Wind turbine generator transmission chain fault monitoring method based on similarity measurement and trend sensing network

Abstract

The invention provides a wind turbine generator transmission chain fault monitoring method based on similarity measurement and a trend perception network, which comprises the following steps: acquiring data, fusing a trend perception algorithm and an attention mechanism, and embedding the data into a long short-term memory network to form a trend perception BiLSTM model with network units arranged bidirectionally; fusing sequence similarity measurement and spatial distribution similarity measurement, screening out highly similar units according to the data, establishing a single unit prediction model for each unit in the highly similar units based on the trend perception BiLSTM model, and forming an integrated prediction model; and obtaining a prediction residual error of a monitoring variable through the integrated prediction model, obtaining a monitoring index through KL divergence, and achieving fault monitoring. According to the method, the problem of insufficient accuracy of the existing method can be solved, and the accuracy and reliability of fault monitoring of the transmission chain of the wind turbine generator are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of wind turbine fault monitoring, and more particularly, to a method for monitoring faults in the drive train of a wind turbine based on similarity measurement and trend perception network. Background Art

[0002] A wind turbine is a complex electro-mechanical hybrid system designed to convert wind energy into electrical energy. Since it usually operates in areas with extremely harsh natural environments, the unit is very prone to faults. Frequent faults not only reduce the operating reliability but also increase the operation and maintenance costs. Therefore, it is of great significance to study fault monitoring methods that can adapt to complex conditions and accurately monitor the abnormal operating state of wind turbines.

[0003] According to different analysis methods used, the monitoring methods can be divided into three categories: (1) knowledge-based, (2) signal processing-based, and (3) data-driven; knowledge-based methods rely on expert knowledge or experience, which are often subjective and have limited accessibility, so it is difficult to popularize; signal processing-based methods usually have reliable monitoring results, but different monitoring signals require different signal analysis techniques, which have high technical requirements for users; data-driven methods estimate the values of state variables through learning and inference of a large amount of data, and then judge whether the operating state of the unit components or subsystems is normal by comparing the actual values with the estimated values. Such methods have greater flexibility and practicality; the present invention belongs to one of the data-driven methods.

[0004] Currently, the data-driven methods have the following deficiencies: (1) When constructing the state estimation model, the time correlation of variables is ignored. The current states of many variables (such as temperature, pressure, and rotational speed) are usually highly correlated with the previous states and evolve over time; (2) Most methods only study single units, and due to the limitation of poor data completeness, the generalization ability of the model is insufficient; (3) The existing similarity measurement methods for wind turbines only rely on statistical features to determine the similarity between operating conditions, ignoring the time characteristics of state variables. In addition, it is difficult to reflect the overall operating conditions and states of the units by comparing only a specific variable to evaluate the similarity of the operating conditions of different wind turbines. The above three problems lead to poor reliability and low accuracy of the current wind turbine drive train fault monitoring methods.

[0005] In view of this, the present invention is specifically proposed. Summary of the Invention

[0006] In view of this, the present invention proposes a fault monitoring method for the drive train of a wind turbine based on similarity measurement and trend perception network, which can solve the problem of insufficient accuracy of existing methods and effectively improve the accuracy and reliability of fault monitoring for the drive train of a wind turbine.

[0007] Specifically, the present invention is realized through the following technical solutions:

[0008] The present invention provides a fault monitoring method for the drive train of a wind turbine based on similarity measurement and trend perception network, including the following steps:

[0009] Obtain data, fuse the trend perception algorithm and the attention mechanism, and embed them into the long short-term memory network to form a trend perception BiLSTM model with a two-way layout of network units;

[0010] Fuse the sequence similarity measurement and the spatial distribution similarity measurement, screen out highly similar units according to the data, and establish a single-unit prediction model for each unit in the highly similar units based on the trend perception BiLSTM model to form an integrated prediction model;

[0011] Obtain the prediction residual of the monitoring variable through the integrated prediction model, and obtain the monitoring index through the KL divergence to realize fault monitoring.

[0012] In addition to providing a fault monitoring method for the drive train of a wind turbine based on similarity measurement and trend perception network, the present invention also provides a fault monitoring system, including:

[0013] Perception model establishment module: used to obtain data, fuse the trend perception algorithm and the attention mechanism, and embed them into the long short-term memory network to form a trend perception BiLSTM model with a two-way layout of network units;

[0014] Prediction module: used to fuse the sequence similarity measurement and the spatial distribution similarity measurement, screen out highly similar units according to the data, and establish a single-unit prediction model for each unit in the highly similar units based on the trend perception BiLSTM model to form an integrated prediction model;

[0015] Monitoring module: used to obtain the prediction residual of the monitoring variable through the integrated prediction model, and obtain the monitoring index through the KL divergence to realize fault monitoring.

[0016] In summary, the solution of the present invention has the following beneficial effects:

[0017] (1) The present invention designs an interval quantile method for removing outliers in SCADA raw data, overcoming the problem that existing data preprocessing methods are difficult to adapt to the frequently fluctuating operating conditions of wind turbines and cannot effectively remove outliers, and outputs a normal data set that can be used for subsequent unit similarity analysis and normal behavior modeling.

[0018] (2) By combining sequence similarity measurement and multi-dimensional space distribution similarity measurement, the present invention can comprehensively compare the similarity between units from multiple perspectives such as operating conditions and unit states, overcoming the problems of existing methods with few parameter variables and difficulty in measuring the similarity of non-linear high-dimensional data, and can effectively find similar wind turbines.

[0019] (3) By combining the trend perception algorithm with the attention mechanism, the problem that conventional machine learning models cannot effectively extract trend features in time series data is solved, effectively enhancing the hidden feature extraction ability of the model; the BiLSTM network is used to effectively mine the spatio-temporal features in multi-variable input data, improving the prediction accuracy of the model.

[0020] (4) Compared with the traditional single-unit model, the present invention adopts a model integration framework to combine the models of multiple similar units into an overall integrated prediction model, further improving the prediction accuracy and stability; finally, the KL divergence is used to transform the residual sequence to construct a new fault monitoring index, effectively improving the reliability of the monitoring index. Description of the Drawings

[0021] By reading the following detailed description of the preferred embodiments, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the present invention. Throughout the drawings, the same reference numerals are used to represent the same components. In the drawings:

[0022] Figure 1 is the flowchart of the wind turbine drive train fault monitoring method of the present invention;

[0023] Figure 2 is the schematic diagram of the structure of the trend perception module of the present invention;

[0024] Figure 3 is the schematic diagram of the method of the present invention;

[0025] Figure 4 is the schematic diagram of the structure of the MSBiLSTM network model of the present invention;

[0026] Figure 5 is the actual case demonstration of the method of the present invention;

[0027] Figure 6A flowchart of a computer device provided by an embodiment of the present invention. Detailed implementation manners

[0028] Here, exemplary embodiments will be described in detail, and examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the present disclosure. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0029] The terms used in the present disclosure are only for the purpose of describing specific embodiments and are not intended to limit the present disclosure. The singular forms "a", "the", and "said" used in the present disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0030] It should be understood that although the terms first, second, third, etc. may be used in the present disclosure to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present disclosure, the first information may also be referred to as the second information. Similarly, the second information may also be referred to as the first information, depending on the context. As used herein, the word "if" may be interpreted as "when" or "while" or "in response to determining".

[0031] Embodiment

[0032] As Figure 1 shown, a fault monitoring method for a wind turbine drive train based on similarity measurement and trend perception network is provided, including the following steps:

[0033] Obtain data, fuse the trend perception algorithm and the attention mechanism, and embed them into a long short-term memory network to form a trend perception BiLSTM model with a two-way layout of network units;

[0034] Fuse sequence similarity measurement and spatial distribution similarity measurement, screen out highly similar units according to the data, and establish a single-unit prediction model for each unit in the highly similar units based on the trend perception BiLSTM model to form an integrated prediction model;

[0035] Obtain the prediction residuals of the monitoring variables through the integrated prediction model, and obtain the monitoring indicators through the KL divergence to achieve fault monitoring.

[0036] In the above method, first, the original data set is collected, and the interval quantile method is used to remove abnormal data. The data set X after outlier cleaning is normalized by the Max-min method to make the dimensions of each parameter consistent. Then, the structure of the trend-aware deep network model is designed as the basis for establishing the normal behavior model. The trend-aware algorithm is combined with the attention mechanism to form a trend-aware module, and the trend-aware module is embedded in the BiLSTM unit. The trend-aware BiLSTM units are connected to form an MSBiLSTM network model. At the same time, the similarity between the target wind turbine and other turbines in the wind farm is calculated to find turbines with highly similar operating characteristics. The similarity weights corresponding to each parameter are determined, the data sequence similarity between the target turbine and other turbines is calculated, and then the multi-dimensional space distribution similarity of the data between the target turbine and other turbines is calculated. The sequence similarity and the space distribution similarity are fused to calculate the total similarity. Then, for the selected similar turbines, based on the MSBiLSTM network model structure, a single-turbine normal behavior model is established, and further, multiple single-turbine models are integrated into a prediction model. Finally, based on the integrated model, the target parameter variable is predicted to obtain a residual sequence, and the KL divergence is calculated as an index for fault monitoring. The 3-sigma algorithm is used to calculate the threshold of the fault monitoring index, and whether there is an abnormality is judged according to whether the monitoring index exceeds the threshold.

[0037] A fault monitoring method for a wind turbine drive train based on similarity measurement and trend-aware network combines time series similarity measurement and multi-dimensional data space distribution similarity measurement, and proposes a similarity measurement method considering sequence similarity and space distribution similarity for screening similar wind turbines. In addition, the trend-aware mechanism is combined with the long short-term memory neural network for normal behavior modeling of the wind turbine drive train.

[0038] The steps of the fault monitoring method for a wind turbine drive train based on similarity measurement and trend-aware network during use are as follows:

[0039] Step (1): Through the Supervisory Control And Data Acquisition (SCADA) system of the wind farm, the environmental parameters, status parameters, and operating parameters of each wind turbine are collected, and the original data set X=(x 1 ,x 2 ,…,x k); k represents the number of parameter variables; due to various factors such as environmental interference, sensor anomalies, and communication failures, there are a small number of abnormal data in the original dataset; the present invention uses the interval quantile method to remove abnormal data. First, the dataset is divided into multiple intervals according to wind speed, and then for each parameter variable within a single interval, the quantile method is used to remove outliers; the calculation process is as follows:

[0040] Arrange the parameter variable data in descending order of value, and evenly divide it into four equal parts according to the total number of samples. The boundary positions are denoted as Q 1 、Q 2 and Q 3 , where Q 1 represents the lower 25% quantile value; Q 2 represents the median value, that is, the value at the 50% position; Q 3 represents the upper 25% quantile value; the interquartile range IQR represents the degree of concentration of the numerical distribution, and the calculation is as follows:

[0041] IQR = Q 3 -Q 1 (1)

[0042] Further calculate the upper bound B h and the lower bound B l :

[0043]

[0044] where α is the normal interval coefficient. Samples with values greater than B h or less than B l are identified as abnormal data and removed;

[0045] Normalize the dataset X after outlier cleaning through the Max - min method, and normalize the parameter variable x i to the range [0, 1] to make the dimensions of each parameter consistent, obtaining the normalized dataset X'; the calculation process is:

[0046]

[0047] where, e represents a sample in the parameter variable x i ; e' represents the normalized sample value; min(x i ) represents the minimum value of the parameter variable x i ; max(x i ) represents the maximum value of the parameter variable x i .

[0048] Step (2): Design the structure of the trend - aware deep network model as the basis for establishing a normal behavior model.

[0049] 2.1) Combine the trend perception algorithm with the attention mechanism to form a trend perception module; this module mainly includes a dynamic trend perception part, an attention part, and a fusion part. Among them, dynamic perception is used to extract local spatio-temporal relationships; attention is used to assign different weights to the extracted local spatio-temporal features; the fusion part is used to fuse local and global information to obtain the hidden information most relevant to the current local. The operation process of the trend perception module can be expressed by the following formula:

[0050]

[0051] q = Norm(S' next ) (5)

[0052]

[0053] Among them, S prev and S next are the previous and next local spatio-temporal hidden states respectively; S' next is the predicted value of S next ; Norm is the normalization function; q is the value after normalization; Softmax is the activation function; is the output of BiLSTM; are the learnable parameters of this module; is the memory information output by the previous network unit; is the feature vector at the current time point; D h is the hidden state dimension.

[0054] The specific structure of the trend perception module is as Figure 2 shown. Take the memory information provided by the previous network unit as the input of the Norm&Drop layer, and the output information of this layer enters the Softmax layer for feature extraction to obtain the hidden layer features, and then input them into the memory gating unit; at the same time, the trend variable of the current data is also sent to the memory gating unit to update the memorized trend information.

[0055] 2.2) Embed the trend perception module into the bidirectional long short-term memory neural network unit (BiLSTM); use the trend perception module to replace the linear layer in BiLSTM to achieve efficient extraction of spatio-temporal information. The processing process of the network unit is described as follows:

[0056]

[0057] Among them, represents the output variable of the first-layer trend perception network; σ represents the activation function; represents the input data at the current time; is the hidden state at the previous time step; is the hidden state at the current time; The memory information output by the previous network unit; Represents the learning parameters of the first layer; Represents the bias related to trend perception in the first layer;

[0058]

[0059] Among them, Represents the output variable of the second-layer trend perception network; Represents the learning parameters of the second layer; Represents the bias related to trend perception in the second layer; [·‖·] represents the parallel connection operation; other parameters are the same as those in formula (7);

[0060]

[0061] Tanh is the hyperbolic tangent activation function; Is the high-order trend perception parameter; Is the high-order basis vector.

[0062] The structure of the trend perception LSTM module is specifically as Figure 3 Shown. First, the memory information of the previous unit is passed into the LSTM unit structure for memory and discard processing, and the output effective information enters the normalization layer and the weighted part respectively; the normalized information is sent into the attention mechanism to assign weights according to the importance of the feature information; the current data is input into the attention mechanism and the trend perception module respectively for learning new information and trends; finally, the output information of the LSTM unit, the attention mechanism, and the trend perception module is fused through the weighted part and output.

[0063] 2.3) Connect the trend perception BiLSTM units to form an MSBiLSTM deep network model; each network unit processes information for k time steps, which not only reduces the number of iterations but also fully considers the local spatio-temporal information. The function of the model can be expressed by the following formula:

[0064]

[0065] Among them, Represents the i-th local spatio-temporal hidden state; Represents the model learning parameters; Represents the (i - 1)-th local spatio-temporal hidden state; Represents the i-th global state in the l-th layer; Represents the input signal of this module; Represents the spatio-temporal embedding vector; using the extracted local state Update the global information MSBiLSTM is the function of the deep learning model.

[0066] The structural schematic diagram of the specific model is as shown in Figure 4 Figure [3]. This network mainly consists of three layers. The upper layer connects the trend-aware LSTM units in the forward order, and the information output by each unit enters the next unit and the weighting part respectively; the middle layer consists of multiple weighting parts arranged in the forward order. Each weighting part receives the information from the forward unit and the backward unit, and the output information is transmitted to the next weighting part; the lower layer connects the trend-aware LSTM units in the reverse order. Similarly, the information output by each unit enters the next unit and the weighting part respectively; the currently input data enters the start of the forward layer and the start of the reverse layer respectively. Thus, the network model can extract the forward features and reverse features of the data simultaneously, effectively improving the prediction accuracy of the model.

[0067] Step (3): Calculate the similarity between the target wind turbine and other turbines in the wind farm, and find the turbines with highly similar operating characteristics.

[0068] 3.1) Calculate the mutual information MI between the target parameter and the candidate parameter i , and find the important parameter variables. Calculate the similarity weight corresponding to the parameter through formula (13);

[0069]

[0070] where k represents the number of input parameters; w i represents the weight corresponding to parameter i.

[0071] 3.2) Calculate the data sequence similarity between the target turbine and the candidate turbines through formula (14); First, construct the distance matrix Then calculate the cumulative distance matrix L:

[0072]

[0073] L(s, t) represents the cumulative distance at state s and time point t; s and t respectively represent the current index positions of the two time series; dis(x is , x jt ) represents the distance between the sth element of time series 1 and the tth element of time series 2; min represents the operation of taking the minimum value;

[0074] The sequence similarity between two variables is defined as D DTW (x i , x j ) = L(m, n), where m represents the variable of the target turbine and n represents the variable of the candidate turbine.

[0075] 3.3) Calculate the similarity of the multi-dimensional space distribution between the target unit and the candidate units; X c and Y c represent the centers of the data sets of the target unit and the candidate units respectively, and are calculated through formula (15).

[0076]

[0077] where x and y represent the average values of the parameter variable x and the parameter variable y respectively.

[0078] Based on the centers of the above-mentioned data sets, calculate the covariance matrix ε:

[0079]

[0080] The calculation of the multi-dimensional space distribution similarity is as follows:

[0081]

[0082] ε -1 represents the inverse covariance matrix;

[0083] Finally, fuse the sequence similarity and the space distribution similarity to calculate the total similarity S:

[0084]

[0085] w i represents the weight corresponding to the parameter variable i; D DTW(i) represents the sequence similarity of the parameter variable i; D MD(i) represents the space distribution similarity of the parameter variable i.

[0086] Step (4): For the selected similar units, based on the above-mentioned deep network model structure, train the corresponding single-unit normal behavior model with the data of each unit, and then based on the model integration framework, integrate multiple single-unit models into one model.

[0087] 4.1) Based on the above-mentioned integrated model, predict the target parameter variable, and the difference between the predicted value and the actual value can obtain the residual sequence.

[0088] 4.2) Using the residual sequence in the normal stage as a benchmark, calculate the KL divergence between the real-time residual and the benchmark residual as an index for fault monitoring.

[0089] 4.3) Based on the above-mentioned fault monitoring index, use the 3-sigma algorithm to calculate the threshold of the fault monitoring index; during the monitoring process, if the monitoring index exceeds the threshold, it is judged that there is an abnormality. If the monitoring index does not exceed the threshold, it is judged to be normal.

[0090] To verify the effectiveness of the fault monitoring method of the present invention, a specific practical case is provided as follows Figure 5 As shown, SCADA data from a wind farm in Inner Mongolia is used. In the case, a fault of too high gearbox oil temperature occurred in the wind turbine; at the position of 2000 points in the figure, the SCADA system detected an abnormality, and then the unit stopped running. Figure 5 (a) is the prediction of the gearbox oil temperature by the method of the present invention, and then the prediction residual is obtained; the horizontal line in the figure is the alarm threshold calculated by the 3-sigma method; Figure 5 (b) is the monitoring index calculated by KL divergence. The horizontal line is the alarm threshold, and the vertical line is the time when the fault is detected, which is about 16 hours earlier than the alarm time of the SCADA system. It can be seen from the figure that the KL divergence index can significantly reflect the abnormal state of the unit, indicating that the fault monitoring method of the present invention can indeed effectively improve the fault monitoring effect.

[0091] In addition to providing a fault monitoring method for the drive train of a wind turbine based on similarity measurement and trend perception network, the present invention also provides a fault monitoring system, including:

[0092] Perception model establishment module: used to obtain data, fuse the trend perception algorithm and the attention mechanism, and embed them into the long short-term memory network to form a trend perception BiLSTM model with a two-way layout of network units;

[0093] Prediction module: used to fuse sequence similarity measurement and spatial distribution similarity measurement, screen out highly similar units according to the data, and establish a single-unit prediction model for each unit in the highly similar units based on the trend perception BiLSTM model to form an integrated prediction model;

[0094] Monitoring module: used to obtain the prediction residual of the monitoring variable through the integrated prediction model, and obtain the monitoring index through KL divergence to achieve fault monitoring.

[0095] In specific implementation, the above-mentioned modules can be implemented as independent entities, or can be combined arbitrarily to be implemented as the same or several entities. For the specific implementation of the above-mentioned units, reference can be made to the method embodiments described above, which will not be elaborated here.

[0096] Figure 6 It is a schematic structural diagram of a computer device disclosed by the present invention. Refer to Figure 6 As shown, the computer device 400 includes at least a memory 402 and a processor 401; the memory 402 is connected to the processor through a communication bus 403, and is used to store computer instructions executable by the processor 401. The processor 401 is used to read computer instructions from the memory 402 to implement the steps of the method described in any of the above embodiments.

[0097] For the above device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial descriptions of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present disclosure solution. A person of ordinary skill in the art can understand and implement it without creative work.

[0098] Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media, and memory devices, such as semiconductor memory devices (e.g., EPROM, EEPROM, and flash memory devices), magnetic disks (e.g., internal disks or removable disks), magneto-optical disks, and CD-ROM and DVD-ROM disks. The processor and the memory can be supplemented by or incorporated into dedicated logic circuitry.

[0099] Finally, it should be noted that: Although this specification contains many specific implementation details, these should not be construed as limiting the scope of any invention or the scope of what is claimed, but are mainly used to describe the features of specific embodiments of a particular invention. Certain features described in multiple embodiments in this specification can also be combined and implemented in a single embodiment. On the other hand, the various features described in a single embodiment can also be separately implemented in multiple embodiments or implemented in any suitable sub-combination. In addition, although the features may function in certain combinations as described above and are even initially claimed as such, one or more features from the claimed combination can in some cases be removed from the combination, and the claimed combination can be directed to a sub-combination or a variant of the sub-combination.

[0100] Similarly, although the operations are depicted in a specific order in the drawings, this should not be construed as requiring the operations to be performed in the specific order shown or sequentially, or requiring all of the illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. In addition, the separation of the various system modules and components in the above embodiments should not be construed as required in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0101] Accordingly, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the acts recited in the claims can be performed in a different order and still achieve the desired result. In addition, the processes depicted in the figures are not necessarily in the particular order or sequential order shown to achieve the desired result. In some implementations, multitasking and parallel processing may be advantageous.

[0102] The foregoing is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present disclosure shall be included within the scope of protection of the present disclosure.

Claims

1. A wind turbine transmission chain fault monitoring method based on similarity measurement and trend perception network, characterized in that: The steps include: Acquire data, integrate trend perception algorithm and attention mechanism, and embed it into long short-term memory network to form trend perception BiLSTM model with bidirectional arrangement of network units; By integrating the sequence similarity metric and the spatial distribution similarity metric, highly similar units are screened out according to the data, and a single unit prediction model is established for each of the highly similar units based on the trend-aware BiLSTM model to form an integrated prediction model; The prediction residual of the monitoring variable is obtained through the integrated prediction model, and the monitoring index is obtained through the KL divergence to realize fault monitoring.

2. The wind turbine transmission chain fault monitoring method according to claim 1, characterized in that: The acquired data were pre-processed to remove outliers using the interval quantile method and then normalized using the Max-min method.

3. The wind turbine transmission chain fault monitoring method according to claim 2, characterized in that: The acquired data is preliminarily subjected to the interval quantile method to remove outliers, and the method of normalizing by the Max-min method includes the following steps: The environmental parameters, state parameters and operating parameters of each wind turbine are collected to obtain the original data set X = (x1, x2, ..., x k ); k represents the number of parameter variables; Arrange the parameter variable data in descending order according to the value, and divide them into four equal parts according to the total number of samples. The dividing positions are recorded as Q1, Q2 and Q3, where Q1 represents the quantile value of the lower 25%; Q2 represents the median value, that is, the value at the 50% position; Q3 represents the quantile value of the upper 25%; the interquartile range IQR represents the concentration of the value distribution, which is calculated as follows: IQR=Q3-Q1 (1) Further calculate the numerical upper bound B h And the numerical lower bound B l : Among them, α is the normal interval coefficient, and its value is greater than B h or less than B l The samples are identified as abnormal data and removed; The parameter variable x i Normalize to the range of [0,1] to make the dimensions of each parameter consistent and obtain the normalized data set X'; the calculation process is: Among them, e represents the parameter variable x i A sample in ; e' represents the normalized sample value; min(x i ) represents the parameter variable x i The minimum value of max(x i ) represents the parameter variable x i The maximum value of .

4. The method for monitoring wind turbine transmission chain fault according to any one of claims 1 to 2, characterized in that: The trend perception algorithm and attention mechanism are integrated to form the trend perception module, and its operation process is as follows: q=Norm(S' next ) (5) Among them, S prev and S next are the previous and next local spatiotemporal hidden states respectively; S' next YesS next The predicted value; Norm is the normalization function; q is the value after normalization; Softmax is the activation function; is the output of BiLSTM; is the learnable parameter of this module; Memory information output by the previous network unit; is the feature vector at the current time point; D h is the hidden state dimension.

5. The wind turbine transmission chain fault monitoring method according to claim 4, characterized in that: The trend perception module is embedded in the long short-term memory neural network unit, and the processing process is described as follows: in, represents the output variable of the first layer trend perception network; σ represents the activation function; Represents the current time input data; is the hidden state at the previous time step; is the hidden state at the current time; Memory information output by the previous network unit; Represents the first layer learning parameters; represents the first-level trend-aware correlation bias; in, represents the output variable of the second layer trend-aware network; represents the second layer learning parameters; represents the trend-aware correlation bias of the second layer; [·‖·] represents the parallel connection operation; the other parameters are consistent with formula (7); Tanh is the hyperbolic tangent activation function; is a high-order trend perception parameter; are high-order basis vectors.

6. The wind turbine transmission chain fault monitoring method according to claim 5, characterized in that: The function of the trend-aware BiLSTM model with bidirectional arrangement of the network units is expressed as follows: in, represents the i-th local spatiotemporal hidden state; Represents model learning parameters; represents the i-1th local spatiotemporal hidden state; It represents the i-th global state in the l-th layer; Represents the input signal of this module; Represents a spatiotemporal embedding vector; using the extracted local state Update global information MSBiLSTM is a function of the model.

7. The method for monitoring wind turbine transmission chain fault according to any one of claims 1 to 2, characterized in that: The method of screening out highly similar units includes the following steps: Calculate the mutual information MI between the target parameter and the candidate parameter i , find the important parameter variables, and calculate the similarity weights corresponding to the parameters through formula (13); Where k represents the number of input parameters; w i Represents the weight corresponding to parameter i; The similarity of the data sequences of the target unit and the candidate unit is calculated by formula (14). First, the distance matrix is ​​constructed. Then calculate the cumulative distance matrix L: L(s,t) represents the cumulative distance between state s and time point t; s and t represent the current index positions of the two time series respectively; dis(x is ,x jt ) represents the distance between the sth element of time series 1 and the tth element of time series 2; min represents the operation of taking the minimum value; The sequence similarity between two variables is defined as D DTW (x i ,x j )=L(m,n), where m represents the variable of the target unit and n represents the variable of the candidate unit; Calculate the multi-dimensional spatial distribution similarity between the target unit and the candidate unit data; X c and Y c Representing the center of the data set of the target unit and the candidate unit respectively, it is calculated by formula (15): in and Represent the average values ​​of parameter variables x and y respectively; Based on the center of the data set, calculate the covariance matrix ε: The multidimensional space distribution similarity is calculated as follows: ε -1 represents the inverse covariance matrix; Finally, the sequence similarity and spatial distribution similarity are combined to calculate the total similarity S: w i represents the weight corresponding to parameter variable i; D DTW(i) represents the sequence similarity of parameter variable i; D MD(i) Indicates the spatial distribution similarity of parameter variable i.

8. A fault monitoring system using the wind turbine transmission chain fault monitoring method according to any one of claims 1 to 7, characterized in that: include: Perception model building module: used to acquire data, integrate trend perception algorithm and attention mechanism, and embed it into the long short-term memory network to form a trend perception BiLSTM model with bidirectional arrangement of network units; Prediction module: used for fusing sequence similarity measurement and spatial distribution similarity measurement, screening out highly similar units according to the data, establishing a single unit prediction model for each of the highly similar units based on the trend-aware BiLSTM model, and forming an integrated prediction model; Monitoring module: used to obtain the prediction residual of the monitoring variable through the integrated prediction model, obtain the monitoring index through KL divergence, and realize fault monitoring.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method for monitoring a wind turbine transmission chain fault according to any one of claims 1 to 7 are executed.

10. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the wind turbine transmission chain fault monitoring method according to any one of claims 1 to 7 are implemented.

Images