Detection Method for Planetary Gearbox of Nuclear Power Circulating Water Pump Based on Space-Time Diagram

Abstract

A detection method for the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph is disclosed. In the method, monitoring data of the planetary gearbox of the nuclear power circulating water pump is collected, and the short-time Fourier transform is used to link the time domain and the frequency domain of the vibration signal; the frequency domain of the vibration signal after short-time Fourier transform processing is divided into K parts, and a spatio-temporal graph and a spatio-temporal graph weight matrix are constructed; furthermore, a Laplacian matrix is constructed for feature extraction; the spatio-temporal graph is converted into a feature matrix and input into a time-domain graph convolutional network model to train the model, and the model parameters are obtained through the argmax function; the optimal hyperparameters are selected using the validation set, and a confusion matrix is constructed to evaluate the model; the optimal model is tested using the test set.

Description

Technical Field

[0001] The present invention belongs to the technical field of nuclear power fault diagnosis, and particularly relates to a detection method for a planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph. Background Technique

[0002] Nuclear power investment is huge and involves many industries. Nuclear safety is related to national security and is the lifeline of nuclear power development. As a lifting device in the circulating water system of a nuclear power plant, the circulating water pump functions to supply cooling water to the condenser of the steam turbine in the conventional island and the auxiliary cooling system of the conventional island, and has a very important position in the system. The failure of the circulating water pump will directly cause major events such as equipment unavailability, system suspension, unit power reduction, and shutdown and reactor shutdown (about 10 million per single unit per day), affecting the safety and economic benefits during the startup and operation of the nuclear power plant.

[0003] At present, the maintenance methods of the circulating water pump in a nuclear power plant include two forms: corrective maintenance and preventive maintenance. The existing two maintenance methods have potential defects that are difficult to detect and problems such as over-maintenance of equipment. As a key device of the circulating water pump, the planetary gearbox has characteristics such as large bearing capacity, high centering requirements, high maintenance costs, and high risks. Its vibration signal components are complex, the characteristics are time-varying, and it is difficult to identify weak damages. The health state of the planetary gearbox has a significant impact on the normal operation of the circulating water pump. Once a failure occurs, it will trigger a serious chain reaction and lead to pump shutdown. Therefore, the planetary gearbox is selected as the research object to carry out research on an intelligent fault diagnosis model.

[0004] The above information disclosed in the background technique section is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0005] The object of the present invention is to provide a detection method for a planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph. By constructing a fault diagnosis model based on a spatio-temporal graph and combining the short-time Fourier transform with the spatio-temporal graph, the time domain and the frequency domain are linked. The spatio-temporal graph not only contains time-frequency domain information but also contains graph structure information. The features of the graph are extracted through the graph structure and node attributes. The vibration signals of different fault types correspond to different graph structures. The graph is converted into a two-dimensional matrix, and a fault diagnosis is realized through a graph convolutional network model, improving the accuracy of feature extraction, thereby enhancing the accuracy of fault diagnosis.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A detection method for a planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph of the present invention includes:

[0008] The first step is to collect the vibration signals of the planetary gearbox of the nuclear power circulating water pump and perform short-time Fourier transform on the vibration signals to calculate the discrete time domain and frequency domain:

[0009]

[0010] Among them, x[n] is the vibration signal, n ∈ [0, N - 1], with N observation values; m is time, k is frequency, ΔM is the time variation, Δf is the frequency variation; Y(k, m) is the output at frequency kΔf and time mΔM; the operator (*) represents conjugation; ω is the window function; j is the imaginary unit;

[0011] The second step is to construct a spatio-temporal graph. Among them, calculate the short-time periodogram:

[0012]

[0013] Among them, T is the window length; m is time, k is frequency, Y(k, m) is the time domain and frequency domain after short-time Fourier transform; p(k, m) is the output at frequency index 1 ≤ k ≤ K and time index 1 ≤ m ≤ M, K and M are positive integers, K takes values as the spatio-temporal graph nodes, and M is the upper limit of the vibration signal sampling time domain.

[0014] After the vibration signal undergoes short-time Fourier transform, its frequency domain is divided into K parts. Using each frequency F1, F2,..., F k as nodes and connecting them with attributes to construct a spatio-temporal graph;

[0015] The third step is to construct a spatio-temporal graph weight matrix. Among them,

[0016]

[0017] Among them, W i，j is the matrix weight; Dis{S(F i，t ),S(F j，t )} is the Euclidean distance between nodes F i and F j at time point t; i and j are positive integers,

[0018] The fourth step is feature extraction. Calculate the Laplace operator at the point (x i , y j ) in the weight matrix:

[0019]

[0020] Among them, x is the abscissa in the matrix, y is the ordinate, i and j are positive integers, and Δx and Δy are the increments of x and y, respectively, assumed to be 1,

[0021] Calculate the degree matrix D:

[0022]

[0023] Among them, d i，j is the node weight corresponding to the main diagonal element, and W i，j is the matrix weight. i and j are positive integers.

[0024] The Laplacian matrix is the difference between the degree matrix and the weight matrix:

[0025] L = D - W,

[0026] where L, D, and W are the Laplacian matrix, degree matrix, and weight matrix, respectively.

[0027] Orthogonally decompose the Laplacian matrix L:

[0028] L = UΛU T

[0029] where U = [u0, u1, …, u n-1 is the eigenvector; Λ = diag([λ0, λ1, …, λ n-1 ) is the eigenvalue matrix. Each spatio-temporal graph is represented by a one-dimensional eigenvector F, which is composed of eigenvalues and is denoted as F = [λ0, λ1, …, λ n-1 ;

[0030] In the fifth step, construct a graph convolutional network model, which includes three layers of graph neural networks, and the function is convGCN. The number of input channels of each layer of the convGCN function is the dimension of the data entering the graph convolutional layer, and the number of output channels is equal to the number of input channels of the next layer. The number of output channels of the last layer is the number of fault classifications. Set the activation function between two layers of the convGCN function to ReLU, and the output is a SoftMax layer;

[0031] In the sixth step, convert the spatio-temporal graph into a feature matrix and input it into the graph convolutional network model. Train the model until the classification accuracy converges, and find the model parameters through the argmax function;

[0032] In the seventh step, construct a confusion matrix C to evaluate the model. Its vertical coordinate reflects the true health and fault categories, and the horizontal coordinate reflects the health and fault categories predicted by the model. The elements are denoted as c ij , where i represents the i-th element in the predicted category, j represents the j-th element in the true category, and i and j are positive integers;

[0033] In the eighth step, verify the graph convolutional network model with the validation set, adjust the hyperparameters, and select the optimal hyperparameters for the classification effect;

[0034] In the ninth step, test the optimal model with the test set.

[0035] In the described method for detecting the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph, in the first step, a PCB vibration acceleration sensor is bonded to the gearbox housing at the same height as the middle position of the gear core package to collect the vibration signal of the planetary gearbox of the nuclear power circulating water pump, and the sampling frequency is 10240 Hz.

[0036] In the described method for detecting the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph, 48 channels are used to collect the vibration data of the planetary gearbox of the nuclear power circulating water pump, including the healthy data of the normal operation of the planetary gearbox of the nuclear power circulating water pump, moderate spalling data of the sun gear, moderate pitting data of the sun gear, moderate spalling data of the planetary gear, and moderate pitting data of the planetary gear.

[0037] In the described method for detecting the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph, there are 20 groups for each data type, which are combined into one file. There are 655360 data points in each file. After being read in respectively, they are cut with a sample length of 2048 and divided into 320 groups. The first 80 samples are taken from each group, and 60% of them are used as the training set for training the model; 20% are used as the validation set for adjusting and selecting the model; and the last 20% are used as the test set.

[0038] In the described method for detecting the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph,

[0039] The Laplace operator is further written as:

[0040]

[0041] where N(i, j) is the set of neighbor nodes of (x i , y i ).

[0042] In the described method for detecting the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph, the window function is the Hanning window:

[0043]

[0044] where N is the observed value of the vibration signal.

[0045] In the described method for detecting the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph, in the sixth step, after training, a confusion matrix C is constructed. Its vertical coordinate reflects the true health and fault categories, and its horizontal coordinate reflects the health and fault categories predicted by the model. The elements are denoted as c ij , where i represents the i-th element in the predicted category, j represents the j-th element in the true category, and i, j are positive integers.

[0046] In the above technical solution, a detection method for the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph provided by the present invention has the following beneficial effects: By using the Hanning window in the Short-Time Fourier Transform (STFT), the present invention overcomes the problem that the frequency-domain characteristics of vibration signal noise mask the effective information of the signal. The deep learning method based on graph theory has excellent feature extraction characteristics and has a wide application prospect in the field of fault diagnosis. Introducing graph theory into spectrum analysis, the spatio-temporal graph not only contains time-frequency domain information but also contains graph structure information, improving the discrimination of feature extraction. Description of the Drawings

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings.

[0048] Figure 1 It is a block diagram of a detection method for a circulating water pump planetary gearbox based on a spatio-temporal graph;

[0049] Figure 2 It is a schematic diagram of the spatio-temporal graph construction process;

[0050] Figure 3 It is a schematic diagram of the Laplace operator;

[0051] Figure 4 It is a schematic diagram of a graph convolutional network model;

[0052] Figure 5 It is a schematic diagram of obtaining an input matrix;

[0053] Figure 6 It is a schematic diagram of a fault classification accuracy curve;

[0054] Figure 7 It is a schematic diagram of a gearbox status confusion matrix;

[0055] Figure 8 It is a schematic diagram of an accuracy comparison curve at different learning rates;

[0056] Figure 9 It is a schematic diagram of a confusion matrix when the learning rate is 0.005;

[0057] Figure 10 It is a schematic diagram of an accuracy comparison curve under different sample ratios of the training set, validation set, and test set. Detailed Embodiments

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0059] Therefore, the following detailed description of the embodiments of the present invention provided in the appendices Figures 1 to 10 is not intended to limit the scope of the claimed present invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0060] It should be noted that like reference numerals and letters denote like items in the following figures, and thus, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0061] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", etc. indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation of the present invention.

[0062] In addition, the terms "first" and "second" are only used for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality" means two or more unless otherwise specifically defined.

[0063] In the present invention, unless otherwise clearly specified and defined, the terms "mounted", "connected", "coupled", "fixed", etc. should be construed in a broad sense. For example, it may be a fixed connection, a detachable connection, or integrated; it may be directly connected or indirectly connected through an intermediate medium, and it may be the internal communication of two elements or the interaction relationship between two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0064] In the present invention, unless otherwise clearly specified and defined, the first feature being "above" or "below" the second feature may include the direct contact of the first and second features, or may include the situation where the first and second features are not in direct contact but in contact through additional features therebetween. Moreover, the first feature being "above", "over" and "on top of" the second feature includes that the first feature is directly above and obliquely above the second feature, or merely indicates that the horizontal height of the first feature is higher than that of the second feature. The first feature being "below", "beneath" and "underneath" the second feature includes that the first feature is directly below and obliquely below the second feature, or merely indicates that the horizontal height of the first feature is lower than that of the second feature.

[0065] To enable those skilled in the art to better understand the technical solution of the present invention, the following will further introduce the present invention in detail with reference to the appended Figures 1 to 10 A method for detecting a planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph includes

[0066] The first step is to collect the vibration signals of the planetary gearbox of the nuclear power circulating water pump and perform short-time Fourier transform on the vibration signals to calculate the discrete time domain and frequency domain:

[0067]

[0068] where x[n] is the vibration signal, n ∈ [0, N - 1], with N observation values; m is time, k is frequency, ΔM is the time variation, Δf is the frequency variation; Y(k, m) is the output at frequency kΔf and time mΔM; the operator (*) represents conjugation; ω is the window function; j is the imaginary unit;

[0069] The second step is to construct a spatio-temporal graph, where the short-time periodogram is calculated:

[0070]

[0071] where T is the window length; m is time, k is frequency, Y(k, m) is the time domain and frequency domain after short-time Fourier transform; p(k, m) is the output at frequency index 1 ≤ k ≤ K and time index 1 ≤ m ≤ M, and K, M are positive integers.

[0072] After the vibration signal undergoes short-time Fourier transform, its frequency domain is divided into K parts. Using each frequency F1, F2,..., F in the short-time periodogram k as nodes, with their attributes connected to construct a spatio-temporal graph;

[0073] The third step is to construct a spatio-temporal graph weight matrix, where

[0074]

[0075] where W i，jis the matrix weight; Dis{S(F i , t), S(F j , t)} is the Euclidean distance between nodes F i and F j at time point t; i and j are positive integers,

[0076] Fourth step, feature extraction, calculate the Laplacian operator at the point (x i , y j ) in the weight matrix:

[0077]

[0078] where x is the abscissa in the matrix, y is the ordinate, i and j are positive integers, Δx and Δy are the increments of x and y respectively, set to 1,

[0079] Calculate the degree matrix D:

[0080]

[0081] where d i，j is the node weight corresponding to the main diagonal element, W i，j is the matrix weight, i and j are positive integers,

[0082] The Laplacian matrix is the difference between the degree matrix and the weight matrix:

[0083] L = D - W,

[0084] where L, D, and W are the Laplacian matrix, degree matrix, and weight matrix respectively.

[0085] Orthogonally decompose the Laplacian matrix L:

[0086] L = UAU T

[0087] where U = [u0, u1,..., u n-1 is the eigenvector; Λ = diag([λ0, λ1,..., λ n-1 ) is the eigenvalue matrix, and each spatio-temporal graph is represented by a one-dimensional eigenvector F, which is composed of eigenvalues, denoted as F = [λ0, λ1,..., λ n-1 ;

[0088] Fifth step, construct a graph convolutional network model, which includes three layers of graph neural networks, the function is convGCN, the number of input channels of each layer of the convGCN function is the dimension of the data entering the graph convolutional layer, the number of output channels is equal to the number of input channels of the next layer, and the number of output channels of the last layer is the number of fault classifications. Set the activation function between two layers of the convGCN function to ReLU, and the output is a SoftMax layer;

[0089] In the sixth step, convert the spatio-temporal graph into a feature matrix and input it into the graph convolutional network model. Train the model until the classification accuracy converges, and obtain the model parameters through the argmax function.

[0090] In the seventh step, construct a confusion matrix C to evaluate the model. Its vertical coordinate reflects the true health and fault categories, and the horizontal coordinate reflects the health and fault categories predicted by the model. The elements are denoted as c ij , where i represents the i-th element in the predicted category, j represents the j-th element in the true category, and i and j are positive integers.

[0091] In the eighth step, use the validation set to verify the graph convolutional network model, adjust the hyperparameters, and select the hyperparameters that yield the optimal classification effect.

[0092] In the ninth step, test the optimal model with the test set.

[0093] In the preferred implementation of the method for detecting the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph, in the first step, collect the vibration signals of the planetary gearbox of the nuclear power circulating water pump by adhesively mounting a PCB vibration acceleration sensor on the gearbox housing at the same height as the middle position of the gear core package. The sampling frequency is 10240 Hz.

[0094] In the preferred implementation of the method for detecting the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph, collect the vibration data of the planetary gearbox of the nuclear power circulating water pump using 48 channels, including the health data of the normal operation of the planetary gearbox of the nuclear power circulating water pump, moderate spalling data of the sun gear, moderate pitting data of the sun gear, moderate spalling data of the planetary gear, and moderate pitting data of the planetary gear.

[0095] In the preferred implementation of the method for detecting the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph, there are 20 groups of each data type. Combine them into one file. Each file has 655360 data points. After reading them in respectively, cut them with a sample length of 2048, and divide them into 320 groups. Take the first 80 samples of each group. Use the first 60% as the training set to train the model; 20% as the validation set to adjust and select the model; and the last 20% as the test set.

[0096] In the preferred implementation of the method for detecting the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph,

[0097] The Laplace operator is further written as:

[0098]

[0099] where N(i, j) is the set of neighbor nodes of (x i , y i ).

[0100] In the preferred embodiment of the detection method for the planetary gearbox of the nuclear power circulating water pump based on the spatio-temporal graph, the window function is the Hanning window:

[0101]

[0102] Wherein, N is the observed value of the vibration signal.

[0103] In the preferred embodiment of the detection method for the planetary gearbox of the nuclear power circulating water pump based on the spatio-temporal graph, in the sixth step, after training, a confusion matrix C is constructed, the ordinate of which reflects the true health and fault categories, and the abscissa reflects the health and fault categories predicted by the model, where the element is denoted as c ij , i represents the i-th element in the predicted category, j represents the j-th element in the true category, and i and j are positive integers.

[0104] In one embodiment, a fault diagnosis method based on a time-domain graph convolutional network includes the following steps:

[0105] Step 1, short-time Fourier transform of data.

[0106] Collect the vibration signal of the planetary gearbox of the circulating water pump, and perform a short-time Fourier transform on the data. Calculate the discrete time domain and frequency domain:

[0107]

[0108] Wherein, x[n] is the vibration signal, n ∈ [0, N - 1], with N observed values; m is time, k is frequency, ΔM is the time change amount, Δf is the frequency change amount; Y(k, m) is the output at frequency kΔf and time mΔM; the operator (*) represents the conjugate; ω is the window function; j is the imaginary unit.

[0109] Select the window function as the Hanning window:

[0110]

[0111] Wherein,

[0112] Step 2, construct a spatio-temporal graph.

[0113] After the vibration signal undergoes a short-time Fourier transform, calculate the short-time periodogram:

[0114]

[0115] Among them, T is the window length; m is time, k is frequency, Y(k, m) is the time domain and frequency domain after short-time Fourier transform; p(k, m) is the output for frequency index 1 ≤ k ≤ K and time index 1 ≤ m ≤ M; K and M are positive integers.

[0116] After the vibration signal undergoes short-time Fourier transform, its frequency domain is divided into K parts. Using each frequency F1, F2,......, F in the short-time periodogram k as nodes, with their attributes interconnected to construct a spatio-temporal graph.

[0117] Step 3: Construct the weight matrix.

[0118] The constructed spatio-temporal graph is an undirected graph, and its weight matrix is calculated as follows:

[0119]

[0120] Among them, W i，j is the matrix weight; Dis{S(F i , t), S(F j , t)} is the Euclidean distance between nodes F i and F j at time point t; i and j are positive integers.

[0121] Step 4: Feature extraction.

[0122] Calculate the Laplacian operator at the point (x i , y j ) in the matrix:

[0123]

[0124] Among them, x is the abscissa in the matrix, y is the ordinate, and i and j are positive integers. Δx and Δy are the increments of x and y respectively, and are set to 1.

[0125] Furthermore, the Laplacian operator can be written as:

[0126]

[0127] Among them, N(i, j) is the set of neighbor nodes of (x i , y i ).

[0128] Calculate the degree matrix D:

[0129]

[0130] Among them, d i，j is the node weight corresponding to the main diagonal element, W i，j is the matrix weight, and i and j are positive integers.

[0131] The Laplacian matrix is the difference between the degree matrix and the weight matrix:

[0132] L = D - W

[0133] where L, D, and W are the Laplacian matrix, degree matrix, and weight matrix, respectively.

[0134] Orthogonally decompose the Laplacian matrix L:

[0135] L = UΛU T

[0136] where u = [u0, u1, …, u n-1 is the eigenvector; Λ = diag([λ0, λ1, …, λ n-1 ) is the eigenvalue matrix. Each spatio-temporal graph can be represented by a one-dimensional eigenvector F, which is composed of eigenvalues, denoted as F = [λ0, λ1, …, λ n-1 .

[0137] Step Five: Construct a graph convolutional network model.

[0138] Select a three-layer graph neural network with the function convGCN. The number of input channels for each layer of convGCN is the dimension of the data entering the graph convolutional layer, the number of output channels is equal to the number of input channels of the next layer, and the number of output channels of the last layer is the number of fault classifications. Additionally, set the activation function between two layers of the convGCN function to ReLU, and the output is a SoftMax layer.

[0139] Step Six: Train the model.

[0140] Convert the spatio-temporal graph into a feature matrix and input it into the model. Train the model until the classification accuracy converges, and obtain the model parameters through the argmax function;

[0141] Step Seven: Construct the confusion matrix C.

[0142] Use it to evaluate the model. The vertical axis reflects the true health and fault categories, and the horizontal axis reflects the health and fault categories predicted by the model. The elements are denoted as c ij , where i represents the i-th element in the predicted category, j represents the j-th element in the true category, and i, j are positive integers.

[0143] Step Seven: Verify the model.

[0144] Verify the model using the validation set, adjust the hyperparameters, and select the hyperparameters that yield the optimal classification effect.

[0145] Step Eight: Test the model.

[0146] Test the optimal model using the test set.

[0147] In one embodiment, a fault diagnosis method for a circulating water pump planetary gearbox based on a spatio-temporal graph includes the following steps:

[0148] Step 1: Data acquisition.

[0149] Use a 48-channel IXXAT data acquisition system to collect data by bonding PCB vibration acceleration sensors on the gearbox housing at the same height as the middle position of the gear core package. The sampling frequency is 10240 Hz. The data includes healthy data, moderate spalling data of the sun gear, moderate pitting data of the sun gear, moderate spalling data of the planetary gear, and moderate pitting data of the planetary gear, with 20 groups for each data type. These data are combined into one file, and each file contains 655360 data points. After reading them in, they are cut with a sample length of 2048 and divided into 320 groups. The first 80 samples are taken from each group, and 60% of them are used as the training set to train the model; 20% are used as the validation set to adjust and select the model; and the last 20% are used as the test set to evaluate the final model.

[0150] Step 2: Construct a spatio-temporal graph.

[0151] After the vibration signal undergoes short-time Fourier transform, calculate the short-time periodogram:

[0152]

[0153] where T is the window length; m is time, k is frequency, Y(k, m) is the time domain and frequency domain after short-time Fourier transform; p(k, m) is the output for frequency index 1 ≤ k ≤ K and time index 1 ≤ m ≤ M; K and M are positive integers.

[0154] After the vibration signal undergoes short-time Fourier transform, its frequency domain is divided into K parts. Using each frequency F1, F2,......, F k in the short-time periodogram as nodes, with properties connected to each other, construct a spatio-temporal graph, as Figure 2 shown. K = 33, that is, the number of spatio-temporal nodes is 33.

[0155] Step 3: Construct a weight matrix.

[0156] The constructed spatio-temporal graph is an undirected graph, and calculate its weight matrix:

[0157]

[0158] where W i，j is the matrix weight; Dis{S(F i , t), S(F j , t)} is the Euclidean distance between nodes F i and F j at time point t; i and j are positive integers.

[0159] Step 4: Feature extraction.

[0160] As Figure 3 shown, calculate the Laplacian operator at the point (x i , y j ) in the matrix:

[0161]

[0162] where x is the abscissa in the matrix, y is the ordinate, and i, j are positive integers. Δx and Δy are the increments of x and y respectively, and are set to 1.

[0163] Furthermore, the Laplacian operator can be written as:

[0164]

[0165] where N(i, j) is the set of neighbor nodes of (x i , y i ).

[0166] Calculate the degree matrix D:

[0167]

[0168] where d i，j is the node weight corresponding to the main diagonal element, W i，j is the matrix weight, and i, j are positive integers.

[0169] The Laplacian matrix is the difference between the degree matrix and the weight matrix:

[0170] L = D - W

[0171] where L, D, and W are the Laplacian matrix, degree matrix, and weight matrix respectively.

[0172] Orthogonally decompose the Laplacian matrix L:

[0173] L = UΛU T

[0174] where U = [u0, u1, …, u n-1 is the eigenvector; Λ = diag([λ0, λ1, …, λ n-1 ) is the eigenvalue matrix. Each spatio-temporal graph can be represented by a one-dimensional feature vector F, which consists of eigenvalues, denoted as F = [λ0, λ1, …, λ n-1 .

[0175] Step 5: Construct a graph convolutional network model.

[0176] Select a three-layer graph neural network with the function convGCN. The weight decay of the first layer of convGCN is set to 6×10-4 The weight decay of the second - layer convGCN is set to 3×10 -4 The weight decay of the third - layer convGCN is set to 1×10 -4 . The number of input channels of each layer of convGCN is the dimension of the data entering the graph convolutional layer, the number of output channels is equal to the number of input channels of the next layer, and the number of output channels of the last layer is the number of fault classifications. In addition, the activation function between two layers of convGCN functions is set to ReLU, and the output is a SoftMax layer.

[0177] Step 6: Set the initial parameters.

[0178] Set the number of training times to 50 and the learning rate to 0.01. The dimension of the input data is 33, and the number of edges of each graph is 10.

[0179] Step 7: Construct the confusion matrix.

[0180] Define that the ordinate of the confusion matrix reflects the true health and fault categories, and the abscissa reflects the health and fault categories predicted by the model. Denote the confusion matrix as C, and its elements are denoted as c ij , where i represents the i - th element in the predicted category, j represents the j - th element in the true category, and i, j are positive integers.

[0181] Step 8: Train the model.

[0182] As shown in Figure 5 , convert the spatio - temporal graph into a feature matrix and input it into the model. Train the model until the classification accuracy converges, and obtain the model parameters through the argmax function. The model fault diagnosis accuracy is as shown in Figure 6 , and the confusion matrix is as shown in Figure 7 . After about 45 times of training, the accuracy approaches 99%.

[0183] Step 9: Validate the model.

[0184] Validate the model with the validation set, adjust the hyperparameters (such as learning rate, sample length), and select the hyperparameters that make the classification effect optimal.

[0185] With other parameters unchanged, set the learning rate to 0.005, 0.015, 0.02 respectively, and the initial learning rate to 0.005 with a learning rate decay of 80% / 5 training times. The accuracy curve is as shown in Figure 8 . When the learning rate is 0.02, the number of training times required for the accuracy to approach 99% is the least; while when the learning rate is 0.005, the accuracy drops to 80%, and the confusion matrix at this time is as shown in Figure 9 . It can be seen that due to the too - low learning rate setting, the learning process is not sufficient, and the pitting state fault of the planetary gear is not recognized.

[0186] With other parameters unchanged, change the sample ratios of the training set, validation set, and test set, as Figure 10 shown. It can be seen the impact of the training set division on the accuracy.

[0187] With other parameters unchanged, change the sample length, that is, the number of input and output channels (H×I) of the second layer of the model. The changes in the accuracy are shown in Table 1. It can be seen that when the number of input and output channels of the second layer of the model is 48×30, the accuracy approaches 99% fastest.

[0188] Therefore, set the learning rate to 0.02 and the number of input and output channels of the second layer to 48×30.

[0189] Step Ten: Test the model.

[0190] Test the optimal model with the test set, and the accuracy approaches 99%.

[0191] Table 1 Comparison of accuracies under different input and output channels of the convolutional layer

[0192]

[0193] Finally, it should be noted that: the described embodiments are only a part of the embodiments of the present application, rather than all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0194] Some exemplary embodiments of the present invention have been described only by way of illustration. Undoubtedly, for those of ordinary skill in the art, the described embodiments can be modified in various different ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A detection method for the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph, characterized in that, It includes the following steps: The first step is to collect the vibration signals of the planetary gearbox of the nuclear power cycle water pump, and perform short-time Fourier transform on the vibration signals to calculate the discrete time domain and frequency domain: ， Where x[n] is the vibration signal, n ∈ [0, N - 1], with N observation values; m is time, k is frequency, ΔM is the time variation, Δf is the frequency variation; Y(k, m) is the output at frequency kΔf and time mΔM; the operator (*) represents conjugation; ω is the window function; j is the imaginary unit; The second step is to construct a spatio-temporal graph, where the short-time periodogram is calculated: ， Where T is the window length; Y(k, m) is the discrete time domain and frequency domain after short-time Fourier transform; p(k, m) is the output at frequency index 1 ≤ k ≤ K and time index 1 ≤ m ≤ M, K and M are positive integers, K is the value of the spatio-temporal graph node, and M is the upper limit of the sampling time domain of the vibration signal; After the vibration signal undergoes short-time Fourier transform, its frequency domain is divided into K parts. Using each frequency F1, F2, ……, F in the short-time periodogram as nodes, with their attributes connected to each other, a spatio-temporal graph is constructed; k as nodes, with their attributes connected to each other, a spatio-temporal graph is constructed; The third step is to construct the spatio-temporal graph weight matrix, where ， Among them, W i,j is the matrix weight; Dis{S(F i , t), S(F j , t)} is the Euclidean distance between nodes F i and F j at time point t; i, j are positive integers; Fourth step, feature extraction, calculate the Laplacian operator at the point (x i , y j ) in the weight matrix: ， Where x is the abscissa in the weight matrix, y is the ordinate, i and j are positive integers, Δx and Δy are the increments of x and y respectively, and are set to 1; Calculate the degree matrix D: ， where d i,j is the node weight corresponding to the main diagonal element, W i,j is the matrix weight, and i, j are positive integers; The Laplacian matrix is the difference between the degree matrix and the weight matrix: ， Where L, D, and W are the Laplacian matrix, degree matrix, and weight matrix respectively; Orthogonally decompose the Laplacian matrix L: ， Among them, is the eigenvector; is the eigenvalue matrix; each spatio-temporal graph is represented by a one-dimensional eigenvector F, which is composed of eigenvalues and denoted as ; The fifth step is to construct a graph convolutional network model, which includes three layers of graph neural networks, and the function is convGCN. The number of input channels of each layer of the convGCN function is the dimension of the data entering the graph convolutional layer, and the number of output channels is equal to the number of input channels of the next layer. The number of output channels of the last layer is the number of fault classifications. The activation function ReLU is set between two layers of the convGCN function, and the output is a SoftMax layer; The sixth step is to convert the spatio-temporal graph into a feature matrix and input it into the graph convolutional network model, train the graph convolutional network model 1000 times, and find the parameters of the graph convolutional network model through the argmax function; In the seventh step, a confusion matrix C is constructed to evaluate the graph convolutional network model, where the ordinate reflects the true health and fault categories, and the abscissa reflects the health and fault categories predicted by the model, and the elements are denoted as c ij , where i represents the i-th element in the predicted category, j represents the j-th element in the true category, and i and j are positive integers; The eighth step is to verify the graph convolutional network model with the validation set, adjust the hyperparameters, and select the hyperparameters that make the classification effect optimal; The ninth step is to test the optimal model with the test set.

2. The detection method of the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph according to claim 1, wherein, In the first step, the vibration signals of the planetary gearbox of the nuclear power cycle water pump are collected by adhesively mounting a PCB vibration acceleration sensor on the gearbox housing at the same height as the middle position of the gear core package, and the sampling frequency is 10240Hz.

3. A detection method for a planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph according to claim 2, characterized in that 48 channels are used to collect the vibration data of the planetary gearbox of the nuclear power cycle water pump, which includes the healthy data of the normal operation of the planetary gearbox of the nuclear power cycle water pump, the moderate spalling data of the sun gear, the moderate pitting data of the sun gear, the moderate spalling data of the planetary gear, and the moderate pitting data of the planetary gear.

4. The method for detecting a planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph according to claim 3, wherein There are 20 groups of each data type, which are combined into a file. There are 655360 data points in each file. After being read separately, they are cut with a sample length of 2048 and divided into 320 groups. The first 80 samples are taken from each group, and the first 60% is used as the training set to train the model; 20% is used as the validation set to adjust and select the model; the last 20% is used as the test set.

5. The detection method of the planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph according to claim 1, wherein, The Laplace operator is further written as: ， Among them, N(i, j) is the set of neighbor nodes of (x i , y i ).

6. A method for detecting a planetary gearbox of a nuclear power circulating water pump based on a spatio-temporal graph according to claim 1, characterized in that, The window function is a Hanning window: ， Among them, , N is the observed value of the vibration signal.

Images