Wind power generation gearbox fault detection method

Abstract

The invention relates to the technical field of gearbox fault detection, and discloses a wind power generation gearbox fault detection method comprising the following steps: step S101, setting M vibration data acquisition point locations on a wind power generation gearbox, and obtaining vibration data of each vibration data acquisition point location; step S102, processing the vibration data to obtain a time frequency matrix; step S103, constructing graph structure data according to the time-frequency matrix of the M vibration data acquisition point locations; s105, inputting the graph structure data into the trained gearbox fault detection model, wherein an output value represents the fault type of the wind power generation gearbox; according to the method, the plurality of vibration data acquisition points are arranged on the wind power generation gear box, the graph structure data is constructed based on the vibration data of the plurality of vibration data acquisition points, and feature aggregation is performed on the graph structure data through the gear box fault detection model, so that the fault detection precision of the wind power generation gear box is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of gearbox fault detection, and more specifically, to a wind power generation gearbox fault detection method. Background Art

[0002] As an important part of renewable energy, wind power generation plays an important role in energy transformation. Wind turbines generate electricity by converting wind energy into electrical energy. One of the core components in this process is the gearbox. The gearbox is responsible for converting the low-speed, high-torque wind wheel rotation into high-speed, low-torque generator speed in the wind power generation system to match the optimal working point of the generator. The reliability of the gearbox directly affects the operating efficiency and overall economy of the wind turbine. Gearbox failure will not only generate high maintenance costs, but also cause unplanned downtime, affect the continuity and stability of power supply, and thus have a negative impact on the operating income of the wind farm.

[0003] Existing wind turbine gearbox fault detection usually collects vibration data through vibration sensors, extracts the time domain features of the vibration data through LSTM (Long Short-Term Memory Neural Network Model), extracts the frequency domain features of the vibration data through CNN (Convolutional Neural Network Model), and combines the time domain features and frequency domain features of the vibration data into a classifier. The classification space of the classifier represents the gearbox fault type.

[0004] The wind turbine gearbox contains multiple components, such as bearings, planetary gear sets, etc. The health status of each component will affect the overall operating efficiency. However, the above solution only obtains vibration data through a single vibration sensor, which may not be sufficient to reflect the health status of the entire gearbox, resulting in low accuracy in wind turbine gearbox fault detection. Summary of the invention

[0005] The present invention provides a wind power gearbox fault detection method, which solves the technical problem in the related art that vibration data obtained by only a single vibration sensor may not be sufficient to reflect the health status of the entire gearbox, resulting in low accuracy of wind power gearbox fault detection.

[0006] The present invention provides a wind power generation gearbox fault detection method, comprising the following steps:

[0007] Step S101, setting M vibration data collection points on the wind power generation gearbox, and acquiring vibration data of the M vibration data collection points within a preset time period T;

[0008] The vibration data is represented by time domain waveform data, which includes N time windows, each time window includes G time points, and each time point corresponds to an acceleration value;

[0009] Step S102, respectively processing the vibration data of the M vibration data collection points to obtain the time-frequency matrix of the M vibration data collection points;

[0010] The number of rows of the time-frequency matrix is ​​N and the number of columns is K. The element in the i-th row and j-th column of the time-frequency matrix represents the normalized power spectrum density corresponding to the j-th frequency value of the i-th time window, where 1≤i≤N, 1≤j≤K;

[0011] Step S103, constructing graph structure data according to the time-frequency matrix of the M vibration data collection points;

[0012] The graph structure data includes M nodes, initial features of the nodes, and edges between the nodes;

[0013] The M nodes establish mapping relationships with the M vibration data collection points respectively;

[0014] The initial characteristics of the node are represented by the time-frequency matrix of the vibration data collection points that establish a mapping relationship with the node;

[0015] Step S104, constructing and training a gearbox fault detection model;

[0016] Step S105, inputting the graph structure data into the trained gearbox fault detection model, and the output value represents the fault type of the wind power gearbox;

[0017] Fault types include: gear damage, bearing failure, spindle imbalance, gearbox overload and coupling failure.

[0018] Furthermore, M, K and G are all custom parameters, T is set to the duration of the wind turbine gearbox main shaft rotating a preset number of times, where the preset number of times is a custom parameter, and N divides T according to a preset time interval, where the preset time interval is a custom parameter.

[0019] Furthermore, the vibration data is processed to obtain a time-frequency matrix, including the following steps:

[0020] Step S201, calculating the window coefficient of each time window of the vibration data by using a window function;

[0021] The calculation formula of the window function is as follows:

[0022]

[0023] Where 1≤g≤G, 1≤t≤N, represents the window coefficient of the g-th time point of the t-th time window;

[0024] Step S202, converting the time domain waveform data of the vibration data into frequency domain waveform data by short-time Fourier transform according to the window coefficient of each time window;

[0025] The frequency domain waveform data includes K frequency values, each of which corresponds to a power spectrum density;

[0026] Step S203: normalize the power spectrum density corresponding to the K frequency values ​​in each time window by using the Min-Max normalization method.

[0027] Furthermore, the calculation formula of power spectrum density includes:

[0028]

[0029]

[0030] Where 1≤k≤K, represents the power spectral density corresponding to the kth frequency value in the tth time window, The complex value representing the kth frequency value in the tth time window, and They respectively represent the acceleration value and window coefficient at the g-th time point in the t-th time window, e represents a natural constant, and y represents an imaginary unit.

[0031] Furthermore, the edges between nodes include: building edges between all nodes that establish mapping relationships with vibration data collection points within a preset distance threshold; building edges between any two nodes if the similarity value between the initial features of the two nodes is greater than or equal to a preset similarity threshold.

[0032] Further, the gearbox fault detection model includes: N first hidden layers, N extraction layers, a second hidden layer and a classifier;

[0033] N first hidden layers share weight parameters and bias parameters;

[0034] Each first hidden layer includes K hidden units;

[0035] The jth hidden unit of the i-th first hidden layer inputs the element value of the i-th row and j-th column of the time-frequency matrix and outputs the first update vector;

[0036] The mth extraction layer is used to extract the first update vector output by the Kth hidden unit of the Nth first hidden layer, and replace the initial feature of each node with the first update vector representation;

[0037] The second hidden layer inputs the graph structure data and outputs the graph update matrix, wherein the m-th row vector of the graph update matrix represents the second update vector of the m-th node;

[0038] The graph update matrix is ​​input into the classifier, and the classification space of the classifier represents the fault types of the wind turbine gearbox.

[0039] Furthermore, the calculation formula for the jth hidden unit of the i-th first hidden layer includes:

[0040]

[0041]

[0042]

[0043]

[0044] in represents the first update vector output by the jth hidden unit of the i-th first hidden layer, Represents the element value of the i-th row and j-th column of the time-frequency matrix input to the j-th hidden unit of the i-th first hidden layer, and denote the reset gate, update gate and intermediate vector of the jth hidden unit of the i-th first hidden layer, respectively. and denote respectively the first weight parameter, the second weight parameter, the third weight parameter, the fourth weight parameter, the fifth weight parameter and the sixth weight parameter of the j-th hidden unit of the i-th first hidden layer, and denote the first bias parameter, the second bias parameter and the third bias parameter of the jth hidden unit of the i-th first hidden layer, respectively. and denote the first update vector and the number of dimensions of the first update vector output by the jth hidden unit of the i-1th first hidden layer, respectively, and denote the dimensions of the first update vector and the first update vector output by the j-1th hidden unit of the i-th first hidden layer, respectively. and All are assigned 0, MLP represents multi-layer perceptron, ⊙ represents point-by-point multiplication, sigmoid represents sigmoid activation function, and tanh represents hyperbolic tangent activation function.

[0045] Furthermore, the calculation formula of the second hidden layer includes:

[0046]

[0047]

[0049]

[0050] Where 1≤u≤M, 1≤v≤M, u≠v, S represents the graph update matrix of the second hidden layer output, h u and They represent the first update vector and the second update vector of the u-th node, respectively, and h v represents the first update vector of the vth node, Col u Represents the set of nodes that have edges connected to the u-th node, Nattion uv and F uv Represent the attention score and the number of edge connections between the u-th node and the v-th node, respectively. and b uv Respectively represent the first weight parameter, the second weight parameter, the third weight parameter and the bias parameter between the u-th node and the v-th node, W u Indicates the weight parameter corresponding to the u-th node, pile M It indicates that the second update vectors of M nodes are stacked, concat indicates a concatenation operation, T indicates a transposition operation, sigmoid indicates a sigmoid activation function, and softmax indicates a softmax activation function.

[0051] Furthermore, a simulation model of a wind turbine gearbox is constructed through a simulation platform, and one or more fault types are randomly generated on the simulation model as sample labels of a training sample, and training data of a training sample is constructed according to steps S101 to S103 until U training samples are obtained, where U is a custom parameter.

[0052] The present invention provides a storage medium storing non-temporary computer-readable instructions. When the non-temporary computer-readable instructions are executed by a computer, the steps in the above-mentioned wind power generation gearbox fault detection method are executed.

[0053] The beneficial effects of the present invention are as follows: the present invention sets a plurality of vibration data collection points in a wind power gearbox, constructs graph structure data based on the vibration data of the plurality of vibration data collection points, and performs feature aggregation on the graph structure data through a gearbox fault detection model, thereby improving the accuracy of wind power gearbox fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flow chart of a wind power generation gearbox fault detection method of the present invention;

[0055] Figure 2 It is a flow chart of the present invention for processing vibration data to obtain a time-frequency matrix;

[0056] Figure 3 It is a comparison chart of the test results of the present invention. DETAILED DESCRIPTION

[0057] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.

[0058] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in one or more embodiments of the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in one or more embodiments of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0059] like Figure 1 to Figure 3 As shown, a wind power generation gearbox fault detection method comprises the following steps:

[0060] Step S101, setting M vibration data collection points on the wind power generation gearbox, and acquiring vibration data of the M vibration data collection points within a preset time period T;

[0061] The vibration data is represented by time domain waveform data, which includes N time windows, each time window includes G time points, and each time point corresponds to an acceleration value;

[0062] Step S102, respectively processing the vibration data of the M vibration data collection points to obtain the time-frequency matrix of the M vibration data collection points;

[0063] The number of rows of the time-frequency matrix is ​​N and the number of columns is K. The element in the i-th row and j-th column of the time-frequency matrix represents the normalized power spectrum density corresponding to the j-th frequency value of the i-th time window, where 1≤i≤N, 1≤j≤K;

[0064] Step S103, constructing graph structure data according to the time-frequency matrix of the M vibration data collection points;

[0065] The graph structure data includes M nodes, initial features of the nodes, and edges between the nodes;

[0066] The M nodes establish mapping relationships with the M vibration data collection points respectively;

[0067] The initial characteristics of the node are represented by the time-frequency matrix of the vibration data collection points that establish a mapping relationship with the node;

[0068] Step S104, constructing and training a gearbox fault detection model;

[0069] Step S105, inputting the graph structure data into the trained gearbox fault detection model, and the output value represents the fault type of the wind power gearbox;

[0070] Fault types include: gear damage, bearing failure, spindle imbalance, gearbox overload and coupling failure.

[0071] It should be noted that gear damage may include: tooth surface wear, tooth root cracks, tooth surface peeling, gear misalignment, etc.; bearing failure may include: inner ring or outer ring damage, rolling element damage, cage failure, etc.; main shaft imbalance may be the imbalance of the input shaft connecting the wind wheel and the gearbox, which may be caused by blade damage or ice on the blade surface; gearbox overload means overheating caused by overload operation or insufficient heat dissipation. Therefore, temperature data can be collected through temperature sensors while collecting vibration data to improve the accuracy of gearbox overload detection; coupling failure includes: coupling wear or damage.

[0072] In one embodiment of the present invention, M, K and G are all custom parameters. Preferably, M is set to 10, K is set to 1000, and G is set to 10; T is set to the duration of the wind turbine gearbox main shaft rotating a preset number of times, wherein the preset number of times is a custom parameter, and preferably, the preset number of times is set to 2; N divides T according to a preset time interval, wherein the preset time interval is a custom parameter, and preferably, the preset time interval is set to 1 second, then N=T / 1.

[0073] In one embodiment of the present invention, 10 vibration data collection points are set in the wind power gearbox, including 4 vibration data collection points surrounding the surface of the gearbox housing as much as possible, 4 vibration data collection points close to the planetary gear set, 1 vibration data collection point close to the main shaft and 1 vibration data collection point close to the coupling.

[0074] In one embodiment of the present invention, the revolutions per minute (RPM) of the main shaft of the wind power gearbox is converted into revolutions per second (RPS), and then the reciprocal of the RPS is taken to obtain the duration of one rotation of the main shaft of the wind power gearbox. For example: the RPM of the main shaft of the wind power gearbox is equal to 10, and the RPS is equal to 10 / 60=1 / 6, then the duration of one rotation of the main shaft of the wind power gearbox is equal to 6 seconds, then T is equal to 6*2 / 1=12, and N divides T according to the preset time interval, then N=12 / 1=12.

[0075] In one embodiment of the present invention, Figure 2 As shown, the vibration data is processed to obtain a time-frequency matrix, including the following steps:

[0076] Step S201, calculating the window coefficient of each time window of the vibration data by using a window function;

[0077] The calculation formula of the window function is as follows:

[0078]

[0079] Where 1≤g≤G, 1≤t≤N, represents the window coefficient of the g-th time point of the t-th time window;

[0080] Step S202, converting the time domain waveform data of the vibration data into frequency domain waveform data by short-time Fourier transform according to the window coefficient of each time window;

[0081] The frequency domain waveform data includes K frequency values, each of which corresponds to a power spectrum density;

[0082] Step S203: normalize the power spectrum density corresponding to the K frequency values ​​in each time window by using the Min-Max normalization method.

[0083] In one embodiment of the present invention, before the vibration data is processed to obtain the time-frequency matrix, a Gaussian filter may be used to perform noise reduction on the vibration data.

[0084] In one embodiment of the present invention, the calculation formula of the Min-Max normalization method is as follows:

[0085]

[0086] Where 1≤k≤K, represents the normalized power spectral density corresponding to the kth frequency value in the tth time window, represents the power spectral density corresponding to the kth frequency value in the tth time window, and Respectively represent the power spectral density corresponding to the maximum frequency value and the minimum frequency value of the tth time window.

[0087] In one embodiment of the present invention, the power spectrum density corresponding to the K frequency values ​​in each time window may also be normalized by using a z-score (standard score).

[0088] In one embodiment of the present invention, the calculation formula of power spectrum density includes:

[0089]

[0090]

[0091] Where 1≤k≤K, represents the power spectral density corresponding to the kth frequency value in the tth time window, The complex value representing the kth frequency value in the tth time window, and They respectively represent the acceleration value and window coefficient at the g-th time point in the t-th time window, e represents a natural constant, and y represents an imaginary unit.

[0092] In one embodiment of the present invention, the edges between nodes include: building edges between all nodes that establish mapping relationships with vibration data collection points within a preset distance threshold; if the similarity value between the initial features of any two nodes is greater than or equal to the preset similarity threshold, then an edge is built between the two nodes, and the similarity value can be calculated by Euclidean distance or cosine similarity, etc.

[0093] In one embodiment of the present invention, the gearbox fault detection model includes: N first hidden layers, N extraction layers, a second hidden layer and a classifier;

[0094] N first hidden layers share weight parameters and bias parameters;

[0095] Each first hidden layer includes K hidden units;

[0096] The jth hidden unit of the i-th first hidden layer inputs the element value of the i-th row and j-th column of the time-frequency matrix and outputs the first update vector;

[0097] The mth extraction layer is used to extract the first update vector output by the Kth hidden unit of the Nth first hidden layer, and replace the initial feature of each node with the first update vector representation;

[0098] The second hidden layer inputs the graph structure data and outputs the graph update matrix, wherein the m-th row vector of the graph update matrix represents the second update vector of the m-th node;

[0099] The graph update matrix is ​​input into the classifier, and the classification space of the classifier represents the fault types of the wind turbine gearbox.

[0100] In one embodiment of the present invention, the classifier is a softmax classifier, which outputs probability values ​​of different fault types of the wind power gearbox, and takes the fault type with the largest probability value as the fault type of the wind power gearbox.

[0101] For example, the probability values ​​of the fault types of gear damage, bearing failure, main shaft imbalance, gearbox overload and coupling failure are 95%, 1%, 1%, 1% and 1% respectively, then the fault type of the wind turbine gearbox is gear damage.

[0102] In one implementation of the present invention, the calculation formula for the jth hidden unit of the i-th first hidden layer includes:

[0103]

[0104]

[0105]

[0106]

[0107] in represents the first update vector output by the jth hidden unit of the i-th first hidden layer, Represents the element value of the i-th row and j-th column of the time-frequency matrix input to the j-th hidden unit of the i-th first hidden layer, and denote the reset gate, update gate and intermediate vector of the jth hidden unit of the i-th first hidden layer, respectively. and denote respectively the first weight parameter, the second weight parameter, the third weight parameter, the fourth weight parameter, the fifth weight parameter and the sixth weight parameter of the j-th hidden unit of the i-th first hidden layer, and denote the first bias parameter, the second bias parameter and the third bias parameter of the jth hidden unit of the i-th first hidden layer, respectively. and denote the first update vector and the number of dimensions of the first update vector output by the jth hidden unit of the i-1th first hidden layer, respectively, and denote the dimensions of the first update vector and the first update vector output by the j-1th hidden unit of the i-th first hidden layer, respectively. and All are assigned 0, MLP represents multi-layer perceptron, ⊙ represents point-by-point multiplication, sigmoid represents sigmoid activation function, and tanh represents hyperbolic tangent activation function.

[0108] In one implementation of the present invention, the calculation formula of the second hidden layer includes:

[0109]

[0110]

[0111]

[0112] Where 1≤u≤M, 1≤v≤M, u≠v, S represents the graph update matrix of the second hidden layer output, h u and They represent the first update vector and the second update vector of the u-th node, respectively, and h v represents the first update vector of the vth node, Col u Represents the set of nodes that have edges connected to the u-th node, Nattion uv and F uv Represent the attention score and the number of edge connections between the u-th node and the v-th node, respectively. and b uv Respectively represent the first weight parameter, the second weight parameter, the third weight parameter and the bias parameter between the u-th node and the v-th node, W u Indicates the weight parameter corresponding to the u-th node, pile M It indicates that the second update vectors of M nodes are stacked, concat indicates a concatenation operation, T indicates a transposition operation, sigmoid indicates a sigmoid activation function, and softmax indicates a softmax activation function.

[0113] In one embodiment of the present invention, a simulation model of a wind turbine gearbox is constructed through a simulation platform, and one or more fault types are randomly generated on the simulation model as sample labels of a training sample, and training data of a training sample is constructed according to steps S101 to S103 until U training samples are obtained, where U is a custom parameter. Preferably, U is set to 2000.

[0114] It should be noted that in the process of training the gearbox fault detection model, the difference between the predicted value output by the gearbox fault detection model at each iteration and the true value of the sample label of the training sample is specified as the loss function, and the weight parameters and bias parameters of the gearbox fault detection model are updated through back propagation and gradient descent algorithms.

[0115] In one embodiment of the present invention, a storage medium is provided, which stores non-transitory computer-readable instructions. When the non-transitory computer-readable instructions are executed by a computer, the steps in the above-mentioned wind power gearbox fault detection method can be executed.

[0116] like Figure 3 As shown, the fault types of 500 test data are tested by a wind power gearbox fault detection method provided by the present invention, the number of test data for each fault type is 100, and the detection accuracy of the wind power gearbox fault detection method provided by the present invention is 100+100+99+99+100 divided by 500, which is equal to 99.6%.

[0117] The above describes an embodiment of the present embodiment, but the present embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present embodiment, ordinary technicians in this field can also make many forms, all of which are within the protection of the present embodiment.

Claims

1. A wind power generation gearbox fault detection method, characterized in that: The following steps are involved: Step S101, setting M vibration data collection points on the wind power generation gearbox, and acquiring vibration data of the M vibration data collection points within a preset time period T; The vibration data is represented by time domain waveform data, which includes N time windows, each time window includes G time points, and each time point corresponds to an acceleration value; Step S102, respectively processing the vibration data of the M vibration data collection points to obtain the time-frequency matrix of the M vibration data collection points; The number of rows of the time-frequency matrix is ​​N and the number of columns is K. The element in the i-th row and j-th column of the time-frequency matrix represents the normalized power spectrum density corresponding to the j-th frequency value of the i-th time window, where 1≤i≤N, 1≤j≤K; Step S103, constructing graph structure data according to the time-frequency matrix of the M vibration data collection points; The graph structure data includes M nodes, initial features of the nodes, and edges between the nodes; The M nodes establish mapping relationships with the M vibration data collection points respectively; The initial characteristics of the node are represented by the time-frequency matrix of the vibration data collection points that establish a mapping relationship with the node; Step S104, constructing and training a gearbox fault detection model; Step S105, inputting the graph structure data into the trained gearbox fault detection model, and the output value represents the fault type of the wind power gearbox; Fault types include: gear damage, bearing failure, spindle imbalance, gearbox overload and coupling failure.

2. A wind power generation gearbox fault detection method according to claim 1, characterized in that: M, K and G are all custom parameters, T is set to the duration of the wind turbine gearbox main shaft rotating a preset number of times, where the preset number of times is a custom parameter, and N divides T according to a preset time interval, where the preset time interval is a custom parameter.

3. A wind power generation gearbox fault detection method according to claim 1, characterized in that: Processing the vibration data to obtain a time-frequency matrix includes the following steps: Step S201, calculating the window coefficient of each time window of the vibration data by using a window function; The calculation formula of the window function is as follows: Where 1≤g≤G, 1≤t≤N, represents the window coefficient of the g-th time point of the t-th time window; Step S202, converting the time domain waveform data of the vibration data into frequency domain waveform data by short-time Fourier transform according to the window coefficient of each time window; The frequency domain waveform data includes K frequency values, each of which corresponds to a power spectrum density; Step S203: normalize the power spectrum density corresponding to the K frequency values ​​in each time window by using the Min-Max normalization method.

4. A wind power generation gearbox fault detection method according to claim 3, characterized in that: The calculation formula for power spectral density includes: Where 1≤k≤K, represents the power spectral density corresponding to the kth frequency value in the tth time window, The complex value representing the kth frequency value in the tth time window, and They respectively represent the acceleration value and window coefficient at the g-th time point in the t-th time window, e represents a natural constant, and y represents an imaginary unit.

5. A wind power generation gearbox fault detection method according to claim 1, characterized in that: The edges between nodes include: building edges between all nodes that establish mapping relationships with vibration data collection points within a preset distance threshold; building edges between any two nodes if the similarity value between the initial features of the nodes is greater than or equal to the preset similarity threshold.

6. A wind power generation gearbox fault detection method according to claim 1, characterized in that: The gearbox fault detection model includes: N first hidden layers, N extraction layers, a second hidden layer and a classifier; N first hidden layers share weight parameters and bias parameters; Each first hidden layer includes K hidden units; The jth hidden unit of the i-th first hidden layer inputs the element value of the i-th row and j-th column of the time-frequency matrix and outputs the first update vector; The mth extraction layer is used to extract the first update vector output by the Kth hidden unit of the Nth first hidden layer, and replace the initial feature of each node with the first update vector representation; The second hidden layer inputs the graph structure data and outputs the graph update matrix, wherein the m-th row vector of the graph update matrix represents the second update vector of the m-th node; The graph update matrix is ​​input into the classifier, and the classification space of the classifier represents the fault types of the wind turbine gearbox.

7. A wind power generation gearbox fault detection method according to claim 6, characterized in that: The calculation formula for the jth hidden unit of the i-th first hidden layer includes: in represents the first update vector output by the jth hidden unit of the i-th first hidden layer, Represents the element value of the i-th row and j-th column of the time-frequency matrix input to the j-th hidden unit of the i-th first hidden layer, and denote the reset gate, update gate and intermediate vector of the jth hidden unit of the i-th first hidden layer, respectively. and denote respectively the first weight parameter, the second weight parameter, the third weight parameter, the fourth weight parameter, the fifth weight parameter and the sixth weight parameter of the j-th hidden unit of the i-th first hidden layer, and denote the first bias parameter, the second bias parameter and the third bias parameter of the jth hidden unit of the i-th first hidden layer, respectively. and denote the first update vector and the number of dimensions of the first update vector output by the jth hidden unit of the i-1th first hidden layer, respectively, and denote the dimensions of the first update vector and the first update vector output by the j-1th hidden unit of the i-th first hidden layer, respectively. and All are assigned 0, MLP represents multi-layer perceptron, ⊙ represents point-by-point multiplication, sigmoid represents sigmoid activation function, and tanh represents hyperbolic tangent activation function.

8. A wind power generation gearbox fault detection method according to claim 6, characterized in that: The calculation formula for the second hidden layer includes: Where 1≤u≤M, 1≤v≤M, u≠v, S represents the graph update matrix of the second hidden layer output, h u and They represent the first update vector and the second update vector of the u-th node, respectively, and h v represents the first update vector of the vth node, Col u Represents the set of nodes that have edges connected to the u-th node, Nattion uv and F uv Represent the attention score and the number of edge connections between the u-th node and the v-th node, respectively. and b uv Respectively represent the first weight parameter, the second weight parameter, the third weight parameter and the bias parameter between the uth node and the vth node, W u Indicates the weight parameter corresponding to the u-th node, pile M It indicates that the second update vectors of M nodes are stacked, concat indicates a concatenation operation, T indicates a transposition operation, sigmoid indicates a sigmoid activation function, and softmax indicates a softmax activation function.

9. A wind power generation gearbox fault detection method according to claim 1, characterized in that: A simulation model of a wind turbine gearbox is constructed through a simulation platform, and one or more fault types are randomly generated on the simulation model as sample labels of a training sample, and training data of a training sample is constructed according to steps S101 to S103 until U training samples are obtained, where U is a custom parameter.

10. A storage medium, characterized in that: It stores non-transitory computer-readable instructions. When the non-transitory computer-readable instructions are executed by a computer, the steps in the wind power generation gearbox fault detection method as claimed in any one of claims 1 to 9 are executed.

Images