Nuclear power station fault diagnosis method based on multi-gating memory graph convolutional network

Abstract

The invention relates to the technical field of fault diagnosis, in particular to a nuclear power station fault diagnosis method based on a multi-gating memory graph convolutional network, and the method comprises the steps: obtaining and processing fault data; embedding the gated memory unit into the graph convolutional network to construct a gated memory graph convolutional network; then, constructing a corresponding fault transfer graph according to the influence range of each fault, inputting fault data into each gated memory graph convolutional network embedded by the specific fault transfer graph in parallel to perform fault transfer feature extraction, and introducing memory and forgetting control for feature processing of each node in a time dimension through an embedded gating mechanism; extracting time and space features; and connecting and fusing the time and space features extracted from each sub-graph by using an attention mechanism aggregation function to generate final fault vector representation. Aiming at similar fault diagnosis of the same type, the model provided by the invention can effectively distinguish similar faults and obtain a good diagnosis result under the condition that other models cannot well distinguish the similar faults.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a nuclear power plant fault diagnosis method based on a multi-gated memory graph convolutional network. Background Art

[0002] Nuclear power is a clean energy source, but accidents in nuclear power plants are disastrous, and ensuring their safe operation is of vital importance. Fault diagnosis and prediction can improve safety and reliability and gain time for protective deployment. Advances in sensor technology and monitoring systems in nuclear power plants have improved data collection capabilities and provided rich real-time data for fault diagnosis, but the diversity and complexity of data pose challenges to traditional fault diagnosis methods. Deep learning technology, with its advantage of relying on big data to automatically learn potential patterns and features, is suitable for environments with high complexity and high safety requirements in nuclear power plants. Although data-driven intelligent fault diagnosis in nuclear power plants has been successful, it faces two challenges in its application. First, there is a lack of prior knowledge guidance, and it is difficult to capture physical laws and system characteristics when processing complex and scarce data, which limits the application effect and generalization ability. Second, the ability to distinguish between similar fault manifestations is poor, and it is easily interfered by redundant system information, affecting the accuracy of diagnosis.

[0003] The present invention combines the advantages of graph convolutional networks' excellent processing capabilities for graph structures with the clear construction of data screening and fault graph structures, and proposes a deep learning model suitable for similar fault diagnosis. The model can efficiently capture the complex correlations between sensor nodes in devices or systems, and extract key spatiotemporal features through a recursive learning process, thereby achieving accurate prediction and classification of fault occurrence, and is suitable for fault diagnosis in complex systems such as industrial monitoring and nuclear power plants. Summary of the invention

[0004] To this end, the present invention provides a nuclear power plant fault diagnosis method based on a multi-gated memory graph convolutional network. The method solves the problem of nuclear power plant failure by constructing a fault transfer graph of similar faults and embedding it into a gated memory graph convolutional network. The method can effectively extract the feature differences of similar faults. For the diagnosis of similar faults of the same type, when other models cannot distinguish them well, the model proposed by the present invention can effectively distinguish similar faults and obtain good diagnostic results.

[0005] To achieve the above object, the present invention provides a nuclear power plant fault diagnosis method based on a multi-gated memory graph convolutional network, comprising:

[0006] Step S1, acquiring historical state parameters of the nuclear power system and preprocessing the historical state parameters to obtain a fault data set, wherein the historical state parameters include real data of fault conditions and simulated data of fault conditions;

[0007] Step S2, embedding the gated memory unit into the graph convolutional network to construct a gated memory graph convolutional network, and embedding the gated memory mechanism to extract the temporal features and spatial features of the gated memory graph convolutional network;

[0008] Step S3, determining a number of fault transfer graphs according to the impact range and propagation path of each fault type of the nuclear power plant, fault data, sensor data and fault association information, and embedding the fault transfer graphs into the corresponding gated memory graph convolutional network;

[0009] Wherein, the number of the gated memory graph convolutional networks is equal to or less than the number of the fault propagation graphs, and the number of the fault propagation graphs is equal to the number of the fault types;

[0010] Step S4, inputting the fault data in parallel into the gated memory graph convolutional network embedded in each fault transfer graph to extract fault information;

[0011] Step S5, connecting and weighting the temporal features and spatial features of each of the gated memory graph convolutional networks using an aggregation method to determine a fault vector representation;

[0012] Step S6, training the model and saving the model parameters to determine the nuclear power plant fault diagnosis system, pre-processing the data to be diagnosed and inputting it into the nuclear power plant fault diagnosis system to determine the diagnosis result.

[0013] Further, in step S1, the preprocessing method includes:

[0014] Step S11, normalizing the historical state parameters to eliminate dimensional differences;

[0015] Step S12, using a sliding window method to group the normalized historical state parameters to determine an operating condition data sequence;

[0016] Step S13, annotating each group of data with operating condition information and storing the resultant fault data set.

[0017] Further, in step S2, the method of embedding the gated memory unit into the graph convolutional network to construct the gated memory graph convolutional network includes:

[0018] Step S21, traversing each time step of the model input sequence to extract the features of each node at the corresponding time step;

[0019] Step S22, using a gated memory mechanism to update the hidden state and cell state of the node in each time step, where:

[0020] Step S221, updating the cell state according to the forget gate and the input gate;

[0021] Step S222, determining the hidden state according to the cell state and the output gate;

[0022] Step S23, obtaining the hidden state of each node at the last time step and determining the hidden state as the output of the model;

[0023] In step S24, a global average pooling process is performed on the hidden state of each node of the gated memory graph convolutional network to determine the fault propagation characteristics of the gated memory graph convolutional network.

[0024] Further, in step S22, the gated memory mechanism includes:

[0025] The forget gate is used to determine the amount of information to be discarded in the cell state;

[0026] Input gate, which controls the influx of new information;

[0027] Output gate, used to determine the hidden state output at the current moment;

[0028] The forget gate, the input gate and the output gate are all determined according to the input features of the current time step and the hidden state of the previous time step.

[0029] Further, in step S3, the method for determining the fault transfer graph includes:

[0030] Step S31, simulating the relationship between nodes in the nuclear power plant system by graph theory, wherein:

[0031] Components represent nodes, and the edge formed by two components represents the influence relationship between the corresponding two nodes;

[0032] Step S32, determining the propagation mode of the fault information in the nuclear power plant system according to the historical fault data by a causal inference method, wherein the propagation mode includes an impact range and a propagation path;

[0033] Step S33, determining the adjacency matrix and its values ​​according to the sensor node numbers corresponding to each component in the fault transfer graph, wherein:

[0034] If there is an edge between node i and node j, then A ij The value of is 1;

[0035] If there is no edge between node i and node j, then A ij The value of is 0;

[0036] Step S34, embed each adjacency matrix into a gated memory graph convolutional network to construct a gated memory graph convolutional network in which the fault propagation graph is embedded.

[0037] Further, in step S4, the method of inputting the fault data in parallel into the gated memory graph convolutional network embedded in each fault transfer graph to extract the fault information includes:

[0038] Step S41, inputting the fault data set in parallel into each gated memory graph convolutional network structure embedded by the fault transfer graph;

[0039] Step S42: determine the fault transfer feature of each fault transfer graph by feature extraction and output the fault transfer feature r mean ;

[0040] Step S43, determining the corresponding fault type of each gated memory graph convolutional network, and determining the difference between similar fault information according to the fault graph structure corresponding to the fault type.

[0041] Furthermore, in step S5, the aggregation method is a weighted aggregation method function based on an attention mechanism, including:

[0042] Step S51, stacking the fault transfer features of each fault transfer graph to obtain a stacked feature, wherein the stacked feature S T Determined according to the following formula:

[0043] Where M is the number of fault transfer graphs, r k mean is the fault transfer feature of the kth fault transfer graph;

[0044] Step S52, mapping the stacked features to the query space, key space and value space according to the weight matrix to determine the query matrix, key matrix and value matrix, wherein the query matrix Q, key matrix K and value matrix V are determined according to the following formulas, respectively:

[0045] Q=W Q S T ,K=W K S T ,V=W V S T ,

[0046] Where W Q , W K and W V They are different weight matrices respectively;

[0047] Step S53, determining an attention score according to the query matrix and the key matrix, and normalizing the attention score according to a softmax function to obtain an attention weight;

[0048] Step S54, using the attention weights to perform weighted summation on the value matrix to determine the output attention, and the output attention is determined according to the following formula:

[0049] Where B is the dimension of the bond matrix K;

[0050] Step S55, the output attention is passed through a linear transformation layer to determine the corresponding fault probability distribution, and the fault probability distribution y is determined according to the following formula:

[0051] y=Linear(T).

[0052] Further, the step S6 includes:

[0053] Step S61, constructing and saving the fault transfer graph input model;

[0054] Step S62, training the fault transfer graph according to the back propagation algorithm and then saving the model parameters to obtain a nuclear power plant fault diagnosis system;

[0055] Step S63, pre-processing the data to be diagnosed and inputting it into the nuclear power plant fault diagnosis system to perform fault diagnosis and determine the diagnosis result.

[0056] Furthermore, in the step S61, the method for constructing the fault transfer graph transfer model includes converting each fault transfer graph into a corresponding adjacency matrix and then inputting each adjacency matrix into a plurality of gated memory graph convolutional networks of the model.

[0057] Furthermore, in step S51, the fault transfer features of each fault transfer graph are stacked according to a pre-stored stacking rule to obtain a stacking feature S T ;

[0058] The stacking rule is stacking along rows.

[0059] Compared with the prior art, the beneficial effect of the present invention is that, by combining the construction of a multi-fault transfer graph with the application of a graph convolutional neural network, the present invention reduces the influence of redundant data and structural information on the information extraction of the graph neural network. Each graph neural network embedded in a specific fault graph focuses on extracting the difference features of corresponding similar faults, thereby improving the fault diagnosis accuracy of the model for more detailed features of similar faults.

[0060] Furthermore, the present invention constructs a gated memory graph convolutional network unit, which enhances the operation of the time dimension of the input variable and improves the effectiveness of fault diagnosis;

[0061] Furthermore, the present invention adopts a self-attention mechanism to perform weighted summation on the information extracted by each gated memory graph convolutional network, which simplifies the redundant part of the information compared to simply aggregating the output using a connection method;

[0062] Furthermore, the method of embedding the fault transfer graph into the graph neural network in the present invention combines prior knowledge with deep learning. Compared with edge construction methods such as aggregation and correlation analysis, it is more explainable and reasonable, and also improves the ability to distinguish similar faults.

[0063] Furthermore, the multi-gated memory graph convolutional network ultimately constructed by the present invention can efficiently diagnose nuclear power plant faults, especially for similar faults of the same type that are difficult to distinguish with ordinary models, and has good diagnostic results. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A step diagram of a nuclear power plant fault diagnosis method based on a multi-gated memory graph convolutional network according to an embodiment of the present invention;

[0065] Figure 2 This is a fault transfer diagram obtained by reasoning when a LOCA accident occurs in a loop heat pipe section in the first embodiment of the present invention;

[0066] Figure 3 An architecture diagram of a multi-gated memory graph convolution model provided by an embodiment of the present invention;

[0067] Figure 4 This is a model training flowchart of a nuclear power plant fault diagnosis method based on a multi-gated memory graph convolutional network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0069] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0070] It should be noted that, in the description of the present invention, terms such as "up", "down", "left", "right", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the drawings. This is merely for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present invention.

[0071] In addition, it should be noted that in the description of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0072] See also Figure 1 The present invention provides a method for diagnosing nuclear power plant faults based on a multi-gated memory graph convolutional network, comprising:

[0073] Step S1, acquiring historical state parameters of a nuclear power system and preprocessing the historical state parameters to obtain a fault data set;

[0074] Wherein, the historical state parameters include real data of fault conditions and simulated data of fault conditions;

[0075] Step S2, embedding a gated memory unit into a graph convolutional network to construct a gated memory graph convolutional network, and embedding a gated memory mechanism to introduce a memory and forget control mechanism into the feature processing of each node of the gated memory graph convolutional network in the time dimension, so as to extract the time features and spatial features of the gated memory graph convolutional network;

[0076] Step S3, determining a number of fault transfer graphs (determining a spatial relationship graph based on each fault transfer graph) according to the impact range and propagation path of each fault type of the nuclear power plant, fault data, sensor data and fault association information, and embedding the fault transfer graph into a corresponding gated memory graph convolutional network;

[0077] Wherein, the number of the gated memory graph convolutional networks is equal to or less than the number of the fault propagation graphs, and the number of the fault propagation graphs is equal to the number of the fault types;

[0078] Step S4, the fault data in the fault data set are input in parallel to the gated memory graph convolutional network embedded in each fault transfer graph to extract fault information; it can be understood that in this way, the features of different fault graphs can be fully utilized, thereby improving the accuracy of fault transfer feature extraction, reducing the influence of irrelevant structures and sensor information on the distinction of similar faults, and improving the ability to distinguish similar faults;

[0079] Step S5, using an aggregation method to connect and weight the temporal features and spatial features of each of the gated memory graph convolutional networks to determine a fault vector representation; it can be understood that the temporal features and spatial features are represented by hidden features h in the present invention. t express;

[0080] Step S6, training the model and saving the model parameters to determine the nuclear power plant fault diagnosis system, pre-processing the data to be diagnosed and inputting it into the nuclear power plant fault diagnosis system to determine the diagnosis result.

[0081] Specifically, in step S1, the preprocessing method includes:

[0082] Step S11, normalizing the historical state parameters to eliminate dimensional differences;

[0083] Step S12, using a sliding window method to group the normalized historical state parameters to determine a working condition data sequence;

[0084] Step S13, annotating each group of data (ie each group of data after data grouping) with operating condition information and storing it to obtain a fault data set.

[0085] Specifically, in step S2, the method of embedding the gated memory unit into the graph convolutional network to construct the gated memory graph convolutional network includes:

[0086] Step S21, traverse each time step of the model input sequence to extract the spatial features g of each node at the corresponding time step t ;

[0087] In the implementation, (1) define the input variable X; where X∈R N×T , T represents the time step, N represents the number of nodes, and the input vector of the tth time step is recorded as x t , that is, x t ∈R N×1 , R is a set of real numbers; (2) for each time step input data x t Use graph convolution operation to determine the feature g of each node t , where g t is the node feature at the tth time step, Among them, A represents the adjacency matrix (used to represent the connection relationship between nodes) of the graph structure of the graph convolutional network (it can be understood that the graph convolutional network needs to be embedded in the graph structure, and the adjacency matrix A is used to pass into the graph convolutional network during the implementation process), For the adjacency matrix with self-loops added ( I is the identity matrix), yes The degree matrix Where i is the i-th node, j is the j-th node, N is the number of nodes), σ is the activation function, expressed as σ(x)=max(0,x), W g is a learnable weight parameter, the specific dimension is related to the input and output dimensions of the model, expressed as Among them, F inand F out The input feature dimension and output feature dimension of the node are determined when the model is initialized; (3) The output of the graph convolutional network forms a new feature representation with the shape of g t ∈R N×H , H is the output feature dimension of each node;

[0088] Step S22, using a gated memory mechanism to update the hidden state and cell state of the graph convolutional network nodes in each time step, where:

[0089] Step S221, updating the cell state according to the forget gate and the input gate;

[0090] It is understandable that the cell state C t The calculation is performed according to the following formula:

[0091] C t =f t ×C t-1 +i t ×tanh(W C [h t-1 ,g t ]+b C ), where f t is the forget gate, C t-1 is the cell state at the t-1th time step, i t is the input gate, tanh is the hyperbolic tangent function, W C is the weight matrix of the memory cell, b C is the bias term of the memory cell, h t-1 is the hidden state at the t-1th time step, g t is the spatial feature of the tth time step;

[0092] Step S222, determining the hidden state according to the cell state and the output gate;

[0093] It is understandable that the hidden state h t The calculation is performed according to the following formula:

[0094] h t =O t ×tanh(C t ), where O t is the output gate, C t is the cell state at the tth time step;

[0095] Understandably, this mechanism allows the model to maintain long-term dependency information at each time step and control the flow of information;

[0096] Step S23, obtaining the hidden state of each node at the last time step and determining the hidden state as the output of the model;

[0097] Step S24, performing global average pooling processing on the hidden state of each node of the gated memory graph convolutional network to determine the fault propagation characteristics of the gated memory graph convolutional network;

[0098] In implementation, the fault propagation feature r mean The calculation is performed according to the following formula: Y is the number of nodes.

[0099] Specifically, in step S22, the gated memory mechanism includes:

[0100] The forget gate is used to determine the amount of information to be discarded in the cell state;

[0101] Input gate, which controls the influx of new information;

[0102] Output gate, used to determine the hidden state output at the current moment;

[0103] The forget gate, the input gate and the output gate are all determined according to the input features of the current time step and the hidden state of the previous time step.

[0104] It is understandable that (1) the forget gate f t Determines how much information is discarded from the cell state, based on the input feature g at the current time step t and the hidden state h at the previous time step t-1 Determine that the calculation formula of the forget gate is: t =σ(W f [h t-1 ,g t ]+b f ), where [h t-1 ,g t ] is the hidden state h of the previous moment t-1 and the input feature g at the current moment t After concatenation, W f is the weight matrix of the forget gate, b f is the bias term of the forget gate; (2) input gate i t Controls the inflow of new information, which is determined by the input feature g of the current time step t and the hidden state h at the previous time step t-1 OK, the calculation formula for the input gate is: t =σ(W i [h t-1 ,g t ]+b i ), where W iis the weight matrix of the input gate, b i is the bias term of the input gate; (3) the output gate O t Determines the hidden state at the current moment, which is determined by the input feature g of the current time step t and the hidden state h at the previous time step t-1 OK, the calculation formula for the output gate is O t =σ(W o [h t-1 ,g t ]+b o ), where W o is the weight matrix of the output gate, b o is the bias term of the output gate.

[0105] In implementation, the weight matrix W and the bias term b are learnable weight parameters. The dimension of the weight matrix usually matches the feature dimension of the input and output, while the dimension of the bias term is usually equal to the number of neurons in the output layer. Before training begins, they are generally assigned values ​​through random initialization. Later, during the training process, these values ​​will be continuously adjusted through the back-propagation algorithm to adjust the weight matrix and bias, thereby changing the values ​​of the forget gate, input gate, output gate and memory cells.

[0106] It can be understood that the weight matrix and bias items are both trainable model parameters. By adjusting the weight matrix and bias items, the neural network can continuously optimize these trainable parameters so that the prediction results gradually approach the true values. During the training process, the values ​​of the weight matrix and bias items are continuously adjusted through the back propagation algorithm to change the output values ​​of the forget gate, input gate, output gate and memory cells, thereby achieving the desired effect.

[0107] Specifically, in step S3, the method for determining the fault transfer graph includes:

[0108] Step S31, using graph theory to simulate the relationship between nodes in the nuclear power plant system, including the edges formed between components,

[0109] Among them, components represent nodes, and the edge formed by two components represents the influence relationship between the corresponding two nodes;

[0110] Step S32, determining the propagation mode of the fault information in the nuclear power plant system according to the historical fault data in the fault data set by a causal inference method, wherein the propagation mode includes an impact range and a propagation path; it can be understood that, for the faults in the target fault set, by analyzing the impact relationship between the faults and the sensors, the propagation mode of the fault information in the nuclear power plant system is determined by using a causal inference method, and all faults in the fault data set are traversed according to this method to construct a fault transfer graph until the traversal is complete;

[0111] In implementation, the method pre-stores data of known fault types that have occurred in the past, and infers a fault transmission diagram based on the fault type and data fluctuation performance.

[0112] Step S33, determining the adjacency matrix and its values ​​according to the sensor node numbers corresponding to each component in the fault transfer graph, wherein:

[0113] If there is an edge between node i and node j, then A ij The value of is 1, that is, A ij =1;

[0114] If there is no edge between node i and node j, then A ij The value of is 0, that is, A ij =0;

[0115] Step S34, embed each adjacency matrix into a gated memory graph convolutional network to construct a gated memory graph convolutional network in which the fault propagation graph is embedded.

[0116] Specifically, in step S4, the method of inputting the fault data in parallel into the gated memory graph convolutional network embedded in each fault transfer graph to extract the fault information includes:

[0117] Step S41, inputting the fault data set in parallel into each gated memory graph convolutional network structure embedded by the fault transfer graph;

[0118] Step S42: determine the fault transfer feature of each fault transfer graph by feature extraction and output the fault transfer feature r mean ;

[0119] Step S43, determining the corresponding fault type of each gated memory graph convolutional network, and determining the difference between similar fault information according to the fault graph structure corresponding to the fault type.

[0120] Specifically, in step S5, the aggregation method is a weighted aggregation method function based on the attention mechanism, including:

[0121] Step S51: The fault transmission characteristics r of each fault transmission graph are mean The stacking feature S is obtained by stacking. T Determined according to the following formula:

[0122] Where M is the number of fault transfer graphs, r k mean is the fault transfer feature of the kth fault transfer graph;

[0123] It can be understood that the stacking feature S TIt represents the intermediate term in the model operation process, which is the stack of features extracted by the graph convolutional network of each fault transfer graph;

[0124] In implementation, STACK{} is used to stack multiple data sets together according to certain rules, similar to splicing multiple matrices or vectors in a certain dimension to form a new, larger data structure; step S52, stacking feature S according to the weight matrix T Map to the query space, key space and value space to determine the query matrix, key matrix and value matrix, wherein the query matrix Q, key matrix K and value matrix V are determined according to the following formulas, respectively:

[0125] Q=W Q S T ,K=W K S T ,V=W V S T ,

[0126] Where W Q , W K and W V They are different weight matrices respectively;

[0127] Step S53, determining an attention score according to the query matrix and the key matrix, and normalizing the attention score according to a softmax function to obtain an attention weight;

[0128] In implementation, the similarity between the query matrix Q and the key matrix K is calculated (the dot product is used to measure the similarity in the present invention), and a normalization factor (one-half the dimension of the key vector) is determined according to the dimension of the key vector to avoid the dot product value being too large;

[0129] In implementation, the attention score is The attention weight is Among them, K T is the transposed matrix of matrix K;

[0130] Step S54, using the attention weights to perform weighted summation on the value matrix to determine the output attention corresponding to the query matrix, wherein the output attention is determined according to the following formula:

[0131] Where B is the dimension of the bond matrix K;

[0132] Step S55, the output attention is passed through a linear transformation layer to determine the corresponding fault probability distribution, and the fault probability distribution y is determined according to the following formula:

[0133] y=Linear(T); where, Linear(T)=WT+b, W∈Rout×in , b∈R out , W is the learnable weight matrix, b is the bias term, out is the output dimension (i.e., the output fault category), and in is the row dimension of T. After initialization, W and b will be updated during the model training process.

[0134] Specifically, the step S6 includes:

[0135] Step S61, construct and save the fault transfer graph input model; it can be understood that the constructed fault transfer graph input model is a multi-gated memory graph convolutional network, which is composed of multiple gated memory graph convolutional networks, and each gated memory graph convolutional network receives a fault transfer graph; therefore, the number of fault transfer graphs is equal to the number of gated memory graph convolutional networks;

[0136] In the implementation, the graph structure positions in multiple gated graph convolutional networks in the model are empty and have no graph structure information. The fault propagation graph needs to be input for initialization to form a complete gated memory graph convolutional network with a graph structure. The number of gated memory graph convolutional networks in the model is set to be determined by the number of fault propagation graphs input;

[0137] Step S62, according to the back propagation algorithm, the training data and the fault label are input into the model, the model parameters are saved after the training fault transfer diagram is input into the model to obtain the nuclear power plant fault diagnosis system; it can be understood that the training is performed according to the training data and the fault label input into the model, the model calculates the predicted output through forward propagation, and according to the error between the predicted output and the fault label, the model parameters are adjusted by the back propagation algorithm, so as to gradually minimize the loss function;

[0138] Step S63, pre-processing the data to be diagnosed and inputting it into the nuclear power plant fault diagnosis system to perform fault diagnosis and determine the diagnosis result.

[0139] Specifically, in step S61, the method for constructing the fault transfer graph transfer model includes converting each fault transfer graph into a corresponding adjacency matrix and then inputting each adjacency matrix into several gated memory graph convolutional networks of the model.

[0140] It can be understood that each fault propagation graph is constructed in the form of an adjacency matrix (i.e., each adjacency matrix represents the topological structure of a fault propagation graph) to describe the connection relationship between nodes, and these adjacency matrices are input into the model and received one by one by multiple gated memory graph convolutional networks in the model.

[0141] Specifically, in step S51, the fault transfer features of each fault transfer graph are stacked according to a pre-stored stacking rule to obtain a stacking feature S T ;

[0142] The stacking rule is to stack along the rows, that is, the output fault transfer feature r of each fault transfer graph is mean As a row, they are stacked together in order, and each row after stacking represents the output r of a fault transfer graph k mean .

[0143] It can be understood that since the output r of the kth fault transfer graph k mean It is one-dimensional data (figuratively speaking, it is a row of data), so the stacking method for the output of all fault transfer graphs is stacking along the rows. After stacking, each row represents the output r of a fault transfer graph. k mean .

[0144] Embodiment 1: Since the probability of nuclear power plant failure is extremely low in real scenarios, it is difficult to obtain fault data in the case of nuclear power plant failure, and it is difficult to have enough actual data to support model training. In this embodiment of the present invention, a simulation data set is used to train and verify the model;

[0145] Step A101, acquiring fault data and dividing it into a training set and a test set in the form of a moving window;

[0146] In this embodiment, 13 nuclear power plant fault data sets were obtained, including 64 feature variables (such as flow, temperature, pressure, etc.). Due to the different dimensions and numerical ranges of the feature variables, the scales of the features are inconsistent, which leads to some features dominating the model training and affecting the performance and stability of the model. Therefore, it is necessary to use normalization processing to eliminate the scale differences between different features.

[0147] The normalization method is: (i in this formula represents the i-th characteristic variable, i = 1, 2, ..., n), x i is the value of the i-th characteristic variable after normalization, and in this embodiment, n=64;

[0148] For the fault diagnosis model used subsequently, since the time series operation method is adopted, the data needs to be segmented and divided using a sliding window method; in implementation, the original data is divided into a time window of 5 seconds each, and multiple time series samples are generated from it, which can enable the model to better capture the dynamic characteristics of data changing over time; in implementation, the training set and the test set are usually divided in a ratio of 7:3.

[0149] Step A102, constructing a gated memory graph convolutional network, and introducing memory and forgetting control for feature processing of each node in the time dimension by embedding a gated memory mechanism;

[0150] The gated memory graph convolutional network is an innovative deep learning network that combines time and space operation processing. It is mainly composed of single-step graph convolution node information aggregation and gated memory and forgetting units.

[0151] Among them, the operation of graph convolution node information aggregation is:

[0152] Among them, the operations of the gated memory and forgetting unit include:

[0153] (1)f t =σ(W f [h t-1 ,g t ]+b f ),

[0154] (2)i t =σ(W i [h t-1 ,g t ]+b i ),

[0155] (3)O t =σ(W o [h t-1 ,g t ]+b o ),

[0156] (4)C t =f t ×C t-1 +i t ×tanh(W C [h t-1 ,g t ]+b C ),

[0157] (5)h t =O t ×tanh(C t );

[0158] After processing all time steps, the node hidden states of each graph convolutional network are averaged and pooled using a global average pooling operation to form the fault transfer feature r of each graph. mean .

[0159] Step A103, combining the fault data, analyzing the impact range and propagation path of each fault type in the nuclear power plant to construct a fault transmission graph and embed it into the gated memory graph convolutional network;

[0160] First, the relationship between the components in the system is modeled through graph theory methods, including establishing nodes (representing each component) and edges (representing the influence relationship between components);

[0161] Then, causal inference methods are applied to determine how fault transfer characteristics propagate through the system under fault conditions;

[0162] like Figure 2 As shown, it is a fault transfer diagram obtained by reasoning when a LOCA accident occurs in a loop heat pipe section. In one implementation, when a LOCA accident occurs in a loop heat pipe section, the inference process based on sensor data includes: (1) the rapid ejection of the coolant in the loop causes a decrease in the amount of water in the nuclear power plant system and a sharp drop in the water level of the pressurizer. As the coolant is lost, the internal pressure of the pressurizer drops rapidly. In order to compensate for the pressure drop, the electric heater increases power to maintain the system pressure stable; (2) the ejection of the coolant causes the temperature in the containment to rise (especially in the area where the coolant contacts the hot surface), and as the coolant is lost and the steam is ejected, the internal pressure of the containment rises, increasing the stress of the system, which may lead to the release of radioactive materials and thus increase the internal radioactivity level; (3) the rapid drop in the coolant pressure in the loop will also affect the coolant pressure in the second and third loops, although to a lesser extent, it will still show a certain degree of drop. In response to the loss of coolant, the coolant injection flow rate is increased by the supplementary pump to maintain the system pressure and water level. The newly injected coolant reduces the temperature of the regeneration downflow outlet. Due to the decrease in the coolant pressure in the system, the downflow pipeline pressure also decreases, resulting in a decrease in the downflow flow rate.

[0163] This embodiment constructs four fault transmission diagrams based on 13 types of faults, namely "RCV fault transmission diagram", "LOCA breach fault transmission diagram", "steam generator fault transmission diagram" and "main pump fault transmission diagram"; and selects 39 sensor nodes from 64 sensor nodes for diagram construction, which greatly simplifies...

[0164] The adjacency matrix is ​​constructed for the fault transfer graph according to the sensor number. The construction method includes: if there is an edge between node i and node j, then A ij The value of is 1, that is, A ij =1; otherwise A ij The value of is 0, that is, A ij =0; embed the constructed adjacency matrix into the gated memory graph convolutional network to construct the gated memory graph convolutional network structure of the fault transfer graph.

[0165] See also Figure 3 As shown, it is an architecture diagram of the multi-gated memory graph convolution model provided by an embodiment of the present invention.

[0166] Step A104, the fault data is input in parallel into each gated memory graph convolutional network embedded by a specific fault propagation graph, and the features extracted by each fault propagation graph gated memory graph convolutional network are connected and weighted using the attention mechanism aggregation function (such as Figure 3 shown);

[0167] Based on the gated memory graph convolutional network embedded in the fault propagation graph constructed in S103, the fault data is input in parallel into the gated memory graph convolutional network embedded in the fault propagation graph. Each network structure extracts the temporal and spatial feature outputs for a similar fault of the same type, where the spatial output r k mean is the output fault propagation feature r of the fault propagation graph embedded by the mth gated memory graph convolutional network mean , A k is the adjacency matrix of the fault transfer graph embedded by the kth gated memory graph convolutional network;

[0168] Input variable X, filter the required nodes according to the adjacency matrix Ak, discard the isolated nodes not included in the adjacency matrix, and ensure that the sensor information aggregation is not affected by useless sensors; use the attention mechanism to perform weighted summation on the extracted temporal and spatial features to obtain the final output attention T of the fault vector; the output r of all M subgraphs k mean The attention mechanism is used for weighted summation, and the fault probability distribution y is obtained through linear transformation;

[0169] in:

[0170]

[0171] Q=W Q S T ,K=W K S T ,V=W V S T ,

[0172]

[0173] y=Linear(T).

[0174] See also Figure 4 As shown, it is a model training flow chart of a nuclear power plant fault diagnosis method based on a multi-gated memory graph convolutional network according to an embodiment of the present invention.

[0175] Step A105, training the model and saving the model parameters to obtain the final nuclear power plant fault diagnosis system, pre-processing the data to be diagnosed and inputting it into the fault diagnosis system to obtain the fault diagnosis result;

[0176] In this implementation, the input parameter is set to 39 on the basis of building a training model and a fault transfer graph. The specific network input parameters are determined according to the number of sensors involved in the fault transfer graph; the number of neurons is 64 (the number of feature variables), the output parameter is 13 (the number of fault data set types), the batch size is set to 32, the optimizer selects the Adam algorithm, the learning rate is set to 0.001, the number of training iterations is set to 300, the loss function selects the cross entropy loss function, and the weights of each layer of the model are randomly initialized;

[0177] In implementation, the training data is put into the model for training. Before the model reaches the number of iterations or the expected loss value, the adaptive moment estimation algorithm is used to iteratively optimize the weights of each target category in the initial model along the negative gradient direction of the loss function until the model converges. The diagnostic model of the model fault with the highest accuracy based on the training set during the training process is saved. The test set data is input to calculate the test set fault diagnosis accuracy, and analysis is performed based on the diagnosis results.

[0178] like Figure 4 As shown in the figure, the number of input nodes is set to 39, the node feature dimension is 5, so the number of input neurons is 5, the number of hidden layer neurons is 64, the number of output neurons is 13, the batch size is set to 32, the optimizer selects the Adam algorithm (adaptive moment estimation algorithm), the number of training iterations is set to 300, the loss function selects the cross entropy loss function, and the learning rate is set to 0.001;

[0179] Randomly initialize the model weights and input the normalized training data into the model for training;

[0180] After the model outputs the predicted probability categories, the category with the highest probability is selected as the fault diagnosis result;

[0181] Use the cross entropy loss function to determine the difference between the predicted category and the true category;

[0182] Before the model reaches the number of iterations (300 times), the adaptive moment estimation algorithm is used to iteratively optimize the weights of each target category in the initial model along the negative gradient direction of the loss function until the model converges.

[0183] The fault diagnosis system of the present invention can have a good distinguishing effect on similar faults, such as main pump failure shutdown and loss of flow accident-main pump shaft stuck failure. Although the causes of these two faults are different, the result is that the main pump cannot work and the system loses flow. Therefore, the sensor parameter trends are similar when these two faults occur, and the fault impact range and propagation path are basically the same. Therefore, it is extremely easy to confuse each other in ordinary fault diagnosis methods, but the multi-gated memory graph convolutional network diagnosis system proposed in the present invention can effectively distinguish such similar faults of the same type.

[0184] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0185] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0186] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0187] The above is a detailed introduction to the nuclear power plant fault diagnosis method based on a multi-gated graph convolutional network provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, any technician familiar with the technical field can easily think of various equivalent modifications or substitutions within the technical scope disclosed by the present invention, and these modifications or substitutions should be covered within the protection scope of the present invention.

[0188] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

[0189] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A nuclear power plant fault diagnosis method based on a multi-gated memory graph convolutional network, characterized in that: include: Step S1, acquiring historical state parameters of the nuclear power system and preprocessing the historical state parameters to obtain a fault data set, wherein the historical state parameters include real data of fault conditions and simulated data of fault conditions; Step S2, embedding the gated memory unit into the graph convolutional network to construct a gated memory graph convolutional network, and embedding the gated memory mechanism to extract the temporal features and spatial features of the gated memory graph convolutional network; Step S3, determining a number of fault transfer graphs according to the impact range and propagation path of each fault type of the nuclear power plant, fault data, sensor data and fault association information, and embedding the fault transfer graphs into the corresponding gated memory graph convolutional network; Wherein, the number of the gated memory graph convolutional networks is equal to or less than the number of the fault propagation graphs, and the number of the fault propagation graphs is equal to the number of the fault types; Step S4, inputting the fault data in parallel into the gated memory graph convolutional network embedded in each fault transfer graph to extract fault information; Step S5, connecting and weighting the temporal features and spatial features of each of the gated memory graph convolutional networks using an aggregation method to determine a fault vector representation; Step S6, training the model and saving the model parameters to determine the nuclear power plant fault diagnosis system, pre-processing the data to be diagnosed and inputting it into the nuclear power plant fault diagnosis system to determine the diagnosis result.

2. The method for nuclear power plant fault diagnosis based on multi-gated memory graph convolutional network according to claim 1, characterized in that: In step S1, the preprocessing method includes: Step S11, normalizing the historical state parameters to eliminate dimensional differences; Step S12, using a sliding window method to group the normalized historical state parameters to determine an operating condition data sequence; Step S13, annotating each group of data with operating condition information and storing the resultant fault data set.

3. The method for nuclear power plant fault diagnosis based on multi-gated memory graph convolutional network according to claim 1, characterized in that: In step S2, the method of embedding the gated memory unit into the graph convolutional network to construct the gated memory graph convolutional network includes: Step S21, traversing each time step of the model input sequence to extract the features of each node at the corresponding time step; Step S22, using a gated memory mechanism to update the hidden state and cell state of the node in each time step, where: Step S221, updating the cell state according to the forget gate and the input gate; Step S222, determining the hidden state according to the cell state and the output gate; Step S23, obtaining the hidden state of each node at the last time step and determining the hidden state as the output of the model; In step S24, a global average pooling process is performed on the hidden state of each node of the gated memory graph convolutional network to determine the fault propagation characteristics of the gated memory graph convolutional network.

4. The method for nuclear power plant fault diagnosis based on multi-gated memory graph convolutional network according to claim 3 is characterized in that: In step S22, the gated memory mechanism includes: The forget gate is used to determine the amount of information to be discarded in the cell state; Input gate, which controls the influx of new information; Output gate, used to determine the hidden state output at the current moment; The forget gate, the input gate and the output gate are all determined according to the input features of the current time step and the hidden state of the previous time step.

5. The method for nuclear power plant fault diagnosis based on multi-gated memory graph convolutional network according to claim 1, characterized in that: In step S3, the method of determining the fault transfer graph includes: Step S31, simulating the relationship between nodes in the nuclear power plant system by graph theory, wherein: Components represent nodes, and the edge formed by two components represents the influence relationship between the corresponding two nodes; Step S32, determining the propagation mode of the fault information in the nuclear power plant system according to the historical fault data by a causal inference method, wherein the propagation mode includes an impact range and a propagation path; Step S33, determining the adjacency matrix and its values ​​according to the sensor node numbers corresponding to each component in the fault transfer graph, wherein: If there is an edge between node i and node j, then A ij The value of is 1; If there is no edge between node i and node j, then A ij The value of is 0; Step S34, embed each adjacency matrix into a gated memory graph convolutional network to construct a gated memory graph convolutional network in which the fault propagation graph is embedded.

6. The method for nuclear power plant fault diagnosis based on multi-gated memory graph convolutional network according to claim 1, characterized in that: In step S4, the method of inputting the fault data in parallel into the gated memory graph convolutional network embedded in each fault transfer graph to extract the fault information includes: Step S41, inputting the fault data set in parallel into each gated memory graph convolutional network structure embedded by the fault transfer graph; Step S42: determine the fault transfer feature of each fault transfer graph by feature extraction and output the fault transfer feature r mean ; Step S43, determining the corresponding fault type of each gated memory graph convolutional network, and determining the difference between similar fault information according to the fault graph structure corresponding to the fault type.

7. The method for nuclear power plant fault diagnosis based on multi-gated memory graph convolutional network according to claim 1, characterized in that: In step S5, the aggregation method is a weighted aggregation method function based on an attention mechanism, including: Step S51, stacking the fault transfer features of each fault transfer graph to obtain a stacked feature, wherein the stacked feature S T Determined according to the following formula: Where M is the number of fault transfer graphs, r k mean is the fault transfer feature of the kth fault transfer graph; Step S52, mapping the stacked features to the query space, key space and value space according to the weight matrix to determine the query matrix, key matrix and value matrix, wherein the query matrix Q, key matrix K and value matrix V are determined according to the following formulas, respectively: Q=W Q S T ,K=W K S T ,V=W V S T , Where W Q , W K and W V They are different weight matrices respectively; Step S53, determining an attention score according to the query matrix and the key matrix, and normalizing the attention score according to a softmax function to obtain an attention weight; Step S54, using the attention weights to perform weighted summation on the value matrix to determine the output attention, and the output attention is determined according to the following formula: Where B is the dimension of the bond matrix K; Step S55, the output attention is passed through a linear transformation layer to determine the corresponding fault probability distribution, and the fault probability distribution y is determined according to the following formula: y=Linear(T).

8. The method for nuclear power plant fault diagnosis based on multi-gated memory graph convolutional network according to claim 1, characterized in that: The step S6 comprises: Step S61, constructing and saving the fault transfer graph input model; Step S62, training the fault transfer graph according to the back propagation algorithm and then saving the model parameters to obtain a nuclear power plant fault diagnosis system; Step S63, pre-processing the data to be diagnosed and inputting it into the nuclear power plant fault diagnosis system to perform fault diagnosis and determine the diagnosis result.

9. The method for nuclear power plant fault diagnosis based on multi-gated memory graph convolutional network according to claim 8, characterized in that: In step S61, the method for constructing the fault transfer graph transfer model includes converting each fault transfer graph into a corresponding adjacency matrix and then inputting each adjacency matrix into several gated memory graph convolutional networks of the model.

10. The method for nuclear power plant fault diagnosis based on multi-gated memory graph convolutional network according to claim 7, characterized in that: In step S51, the fault transfer features of each fault transfer graph are stacked according to a pre-stored stacking rule to obtain a stacking feature S T ; The stacking rule is stacking along rows.

Images