Intelligent fault diagnosis method for wind turbine gearbox, electronic device and medium

Abstract

This invention provides an intelligent fault diagnosis method, electronic device, and medium for wind turbine gearboxes. The method includes: acquiring real-time operational data of the wind turbine gearbox's state variables, including environmental parameters, grid parameters, and component state parameters; the real-time operational data includes operational data within a preset time period; repeatedly inputting the real-time operational data of the state variables into a pre-trained fault diagnosis model to obtain multiple initial fault diagnosis results, wherein the fault diagnosis model is constructed based on a Bayesian graph convolutional network, and the initial fault diagnosis results include the probability of each fault type among multiple fault types in the wind turbine gearbox; and fusing the multiple initial fault diagnosis results using cooperative game theory to obtain a final fault diagnosis result. This method can quickly, efficiently, and accurately diagnose faults in wind turbine gearboxes.

Description

Technical Field

[0001] This invention relates to the field of wind turbine fault diagnosis technology, and more specifically to an intelligent fault diagnosis method, electronic equipment and medium for wind turbine gearboxes. Background Technology

[0002] As one of the most economical clean energy sources, wind power plays an increasingly important foundational role in promoting the green and low-carbon transformation of the energy system, addressing climate change, and ensuring energy supply security. The gearbox, a key component of large wind turbines, is typically installed in the nacelle, which is tens or even hundreds of meters high. Due to long-term operation, variable load conditions, and external environmental factors, it is prone to problems such as gear fatigue, tooth surface wear, gear fracture, and bearing failure. If these faults are not detected and addressed in a timely manner, they can cause irreversible internal failures in the gearbox, leading to wind turbine shutdowns. This not only increases the cost of turbine operation and maintenance but also results in substantial economic losses from turbine downtime.

[0003] In related technologies, vibration signals are typically used as the signal source to measure faults in wind turbine gearboxes. While this traditional fault diagnosis method can identify faults to some extent, its diagnostic accuracy and robustness are often limited because vibration signals from wind turbine gearboxes are often affected by noise interference under actual operating conditions, especially complex conditions, making it difficult to extract and identify fault characteristics. Therefore, a solution that can accurately diagnose wind turbine gearbox faults is currently lacking.

[0004] In view of this, the present invention is hereby proposed. Summary of the Invention

[0005] The present invention is proposed in view of the above-mentioned problems. According to one aspect of the present invention, an intelligent fault diagnosis method for wind turbine gearboxes is provided, comprising:

[0006] The real-time operating data of the state variables of the wind turbine gearbox are obtained. The state variables include environmental parameters, power grid parameters and component state parameters of the wind turbine gearbox. The real-time operating data includes operating data within a preset time period.

[0007] The real-time running data of the state variables are repeatedly input into the pre-trained fault diagnosis model to obtain multiple initial fault diagnosis results. The fault diagnosis model is constructed based on a Bayesian graph convolutional network, and the initial fault diagnosis results include the probability of each fault type among multiple fault types in the wind turbine gearbox.

[0008] The multiple initial fault diagnosis results are fused using cooperative game theory to obtain the final fault diagnosis result.

[0009] For example, before repeatedly inputting the real-time running data of the state variables into the pre-trained fault diagnosis model, the method further includes:

[0010] Based on the real-time operational data of the state variables, the state variables are filtered to determine the state variables with higher importance.

[0011] The step of repeatedly inputting the real-time running data of the state variables into the pre-trained fault diagnosis model includes: repeatedly inputting the real-time running data of the state variables with higher importance into the pre-trained fault diagnosis model.

[0012] For example, the filtering of the state variables based on real-time operational data includes:

[0013] Using LightGBM Gini Exponential and / or Spearman correlation analysis is used to determine the importance of each state variable in order to filter the state variables.

[0014] For example, the one utilizing LightGBM Gini Exponential and / or Spearman correlation analysis determines the importance of each state variable in order to screen the state variables, including:

[0015] Based on LightGBM and the real-time operational data of the aforementioned state variables, the relative status variables of each state variable to the various fault types are determined. Gini Index, wherein Gini The index is positively correlated with the importance of the corresponding state variable;

[0016] choose Gini The first preset number of state variables with larger exponents are considered as higher-importance state variables;

[0017] or,

[0018] The use of LightGBM Gini Exponential and / or Spearman correlation analysis determines the importance of each state variable in order to screen the state variables, including:

[0019] Based on the real-time running data of the state variables, calculate the Spearman correlation between each state variable and other state variables.

[0020] For any of the state variables, if the Spearman correlation between the state variable and at least some of the other state variables is greater than or equal to the correlation threshold, the state variable is determined to be a highly important state variable.

[0021] For example, the one utilizing LightGBM Gini Exponential and / or Spearman correlation analysis determines the importance of each state variable in order to screen the state variables, including:

[0022] Based on LightGBM and the real-time operational data of the aforementioned state variables, the relative status variables of each state variable to the various fault types are determined. Gini Index, wherein Gini The index is positively correlated with the importance of the corresponding state variable;

[0023] choose Gini The first set of state variables consists of a pre-defined number of state variables with a larger exponent.

[0024] Based on the real-time running data of the state variables, calculate the Spearman correlation between each state variable and other state variables.

[0025] For any state variable among the state variables, if the Spearman correlation between the state variable and at least some of the other state variables is greater than or equal to the correlation threshold, the state variable is added to the second set of state variables.

[0026] The intersection of the first set of state variables and the second set of state variables is taken as the state variable with higher importance;

[0027] or,

[0028] The use of LightGBM Gini Exponential and / or Spearman correlation analysis determines the importance of each state variable in order to screen the state variables, including:

[0029] Based on the real-time running data of the state variables, calculate the Spearman correlation between each state variable and other state variables.

[0030] For any of the state variables, if the Spearman correlation between the state variable and at least some of the other state variables is greater than or equal to the correlation threshold, the state variable is added to the third set of state variables.

[0031] Based on LightGBM and the real-time operational data of each state variable in the third state variable set, the relative status of each state variable in the third state variable set to the various fault types is determined. Gini Index, wherein Gini The index is positively correlated with the importance of the corresponding state variable;

[0032] choose GiniThe first preset number of state variables with larger exponents are considered as higher-importance state variables;

[0033] or,

[0034] Based on LightGBM and the real-time operational data of the aforementioned state variables, the relative status variables of each state variable to the various fault types are determined. Gini Index, wherein Gini The index is positively correlated with the importance of the corresponding state variable;

[0035] choose Gini The first set of state variables with larger exponents constitutes the fourth set of state variables.

[0036] Based on the real-time running data of each state variable in the fourth set of state variables, calculate the Spearman correlation between each state variable in the fourth set of state variables and other state variables in the fourth set of state variables.

[0037] For any state variable in the fourth set of state variables, if the Spearman correlation between the state variable and at least some of the other state variables in the fourth set of state variables is greater than or equal to the correlation threshold, the state variable is determined to be a highly important state variable.

[0038] For example, before repeatedly inputting the real-time operational data of the highly important state variables into the pre-trained fault diagnosis model, the method further includes:

[0039] Normalize the real-time running data of the aforementioned highly important state variables.

[0040] For example, the step of fusing the multiple initial fault diagnosis results using cooperative game theory to obtain the final fault diagnosis result includes:

[0041] For any one of the plurality of initial fault diagnosis results, calculate the consistency correlation coefficient between the initial fault diagnosis result and the remaining fault diagnosis results, wherein the remaining fault diagnosis results are the sum of the other initial fault diagnosis results among the plurality of initial fault diagnosis results excluding the initial fault diagnosis result;

[0042] The final fault diagnosis result is determined by summing the products of the multiple initial fault diagnosis results and their respective consistency correlation coefficients.

[0043] For example, before repeatedly inputting the real-time running data of the state variables into the pre-trained fault diagnosis model, the method further includes:

[0044] The real-time operating data of the state variables are processed using linear interpolation.

[0045] According to another aspect of the present invention, an electronic device is provided, including a processor and a memory, wherein the memory stores a computer program, and the processor is used to execute the computer program to implement the method as described above.

[0046] According to another aspect of the present invention, a computer-readable storage medium is provided, which stores a computer program / instructions that, when executed by a processor, implement the method described above.

[0047] In the above technical solution, combining Bayesian graph convolutional networks significantly enhances the ability to extract nonlinear features of wind turbine gearbox faults. By repeatedly inputting the data and fusing multiple initial fault diagnosis results through cooperative game theory, the impact of random noise can be reduced, avoiding random errors caused by random model parameters. This helps improve the robustness and reliability of the method, and increases the accuracy of prediction results. Furthermore, this solution eliminates the need for fault feature extraction and screening based on engineering experience, reducing human intervention and enabling a shift from an experience-driven, artificial feature paradigm to a data-driven representation learning paradigm. This not only improves the accuracy of the results but also increases the speed of fault diagnosis. In summary, this method can quickly, efficiently, and accurately diagnose faults in wind turbine gearboxes.

[0048] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0049] The above and other objects, features, and advantages of the present invention will become more apparent from the more detailed description of the embodiments of the invention in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same parts or steps.

[0050] Figure 1 A schematic flowchart of a wind turbine gearbox intelligent fault diagnosis method according to an embodiment of the present invention is shown.

[0051] Figure 2 A schematic block diagram of an electronic device according to an embodiment of the present invention is shown. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the present invention more apparent, exemplary embodiments according to the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely a part of the embodiments of the present invention, and not all of the embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein. Based on the embodiments of the present invention described herein, all other embodiments obtained by those skilled in the art without inventive effort should fall within the protection scope of the present invention.

[0053] Wind turbines are complex electromechanical systems, typically located in remote areas and exposed to harsh environments for extended periods, resulting in extremely difficult operating conditions and high maintenance costs. Due to these harsh conditions, the failure rate is relatively high. Gearboxes are often installed in confined spaces at high altitudes, and common gearbox failures include gear wear, broken teeth, and poor lubrication. Bearing failures such as pitting, insufficient lubrication, fatigue damage, and cracks may also occur. Therefore, researching intelligent diagnostic methods for typical wind turbine gearbox failures is of significant practical importance for improving the safe and stable operation of wind turbines, reducing maintenance costs, and promoting the safe, reliable, and efficient operation of wind turbines in new power systems. In view of this, this invention provides an intelligent fault diagnosis method for wind turbine gearboxes, which can quickly, efficiently, and accurately diagnose gearbox failures.

[0054] According to one aspect of the present invention, an intelligent fault diagnosis method for wind turbine gearboxes is provided. Figure 1 A schematic flowchart illustrating an intelligent fault diagnosis method for wind turbine gearboxes according to an embodiment of the present invention is shown. Figure 1 As shown, the method may include the following steps S110, S150 and S160.

[0055] In step S110, real-time operating data of the state variables of the wind turbine gearbox are obtained. The state variables include environmental parameters, grid parameters and component state parameters of the wind turbine gearbox. The real-time operating data includes operating data within a preset time period.

[0056] Real-time operating data of various state variables of the wind turbine gearbox can be collected by sensors and stored in the wind turbine's data acquisition and monitoring control system. In this embodiment, state variables include environmental parameters, grid parameters, and component state parameters of the wind turbine gearbox. Specifically, environmental parameters may include wind speed, ambient temperature, and wind direction. Environmental parameters directly affect the operating status of various components of the wind turbine. Grid parameters may include the active power of the wind turbine, the magnitude of the current generated by the wind turbine, etc. Grid parameters can reflect the dispatch status of the grid where the wind turbine is located and are of great significance for predictive maintenance of the wind turbine. Component state parameters of the wind turbine gearbox may include blade pitch angle, pitch motor temperature (also known as blade motor temperature), pitch rate (also known as blade pitch speed), gearbox bearing temperature, generator bearing temperature, gearbox and generator oil temperature, and rotor speed, etc. In a specific embodiment, the state variables of the wind turbine gearbox are shown in Table 1 below.

[0057] Table 1 State variables of wind turbine units

[0058]

[0059] In this embodiment, the real-time operating data includes operating data within a preset duration. That is, for any state variable, the real-time operating data of that state variable includes all operating data collected within a preset duration ending at the current time. For example, if the preset duration is 3 seconds, the current time is 12:13:28, and the state variable is the temperature of the rear end of the high-speed shaft of the gearbox, then the real-time operating data of this state variable includes multiple temperatures of the rear end of the high-speed shaft of the gearbox collected from 12:13:25 to 12:13:28.

[0060] In step S150, the real-time running data of the state variables are repeatedly input into the pre-trained fault diagnosis model to obtain multiple initial fault diagnosis results. The fault diagnosis model is constructed based on a Bayesian graph convolutional network, and the initial fault diagnosis results include the probability of each fault type among multiple fault types in the wind turbine gearbox.

[0061] In this embodiment, various fault types include common faults such as tooth root cracks, gear fractures, missing teeth, gear pitting, gearbox bearing inner / outer ring faults, and gearbox bearing rolling element faults.

[0062] In some implementation schemes of this example, the fault diagnosis model is trained in the following way: acquiring historical operating data of the wind turbine's state variables; dividing the historical operating data into multiple sets of data using a sliding window method, with the length of the sliding window being a preset duration; adding fault labels to each set of data based on user input; dividing the dataset into training, validation, and test sets in a 6:2:2 ratio, and using the training, validation, and test sets to train the fault diagnosis model. Fault labels can be obtained by combining expert experience and are used to label fault types. Those skilled in the art will understand the specific implementation methods for training the fault diagnosis model using training, validation, and test sets, which will not be elaborated upon here.

[0063] In this example, the fault diagnosis model is constructed based on a Bayesian Graph Convolutional Neural Network (Bayesian-GCNN). Bayesian-GCNN is a graph convolutional neural network that incorporates Bayesian methods, treating the observed graph as a family of parameters of a random graph. Then, inference is performed based on the joint posterior of the random graph parameters and the weights and node labels in the GCNN, with the posterior probability of the label being:

[0064] ;

[0065] in, W Let be a random variable, representing a random graph. The weights of the Bayesian-GCNN network are given, where λ is the parameter of the random graph.

[0066] In Bayesian-GCNN, a combined mixed membership random block model (a-MMSBM) is used for posterior inference on the graph, and its parameters λ={π, β} are learned using a stochastic optimization method.

[0067] MMSBM can capture complex relationships in networks and is often used to model graphs with relatively strong community structures. Its core idea is that each node belongs to a different community with a certain probability, and the random block model is generalized through the classification behavior it exhibits.

[0068] because Typically noisy, it may be difficult to fit the adopted parametric block model well; therefore, maximum a posteriori estimation is used instead. π and β The integral over is given by the following formula:

[0069] ;

[0070] pass π and β Add appropriate priors and use approximations:

[0071] ;

[0072] in, W s,i From the diagram The corresponding Bayesian-GCNN network uses the Monte Carlo method. It was obtained by approximate sampling. from Obtained from sampling.

[0073] In some embodiments, when training a fault diagnosis model based on a Bayesian-GCNN network, the training set data can be input first, and the model input can be set. X Output Y and observation map And train GCNN to initialize the inference in MMSBM and the weights in Bayesian-GCNN. Then, iteratively train MMSBM to obtain and from Mid-sampling Finally, through the diagram... The corresponding Bayesian-GCNN network uses Monte Carlo sampling to obtain the weights. W i Then calculate an approximate value of the test probability. The model that achieves the highest accuracy on the validation set after an appropriate number of training rounds on the training set is the final fault diagnosis model.

[0074] The initial fault probability diagnosis result of this embodiment can be represented in the form of a two-dimensional matrix, which shows the probability of each of the various faults that the current wind turbine has.

[0075] In step S160, cooperative game theory is used to fuse multiple initial fault diagnosis results to obtain the final fault diagnosis result.

[0076] In step S150, multiple initial fault diagnosis results are obtained by repeatedly inputting real-time running data of the state variables. In this example scheme, considering the randomness of Bayesian inference, the weights in the model are adjusted each time during training or inference. W Random sampling will be performed. To reduce the impact of chance, in this embodiment, the input is repeated multiple times, resulting in multiple results. The specific number of repetitions can be determined based on the user's real-time requirements and result accuracy requirements. In a specific embodiment, the number of repetitions can be three. Due to the weights in the model... W Random sampling will be performed, so there will be differences between the multiple initial fault diagnosis results obtained.

[0077] After obtaining multiple initial fault diagnosis results, this paper considers fusing them through cooperative game theory. Cooperative game theory refers to a game played in a cooperative manner, where at least one participant's benefit increases, thus increasing the overall benefit. This example uses cooperative game theory to fuse multiple initial fault diagnosis results. The participants treat multiple different initial fault probability diagnosis results as a whole and engage in cooperative game theory to output the final fault diagnosis result. Fusing multiple initial fault diagnosis results through cooperative game theory helps improve the accuracy, robustness, and reliability of the final result.

[0078] In this invention, the filtered state variables are used as the input feature matrix of the Bayesian graph convolutional neural network. X A graph structure built using the correlations between state variables G The inference is performed within a Bayesian inference framework, outputting the posterior probability for each fault type. Due to the randomness of Bayesian networks, the inference is repeated three times, and the results from the three inferences are fused to improve the robustness and reliability of the diagnosis.

[0079] In the above technical solution, the ability to extract nonlinear features of wind turbine gearbox faults can be significantly enhanced by combining Bayesian graph convolutional networks. By repeatedly inputting the data and fusing multiple initial fault diagnosis results obtained from these inputs through cooperative game theory, the influence of random noise can be reduced, and random errors caused by random model parameters can be avoided. This helps to improve the robustness and reliability of the method and the accuracy of the prediction results. In addition, this solution does not require fault feature extraction and screening based on engineering practice experience, reducing human intervention and enabling a shift from an experience-driven, artificial feature paradigm to a data-driven representation learning paradigm. This not only helps to improve the accuracy of the results but also increases the speed of fault diagnosis. In summary, this method can quickly, efficiently, and accurately diagnose faults in wind turbine gearboxes.

[0080] For example, before repeatedly inputting the real-time running data of the state variables into the pre-trained fault diagnosis model, the method further includes: step S130, filtering the state variables based on the real-time running data of the state variables to determine the state variables with higher importance; wherein, step S150, repeatedly inputting the real-time running data of the state variables into the pre-trained fault diagnosis model includes: repeatedly inputting the real-time running data of the state variables with higher importance into the pre-trained fault diagnosis model.

[0081] In step S110, real-time operational data of various state variables of the wind turbine were acquired. The inventors discovered that different state variables contribute differently to the target parameter (i.e., fault type). Too many state variables increase the redundancy in the feature space, leading to increased computational load and negatively impacting the accuracy of fault diagnosis, thus reducing model performance. Therefore, in this example, before inputting the real-time operational data of multiple state variables into the pre-trained fault diagnosis model, the state variables are screened, selecting those with higher importance as input. Specific screening methods include, but are not limited to, correlation coefficients, random forests, game theory, and LightGBM. Gini This example does not limit the use of one or more of the following: indices, Spearman correlation analysis, etc.

[0082] The above approach reduces the dimensionality of the data by selecting real-time running data of highly important state variables as input to the model, thereby alleviating the computational burden during model training and practical applications. At the same time, it avoids the negative impact of irrelevant state variables on the results and improves the accuracy of the model.

[0083] For example, step S130, filtering state variables based on real-time operational data, includes: utilizing LightGBM's... Gini Exponential and / or Spearman correlation analysis determines the importance of each state variable in order to screen for state variables.

[0084] In this example, the solution is limited to using LightGBM. Gini Exponential and / or Spearman correlation analysis. LightGBM, a high-efficiency machine learning algorithm based on the gradient boosting framework, is widely used for feature selection. This algorithm employs an innovative histogram algorithm to find the optimal split point, quantifying feature importance by counting the number of times each feature is selected as a split point, thus intuitively reflecting the contribution of features in model construction. Its core advantages are reflected in three aspects: First, its unique feature splitting strategy combined with efficient parallel computing mechanisms significantly improves model training speed, making it particularly suitable for handling large-scale datasets and high-dimensional feature spaces; second, it continuously optimizes the model's predictive ability based on gradient boosting technology, maintaining high accuracy in both classification and regression tasks; finally, the algorithm supports multiple data input formats, has good compatibility, and can flexibly adapt to the needs of different application scenarios. In summary, LightGBM... GiniThe index, through the dynamic splitting mechanism of gradient boosting trees, can efficiently capture the nonlinear relationships and interaction effects between features and target variables. Its histogram algorithm and leaf-wise growth strategy significantly reduce computational complexity (several times higher than traditional methods), making it particularly suitable for high-dimensional data scenarios. Compared to other correlation analysis algorithms, this algorithm is more suitable for wind turbine gearbox fault analysis scenarios, helping to screen out more important state variables. Spearman correlation analysis, a nonparametric statistical method, also has unique advantages in feature selection. This method effectively identifies nonlinear but stable feature associations by calculating the strength of the monotonic relationship between features and target variables, i.e., the Spearman correlation coefficient. The core advantages of Spearman correlation analysis are mainly reflected in three aspects: First, as a rank correlation analysis method, it is insensitive to outliers and does not require data to follow a normal distribution, thus exhibiting stronger robustness; second, it can capture both linear and nonlinear monotonic relationships, making it more widely applicable than the traditional Pearson correlation coefficient; finally, it has high computational efficiency and intuitive results, allowing for rapid evaluation of the correlation strength between a large number of features and target variables, providing a reliable basis for feature selection. Compared to other methods, these two approaches can avoid the drawbacks of random forests, such as feature importance depending on the performance of the base model, high computational cost of game theory (e.g., SHAP values), and the need to discretize continuous features for information gain. They also maintain the global feature importance assessment... Gini It combines the dual capabilities of variable correlation analysis (Spilman) with the ability to balance computational efficiency and model interpretability, making it suitable for rapid feature optimization needs in industrial-scale data scenarios (such as SCADA systems).

[0085] In one implementation of this example, LightGBM is utilized. Gini Index and Spearman correlation analysis determine the importance of each state variable for screening. By combining the two methods, their advantages can be complemented, thereby further improving screening efficiency and effectiveness.

[0086] For example, using LightGBM Gini Exponential and / or Spearman correlation analysis is used to determine the importance of each state variable in order to filter state variables, including: determining the relative importance of each state variable to various fault types based on LightGBM and real-time operational data of the state variables. Gini index, of which Gini The index is positively correlated with the importance of the corresponding state variable; selection Gini A pre-defined number of state variables with larger exponents are considered as more important state variables.

[0087] In some embodiments, based on LightGBM and real-time operational data of state variables, the relative status variables of each state variable to multiple fault types are determined. Gini The index, including the method by which each state variable is determined, Gini index:

[0088] Assume there is n One state variable { X 1, X 2…… X n Input LightGBM, and these state variables have K Categories (i.e.) K (Types of failure). Assumptions V As for the importance of state variables, then the first j The importance of each state variable is Then the node m of Gini The index is:

[0089] ;

[0090] in, p mk Represents a node m The selected samples belong to the category k The probability of.

[0091] When state variables X j Selected as a node m When defining the partitioning features, nodes m If a node is divided into two child nodes (denoted as the left node and the right node), then the node... m Before and after Gini Exponential change (i.e.) Gini The gain is:

[0092] ;

[0093] in, G l and G r They represent nodes respectively m The two new nodes that split Gini index.

[0094] State variables X j In the i The importance of a variable in a tree is defined as the contribution of that variable to all nodes. Gini Sum of gains:

[0095] ;

[0096] State variables X j Importance is defined as follows within the entire forest of LightGBM:

[0097] ;

[0098] in, C This indicates the number of trees in the LightGBM forest.

[0099] For ease of comparison, the importance of all state variables is normalized. State variables X j The normalization importance is defined as:

[0100] ;

[0101] in, n The total number of state variables. State variables X j The Gini index, for n The sum of the gains of each state variable.

[0102] In the scheme of this example, after obtaining each state variable Gini After indexing, the state variables can be sorted from largest to smallest, and a preset number of state variables with the highest ranking can be selected as the most important state variables. This preset number can be chosen according to actual needs; for example, it can be n-2, where n is the total number of state variables. Of course, the preset number can also be adjusted based on user input, which will not be elaborated further.

[0103] The above technical solution utilizes LightGBM's... Gini The index measures the importance of each state variable, which can efficiently capture the nonlinear relationship and interaction effect between each state variable and the fault of the wind turbine. This allows for the selection of more important state variables while reducing dimensionality, which helps to improve the efficiency of fault diagnosis while ensuring the accuracy of fault diagnosis.

[0104] For example, using LightGBM GiniExponential and / or Spearman correlation analysis determines the importance of each state variable in order to screen state variables, including: calculating the Spearman correlation between each state variable and other state variables based on real-time running data of the state variables; for any state variable, if the Spearman correlation between the state variable and at least some of the other state variables is greater than or equal to the correlation threshold, the state variable is determined to be a highly important state variable.

[0105] The correlation threshold can be a preset value or a real-time value input by the user based on actual conditions; this document does not impose any restrictions on this. It is understood that Spearman correlation can have positive and negative values; the Spearman correlation values ​​in the above schemes described in this document are all absolute values. In one specific embodiment, the correlation threshold can be 0.1.

[0106] In some implementation schemes, based on real-time runtime data of the state variables, the Spearman correlation between each state variable and other state variables is calculated, including: for any given state variable, the Spearman correlation between that state variable and any other state variable (for ease of distinction, this state variable is called the target variable):

[0107] ;

[0108] In the formula, ρ XY State variables X With target variable Y The Spearman correlation coefficient between them. , They are X , Y The i Data in sample size h Rank in the data, d i It is the difference in the ranking of each pair of variables.

[0109] After obtaining the Spearman correlation between each state variable and other state variables, for any given state variable, it can be determined as a highly important state variable if the Spearman correlation between that state variable and at least a portion of the other state variables is greater than or equal to a correlation threshold. This "at least a portion" can be one or more; for example, when the total number of state variables is n, the "at least a portion" can be n-1, n-2, n-3, etc.

[0110] The above scheme determines the importance of each state variable through Spearman correlation analysis, thus accurately identifying and removing state variables that are weakly correlated with most other state variables. These variables, failing to form stable association structures with other variables in the system, may be isolated variables and difficult to participate in the collaborative representation during the modeling process. Excluding them achieves dimensionality compression and noise filtering of the feature space, enhancing the collaborative representation capability of the various state variables in the input model, which helps to further improve the accuracy of fault diagnosis.

[0111] For example, using LightGBM Gini Exponential and / or Spearman correlation analysis is used to determine the importance of each state variable in order to filter state variables, including: determining the relative importance of each state variable to various fault types based on LightGBM and real-time operational data of the state variables. Gini index, of which Gini The index is positively correlated with the importance of the corresponding state variable; selection Gini A first set of state variables consists of a pre-set number of state variables with larger exponents; based on the real-time running data of the state variables, the Spearman correlation between each state variable and other state variables is calculated; for any state variable, if the Spearman correlation between the state variable and at least some of the other state variables is greater than or equal to the correlation threshold, the state variable is added to the second set of state variables; the intersection of the first set of state variables and the second set of state variables is taken as the state variable with higher importance.

[0112] Determine state variables Gini The specific implementation of the Spearman correlation of the index and state variables has been described in detail above and will not be repeated here.

[0113] In this example, we consider using LightGBM separately. Gini The index and Spearman correlation analysis screened the original multiple state variables and took the union of the results obtained from each screening. In this way, a combination of state variables with high correlation to wind turbine faults and strong synergistic characterization ability can be obtained, which helps to further improve the accuracy of fault diagnosis.

[0114] For example, using LightGBM GiniExponential and / or Spearman correlation analysis determines the importance of each state variable to screen state variables, including: calculating the Spearman correlation between each state variable and other state variables based on real-time operational data of the state variables; adding any state variable to a third set of state variables if its Spearman correlation with at least some of the other state variables is greater than or equal to a correlation threshold; and determining the relative importance of each state variable in the third set of state variables to various fault types based on LightGBM and real-time operational data of each state variable in the third set of state variables. Gini index, of which Gini The index is positively correlated with the importance of the corresponding state variable; selection Gini A pre-defined number of state variables with larger exponents are considered as more important state variables.

[0115] Determine state variables Gini The specific implementation of the Spearman correlation of the index and state variables has been described in detail above and will not be repeated here.

[0116] In this example, we first use Spearman correlation analysis to remove isolated variables from the original multiple state variables, and then use LightGBM to filter state variables that are highly correlated with wind turbine faults. In this way, we can obtain a combination of state variables that are highly correlated with wind turbine faults and have strong collaborative characterization capabilities, which helps to further improve the accuracy of fault diagnosis.

[0117] For example, based on LightGBM and real-time operational data of state variables, the relative status variables of each state variable to multiple fault types are determined. Gini index, of which Gini The index is positively correlated with the importance of the corresponding state variable; selection Gini A fourth set of state variables is formed by a pre-defined number of state variables with larger exponents. Based on the real-time running data of each state variable in the fourth set of state variables, the Spearman correlation between each state variable in the fourth set of state variables and other state variables in the fourth set of state variables is calculated. For any state variable in the fourth set of state variables, if the Spearman correlation between the state variable and at least some of the other state variables in the fourth set of state variables is greater than or equal to the correlation threshold, the state variable is determined to be a highly important state variable.

[0118] Determine state variables Gini The specific implementation of the Spearman correlation of the index and state variables has been described in detail above and will not be repeated here.

[0119] In this example solution, LightGBM is first utilized.Gini The system employs a coarse selection of indices followed by a finer selection using Spearman correlation analysis. Specifically, firstly, the original state variables are assessed for feature importance based on LightGBM. Based on the obtained importance scores, they are sorted in descending order to identify state variables exhibiting low feature contributions across multiple fault types. These state variables can be considered redundant features contributing limitedly to the overall system modeling and can be eliminated. Subsequently, Spearman correlation analysis is performed on the remaining set of state variables after the first step, identifying variables with weak correlations to most other variables—that is, state variables with generally low absolute values ​​of Spearman correlation coefficients. These state variables, failing to form stable association structures with other state variables, may be isolated variables and difficult to participate in collaborative representation; therefore, they are further excluded. This yields a combination of state variables with high correlation to wind turbine faults and strong collaborative representation capabilities, which helps to further improve the accuracy of fault diagnosis. This screening scheme can be considered the optimal state variable screening scheme in this paper.

[0120] The "LightGBM followed by Spearman" screening strategy proposed in this invention has significant advantages and is irreplaceable in terms of the accuracy, stability, and final model performance of variable selection. First, LightGBM can identify key variables that can distinguish between multiple types of faults from a model perspective, possessing the ability to handle nonlinear relationships and complex interactions between variables. This initial screening effectively compresses the feature space and eliminates low-value variables that are insensitive to multiple faults. Subsequently, based on Spearman correlation analysis of pairwise relationships between variables, variables with correlation coefficients below a preset threshold are eliminated, thereby removing "isolated variables" or noise variables that lack a coordinated trend with other variables. This retains a set of variables with structural correlation and potential synergistic information, which helps improve the model's stability, structural integrity, and fault characterization ability. In contrast, if Spearman correlation analysis is performed first and then LightGBM modeling is used, it is easy to incorrectly eliminate some features that, although weakly correlated with other variables, can form nonlinear interactions with fault variables in the model and have important modeling value, thus weakening the model's effectiveness. However, if the union of the results is selected using both methods, it is difficult to effectively control the total number of variables, which can easily introduce noisy features and redundant information, reducing the model training efficiency and prediction robustness.

[0121] For example, before repeatedly inputting the real-time running data of the more important state variables into the pre-trained fault diagnosis model, the method further includes: step S140, normalizing the real-time running data of the more important state variables.

[0122] In some implementation schemes, for any real-time running data (which can be called specific running data) of any state variable (which can be called specific state variable), the following formula can be used for normalization:

[0123] ;

[0124] In the formula, x i For specific runtime data, x min and x max These are the minimum and maximum values ​​in the real-time running data of a specific state variable.

[0125] In this example, normalization is performed before inputting real-time running data into the model, which can eliminate the interference of feature scale differences on the model, reduce the model's predictive error, and improve the model's accuracy and robustness.

[0126] For example, a cooperative game is used to fuse multiple initial fault diagnosis results to obtain a final fault diagnosis result, including: for any one of the multiple initial fault diagnosis results, calculating the consistency correlation coefficient between the initial fault diagnosis result and the remaining fault diagnosis results, wherein the remaining fault diagnosis results are the sum of the other initial fault diagnosis results excluding the initial fault diagnosis result; and determining the final fault diagnosis result based on the sum of the products of the multiple initial fault diagnosis results and their respective corresponding consistency correlation coefficients.

[0127] Optionally, the consistency correlation coefficient between the initial fault diagnosis result and the remaining fault diagnosis results is calculated, including calculating the consistency correlation coefficient using the following formula:

[0128] ;

[0129] In the formula, The correlation coefficient represents consistency. This indicates the initial fault diagnosis result; This indicates the first [number] in the initial fault diagnosis result. i The value (i.e., the nth value) i (probability of each type of failure); This indicates the first of the remaining fault diagnosis results. i One value; express The average probability of each fault in the process; This represents the average probability of each fault in the remaining fault diagnosis results; m This indicates the total number of fault types.

[0130] Optionally, the final fault diagnosis result is determined based on the sum of the products of multiple initial fault diagnosis results and their respective corresponding consistency correlation coefficients, including: calculating the sum of the products of multiple initial fault diagnosis results and their respective corresponding consistency correlation coefficients; and normalizing each element in the sum of the products to obtain the final fault diagnosis result.

[0131] In some embodiments, the sum of the products of multiple initial fault diagnosis results and their respective corresponding consistency correlation coefficients is calculated, including by using the following formula:

[0132] .

[0133] Normalize the elements in the sum of the product, including by using the following formula:

[0134] .

[0135] The above technical solution can effectively improve the accuracy of fault diagnosis by fusing multiple fault probability diagnosis matrices through cooperative game theory.

[0136] For example, before repeatedly inputting the real-time operational data of the state variables into the pre-trained fault diagnosis model, the method further includes: step S120, processing the real-time operational data of the state variables using linear interpolation. Specifically, linear interpolation can be used to process missing values ​​and outliers in the real-time operational data of each state variable. It is understood that due to reasons such as abnormal data transmission and storage, missing values ​​and outliers may exist in the directly exported system data, and the presence of missing values ​​and outliers will affect the fitting effect of the model. In this embodiment, missing values ​​and outliers can be identified first, and outliers can be directly deleted. Those skilled in the art will understand the method of outlier identification, which will not be elaborated here. For the deleted outliers and missing values, linear interpolation can be used to supplement the data to ensure data integrity.

[0137] In some embodiments, linear interpolation can be expressed by the following formula:

[0138] ;

[0139] in, y i It is the first i The nth point (i.e., the nth point in the real-time running data of a certain state variable) i The value after which the original value is replaced. y i-1 , y i-2 It is the first i The two closest normal points (i.e., the points in the real-time running data of a certain state variable) are before the current point.i -1 value and the first i -2 values).

[0140] The above technical solution uses linear interpolation to process the real-time operating data of state variables, which can ensure the continuity of data and avoid the adverse effects of abnormal data on the results, thus ensuring the accuracy of fault diagnosis results.

[0141] In one specific embodiment of this paper, the method includes steps S110, S120, S130, S140, S150, and S160. Step S150 employs the aforementioned preferred state variable selection scheme.

[0142] According to another aspect of the present invention, an electronic device is also provided. Figure 2 A schematic block diagram of an electronic device according to an embodiment of the present invention is shown. Figure 2 As shown, the electronic device 200 includes a processor 210 and a memory 220. The memory 220 stores a computer program, which the processor 210 executes to implement the method described above.

[0143] According to another aspect of the present invention, a computer-readable storage medium is also provided. The storage medium stores a computer program / instructions that, when executed by a processor, implement the method described above. The storage medium may, for example, include a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a portable compact disc read-only memory (CD-ROM), a USB memory, or any combination of the above storage media. The computer-readable storage medium may be any combination of one or more computer-readable storage media.

[0144] Those skilled in the art will readily understand the implementation structure, working principle, and beneficial effects of electronic devices and computer-readable storage media by reading the above methods. For the sake of brevity, further details will not be elaborated here.

[0145] Although exemplary embodiments have been described herein with reference to the accompanying drawings, it should be understood that the above exemplary embodiments are merely illustrative and are not intended to limit the scope of the invention. Various changes and modifications can be made therein by those skilled in the art without departing from the scope and spirit of the invention.

[0146] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0147] In the several 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 units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed.

[0148] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0149] Furthermore, those skilled in the art will understand that although some embodiments described herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of the invention and form different embodiments.

[0150] The various component embodiments of the present invention can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some modules in the electronic device according to embodiments of the present invention. The present invention can also be implemented as an apparatus program (e.g., a computer program and computer program product) for performing some or all of the methods described herein. Such programs implementing the present invention can be stored on a computer-readable medium or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0151] The above description is merely a specific embodiment of the present invention or an explanation of the specific embodiment. The scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent fault diagnosis of wind turbine gearboxes, characterized in that, include: The real-time operating data of the state variables of the wind turbine gearbox are obtained. The state variables include environmental parameters, power grid parameters and component state parameters of the wind turbine gearbox. The real-time operating data includes operating data within a preset time period. The real-time running data of the state variables are repeatedly input into the pre-trained fault diagnosis model to obtain multiple initial fault diagnosis results. The fault diagnosis model is constructed based on a Bayesian graph convolutional network, and the initial fault diagnosis results include the probability of each fault type among multiple fault types in the wind turbine gearbox. The multiple initial fault diagnosis results are fused using cooperative game theory to obtain the final fault diagnosis result; Before repeatedly inputting the real-time operational data of the state variables into the pre-trained fault diagnosis model, the method further includes: Based on the real-time operational data of the state variables, the state variables are filtered to determine the state variables with higher importance. The step of repeatedly inputting the real-time running data of the state variables into the pre-trained fault diagnosis model includes: repeatedly inputting the real-time running data of the state variables with higher importance into the pre-trained fault diagnosis model. The filtering of the state variables based on the real-time operating data of the state variables includes: Using LightGBM Gini The importance of each state variable was determined by the index and Spearman correlation analysis in order to filter the state variables; The use of LightGBM Gini Exponential and Spearman correlation analysis determined the importance of each state variable in order to screen the state variables, including: Based on LightGBM and the real-time operational data of the aforementioned state variables, the relative status variables of each state variable to the various fault types are determined. Gini Index, wherein Gini The index is positively correlated with the importance of the corresponding state variable; In obtaining the state variables described above Gini After the exponentiation, the state variables are sorted in descending order, and the first preset number of state variables are selected to form the fourth set of state variables. Based on the real-time running data of each state variable in the fourth set of state variables, calculate the Spearman correlation between each state variable in the fourth set of state variables and other state variables in the fourth set of state variables. For any state variable in the fourth set of state variables, if the Spearman correlation between the state variable and at least some of the other state variables in the fourth set of state variables is greater than or equal to the correlation threshold, the state variable is determined to be a highly important state variable.

2. The method according to claim 1, characterized in that, Before repeatedly inputting the real-time operational data of the highly important state variables into the pre-trained fault diagnosis model, the method further includes: Normalize the real-time running data of the aforementioned highly important state variables.

3. The method according to claim 1, characterized in that, The step of fusing the multiple initial fault diagnosis results using cooperative game theory to obtain the final fault diagnosis result includes: For any one of the plurality of initial fault diagnosis results, calculate the consistency correlation coefficient between the initial fault diagnosis result and the remaining fault diagnosis results, wherein the remaining fault diagnosis results are the sum of the other initial fault diagnosis results among the plurality of initial fault diagnosis results excluding the initial fault diagnosis result; The final fault diagnosis result is determined by summing the products of the multiple initial fault diagnosis results and their respective consistency correlation coefficients.

4. The method according to claim 1, characterized in that, Before repeatedly inputting the real-time operational data of the state variables into the pre-trained fault diagnosis model, the method further includes: The real-time operating data of the state variables are processed using linear interpolation.

5. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor is used to execute the computer program to implement the method as described in any one of claims 1-4.

6. A computer-readable storage medium, characterized in that, The system stores a computer program / instructions that, when executed by a processor, implement the method as described in any one of claims 1-4.

Images