data perspective and x-dimensional vector node from the mathematical model perspective.

[0101] In this embodiment, real-time operating data is collected at key locations of the nuclear power plant using temperature, pressure, flow and water level sensors.

[0102] It should be noted that the existing fault diagnosis technologies include expert knowledge-based rule knowledge base, fault tree, symbolic directed graph and other methods, as well as data-driven principal component analysis, support vector machine, cluster analysis, artificial neural network and other fault diagnosis methods. However, expert knowledge-based diagnosis methods generally have the problem of high requirements on the quality of expert experience and difficulty in manually constructing complex models, while data-driven methods have the problem of high requirements on the volume and quality of training data and insufficient interpretability of the model.

[0103] In this embodiment, the fault diagnosis method is implemented through Bayesian reasoning (network), which can solve the shortcomings of using a single diagnosis method (i.e., a single diagnosis method based on expert knowledge or data-driven) by combining expert knowledge and data-driven. However, most of the existing Bayesian models only judge the node status based on the data threshold, resulting in unsatisfactory diagnostic accuracy of the model; therefore, in this embodiment, on the basis of the ordinary Bayesian network, the single data threshold information is expanded to multi-feature data vector information (data feature vector), and the model structure learning algorithm is improved at the same time, so that the Bayesian model supports the learning and reasoning calculation of vector situation data. Finally, an adaptive update strategy for the Bayesian network fault reasoning model is proposed to improve the robustness of the model and the fault diagnosis accuracy.

[0104] In addition, in one embodiment, the data preprocessing step includes the following steps:

[0105] Feature selection step: used to perform feature selection on the data of k vector nodes based on the sliding time window algorithm to obtain multi-dimensional discretized feature data of the k vector nodes: the feature selection step includes the following steps:

[0106] Normalization step: used to normalize the data of each vector node to obtain one-dimensional time series data of each vector node;

[0107] Sliding window step: used to select data of n time lengths from the one-dimensional time series data of each vector node using a sliding time window algorithm;

[0108] Feature extraction step: for extracting features from data of n time lengths using p feature extraction algorithms, each feature extraction algorithm extracts a data feature, and combines the data of n time lengths to obtain p+1-dimensional feature data of each vector node;

[0109] Discretization step: for dividing the p+1-dimensional feature data of each vector node into integer values ​​of 1-10 by using the max-min discretization method, and obtaining the p+1-dimensional discretized feature data of each vector node;

[0110] Feature screening step: for performing feature screening on the p+1-dimensional discretized feature data of each vector node by using principal component analysis to obtain the data feature vector of each vector node; the feature screening step includes the following steps:

[0111] Statistic calculation steps: used to calculate the T2 statistic (Hotelling statistic) of p+1-dimensional discretized feature data respectively

[0112] and the contribution rate of the Q statistic (squared prediction error);

[0113] Dimensionality reduction step: It is used to screen the parameters used to characterize the fault characteristics according to the contribution rate of T2 statistics and Q statistics, and obtain the discretized characteristic data of each vector node; wherein, the characteristic dimension of the i-th vector node is reduced from p+1 dimension to

[0114] x i Dimension, get the x of the i-th vector node i dimensional discretized feature data, x i <=p+1;

[0115] The step of obtaining vector data is used to use the vector form data of the discretized feature data of each vector node as a data feature vector, and the data feature vector is used as the input of the Bayesian model or the Bayesian vector reasoning model.

[0116] In this implementation manner, the fault history operation data is obtained based on the collected previous real-time operation data.

[0117] It should be noted that nuclear power plants generally have hundreds or even thousands of sensor parameters, which collect measurement values ​​of various systems and equipment. However, when a certain device fails, most of the parameters related to the failure are only the measurement values ​​of adjacent devices and systems. The measurement values ​​in the system that are not closely related to the device are not very volatile, and there is no need to use them for model training or diagnosis of this fault. The purpose of the data screening step is to filter out unnecessary data parameters and reduce the time and space complexity of data processing during model training or fault diagnosis.

[0118] In this implementation, the sliding time window length of the sliding time window algorithm is n.

[0119] It should be noted that signal feature extraction can also be performed through methods such as short-time Fourier transform and wavelet transform; among them, wavelet transform and wavelet packet transform are relatively mature signal processing algorithms in theory, but there is a problem that the extracted features are not interpretable.

[0120] In addition, in one implementation, in the prior probability setting step, the prior probability of the faulty node is obtained in the following manner:

[0121] Obtain the prior probability of faulty nodes through expert experience knowledge;

[0122] Alternatively, the training set is used to train the Bayesian model to obtain the prior probability of the faulty node.

[0123] In addition, in one embodiment, the structure learning step includes the following steps:

[0124] The data feature vectors of k vector nodes corresponding to the fault history operation data are divided into One-dimensional feature data, where x i is the number of dimensions of the i-th vector node;

[0125] Learning using the max-min hill climbing algorithm The connection relationship between the 1-dimensional feature data is determined, and the feature data belonging to the same vector node are prohibited from being connected to each other;

[0126] Design a node relationship storage matrix to store the association relationship of vector nodes in a two-dimensional matrix, and use an index algorithm to query the association relationship of each feature data under the vector node;

[0127] Complete the structural learning step and obtain the Bayesian model after structural learning.

[0128] In this embodiment, one-dimensional feature data is also referred to as one-dimensional feature nodes.

[0129] In this embodiment, the sensor nodes are further divided into secondary nodes (i.e. 1-dimensional feature data).

[0130] In addition, in one embodiment, the fault reasoning and adaptive updating steps include the following steps:

[0131] Real-time data preprocessing step: used to preprocess the collected real-time operation data by using the data screening step and the data preprocessing step, and obtain the data feature vectors of k vector nodes corresponding to the real-time operation data as the data to be diagnosed;

[0132] Fault reasoning diagnosis step: used to input the data to be diagnosed into the Bayesian vector reasoning model for fault reasoning to obtain a fault diagnosis result; the fault diagnosis result includes the fault occurrence result and probability corresponding to the data to be diagnosed at each moment;

[0133] Update judgment step: used to obtain verification data of the fault occurrence result and probability, and compare it with the fault occurrence result and probability obtained in the fault reasoning diagnosis step:

[0134] If the result accuracy of the fault reasoning diagnosis step is greater than a given threshold, then continue to execute the fault reasoning diagnosis step;

[0135] Otherwise, perform the adaptive update step;

[0136] Adaptive update step: used to use the verification data of the fault occurrence results and probabilities as a training set, execute the parameter learning step in the model learning step, and establish new conditional probability information of the Bayesian model structure; also used to fuse the new conditional probability information of the Bayesian model structure and the original conditional probability information of the Bayesian model structure to obtain updated conditional probability information and the Bayesian vector inference model:

[0137] a2=(1-y)*a1+y*b;

[0138] Among them, a1 is the original conditional probability information; b is the new conditional probability information; a2 is the updated conditional probability information; y is the learning rate, and its value range is 0 to 1;

[0139] The updated Bayesian vector reasoning model is used to perform the fault reasoning diagnosis step.

[0140] In this embodiment, the original conditional probability information refers to the conditional probability information of the Bayesian model structure before the parameter learning step in the model learning step is performed using the verification data of the fault occurrence results and probabilities as a training set in the adaptive update step.

[0141] In this implementation, the default value of the learning rate y is 1 / the number of fault types, and can also be manually set based on expert experience.

[0142] In addition, in one embodiment, in the normalization step, a z-score regularization method is used to normalize the data of each vector node.

[0143] In addition, in one embodiment, the p feature extraction algorithms are the following 15 feature extraction algorithms:

[0144] Mean value extraction algorithm, variance extraction algorithm, standard deviation extraction algorithm, skewness extraction algorithm, kurtosis extraction algorithm, maximum value extraction algorithm, minimum value extraction algorithm, peak-to-peak value extraction algorithm, root mean square extraction algorithm, peak factor extraction algorithm, waveform factor extraction algorithm, pulse factor extraction algorithm, margin factor extraction algorithm, kurtosis factor extraction algorithm and energy extraction algorithm.

[0145] In this embodiment, after the 15 feature extraction algorithms are used to extract features, 15 data features are obtained, and then the original data of n time lengths are added to obtain 16-dimensional feature data of each vector node.

[0146] In one embodiment, a nuclear power plant fault diagnosis device based on feature vectors and Bayesian reasoning is provided, the device comprising the following modules:

[0147] Model learning and real-time monitoring module: used to obtain the Bayesian model and determine whether the Bayesian model has executed the model learning module:

[0148] If the model learning module has not been executed, execute the model learning module;

[0149] If the model learning module has been executed, the fault reasoning and adaptive update module will be executed;

[0150] Data acquisition module: used to obtain the fault history operation data of the nuclear power plant and the collected real-time operation data;

[0151] Data screening module: for the historical fault operation data of the nuclear power plant or the collected real-time operation data, select the data of k sensor nodes of the fault location equipment, associated equipment, subsystems and related systems according to the specific fault, and each sensor node is regarded as a vector node;

[0152] Data preprocessing module: used to preprocess the data of k vector nodes to obtain the data feature vectors of k vector nodes; the data feature vector of the i-th vector node is x i dimensional vector data;

[0153] Model learning module: used to perform structure learning and parameter learning on the Bayesian model according to the fault history operation data to obtain the Bayesian vector reasoning model; the model learning module includes the following submodules:

[0154] The historical data preprocessing submodule is used to preprocess the fault historical operation data using the data screening module and the data preprocessing module, and obtain the data feature vectors of k vector nodes corresponding to the fault historical operation data as a training set;

[0155] Prior probability setting submodule: used to set the prior probability of the faulty node;

[0156] Structural learning submodule: used to perform structural learning on the Bayesian model using the maximum and minimum hill climbing algorithm for the data feature vector, and obtain the Bayesian model after structural learning;

[0157] Parameter learning submodule: Based on the Bayesian model after structure learning and the training set, the maximum likelihood estimation algorithm is used to establish the conditional probability information of the Bayesian model structure, find the parameter value that maximizes the probability of data appearing in the training set, and learn the parameters of the Bayesian model;

[0158] Learning completion submodule: The Bayesian model after the structure learning submodule and the parameter learning submodule is used as the Bayesian vector inference model;

[0159] Fault reasoning and adaptive update module: used to perform fault reasoning diagnosis based on the collected real-time operation data using the Bayesian vector reasoning model to obtain fault diagnosis results; also used to adaptively update the Bayesian vector reasoning model.

[0160] In one embodiment, a computer storage medium is provided, wherein a computer program is stored in the storage medium, and when the computer program is run, any one of the above-mentioned nuclear power plant fault diagnosis methods based on feature vectors and Bayesian reasoning is executed.

[0161] In one embodiment, a computer program product is provided, comprising a computer program / instruction, which, when executed by a processor, implements the steps of any one of the above-mentioned nuclear power plant fault diagnosis methods based on feature vectors and Bayesian reasoning.

[0162] The technical solution provided by the present invention is further described in detail above through several specific implementation modes in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the several specific implementation modes described above are not intended to be used as limitations on the present invention. Any reasonable changes and improvements to the present invention, reasonable combinations of implementation modes and equivalent substitutions based on the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for fault diagnosis of nuclear power plant based on feature vector and Bayesian reasoning, characterized in that: The method comprises the following steps: Model learning and real-time monitoring steps: used to obtain the Bayesian model and determine whether the Bayesian model has executed the model learning step: If the model learning step has not been executed, execute the model learning step; If the model learning step has been performed, the fault reasoning and adaptive update steps are performed; Data acquisition step: used to obtain the fault history operation data of the nuclear power plant and the collected real-time operation data; Data screening step: for the historical fault operation data or the collected real-time operation data of the nuclear power plant, the data of k sensor nodes of the fault location equipment, associated equipment, subsystems and related systems are selected as needed according to the specific fault, and each sensor node is regarded as a vector node; Data preprocessing step: used to preprocess the data of k vector nodes to obtain the data feature vectors of k vector nodes; the data feature vector of the i-th vector node is x i dimensional vector data; Model learning step: used to perform structure learning and parameter learning on the Bayesian model according to the fault history operation data to obtain the Bayesian vector reasoning model; the model learning step includes the following steps: Historical data preprocessing step: used to preprocess the fault historical operation data by using the data screening step and the data preprocessing step, and obtain the data feature vectors of k vector nodes corresponding to the fault historical operation data as a training set; Prior probability setting step: used to set the prior probability of the faulty node; Structural learning step: It is used to perform structural learning on the Bayesian model by using the maximum and minimum hill climbing algorithm for the data feature vector to obtain the Bayesian model after structural learning; Parameter learning step: Based on the Bayesian model after structure learning and the training set, the maximum likelihood estimation algorithm is used to establish the conditional probability information of the Bayesian model structure, find the parameter value that maximizes the probability of data appearing in the training set, and learn the parameters of the Bayesian model; Learning completion step: the Bayesian model that has undergone the structure learning step and the parameter learning step is used as the Bayesian vector inference model; Fault reasoning and adaptive updating steps: used to perform fault reasoning diagnosis using a Bayesian vector reasoning model based on the collected real-time operating data to obtain a fault diagnosis result; and also used to adaptively update the Bayesian vector reasoning model.

2. The method for nuclear power plant fault diagnosis based on feature vector and Bayesian reasoning according to claim 1 is characterized in that: The data preprocessing step The following steps are involved: Feature selection step: used to perform feature selection on the data of k vector nodes based on the sliding time window algorithm to obtain multi-dimensional discretized feature data of the k vector nodes: the feature selection step includes the following steps: Normalization step: used to normalize the data of each vector node to obtain one-dimensional time series data of each vector node; Sliding window step: used to select data of n time lengths from the one-dimensional time series data of each vector node using a sliding time window algorithm; Feature extraction step: for extracting features from data of n time lengths using p feature extraction algorithms, each feature extraction algorithm extracts a data feature, and combines the data of n time lengths to obtain p+1-dimensional feature data of each vector node; Discretization step: for dividing the p+1-dimensional feature data of each vector node into integer values ​​of 1-10 by using the max-min discretization method, and obtaining the p+1-dimensional discretized feature data of each vector node; Feature screening step: for performing feature screening on the p+1-dimensional discretized feature data of each vector node by using principal component analysis to obtain the data feature vector of each vector node; the feature screening step includes the following steps: Statistic calculation steps: used to calculate the contribution rates of T2 statistics and Q statistics of p+1-dimensional discretized feature data respectively; Dimensionality reduction step: It is used to screen the parameters used to characterize the fault characteristics according to the contribution rate of T2 statistics and Q statistics, and obtain the discretized feature data of each vector node; wherein, the feature dimension of the i-th vector node is reduced from p+1 to x i Dimension, get the x of the i-th vector node i dimensional discretized feature data, x i <=p+1; The step of obtaining vector data is used to use the vector form data of the discretized feature data of each vector node as a data feature vector, and the data feature vector is used as the input of the Bayesian model or the Bayesian vector reasoning model.

3. The method for nuclear power plant fault diagnosis based on feature vector and Bayesian reasoning according to claim 1 is characterized in that: In the prior probability setting step, the prior probability of the faulty node is obtained in the following manner: Obtain the prior probability of faulty nodes through expert experience knowledge; Alternatively, the training set is used to train the Bayesian model to obtain the prior probability of the faulty node.

4. The method for nuclear power plant fault diagnosis based on feature vector and Bayesian reasoning according to claim 1, characterized in that: The structure learning step includes the following steps: The data feature vectors of k vector nodes corresponding to the fault history operation data are divided into One-dimensional feature data; Learning using the max-min hill climbing algorithm The connection relationship between the 1-dimensional feature data is determined, and the feature data belonging to the same vector node are prohibited from being connected to each other; Design a node relationship storage matrix to store the association relationship of vector nodes in a two-dimensional matrix, and use an index algorithm to query the association relationship of each feature data under the vector node; Complete the structural learning step and obtain the Bayesian model after structural learning.

5. The method for nuclear power plant fault diagnosis based on feature vector and Bayesian reasoning according to claim 1 is characterized in that: The fault reasoning and adaptive updating steps include the following steps: Real-time data preprocessing step: used to preprocess the collected real-time operation data by using the data screening step and the data preprocessing step, and obtain the data feature vectors of k vector nodes corresponding to the real-time operation data as the data to be diagnosed; Fault reasoning diagnosis step: used to input the data to be diagnosed into the Bayesian vector reasoning model for fault reasoning to obtain a fault diagnosis result; the fault diagnosis result includes the fault occurrence result and probability corresponding to the data to be diagnosed at each moment; Update judgment step: used to obtain verification data of the fault occurrence result and probability, and compare it with the fault occurrence result and probability obtained in the fault reasoning diagnosis step: If the result accuracy of the fault reasoning diagnosis step is greater than a given threshold, then continue to execute the fault reasoning diagnosis step; Otherwise, perform the adaptive update step; Adaptive update step: used to use the verification data of the fault occurrence results and probabilities as a training set, execute the parameter learning step in the model learning step, and establish new conditional probability information of the Bayesian model structure; also used to fuse the new conditional probability information of the Bayesian model structure and the original conditional probability information of the Bayesian model structure to obtain updated conditional probability information and the Bayesian vector inference model: a2=(1-y)*a1+y*b; Among them, a1 is the original conditional probability information; b is the new conditional probability information; a2 is the updated conditional probability information; y is the learning rate, and its value range is 0 to 1; The updated Bayesian vector reasoning model is used to perform the fault reasoning diagnosis step.

6. The method for nuclear power plant fault diagnosis based on feature vector and Bayesian reasoning according to claim 2 is characterized in that: In the normalization step, the z-score regularization method is used to normalize the data of each vector node.

7. The method for nuclear power plant fault diagnosis based on feature vector and Bayesian reasoning according to claim 2 is characterized in that: The p feature extraction algorithms are the following 15 feature extraction algorithms: Mean value extraction algorithm, variance extraction algorithm, standard deviation extraction algorithm, skewness extraction algorithm, kurtosis extraction algorithm, maximum value extraction algorithm, minimum value extraction algorithm, peak-to-peak value extraction algorithm, root mean square extraction algorithm, peak factor extraction algorithm, waveform factor extraction algorithm, pulse factor extraction algorithm, margin factor extraction algorithm, kurtosis factor extraction algorithm and energy extraction algorithm.

8. A nuclear power plant fault diagnosis device based on feature vector and Bayesian reasoning, characterized in that: The device comprises the following modules: Model learning and real-time monitoring module: used to obtain the Bayesian model and determine whether the Bayesian model has executed the model learning module: If the model learning module has not been executed, execute the model learning module; If the model learning module has been executed, the fault reasoning and adaptive update module will be executed; Data acquisition module: used to obtain the fault history operation data of the nuclear power plant and the collected real-time operation data; Data screening module: for the historical fault operation data of the nuclear power plant or the collected real-time operation data, select the data of k sensor nodes of the fault location equipment, associated equipment, subsystems and related systems according to the specific fault, and each sensor node is regarded as a vector node; Data preprocessing module: used to preprocess the data of k vector nodes to obtain data feature vectors of k vector nodes; The data feature vector of the i-th vector node is x i dimensional vector data; Model learning module: used to perform structure learning and parameter learning on the Bayesian model according to the fault history operation data to obtain the Bayesian vector reasoning model; the model learning module includes the following submodules: The historical data preprocessing submodule is used to preprocess the fault historical operation data using the data screening module and the data preprocessing module, and obtain the data feature vectors of k vector nodes corresponding to the fault historical operation data as a training set; Prior probability setting submodule: used to set the prior probability of the faulty node; Structural learning submodule: used to perform structural learning on the Bayesian model using the maximum and minimum hill climbing algorithm for the data feature vector, and obtain the Bayesian model after structural learning; Parameter learning submodule: Based on the Bayesian model after structure learning and the training set, the maximum likelihood estimation algorithm is used to establish the conditional probability information of the Bayesian model structure, find the parameter value that maximizes the probability of data appearing in the training set, and learn the parameters of the Bayesian model; Learning completion submodule: the Bayesian model after the structure learning submodule and the parameter learning submodule is used as the Bayesian vector inference model; Fault reasoning and adaptive update module: used to perform fault reasoning diagnosis based on the collected real-time operation data using the Bayesian vector reasoning model to obtain fault diagnosis results; also used to adaptively update the Bayesian vector reasoning model.

9. A computer storage medium, characterized in that The storage medium stores a computer program, and when the computer program is run, the method for diagnosing nuclear power plant faults based on feature vectors and Bayesian reasoning as described in any one of claims 1 to 7 is executed.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the nuclear power plant fault diagnosis method based on feature vectors and Bayesian reasoning as described in any one of claims 1 to 7 are implemented.

Images