Wind turbine fault early warning method and device based on dynamic weight and network flow entropy

Abstract

The application discloses a wind turbine fault early warning method and device based on dynamic weight and network flow entropy, and relates to the field of data processing, which comprises the following steps: according to the variable data matrix of each window, using a trained hidden Markov model to divide the working condition of each sampling time in the window, counting the number of working condition changes in each window and calculating the dynamic weight of each window; according to the variable data matrix of each window, using the maximum information coefficient and sparse inverse covariance estimation to establish the correlation network of each window, and using the network flow entropy method to calculate the network flow entropy value of each variable in each window, then using the dynamic weight of each window to correct the network flow entropy value of each variable in the corresponding window, and calculating the average value of the corrected network flow entropy values of all variables in each window; early warning according to the average value of the corrected network flow entropy values of all variables in each window, solving the problem of false alarm under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of data processing, specifically to a method and device for early warning of wind turbine faults based on dynamic weights and network flow entropy. Background Technology

[0002] Large wind turbines often operate in harsh environments and under complex conditions, resulting in a high failure rate and consequently high operation and maintenance costs, accounting for approximately 25% of total investment and severely impacting production efficiency. To reduce costs and increase efficiency, wind power operation and maintenance is shifting from traditional reactive maintenance to predictive maintenance. Therefore, conducting research on efficient and accurate early fault warning for wind turbines is of great significance for ensuring reliable turbine operation and reducing operation and maintenance costs.

[0003] Currently, the diagnostic methods for wind turbines mainly include physical model-driven and data-driven methods. However, traditional physical model-driven methods are difficult to accurately describe the nonlinear and strongly coupled characteristics of wind turbines. Due to the complex operating conditions and harsh and variable operating environment of wind turbines, and the interference of strong background noise on their monitoring data, the degradation process of each turbine has individual heterogeneous characteristics. The robustness of data-driven methods depends on their training dataset, and is limited by the quantity and quality of the training dataset, resulting in significant limitations in the robustness of data-driven methods.

[0004] Early warning methods based on complex network analysis only require SCADA data from a single wind turbine to model, eliminating the reliance on massive amounts of historical fault samples. However, under varying operating conditions, the algorithm's ability to accurately distinguish whether changes in network structure stem from faults or changes in operating conditions has become a core challenge restricting its engineering application. Summary of the Invention

[0005] In response to the aforementioned problems such as the difficulty in accurately extracting early weak fault characteristics of wind turbines, the excessive reliance on large amounts of high-quality data and insufficient generalization ability of data-driven algorithm-based early fault prediction, and the existence of false alarms under complex operating conditions, the purpose of this application is to propose a wind turbine fault early warning method and device based on dynamic weights and network flow entropy.

[0006] In a first aspect, the present invention provides a wind turbine fault early warning method based on dynamic weights and network flow entropy, comprising the following steps:

[0007] The variable data corresponding to all variables at N sampling times continuously monitored in the SCADA system of the wind turbine are obtained and preprocessed to obtain preprocessed variable data; the preprocessed variable data is then subjected to right-expanded window processing to obtain the variable data matrix for each window;

[0008] Construct and train a Hidden Markov Model (HMM) to obtain a trained HMM; based on the variable data matrix of each window, use the trained HMM to divide the working conditions at each sampling time within the window, count the number of working condition changes in each window, and calculate the dynamic weight of each window.

[0009] Based on the variable data matrix of each window, the association network of each window is established using the maximum information coefficient and sparse inverse covariance estimation. The network flow entropy value of each variable in each window is calculated using the network flow entropy method. Then, the network flow entropy value of each variable in the corresponding window is corrected using the dynamic weight of each window to obtain the corrected network flow entropy value of each variable in each window. Finally, the average value of the corrected network flow entropy values ​​of all variables in each window is calculated.

[0010] The early warning condition for a fault is determined based on the average of the corrected network flow entropy values ​​of all variables in each window. If the condition is met, an early warning is issued; otherwise, no warning is issued.

[0011] Preferably, during the training process of the Hidden Markov Model, a physical mapping relationship is established between each hidden state in the hidden state space of the Hidden Markov Model and the operating conditions: the first hidden state Corresponding to the start, the second hidden state Corresponding to maximum power point tracking, the third hidden state Corresponding to constant speed, the 4th hidden state Corresponding to constant power;

[0012] The following physical constraints are imposed during the iterative computation of the optimal parameters of the Hidden Markov Model:

[0013] The denormalized power generation for each hidden state's corresponding operating condition The following requirements must be met: ,and Approaching rated power;

[0014] If the operating condition is startup, then the rotational speed V≈0, the change in rotational speed ΔV>0, and the blade tilt angle β at the t-th sampling time. t maximum;

[0015] If the operating condition is maximum power point tracking, then the change in rotational speed ΔV > 0, and the difference between the blade tilt angle at the t-th sampling time and the blade tilt angle at the (t-1)-th sampling time (β) t -β t-1 ) < 0;

[0016] If the operating condition is constant speed, then the speed V≈V 额定 The change in rotational speed ΔV≈0, and the blade tilt angle β at the t-th sampling time.t ≈0; V 额定 Indicates the rated speed;

[0017] If the operating condition is constant power, then the speed V≈V 额定 The change in rotational speed ΔV≈0, and the blade tilt angle β at the t-th sampling time. t ≈0;

[0018] The Hidden Markov Model with optimal parameters is used as the trained Hidden Markov Model.

[0019] Preferably, based on the variable data matrix of each window, a trained Hidden Markov Model is used to divide the working conditions at each sampling time point within the window, count the number of working condition changes in each window, and calculate the dynamic weight of each window, specifically including:

[0020] The current window's variable data matrix contains the rotational speed v, blade tilt angle β, and power generation. As input variables to the Hidden Markov Model, the initial observation vector at the t-th sampling time within the current window is constructed. Where t=1,2,…,T', and T' represents the total number of sampling times within the current window. Represents the transpose of a matrix; These represent the rotational speed, blade tilt angle, and power generation at the t-th sampling time within the current window, respectively. These represent the changes in rotational speed and power generation at the t-th sampling time within the current window, respectively. The change in rotational speed at the t-th sampling time is the difference between the rotational speed at the t-th sampling time and the rotational speed at the (t-1)-th sampling time. The change in power generation at the t-th sampling time is the difference between the power generation at the t-th sampling time and the power generation at the (t-1)-th sampling time.

[0021] Normalize each element of the initial observation vector at each sampling time within the current window to obtain the final observation vector at each sampling time within the current window;

[0022] The final observation vector at each sampling time within the current window is input into the trained Hidden Markov Model. The hidden state at each sampling time within the current window is searched in the hidden state space. The working condition at each sampling time within the current window is determined based on the hidden state at each sampling time within the current window and the physical mapping relationship.

[0023] Count the number of times the working conditions change within the current window and calculate the dynamic weight of the current window. :

[0024] ;

[0025] in, The tolerance level of the current window is expressed as: , This is the tolerance coefficient. This represents the number of times the operating conditions within the current window have changed.

[0026] Preferably, the association network for each window is constructed using the maximum information coefficient and sparse inverse covariance estimation based on the variable data matrix of each window, specifically including:

[0027] The mutual information between any two variables in the variable data matrix of the current window is calculated using the grid partitioning method, as shown in the following formula:

[0028] ;

[0029] in, and These represent the first variable in the variable data matrix of the current window. The first variable and the second Variable data for each variable, , This represents the total number of variables in the variable data matrix of the current window, where lg represents the logarithm to the base 10. Indicates falling along the first The first variable and the second The joint probability within the grid partitioned by the axes of the variables. Indicates falling on the 1st The marginal probability of the row containing each variable Indicates falling on the 1st Marginal probabilities of each variable in its column; The first variable in the variable data matrix of the current window The first variable and the second Mutual information of the variables;

[0030] The maximum information coefficient between any two variables in the variable data matrix of the current window is calculated based on their mutual information, as shown in the following formula:

[0031] ;

[0032] in, The first variable in the variable data matrix of the current window The first variable and the second The maximum information coefficient of each variable max This indicates taking the maximum value. min This indicates taking the minimum value. and For along the first The first variable and the second The number of grid rows and columns for each variable along its axis. , Indicates the maximum number of grid cells threshold;

[0033] Construct the maximum information matrix based on the maximum information coefficient between any two variables in the variable data matrix of the current window, as shown in the following formula:

[0034] ;

[0035] If the maximum information coefficient between two variables in the variable data matrix of the current window is less than the threshold, it means that the correlation between the two variables in the variable data matrix of the current window is not significant, and therefore no edge is established between the two variables in the variable data matrix of the current window; otherwise, an edge is established between the two variables in the variable data matrix of the current window.

[0036] The sparse inverse covariance estimation is calculated for the variable data matrix of the current window. The specific process is as follows:

[0037] Map the variable data matrix of the current window from simplex space to Euclidean space, and perform a CLR transformation on each variable data in the variable data matrix of the current window:

[0038] ;

[0039] in, This represents the element in the a-th row and b-th column of the variable data matrix in the current window. This represents the transformed value corresponding to the element in the a-th row and b-th column of the variable data matrix in the current window; Indicates CLR transformation, This represents the geometric mean of all elements in the a-th row of the variable data matrix in the current window;

[0040] Construct the CLR transform vector for the current window at the a-th sampling time by using the transformed values ​​corresponding to all elements in the a-th row of the variable data matrix of the current window. And calculate the mean of the CLR transform vector at all sampling times in the current window. The transformed empirical covariance matrix is ​​calculated as follows:

[0041] ;

[0042] in, Represents the transformed empirical covariance matrix;

[0043] The transformed empirical covariance matrix is ​​input into the Graphical Lasso algorithm, and the following objective function is solved using block coordinate descent to obtain an accurate estimate of the matrix:

[0044] ;

[0045] in, Representing an exact matrix The estimated value, constraint It must be a positive definite matrix. express and The trace of the product. express The determinant, Indicates the penalty parameter. Represents the L1 norm;

[0046] After the above calculations, we finally obtain the estimated value of the exact matrix under the current window. The exact matrix estimate The elements in the table are used as the weights of the edges between all nodes that have edges established under the current window, thus establishing the connection network of the current window;

[0047] By repeating the above steps, the association network of each window can be obtained.

[0048] Preferably, the network flow entropy value of each variable in each window is calculated using the network flow entropy method for the association network of each window. Then, the network flow entropy value of each variable in the corresponding window is corrected using the dynamic weights of each window, resulting in the corrected network flow entropy value of each variable in each window. Specifically, this includes:

[0049] For each window's associated network, extract the local network. The local network is then divided into several parts. Variables As the central variable, its neighboring variables These are the remaining variables in the network that have edges with the central variable, excluding the central variable. ', d' represents the total number of neighboring variables. ;

[0050] For each local network within each window, the variable data are based on the central variable at the t-th sampling time. and neighbor variables' variable data joint probability Defined as:

[0051] ;

[0052] in, This represents the variable data based on the central variable at the t-th sampling time. and neighbor variables' variable data The Aitchison distance is calculated as follows:

[0053] ;

[0054] in, Representing the central variable The geometric mean of is expressed as: ; Representing neighbor variables The geometric mean of is expressed as: ;

[0055] The neighbor variable is calculated using the following formula. Relative to the central variable conditional probability :

[0056] ;

[0057] in, This represents the central variable in the association network of the current window. and neighbor variables The weight of the edge;

[0058] For the central variable in each local network, the network flow entropy value based on T' sampling times is calculated using the following formula:

[0059] ;

[0060] ;

[0061] in, The central variable at the t-th sampling time of the current window. The probability, The central variable representing the current window Network flow entropy value;

[0062] The following formula is used to apply to the central variable of the current window. The network flow entropy value is corrected:

[0063] ;

[0064] in, The central variable in the current window The corrected network flow entropy value, The dynamic weight of the current window;

[0065] By repeating the above steps, the corrected network flow entropy values ​​of each variable in each window are obtained.

[0066] Preferably, the early warning condition for faults is determined based on the average of the corrected network flow entropy values ​​of all variables in each window. If the condition is met, a fault is identified and an early warning is issued; otherwise, no warning is issued. Specifically, this includes:

[0067] Calculate the absolute difference between the average of the corrected network flow entropy values ​​of all variables in the current window and the average of the corrected network flow entropy values ​​of all variables in the previous window, and obtain the change in network flow entropy value of the current window ΔWNFE.

[0068] Starting from the current window, if three or more windows out of five consecutive windows have ΔWNFE values ​​that simultaneously satisfy the following two formulas, then a fault is determined to exist in the current window, and an early warning is issued:

[0069] ;

[0070] ;

[0071] in, This represents the change in network flow entropy value at the nth window. This represents the change in network flow entropy value at the (n+1)th window. For all windows The upper quartile, For all windows The lower quartile, The threshold parameter ΔWNFE is set based on the normal operation of the wind turbine. `max` indicates taking the maximum value. This indicates the change ratio to the threshold.

[0072] Secondly, the present invention provides a wind turbine fault early warning device based on dynamic weights and network flow entropy, characterized in that it includes:

[0073] The data processing module is configured to acquire variable data corresponding to all variables at N sampling times continuously monitored in the SCADA system of the wind turbine and preprocess them to obtain preprocessed variable data; and to perform right expansion window processing on the preprocessed variable data to obtain the variable data matrix for each window.

[0074] The weight calculation module is configured to build and train a Hidden Markov Model (HMM) to obtain a trained HMM; based on the variable data matrix of each window, the trained HMM is used to divide the working conditions at each sampling time within the window, count the number of working condition changes in each window, and calculate the dynamic weight of each window.

[0075] The correction module is configured to establish the association network of each window based on the variable data matrix of each window using the maximum information coefficient and sparse inverse covariance estimation; calculate the network flow entropy value of each variable of each window using the network flow entropy method; then correct the network flow entropy value of each variable of the corresponding window using the dynamic weight of each window to obtain the corrected network flow entropy value of each variable of each window; and calculate the average value of the corrected network flow entropy values ​​of all variables of each window.

[0076] The early warning module is configured to determine whether the early warning conditions for a fault are met based on the average of the corrected network flow entropy values ​​of all variables in each window. If so, an early warning is issued and a fault is identified; otherwise, no warning is issued.

[0077] Thirdly, the present invention provides an electronic device including one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any implementation of the first aspect.

[0078] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the implementations of the first aspect.

[0079] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method as described in any of the implementations in the first aspect.

[0080] Compared with the prior art, the present invention has the following beneficial effects:

[0081] (1) The wind turbine fault early warning method based on dynamic weight and network flow entropy mentioned in this invention uses the maximum information coefficient and sparse inverse covariance estimation to establish the correlation network of each window to capture the changes of the correlation network in the critical stage of the fault, without the need for complex early fault feature extraction.

[0082] (2) The wind turbine fault early warning method based on dynamic weight and network flow entropy mentioned in this invention utilizes the network flow entropy method based on dynamic weight. It is a model-free method and does not require a large amount of high-quality data for modeling. It only requires the historical operation data of a single wind turbine.

[0083] (3) The wind turbine fault early warning method based on dynamic weight and network flow entropy mentioned in this invention divides the operating conditions at each sampling time within the window, counts the number of operating condition changes in each window and calculates the dynamic weight of each window. The dynamic weight takes into account the operation of the wind turbine under multiple complex operating conditions, improves the accuracy of the early warning and reduces the false alarm problem caused by multiple operating condition jumps. Attached Figure Description

[0084] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0085] Figure 1 This is a flowchart illustrating a wind turbine fault early warning method based on dynamic weights and network flow entropy, as an embodiment of this application.

[0086] Figure 2 The evolution and early warning diagram of ΔWNFE for bearing failure of wind turbine generator in embodiment of this application;

[0087] Figure 3 This is a schematic diagram of a wind turbine fault early warning device based on dynamic weights and network flow entropy, as an embodiment of this application.

[0088] Figure 4 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0089] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0090] Figure 1 This application illustrates an embodiment of a wind turbine fault early warning method based on dynamic weights and network flow entropy, comprising the following steps:

[0091] S1: Obtain variable data corresponding to all variables at N sampling times continuously monitored in the SCADA system of the wind turbine and preprocess them to obtain preprocessed variable data; perform right expansion window processing on the preprocessed variable data to obtain the variable data matrix of each window.

[0092] Specifically, the embodiments of this application first collect variable data of continuously monitored variables covered by the SCADA system of the wind turbine, that is, collect variable data corresponding to M variables at N sampling times of continuous monitoring. The variable data is preprocessed using methods such as standardization and missing value handling to obtain preprocessed variable data. The preprocessed variable data is then subjected to right-expanding window processing, setting the initial window length and the step size for right expansion to obtain the variable data matrix for each window. In one example, the step size is set to 1, the initial window length is L, which contains preprocessed variable data from sampling times 1 to L, the next window length is L+1, which contains preprocessed variable data from sampling times 1 to L+1, the nth window is taken as the current window, and the nth window length is T', which contains preprocessed variable data from sampling times 1 to T', where T' = L+n-1; until the last window has a length of N.

[0093] S2, construct and train a Hidden Markov Model to obtain a trained Hidden Markov Model; based on the variable data matrix of each window, use the trained Hidden Markov Model to divide the working conditions at each sampling time within the window, count the number of working condition changes in each window, and calculate the dynamic weight of each window.

[0094] In a specific embodiment, during the training process of the Hidden Markov Model, a physical mapping relationship is established between each hidden state in the hidden state space of the Hidden Markov Model and the operating conditions: the first hidden state Corresponding to the start, the second hidden state Corresponding to maximum power point tracking, the third hidden state Corresponding to constant speed, the 4th hidden state Corresponding to constant power;

[0095] The following physical constraints are imposed during the iterative computation of the optimal parameters of the Hidden Markov Model:

[0096] The denormalized power generation for each hidden state's corresponding operating condition The following requirements must be met: ,and Approaching rated power;

[0097] If the operating condition is startup, then the rotational speed V≈0, the change in rotational speed ΔV>0, and the blade tilt angle β at the t-th sampling time. t maximum;

[0098] If the operating condition is maximum power point tracking, then the change in rotational speed ΔV > 0, and the difference between the blade tilt angle at the t-th sampling time and the blade tilt angle at the (t-1)-th sampling time (β) t -β t-1 ) < 0;

[0099] If the operating condition is constant speed, then the speed V≈V 额定 The change in rotational speed ΔV≈0, and the blade tilt angle β at the t-th sampling time. t ≈0; V 额定 Indicates the rated speed;

[0100] If the operating condition is constant power, then the speed V≈V 额定 The change in rotational speed ΔV≈0, and the blade tilt angle β at the t-th sampling time. t ≈0;

[0101] The Hidden Markov Model with optimal parameters is used as the trained Hidden Markov Model.

[0102] In a specific embodiment, based on the variable data matrix of each window, a trained Hidden Markov Model is used to divide the working conditions at each sampling time point within the window, count the number of working condition changes in each window, and calculate the dynamic weight of each window. Specifically, this includes:

[0103] The current window's variable data matrix contains the rotational speed v, blade tilt angle β, and power generation. As input variables to the Hidden Markov Model, the initial observation vector at the t-th sampling time within the current window is constructed. Where t=1,2,…,T', and T' represents the total number of sampling times within the current window. Represents the transpose of a matrix; These represent the rotational speed, blade tilt angle, and power generation at the t-th sampling time within the current window, respectively. These represent the changes in rotational speed and power generation at the t-th sampling time within the current window, respectively. The change in rotational speed at the t-th sampling time is the difference between the rotational speed at the t-th sampling time and the rotational speed at the (t-1)-th sampling time. The change in power generation at the t-th sampling time is the difference between the power generation at the t-th sampling time and the power generation at the (t-1)-th sampling time.

[0104] Normalize each element of the initial observation vector at each sampling time within the current window to obtain the final observation vector at each sampling time within the current window;

[0105] The final observation vector at each sampling time within the current window is input into the trained Hidden Markov Model. The hidden state at each sampling time within the current window is searched in the hidden state space. The working condition at each sampling time within the current window is determined based on the hidden state at each sampling time within the current window and the physical mapping relationship.

[0106] Count the number of times the working conditions change within the current window and calculate the dynamic weight of the current window. :

[0107] ;

[0108] in, The tolerance level of the current window is expressed as: , This is the tolerance coefficient. This represents the number of times the operating conditions within the current window have changed.

[0109] Specifically, in the embodiments of this application, the rotational speed v, blade tilt angle β, and power generation are used as variables in the data matrix of each window. As input variables for a hidden Markov model.

[0110] First-order feature calculation is performed on the input variables of rotational speed and power generation:

[0111] ;

[0112] ;

[0113] in, The rotational speed at the t-th sampling time of the current window; The rotational speed at the (t-1)th sampling time of the current window. This represents the change in rotational speed at the t-th sampling time in the current window; The power generation at the t-th sampling time of the current window; Let be the power generation at the t-th sampling time of the current window. This represents the change in power generation at the t-th sampling time of the current window.

[0114] Therefore, the initial observation vector at the t-th sampling time within the current window is constructed. Then, the elements of the initial observation vector at the t-th sampling time within the current window are normalized to obtain the final observation vector. .

[0115] Since the embodiments of this application only have four operating conditions, the hidden state space of the Hidden Markov Model also has only four states. Furthermore, a physical mapping relationship is established between each hidden state in the hidden state space and the operating condition. Under the condition of satisfying physical constraints, the Hidden Markov Model is trained unsupervised to determine the optimal parameters of the Hidden Markov Model. This yields a trained Hidden Markov Model.

[0116] Subsequently, based on the trained Hidden Markov Model, the operating conditions at each sampling time point within each window are further divided. This involves defining the variable data matrix for each window, including rotational speed v, blade tilt angle β, and power generation. The corresponding final observation vector is input into the trained Hidden Markov Model, which outputs the hidden state of each window at each sampling time, and determines the working condition of each window at each sampling time by combining the physical mapping relationship.

[0117] Following the above process, after dividing each window into different operating conditions, the number of times the operating conditions change is counted, and the dynamic weight of each window is calculated. In one example, the tolerance coefficient is set to 0.1.

[0118] S3. Based on the variable data matrix of each window, establish the association network of each window using the maximum information coefficient and sparse inverse covariance estimation. Apply the network flow entropy method to the association network of each window to calculate the network flow entropy value of each variable in each window. Then, use the dynamic weight of each window to correct the network flow entropy value of each variable in the corresponding window to obtain the corrected network flow entropy value of each variable in each window. Finally, calculate the average value of the corrected network flow entropy values ​​of all variables in each window.

[0119] In a specific embodiment, an association network for each window is established based on the variable data matrix of each window using the maximum information coefficient and sparse inverse covariance estimation, specifically including:

[0120] The mutual information between any two variables in the variable data matrix of the current window is calculated using the grid partitioning method, as shown in the following formula:

[0121] ;

[0122] in, and These represent the first variable in the variable data matrix of the current window. The first variable and the second Variable data for each variable, , This represents the total number of variables in the variable data matrix of the current window, where lg represents the logarithm to the base 10. Indicates falling along the first The first variable and the second The joint probability within the grid partitioned by the axes of the variables. Indicates falling on the 1st The marginal probability of the row containing each variable Indicates falling on the 1st Marginal probabilities of each variable in its column; The first variable in the variable data matrix of the current window The first variable and the second Mutual information of the variables;

[0123] The maximum information coefficient between any two variables in the variable data matrix of the current window is calculated based on their mutual information, as shown in the following formula:

[0124] ;

[0125] in, The first variable in the variable data matrix of the current window The first variable and the second The maximum information coefficient of each variable max This indicates taking the maximum value. min This indicates taking the minimum value. and For along the first The first variable and the second The number of grid rows and columns for each variable along its axis. , Indicates the maximum number of grid cells threshold;

[0126] Construct the maximum information matrix based on the maximum information coefficient between any two variables in the variable data matrix of the current window, as shown in the following formula:

[0127] ;

[0128] If the maximum information coefficient between two variables in the variable data matrix of the current window is less than the threshold, it means that the correlation between the two variables in the variable data matrix of the current window is not significant, and therefore no edge is established between the two variables in the variable data matrix of the current window; otherwise, an edge is established between the two variables in the variable data matrix of the current window.

[0129] The sparse inverse covariance estimation is calculated for the variable data matrix of the current window. The specific process is as follows:

[0130] Map the variable data matrix of the current window from simplex space to Euclidean space, and perform a CLR transformation on each variable data in the variable data matrix of the current window:

[0131] ;

[0132] in, This represents the element in the a-th row and b-th column of the variable data matrix in the current window. This represents the transformed value corresponding to the element in the a-th row and b-th column of the variable data matrix in the current window; Indicates CLR transformation, This represents the geometric mean of all elements in the a-th row of the variable data matrix in the current window;

[0133] Construct the CLR transform vector for the current window at the a-th sampling time by using the transformed values ​​corresponding to all elements in the a-th row of the variable data matrix of the current window. And calculate the mean of the CLR transform vector at all sampling times in the current window. The transformed empirical covariance matrix is ​​calculated as follows:

[0134] ;

[0135] in, Represents the transformed empirical covariance matrix;

[0136] The transformed empirical covariance matrix is ​​input into the Graphical Lasso algorithm, and the following objective function is solved using block coordinate descent to obtain an accurate estimate of the matrix:

[0137] ;

[0138] in, Representing an exact matrix The estimated value, constraint It must be a positive definite matrix. express and The trace of the product. express The determinant, Indicates the penalty parameter. Represents the L1 norm;

[0139] After the above calculations, we finally obtain the estimated value of the exact matrix under the current window. The exact matrix estimate The elements in the table are used as the weights of the edges between all nodes that have edges established under the current window, thus establishing the connection network of the current window;

[0140] By repeating the above steps, the association network of each window can be obtained.

[0141] Specifically, embodiments of this application utilize the Maximum Information Coefficient (MIC) and Sparse Inverse Covariance Estimation (SPIEC) to establish a correlation network for each window's variable data matrix. The specific process is as follows: first, the mutual information between any two variables in the current window is calculated using a grid partitioning method; then, the maximum information coefficient (MIC) between any two variables in the current window is calculated using the mutual information, and a maximum information coefficient matrix is ​​constructed. In one example, if the MIC values ​​of two variables are less than 0.2, the two variables are considered to have insignificant correlation, meaning no edge is established between them.

[0142] After filtering out insignificant edges using MIC, SPIEC calculation is performed on the variable data matrix of the current window: first, the variable data matrix of the current window is mapped from simplex space to Euclidean space; then, a CLR transformation is performed on each row of the variable data matrix of the current window; finally, the transformed empirical covariance matrix is ​​calculated; after calculating the transformed empirical covariance matrix S, the exact matrix is ​​then calculated. The estimation of the exact matrix is ​​as follows. In the embodiments of this application, SPEICE utilizes Graphical Lasso optimization to solve for the exact matrix. The input to Graphical Lasso is the transformed empirical covariance matrix S obtained above, and the solution is obtained by minimizing the objective function. This is a convex optimization problem, and in the embodiments of this application, Block Coordinate Descent is used for fast solution. The penalty parameter in the objective function... The larger the network, the sparser the network. The result is the standard inverse covariance matrix.

[0143] After the above calculations, we finally obtain the estimated value of the exact matrix under the current window. The elements in the estimated value of the precise matrix serve as the weights of the edges between nodes in the associated network that are not filtered by the MIC threshold under the current window.

[0144] By repeating the above calculations, the associated networks under all windows can be obtained.

[0145] In a specific embodiment, the network flow entropy value of each variable in each window is calculated using the network flow entropy method for the association network of each window. Then, the network flow entropy value of each variable in the corresponding window is corrected using the dynamic weights of each window, resulting in the corrected network flow entropy value of each variable in each window. Specifically, this includes:

[0146] For each window's associated network, extract the local network. The local network is then divided into several parts. Variables As the central variable, its neighboring variables These are the remaining variables in the network that have edges with the central variable, excluding the central variable. ', d' represents the total number of neighboring variables. ;

[0147] For each local network within each window, the variable data are based on the central variable at the t-th sampling time. and neighbor variables' variable data joint probability Defined as:

[0148] ;

[0149] in, This represents the variable data based on the central variable at the t-th sampling time. and neighbor variables' variable data The Aitchison distance is calculated as follows:

[0150] ;

[0151] in, Representing the central variable The geometric mean of is expressed as: ; Representing neighbor variables The geometric mean of is expressed as: ;

[0152] The neighbor variable is calculated using the following formula. Relative to the central variable conditional probability :

[0153] ;

[0154] in, This represents the central variable in the association network of the current window. and neighbor variables The weight of the edge;

[0155] For the central variable in each local network, the network flow entropy value based on T' sampling times is calculated using the following formula:

[0156] ;

[0157] ;

[0158] in, The central variable at the t-th sampling time of the current window. The probability, The central variable representing the current window Network flow entropy value;

[0159] Use the following formula to apply the variables in the current window The network flow entropy value is corrected:

[0160] ;

[0161] in, Variables in the current window The corrected network flow entropy value, The dynamic weight of the current window;

[0162] By repeating the above steps, the corrected network flow entropy values ​​of each variable in each window are obtained.

[0163] Specifically, in the embodiments of this application, the Network Flow Entropy (NFE) method is applied to each associated network to calculate the network flow entropy value of each variable in the current window. Then, the network flow entropy value of each variable is corrected using the dynamic weight of each window to obtain the corrected network flow entropy value of each variable in the current window. Furthermore, the average of the corrected network flow entropy values ​​of all variables in each window is taken to calculate the average of the corrected network flow entropy values ​​of all variables in each window.

[0164] S4. Determine whether the early warning condition for faults is met based on the average value of the corrected network flow entropy of all variables in each window. If so, determine that a fault exists and issue an early warning; otherwise, do not issue a warning.

[0165] In a specific embodiment, the method determines whether the early fault warning condition is met based on the average of the corrected network flow entropy values ​​of all variables in each window. If so, a fault is identified and an early warning is issued; otherwise, no warning is issued. Specifically, this includes:

[0166] Calculate the absolute difference between the average of the corrected network flow entropy values ​​of all variables in the current window and the average of the corrected network flow entropy values ​​of all variables in the previous window, and obtain the change in network flow entropy value of the current window ΔWNFE.

[0167] Starting from the current window, if three or more windows out of five consecutive windows have ΔWNFE values ​​that simultaneously satisfy the following two formulas, then a fault is determined to exist in the current window, and an early warning is issued:

[0168] ;

[0169] ;

[0170] in, This represents the change in network flow entropy value at the nth window. This represents the change in network flow entropy value at the (n+1)th window. For all windows The upper quartile, For all windows The lower quartile, The threshold parameter ΔWNFE is set based on the normal operation of the wind turbine. `max` indicates taking the maximum value. This indicates the change ratio to the threshold.

[0171] Specifically, in the embodiments of the present application, the absolute difference ΔWNFE between the average value of the corrected network flow entropy of all variables in the current window and the average value of the corrected network flow entropy of all variables in the previous window is calculated first, as shown in the following formula:

[0172] ;

[0173] Starting from the current window, among the ΔWNFE of 5 consecutive windows, when the ΔWNFE of 3 or more windows satisfy the following two early warning formulas at the same time, it indicates that the current window meets the early fault early warning conditions, so an early warning can be issued; otherwise, no early warning is issued.

[0174] Specific embodiments are used below to describe the effect of the present invention.

[0175] A verification experiment is carried out on the early warning method proposed in the embodiments of the present application, taking No. T07 wind turbine of a certain wind farm as the object. The unit has a horizontal-axis three-blade structure, with a rated power of 2MW, a rated wind speed of 12m / s, and the cut-in wind speed and cut-out wind speed are 4m / s and 25m / s respectively. The diameter of the wind rotor is 90m, and its maximum rotational speed can reach 14.9r / min. The gearbox adopts a three-stage planetary gear structure to drive an asynchronous generator, whose maximum rotational speed is 2016r / min, rated voltage is 690V, and the connected grid frequency is 50Hz. The unit tower is a steel pipe structure, and the hub height is 80m. Data acquisition is completed by the SCADA system equipped with the wind turbine, and the sampling period is 10min. A typical case where the generator bearing of T07 unit was damaged at 06:08:00 on August 20, 2017 and the generator was damaged at 14:47:00 on August 21, 2017 is used for analysis and verification. Continuous sampling moments including two faults, that is, continuous SCADA data from 5:40:00 on August 18, 2017 to 17:00:00 on August 21, 2017, are selected to construct verification samples.

[0176] Figure 2 is shown as a trend image of the ΔWNFE value changing with time at continuous sampling moments when the wind turbine has a generator bearing fault. The ΔWNFE value first has 3 or more windows exceeding the early warning threshold in the 235th-239th time windows, which satisfies the early warning criterion, indicating that the network structure of the wind turbine has changed from a stable state to an unstable critical state.

[0177] Checking the fault log of the wind turbine, the generator bearing damage fault occurred at the 245th time window, and the generator fault occurred at the 440th time window which caused shutdown. Comparing with the actual fault occurrence time, the early warning method proposed in the embodiments of the present application issues an early warning about 1 hour earlier than the actual generator bearing damage time, and issues an early warning about 33.5 hours earlier than the actual generator damage time.

[0178] Maximum Information Coefficient (MIC) can effectively quantify complex nonlinear correlations, but it cannot distinguish indirect relationships and is prone to over-dense networks due to spurious correlations. Sparse Inverse Covariance Estimation (SPIEC), on the other hand, can eliminate spurious correlations caused by multivariate coupling through regularization penalties, generating sparse networks, but its ability to capture nonlinearities is weak. To adapt to the high-noise, nonlinear environment of wind turbine sites, this application proposes constructing a correlation network based on MIC-SPIEC: first, MIC is used to initially screen parameter edges to retain nonlinear characteristics; then, SPIEC is used to calculate edge weights and eliminate spurious correlations. This results in a final network with both excellent nonlinear representation and noise resistance. After constructing a baseline network based on health SCADA data, when an early fault occurs, even if a single variable does not trigger an alarm threshold, changes in the local network topology can serve as a sensitive early warning signal. Compared to traditional Pearson correlation coefficient networks, it has a stronger ability to capture early anomalies.

[0179] Wind power operation data is a highly non-stationary time series. Under external wind conditions and control system intervention, the turbine frequently switches between five operating conditions: startup, MPPT (Maximum Power Point Tracking), constant speed, constant power, and shutdown. These normal operating condition switches, like early faults, can cause network topology changes. Failure to accurately distinguish between the two will lead to frequent false alarms. Since the MIC-SPIEC-based interconnected network is constructed based on time windows, the more frequent the operating condition switches within a window, the greater the network fluctuations. To address this problem, the embodiments of this application construct dynamic weights by statistically analyzing the number of operating condition switches within a single window, which are then used to correct network flow entropy calculations, thereby eliminating interference from normal operating condition changes on early warning systems.

[0180] To further illustrate the superiority of this method, the early warning method mentioned in the embodiments of this application is compared with the time network flow entropy method based on PCC and the wind turbine fault early warning method based on the BI-LSTM model. The results are shown in Table 1. Table 1 shows that the early warning method mentioned in the embodiments of this application achieves different degrees of earlier warning time compared to the other two methods.

[0181] Table 1. Comparison of different methods:

[0182] Warning time window 239th time window 258th time window 255th time window

[0183] Further reference Figure 3 As an implementation of the methods shown in the above figures, this application provides an embodiment of a wind turbine fault early warning device based on dynamic weights and network flow entropy. This device embodiment is similar to... Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0184] This application provides a wind turbine fault early warning device based on dynamic weights and network flow entropy, including:

[0185] Data processing module 1 is configured to acquire variable data corresponding to all variables at N sampling times continuously monitored in the SCADA system of the wind turbine and preprocess them to obtain preprocessed variable data; and to perform right expansion window processing on the preprocessed variable data to obtain the variable data matrix of each window.

[0186] The weight calculation module 2 is configured to construct and train a hidden Markov model to obtain a trained hidden Markov model; based on the variable data matrix of each window, the trained hidden Markov model is used to divide the working conditions at each sampling time within the window, count the number of working condition changes in each window, and calculate the dynamic weight of each window.

[0187] The correction module 3 is configured to establish the association network of each window based on the variable data matrix of each window using the maximum information coefficient and sparse inverse covariance estimation; calculate the network flow entropy value of each variable of each window using the network flow entropy method; then correct the network flow entropy value of each variable of the corresponding window using the dynamic weight of each window to obtain the corrected network flow entropy value of each variable of each window; and calculate the average value of the corrected network flow entropy values ​​of all variables of each window.

[0188] The early warning module 4 is configured to determine whether the early warning conditions for faults are met based on the average value of the corrected network flow entropy of all variables in each window. If so, a fault is identified and an early warning is issued; otherwise, no warning is issued.

[0189] Figure 4 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. For example... Figure 4 As shown, the electronic device in this embodiment includes a processor 401 and a memory 402; wherein the memory 402 is used to store computer execution instructions; and the processor 401 is used to execute the computer execution instructions stored in the memory to implement the various steps performed by the electronic device in the above embodiment. For details, please refer to the relevant descriptions in the foregoing method embodiments.

[0190] Alternatively, the memory 402 can be either standalone or integrated with the processor 401.

[0191] When the memory 402 is set up independently, the electronic device also includes a bus 403 for connecting the memory 402 and the processor 401.

[0192] This invention also provides a computer storage medium storing computer execution instructions, which, when executed by processor 401, implement the above method.

[0193] This invention also provides a computer program product, including a computer program, which, when executed by a processor 401, implements the above-described method.

[0194] In the embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0195] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0196] Furthermore, the functional modules in the various embodiments of this invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit formed by the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0197] The integrated modules implemented as software functional modules described above can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor 401 to execute some steps of the methods of the various embodiments of this application.

[0198] It should be understood that the processor 401 described above can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor, or the processor 401 can be any conventional processor 401. The steps of the method disclosed in this invention can be directly manifested as the hardware processor 401 executing the steps, or as a combination of hardware and software modules within the processor 401 executing the steps.

[0199] The memory 402 may include high-speed RAM memory, and may also include non-volatile memory NVM, such as at least one disk storage device, and may also be a USB flash drive, portable hard drive, read-only memory, disk or optical disc, etc.

[0200] Bus 403 can be an Industry Standard Architecture (ISA), a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Bus 403 can be divided into address bus, data bus, control bus, etc. For ease of illustration, the bus 403 in the accompanying drawings of this application is not limited to only one bus 403 or one type of bus 403.

[0201] The aforementioned storage medium can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium accessible to general-purpose or special-purpose computers.

[0202] An exemplary storage medium is coupled to processor 401, enabling processor 401 to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of processor 401. Processor 401 and storage medium can reside in application-specific integrated circuits (ASICs). Alternatively, processor 401 and storage medium can exist as discrete components in an electronic device or host device.

[0203] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0204] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for early warning of wind turbine faults based on dynamic weights and network flow entropy, characterized in that, Includes the following steps: The variable data corresponding to all variables at N sampling times continuously monitored in the SCADA system of the wind turbine are obtained and preprocessed to obtain preprocessed variable data; the preprocessed variable data is then subjected to right-expanded window processing to obtain the variable data matrix for each window; Construct and train a Hidden Markov Model to obtain a trained Hidden Markov Model; based on the variable data matrix of each window, use the trained Hidden Markov Model to divide the working conditions at each sampling time within the window, count the number of working condition changes in each window, and calculate the dynamic weight of each window. Based on the variable data matrix of each window, the association network of each window is established using the maximum information coefficient and sparse inverse covariance estimation. The network flow entropy value of each variable in each window is calculated using the network flow entropy method. Then, the network flow entropy value of each variable in the corresponding window is corrected using the dynamic weight of each window to obtain the corrected network flow entropy value of each variable in each window. Finally, the average value of the corrected network flow entropy values ​​of all variables in each window is calculated. The early warning condition for a fault is determined based on the average of the corrected network flow entropy values ​​of all variables in each window. If the condition is met, an early warning is issued; otherwise, no warning is issued.

2. The wind turbine fault early warning method based on dynamic weights and network flow entropy according to claim 1, characterized in that, During the training of the Hidden Markov Model (HMM), a physical mapping relationship is established between each hidden state in the hidden state space of the HMM and the operating conditions: the first hidden state... Corresponding to the start, the second hidden state Corresponding to maximum power point tracking, the third hidden state Corresponding to constant speed, the 4th hidden state Corresponding to constant power; The following physical constraints are imposed during the iterative computation of the optimal parameters of the Hidden Markov Model: The denormalized power generation for each hidden state's corresponding operating condition The following requirements must be met: ,and Approaching rated power; If the operating condition is startup, then the rotational speed V≈0, the change in rotational speed ΔV>0, and the blade tilt angle β at the t-th sampling time. t maximum; If the operating condition is maximum power point tracking, then the change in rotational speed ΔV > 0, and the difference between the blade tilt angle at the t-th sampling time and the blade tilt angle at the (t-1)-th sampling time (β) t -β t-1 ) < 0; If the operating condition is constant speed, then the speed V≈V 额定 The change in rotational speed ΔV≈0, and the blade tilt angle β at the t-th sampling time. t ≈0; V 额定 Indicates the rated speed; If the operating condition is constant power, then the speed V≈V 额定 The change in rotational speed ΔV≈0, and the blade tilt angle β at the t-th sampling time. t ≈0; The Hidden Markov Model with optimal parameters is used as the trained Hidden Markov Model.

3. The wind turbine fault early warning method based on dynamic weights and network flow entropy according to claim 2, characterized in that, Based on the variable data matrix of each window, the trained Hidden Markov Model is used to divide the working conditions at each sampling time point within the window, count the number of working condition changes in each window, and calculate the dynamic weight of each window, specifically including: The current window's variable data matrix contains the rotational speed v, blade tilt angle β, and power generation. As input variables to the Hidden Markov Model, the initial observation vector at the t-th sampling time within the current window is constructed. Where t=1,2,…,T', and T' represents the total number of sampling times within the current window. Represents the transpose of a matrix; These represent the rotational speed, blade tilt angle, and power generation at the t-th sampling time within the current window, respectively. These represent the changes in rotational speed and power generation at the t-th sampling time within the current window, respectively. The change in rotational speed at the t-th sampling time is the difference between the rotational speed at the t-th sampling time and the rotational speed at the (t-1)-th sampling time. The change in power generation at the t-th sampling time is the difference between the power generation at the t-th sampling time and the power generation at the (t-1)-th sampling time. Normalize each element of the initial observation vector at each sampling time within the current window to obtain the final observation vector at each sampling time within the current window; The final observation vector at each sampling time within the current window is input into the trained Hidden Markov Model. The hidden state at each sampling time within the current window is searched in the hidden state space. The working condition at each sampling time within the current window is determined based on the hidden state at each sampling time within the current window and the physical mapping relationship. Count the number of times the working conditions change within the current window and calculate the dynamic weight of the current window. : ； in, The tolerance level of the current window is expressed as: , This is the tolerance coefficient. This represents the number of times the operating conditions within the current window have changed.

4. The wind turbine fault early warning method based on dynamic weights and network flow entropy according to claim 1, characterized in that, Based on the variable data matrix of each window, an association network for each window is constructed using the maximum information coefficient and sparse inverse covariance estimation, specifically including: The mutual information between any two variables in the variable data matrix of the current window is calculated using the grid partitioning method, as shown in the following formula: ； in, and These represent the first variable in the variable data matrix of the current window. The first variable and the second Variable data for each variable, , This represents the total number of variables in the variable data matrix of the current window, where lg represents the logarithm to the base 10. Indicates falling along the first The first variable and the second The joint probability within the grid partitioned by the axes of the variables. Indicates falling on the 1st The marginal probability of the row containing each variable Indicates falling on the 1st Marginal probabilities of each variable in its column; The first variable in the variable data matrix of the current window The first variable and the second Mutual information of the variables; The maximum information coefficient between any two variables in the variable data matrix of the current window is calculated based on their mutual information, as shown in the following formula: ； in, The first variable in the variable data matrix of the current window The first variable and the second The maximum information coefficient of each variable max This indicates taking the maximum value. min This indicates taking the minimum value. and For along the first The first variable and the second The number of grid rows and columns for each variable along its axis. , Indicates the maximum number of grid cells threshold; Construct the maximum information matrix based on the maximum information coefficient between any two variables in the variable data matrix of the current window, as shown in the following formula: ； If the maximum information coefficient between two variables in the variable data matrix of the current window is less than the threshold, it means that the correlation between the two variables in the variable data matrix of the current window is not significant, and therefore no edge is established between the two variables in the variable data matrix of the current window; otherwise, an edge is established between the two variables in the variable data matrix of the current window. The sparse inverse covariance estimation is calculated for the variable data matrix of the current window. The specific process is as follows: Map the variable data matrix of the current window from simplex space to Euclidean space, and perform a CLR transformation on each variable data in the variable data matrix of the current window: ； in, This represents the element in the a-th row and b-th column of the variable data matrix in the current window. This represents the transformed value corresponding to the element in the a-th row and b-th column of the variable data matrix in the current window; Indicates CLR transformation, This represents the geometric mean of all elements in the a-th row of the variable data matrix in the current window; Construct the CLR transform vector for the current window at the a-th sampling time by using the transformed values ​​corresponding to all elements in the a-th row of the variable data matrix of the current window. And calculate the mean of the CLR transform vector at all sampling times in the current window. The transformed empirical covariance matrix is ​​calculated as follows: ； in, Represents the transformed empirical covariance matrix; The transformed empirical covariance matrix is ​​input into the Graphical Lasso algorithm, and the following objective function is solved using the block coordinate descent method to obtain an estimate of the exact matrix: ； in, Representing an exact matrix The estimated value, constraint It must be a positive definite matrix. express and The trace of the product. express The determinant, Indicates the penalty parameter. Represents the L1 norm; After the above calculations, we finally obtain the estimated value of the exact matrix under the current window. The exact matrix estimate The elements in the table are used as the weights of the edges between all nodes that have edges established under the current window, thus establishing the connection network of the current window; By repeating the above steps, the association network of each window can be obtained.

5. The wind turbine fault early warning method based on dynamic weights and network flow entropy according to claim 1, characterized in that, For the network of each window, the network flow entropy method is used to calculate the network flow entropy value of each variable in each window. Then, the dynamic weights of each window are used to correct the network flow entropy value of each variable in the corresponding window, resulting in the corrected network flow entropy value of each variable in each window. Specifically, this includes: For each window's associated network, extract the local network. The local network is then divided into several parts. Variables As the central variable, its neighboring variables These are the remaining variables in the network that have edges with the central variable, excluding the central variable. ', d' represents the total number of neighboring variables, ; For each local network within each window, the variable data are based on the central variable at the t-th sampling time. and neighbor variables' variable data joint probability Defined as: ； in, This represents the variable data based on the central variable at the t-th sampling time. and neighbor variables' variable data The Aitchison distance is calculated as follows: ； in, Representing the central variable The geometric mean of is expressed as: ; Representing neighbor variables The geometric mean of is expressed as: ; The neighbor variable is calculated using the following formula. Relative to the central variable conditional probability : ； in, This represents the central variable in the association network of the current window. and neighbor variables The weight of the edge; For the central variable in each local network, the network flow entropy value based on T' sampling times is calculated using the following formula: ； ； in, The central variable at the t-th sampling time of the current window. The probability, The central variable representing the current window Network flow entropy value; The following formula is used to apply to the central variable of the current window. The network flow entropy value is corrected: ； in, The central variable in the current window The corrected network flow entropy value, The dynamic weight of the current window; By repeating the above steps, the corrected network flow entropy values ​​of each variable in each window are obtained.

6. The wind turbine fault early warning method based on dynamic weights and network flow entropy according to claim 1, characterized in that, The system determines whether the early fault warning conditions are met based on the average of the corrected network flow entropy values ​​of all variables in each window. If so, a fault is identified and an early warning is issued; otherwise, no warning is issued. Specifically, this includes: Calculate the absolute difference between the average of the corrected network flow entropy values ​​of all variables in the current window and the average of the corrected network flow entropy values ​​of all variables in the previous window, and obtain the change in network flow entropy value of the current window ΔWNFE. Starting from the current window, if three or more windows out of five consecutive windows have ΔWNFE values ​​that simultaneously satisfy the following two formulas, then a fault is determined to exist in the current window, and an early warning is issued: ； ； in, This represents the change in network flow entropy value at the nth window. This represents the change in network flow entropy value at the (n+1)th window. For all windows The upper quartile, For all windows The lower quartile, The threshold parameter ΔWNFE is set based on the normal operation of the wind turbine. `max` indicates taking the maximum value. This indicates the change ratio to the threshold.

7. A wind turbine fault early warning device based on dynamic weights and network flow entropy, characterized in that, include: The data processing module is configured to acquire variable data corresponding to all variables at N sampling times continuously monitored in the SCADA system of the wind turbine and preprocess them to obtain preprocessed variable data; and to perform right expansion window processing on the preprocessed variable data to obtain a variable data matrix for each window. The weight calculation module is configured to construct and train a hidden Markov model to obtain a trained hidden Markov model; based on the variable data matrix of each window, the trained hidden Markov model is used to divide the working conditions at each sampling time within the window, count the number of working condition changes in each window, and calculate the dynamic weight of each window. The correction module is configured to build the association network for each window based on the variable data matrix of each window using the maximum information coefficient and sparse inverse covariance estimation. The network flow entropy value of each variable in each window is calculated using the network flow entropy method. Then, the network flow entropy value of each variable in the corresponding window is corrected using the dynamic weight of each window to obtain the corrected network flow entropy value of each variable in each window. Finally, the average value of the corrected network flow entropy values ​​of all variables in each window is calculated. The early warning module is configured to determine whether the early warning conditions for a fault are met based on the average of the corrected network flow entropy values ​​of all variables in each window. If so, an early warning is issued and a fault is identified; otherwise, no warning is issued.

8. An electronic device, comprising: One or more processors; A storage device for storing one or more programs, characterized in that, when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

Images