Unsupervised fault diagnosis method and related device for wind turbine gearbox based on variable correlation graph

Abstract

The present invention provides an unsupervised fault diagnosis method and related apparatus for wind turbine gearboxes based on variable correlation graphs, comprising the following steps: constructing a variable correlation graph based on real-time multi-source monitoring data of the wind turbine gearbox under test; using the variable correlation graph and its corresponding monitoring variables as input to a graph autoencoder to obtain reconstructed variable features; calculating anomaly scores between the monitoring variables and reconstructed variable features corresponding to the variable correlation graph; comparing the obtained anomaly scores with a discrimination threshold, and determining the fault of the wind turbine gearbox under test based on the comparison results; the present invention fully utilizes the coupling relationship between multiple monitoring variables, improves the robustness to complex operating conditions and the accuracy of fault detection, and has good fault location capabilities, effectively reducing the false alarm rate, and has high engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine operation status monitoring and fault diagnosis technology, specifically involving an unsupervised fault diagnosis method and related device for wind turbine gearboxes based on variable correlation graphs. Background Technology

[0002] As a key transmission component in wind power systems, the operating status of the wind turbine gearbox directly affects the safety and reliability of the entire unit. Because the gearbox operates in a complex and ever-changing environment, it is affected by multiple factors such as wind speed fluctuations, load variations, and environmental conditions. Consequently, its operating data exhibits characteristics of strong nonlinearity, strong coupling, and multi-source heterogeneity.

[0003] Existing gearbox fault diagnosis methods mainly include mechanistic model-based methods and data-driven methods. Mechanism-based methods rely on accurate physical modeling, which makes it difficult to adapt to complex operating conditions. Data-driven methods often rely on supervised learning with labeled data, but in practical engineering, obtaining fault samples is difficult and labeling costs are high, thus limiting their application.

[0004] In addition, traditional unsupervised methods (such as autoencoders) usually process each monitoring variable independently or model them using a simple splicing method, ignoring the inherent relationship between variables. This can easily lead to false alarms when the operating conditions change significantly, resulting in poor stability of the diagnostic results.

[0005] Therefore, there is an urgent need for a fault diagnosis method that can fully utilize multivariate correlation information to improve diagnostic accuracy and robustness under fault-free conditions. Summary of the Invention

[0006] The purpose of this invention is to provide an unsupervised fault diagnosis method and related apparatus for wind turbine gearboxes based on variable correlation graphs, which solves the problems of low accuracy caused by insufficient fault samples and insufficient utilization of variable correlation information in existing unsupervised fault diagnosis methods for wind turbine gearboxes.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the unsupervised fault diagnosis method for wind turbine gearboxes based on variable correlation graphs provided by the present invention includes the following steps: Based on the real-time multi-source monitoring data of the gearbox of the wind turbine under test, a variable correlation diagram is constructed; The variable association graph and its corresponding monitoring variables are used as inputs to the graph autoencoder to obtain the reconstructed variable features; Calculate the anomaly score between the features of the monitored variables and the reconstructed variables corresponding to the variable correlation diagram; The obtained anomaly score is compared with the discrimination threshold, and the fault of the gearbox of the wind turbine under test is determined based on the comparison result.

[0008] Preferably, a variable correlation diagram is constructed based on the real-time multi-source monitoring data of the gearbox of the wind turbine under test. The specific method is as follows: The real-time multi-source monitoring data includes multiple monitoring variables; Calculate the association strength among multiple monitored variables; Construct a variable association matrix based on the obtained association strength; Set association thresholds based on the variable association matrix; Based on the obtained association threshold, the variable association matrix is ​​sparsified to obtain the adjacency matrix; A variable association graph is constructed based on the obtained adjacency matrix.

[0009] Preferably, the graph autoencoder includes an encoder and a decoder, wherein: The encoder expression is:

[0010] in, For the first Layer node features; For the first Layer training weight parameters; It is a non-linear activation function; The expression for the decoder is:

[0011] in, This represents the final latent feature representation of the encoder output; Indicates the decoding function; This represents the reconstructed variable feature matrix.

[0012] Preferably, the anomaly score between the features of the monitored variable and the reconstructed variable corresponding to the variable correlation graph is calculated using the following formula:

[0013] in, Indicates the first in real-time multi-source monitoring data The values ​​of the monitored variables; Indicates the corresponding monitoring variable The reconstructed value; Indicates the dimension of the variable.

[0014] Preferably, the discrimination threshold is obtained according to the following formula:

[0015] in, To determine the threshold; For the set of abnormal ratings; For quantile parameters; This is a quantile function.

[0016] Preferably, the diagnostic method further includes: Based on the characteristics of the reconstructed variables, the fault location of the wind turbine gearbox is determined.

[0017] Preferably, the fault location of the wind turbine gearbox is located based on the characteristics of the reconstructed variables. Specifically, the method is as follows: Based on the reconstructed variable characteristics, the local anomaly score value of each monitored variable is calculated separately; The monitoring variable corresponding to the maximum local anomaly score value is used as the anomaly monitoring variable; The fault location of the wind turbine gearbox is obtained based on anomaly monitoring variables.

[0018] Secondly, the wind turbine gearbox unsupervised fault diagnosis system based on variable correlation graphs provided by the present invention includes: The variable correlation graph construction unit is used to construct a variable correlation graph based on the real-time multi-source monitoring data of the gearbox of the wind turbine under test. The variable feature reconstruction unit is used to take the variable association graph and its corresponding monitoring variables as input to the graph autoencoder to obtain the reconstructed variable features; The anomaly score calculation unit is used to calculate the anomaly score between the features of the monitored variable and the reconstructed variable corresponding to the variable correlation diagram; The anomaly determination unit is used to compare the obtained anomaly score with the discrimination threshold, and to determine the anomaly of the gearbox of the wind turbine under test based on the comparison result.

[0019] Thirdly, the present invention provides an electronic device including a processor and a memory, wherein the memory stores computer instructions, and when the computer instructions are executed by the processor, the electronic device performs the method described thereon.

[0020] Fourthly, the present invention provides a computer program product, the computer program product including computer-executable instructions, which, when executed, implement the method described.

[0021] Compared with the prior art, the beneficial effects of the present invention are: This invention provides an unsupervised fault diagnosis method for wind turbine gearboxes based on variable correlation graphs. By constructing a variable correlation graph, it explicitly introduces the inherent coupling relationships between different variables in multi-source monitoring data into the diagnostic process. This overcomes the shortcomings of traditional unsupervised methods (such as ordinary autoencoders) that process variables independently or simply splice them together, ignoring the correlation information between variables. Therefore, it can accurately capture the gearbox operating status even under complex operating conditions such as wind speed fluctuations and load changes, significantly reducing the false alarm rate caused by changes in operating conditions and improving the stability and reliability of diagnostic results. Furthermore, this invention uses a graph autoencoder to jointly reconstruct the variable correlation graph and its corresponding monitoring variables. It can learn the variable characteristics and correlation patterns under normal operating conditions without any fault labels, solving the problems of difficult fault sample acquisition and high labeling costs in practical engineering. This achieves true unsupervised fault diagnosis and has good engineering practical value. Finally, by calculating the anomaly score between the monitored variables and the reconstructed variable features and comparing it with a discrimination threshold, it can quantitatively and automatically determine whether a gearbox fault has occurred, avoiding the experience and uncertainty of manually setting thresholds.

[0022] Furthermore, since the anomaly score is based on the cumulative reconstruction error of each variable, the contribution of each variable can be further analyzed, thereby identifying the abnormal variables and their corresponding fault locations, which is beneficial for on-site maintenance personnel to respond quickly and perform accurate repairs. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method involved in the embodiments of the present invention; Figure 2 This is a heatmap showing the correlation between variables in Embodiment 4 of the present invention; Figure 3 This is a schematic diagram of the graphic autoencoder model structure of Embodiment 4 of the present invention; Figure 4 This is the convergence curve of the reconstruction loss during model training in Embodiment 4 of the present invention; Figure 5 This is the curve of the abnormal score changing over time in Embodiment 4 of the present invention; Figure 6 This is a score distribution diagram of normal and abnormal samples in Embodiment 4 of the present invention; Figure 7 This is the sorting result of variable-level anomaly contributions in a typical anomaly window of Embodiment 4 of the present invention. Detailed Implementation

[0024] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0025] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0026] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0027] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0028] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0029] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0030] Example 1 The unsupervised fault diagnosis method for wind turbine gearboxes based on variable correlation graphs provided in this embodiment includes the following steps: Based on the real-time multi-source monitoring data of the gearbox of the wind turbine under test, a variable correlation diagram is constructed; The variable association graph and its corresponding monitoring variables are used as inputs to the graph autoencoder to obtain the reconstructed variable features; Calculate the anomaly score between the features of the monitored variables and the reconstructed variables corresponding to the variable correlation diagram; The obtained anomaly score is compared with the discrimination threshold, and the fault of the gearbox of the wind turbine under test is determined based on the comparison result.

[0031] Example 2 Based on Example 1, this example provides an unsupervised fault diagnosis method for wind turbine gearboxes based on variable correlation graphs, and a method for obtaining graph autoencoders, including the following steps: Step 1: Collect historical multi-source monitoring data of the wind turbine gearbox, which includes multiple monitoring variables; Historical multi-source monitoring data is preprocessed to obtain preprocessed multi-source monitoring data; A sliding time window was used to segment the preprocessed multi-source monitoring data, construct a sample dataset, and divide it into a training set and a test set. The training set only included monitoring variables that were operating normally. In this embodiment, the preprocessing includes missing value imputation, outlier removal, and data standardization.

[0032] Step 2: Calculate the correlation between the monitored variables in the training set to obtain the correlation strength between the variables, and construct a variable correlation matrix based on the correlation strength; set a correlation threshold according to the statistical distribution characteristics of the correlation strength in the variable correlation matrix. The correlation strengths in the variable correlation matrix are compared with the correlation thresholds. Compare and retain those greater than or equal to the association threshold. The association strength will be less than the association threshold. The correlation strength is suppressed or set to zero to obtain a sparse adjacency matrix, and a variable correlation graph is constructed based on the adjacency matrix; Step 3: Construct a graph autoencoder model based on graph neural networks. Use the variable association graph of the training set and its corresponding monitoring variables to train the graph autoencoder model and obtain the trained graph autoencoder model. Step 4: Optimize the graph autoencoder model using the variable association graph corresponding to the test set and the corresponding monitoring variables to obtain the optimized graph autoencoder.

[0033] In this embodiment, the model is trained using training set samples. By minimizing the difference between the input and the reconstruction, the model learns a normal operating mode. .

[0034] Example 3 like Figure 1 As shown in the figure, the unsupervised fault diagnosis method for wind turbine gearboxes based on variable correlation graphs provided in this embodiment includes the following steps: Step 1: Collect real-time multi-source monitoring data of the wind turbine gearbox, wherein the real-time multi-source monitoring data includes multiple monitoring variables; The real-time multi-source monitoring data is preprocessed to obtain preprocessed multi-source monitoring data. In this embodiment, the real-time multi-source monitoring data comes from the wind turbine SCADA system and condition monitoring system. The real-time multi-source monitoring data includes operating condition variables, electrical state variables, thermal state variables, and mechanical vibration variables. Assume the system is in The running state at any given time is determined by Composed of variables, its state vector is represented as: in, represent Time of the first d The values ​​of the monitored variables; The monitored variables include: (1) Operating condition variables: wind speed, active power, and rotational speed; (2) Electrical state variables: three-phase voltage, three-phase current, reactive power; (3) Thermal state variables: gearbox temperature, gear bearing temperature, generator temperature, engine compartment temperature; (4) Mechanical vibration variables: amplitude characteristics (root mean square value, peak value, kurtosis, etc.), frequency domain characteristics (spectral energy, characteristic frequency components, etc.), time-frequency characteristics (wavelet energy, short-time Fourier transform characteristics, etc.), acceleration signal characteristics.

[0035] In this embodiment, the preprocessing methods for real-time multi-source monitoring data include: (1) Missing values ​​are filled using linear interpolation or forward filling; (2) Outliers are removed using statistical thresholding or median filtering methods; (3) Standardize each variable to eliminate the influence of dimensions; the standardization method is as follows:

[0036] in, and Variables Mean and standard deviation in normal operating data; The preprocessed multi-source monitoring data is segmented using a sliding time window, with a window length of 1. Step size is Then the sample set is represented as: Each sample This indicates the changes of each variable within the time window.

[0037] The sliding time window is used to convert continuous time series data into multiple local samples, so that each sample can characterize the operating state of the system within a certain time range, and is used to subsequently calculate the statistical characteristics of each variable within the time window and the correlation between variables, thereby improving the stability and accuracy of variable association modeling.

[0038] Step 2: Calculate the correlation between the monitored variables in the training set to obtain the correlation strength between the variables, and construct a variable correlation matrix based on the correlation strength; set a correlation threshold according to the statistical distribution characteristics of the correlation strength in the variable correlation matrix. The correlation strengths in the variable correlation matrix are compared with the correlation thresholds. Compare and retain those greater than or equal to the association threshold. The association strength will be less than the association threshold. The correlation strength is suppressed or set to zero to obtain a sparse adjacency matrix, and a variable correlation graph is constructed based on the adjacency matrix.

[0039] In this embodiment, the Pearson correlation coefficient is used to calculate the association strength between the monitored variables:

[0040] in, and They represent the times respectively. Next The first monitoring variable and the first The values ​​of the monitored variables; and These represent the monitored variables. and monitoring variables The mean value within the sliding time window; and These represent the monitored variables. and monitoring variables The standard deviation within the sliding time window; This indicates the number of sampling points within the sliding time window.

[0041] To uniformly represent the association strength, we take... absolute value: Thus, the variable correlation matrix is ​​obtained.

[0042] In this embodiment, to reduce the noise impact of weak correlations, the correlation matrix is ​​sparsified:

[0043] in, This is a threshold used to filter for valid associations between monitored variables. The value range is 0.2 to 0.6; In this embodiment, a variable association graph is constructed based on the above matrix: Among them, the node set Represents variables; edge set Indicates the strength of the association between variables; Step 3: Use the variable association graph and its corresponding monitoring variables as input to the graph autoencoder to obtain the reconstructed variable features.

[0044] In this embodiment, a self-loop structure is introduced to ensure the stability of the graph neural network:

[0045] And on Perform symmetric normalization:

[0046] in, It is a degree matrix; It is the identity matrix; It is a self-loop structure; This is the normalized self-loop structure.

[0047] The graph autoencoder is as follows: The graph autoencoder consists of two parts: an encoder and a decoder. The encoder adopts a graph neural network structure, and the decoder adopts a fully connected network or a linear mapping structure to reconstruct the features of the input variables. Based on the variable association graph, node feature propagation and aggregation are achieved through graph convolution operations, specifically as follows: in, For the first Layer node features; For the first Layer training weight parameters; It is a non-linear activation function; The decoder performs a reverse mapping on the encoded low-dimensional features to reconstruct the input variable features, specifically as follows:

[0048] in, This represents the final latent feature representation of the encoder output; The decoding function is represented by a fully connected network or a linear transformation. This represents the reconstructed variable feature matrix, whose structure and dimensions are consistent with the input variable features.

[0049] Step 4: Calculate the anomaly score between the features of the monitored variable and the reconstructed variable corresponding to the variable association diagram.

[0050] Define anomaly scoring for:

[0051] in, Indicates the first input sample The values ​​of the monitored variables; Indicates the corresponding monitoring variable The reconstructed value; Indicates the dimension of the variable.

[0052] Step 5: Compare the obtained anomaly score with the discrimination threshold, and determine the fault of the gearbox of the wind turbine under test based on the comparison result.

[0053] In this embodiment, the discrimination threshold is calculated according to the following formula:

[0054] in, This represents the set of abnormal scores corresponding to all normally functioning samples in the training set of the defined abnormal score set; This is a quantile parameter used to control the sensitivity of anomaly detection; its value ranges from 95% to 99%. This is a quantile function.

[0055] In this embodiment, the set of anomaly scores is defined. The method to obtain it is: Acquire normal historical multi-source monitoring data of the gearbox of the wind turbine under test under normal operating conditions; Construct a variable correlation graph based on normal historical multi-source monitoring data; The obtained variable correlation graph and its corresponding monitoring variables are used as input to the graph autoencoder to obtain the reconstructed variable features; Calculate the anomaly scores between the monitored variables and reconstructed variables corresponding to the variable association graph, and then obtain the anomaly score set. .

[0056] In this embodiment, the specific determination rules are as follows: If so, the gearbox of the wind turbine under test has failed.

[0057] Step 7, the diagnostic method further includes: locating the fault location of the wind turbine gearbox based on the reconstructed variable features.

[0058] In this embodiment, after determining the abnormal state, in order to locate the source of the abnormality, a local abnormality score is calculated for each monitored variable:

[0059] Local anomaly scoring Indicates the first The contribution of each monitored variable to the overall degree of abnormality, and its relationship with the overall abnormality score. The relationship is:

[0060] The monitoring variable corresponding to the maximum value of the local anomaly score is used as the anomaly monitoring variable; the fault location of the wind turbine gearbox is obtained based on the anomaly monitoring variable.

[0061] Example 4 This embodiment uses online monitoring data of a wind turbine gearbox in a wind farm as the object to verify the unsupervised fault diagnosis method for wind turbine gearboxes based on variable correlation graphs described in this application. The data used comes from the wind turbine SCADA system and condition monitoring system, with a sampling frequency of 1 Hz. The data time span covers multiple continuous operating cycles, including both normal operation and abnormal operation phases used for subsequent verification. To ensure that this embodiment can comprehensively reflect the operating status of the gearbox, 21 monitoring variables are selected as input variables. These variables include wind speed, active power, rotational speed, voltage A, voltage B, current A, current B, current C, reactive power, generator-side bearing temperature, non-drive-end bearing temperature, generator induction temperature, nacelle temperature, generator temperature 2, generator temperature 3, gearbox temperature, gear bearing temperature, drive-side acceleration, main shaft vibration in the Z direction, first-stage intermediate gear vibration, and second-stage intermediate gear vibration. These variables cover operating condition information, electrical status information, thermal status information, and mechanical vibration information, and can comprehensively characterize the operating status of the gearbox. The 21 variables can be further organized into variable groups as shown in Table 1.

[0062] Table 1

[0063] The fault diagnosis method in this embodiment mainly includes six steps: data acquisition and preprocessing, variable association graph construction, graph autoencoder modeling, model training, anomaly score calculation, and fault diagnosis. Each step is connected to the previous one to form a complete unsupervised fault diagnosis process.

[0064] First, in step 1, multi-source monitoring data of the wind turbine gearbox is collected and preprocessed. Since the actual collected data often contains missing values, outliers, and dimensional differences, it needs to be uniformly processed before entering the model. In this embodiment, short-term missing data points are filled using linear interpolation, and variables with short-term continuous missing periods but unsuitable for interpolation are filled using forward imputation. Outlier data points that significantly deviate from the normal operating range are removed using a combination of statistical thresholding and median filtering to avoid interference from outliers in subsequent variable correlation analysis. After handling missing values ​​and removing outliers, the variables are standardized based on normal operating data to bring variables with different dimensions into a unified numerical range. To transform long-term series data into samples suitable for graph learning model processing, this embodiment further uses a sliding time window to segment the data, setting a fixed window length and a fixed sliding step size to divide the continuous time series into multiple local samples, each representing the multivariate state of the gearbox within a time interval. This process constructs a sample dataset, which is then divided into a training set and a test set. The training set contains only normal operating data, while the test set contains both normal samples and samples to be identified. The specific values ​​for parameters such as the number of samples, window length, window stride, and sampling frequency in the training and test sets are shown in Table 2.

[0065] Table 2

[0066] After completing the sample construction, the process proceeds to step 2, which involves constructing a variable correlation graph based on the normal operating data in the training set. Specifically, firstly, the correlation between each monitored variable is statistically analyzed using normal samples from the training set to obtain the correlation strength between variables, and a variable correlation matrix is ​​constructed accordingly. This matrix reflects the coupling relationship between variables in the gearbox under normal operating conditions. In this embodiment, the variable correlation matrix is ​​obtained based on the statistical analysis of normal samples, thus reflecting the relatively stable variable relationships under normal system conditions, rather than the perturbation relationships under abnormal conditions. Subsequently, a correlation threshold is set according to the statistical distribution characteristics of the correlation strength in the variable correlation matrix. Relationships with weak correlation strength are suppressed or set to zero, retaining only strong correlations above the threshold, thereby obtaining a sparse adjacency matrix, and a variable correlation graph is constructed based on this adjacency matrix. Through threshold processing, the impact of noisy edges on model training can be effectively reduced while retaining the coupling relationships of the main variables. To enhance the training stability of the graph neural network, this embodiment adds node self-loops to the adjacency matrix and normalizes the matrix to obtain the final graph structure input used for subsequent modeling. The variable correlation obtained in this step can be intuitively displayed through a correlation heatmap, such as... Figure 2 As shown. (Through) Figure 2 It can be seen that under normal operating conditions, there is a clear correlation between wind speed, power, and rotational speed. There is also a clear coupling characteristic between gearbox temperature, gear bearing temperature, and vibration signal, indicating that the variable correlation graph can effectively characterize the multivariate coupled operating mechanism of the gearbox. The graph construction parameters in this step, as shown in Table 3, include the number of nodes, the number of edges, the threshold setting method, and the sparsity rate.

[0067] Table 3

[0068] In step 3, a graph autoencoder model based on a graph neural network is constructed. This model takes the variable association graph and corresponding variable features obtained in step 2 as input, and propagates and aggregates the variable features through graph convolution operations to extract high-order association features between variables, thereby achieving feature representation learning for a multivariate system. In this embodiment, the graph autoencoder model includes an encoder and a decoder. The encoder adopts a graph convolutional network structure, and the decoder is used to restore low-dimensional latent features to a reconstruction result consistent with the input data structure. Since the input graph structure originates from the variable relationships obtained from statistical analysis of normal operating data, the model learns not only the changing patterns of individual variables during training, but also the stable association patterns between variables under normal conditions. Figure 3The diagram illustrates the structure of the graph autoencoder model in this embodiment. Specifically, the input includes a variable association graph and node features. The encoder part maps the node features layer by layer through a two-layer graph convolutional network to form a low-dimensional implicit representation. The decoder part then restores this low-dimensional representation to the reconstructed output (as shown in Table 4).

[0069] Table 4

[0070] In step 4, the graph autoencoder model is trained using normal operating data from the training set. Since this embodiment employs an unsupervised diagnostic method, fault labels are not introduced during the training phase; instead, only normal samples are used to enable the model to learn the normal operating modes of the system. During training, the normal samples constructed in step 1 are input into the graph autoencoder model batch by batch. The model outputs corresponding reconstruction results, and the model parameters are continuously adjusted based on the differences between the input samples and the reconstruction results, enabling the model to have a strong reconstruction capability for normal samples.

[0071] In this embodiment, the Adam optimizer was used for model training, with a batch size of 256, a maximum number of training epochs of 200, a learning rate of 1e-3, and an early stopping strategy to prevent overfitting. As the number of training epochs increased, the model reconstruction loss gradually decreased and tended to stabilize, indicating that the model could gradually learn the variable distribution characteristics and correlation structure under normal gearbox conditions. Figure 4 The convergence curve of the reconstruction loss during model training is given. As can be seen from the figure, the loss of the model decreases rapidly in the early training stage and gradually converges in the later stage, indicating that the training process is stable and effective. The training parameters are further organized into Table 5.

[0072] Table 5

[0073] In step 5, the test set data is input into the trained graph autoencoder model to calculate anomaly scores. Specifically, the test samples are first input into the encoder, which extracts low-dimensional latent feature representations. Then, the decoder generates a reconstruction result with the same structure and dimensions as the input data. Since the model only learns the normal operating mode during the training phase, when the input sample is consistent with the normal mode, the model can achieve good reconstruction, and the difference between the input data and the reconstruction result is small. When the input sample deviates from the normal operating mode, the model has difficulty reconstructing accurately, and the difference between the input data and the reconstruction result increases significantly. In this embodiment, the anomaly score of each sample is obtained by comparing the difference between the input data and the reconstruction result, which is used to characterize the degree to which the current operating state deviates from the normal mode. Statistical analysis of the test results shows that the mean abnormal score for the normal operating sample was 0.2839, the standard deviation was 0.0209, the 99th percentile was 0.3307, and the maximum value was 0.3573; the mean abnormal score for the abnormal sample was 0.6382, the standard deviation was 0.1089, the 99th percentile was 0.8446, and the maximum value was 0.9045. Figure 5 The curves showing the change of the anomaly score over time are shown. It can be seen that during the normal operation phase, the anomaly score remains at a low level and fluctuates slightly around the stable range; during the anomaly phase, the anomaly score increases significantly and continues to exceed the threshold, forming a continuous anomaly response range. Figure 6 The score distribution charts for normal and abnormal samples are presented. The results clearly show that the scores of abnormal samples generally shift towards higher values, clearly distinguishing them from normal samples. The corresponding statistical results are summarized in Table 6.

[0074] Table 6

[0075] In step 6, a discrimination threshold is determined based on the abnormal score distribution of normal operating data, and fault diagnosis is performed on the test samples. Specifically, this embodiment uses the high quantile of the abnormal score distribution of normal samples as the discrimination threshold. While considering both false alarm rate and detection sensitivity, an abnormal score threshold is determined to distinguish between normal and abnormal states. According to the implementation results, when the abnormal score of a test sample exceeds the discrimination threshold, the operating state corresponding to that sample is determined to be abnormal. Furthermore, after determining an anomaly, this embodiment also analyzes the variable-level anomaly contribution to enhance the interpretability of the diagnostic results. By calculating and sorting the local reconstruction errors of each monitored variable in a single abnormal sample, the main contributing variables that cause the overall abnormal score to increase can be identified. Figure 7The ranking of variable-level anomaly contributions within a typical anomaly window is presented. The results show that vibration variables such as drive-side acceleration, main shaft vibration, first-stage intermediate gear vibration, and second-stage intermediate gear vibration have the highest anomaly contributions under abnormal conditions, while thermal variables such as gearbox temperature and gear bearing temperature are second. This indicates that the fault evolution process in this embodiment exhibits a clear "vibration-first, temperature-later" characteristic. This is consistent with the actual fault mechanism of the gearbox, thus verifying that the method in this embodiment can not only perform anomaly detection but also provide variable-level fault location results with engineering significance. To further illustrate the effectiveness of the method in this embodiment, it can be compared with traditional graph-less autoencoder methods and dense graph autoencoder methods, forming Table 7, which lists indicators such as false alarm rate, the proportion of consecutive anomaly threshold crossings, and the correlation between anomaly score and power.

[0076] Table 7

[0077] The implementation results show that the method in this embodiment is superior to the comparative method in terms of false alarm rate of normal samples, continuity of abnormal response and robustness under operating conditions.

[0078] Example 5 The wind turbine gearbox unsupervised fault diagnosis system based on variable correlation graphs provided in this embodiment is characterized by including: The variable correlation graph construction unit is used to construct a variable correlation graph based on the real-time multi-source monitoring data of the gearbox of the wind turbine under test. The variable feature reconstruction unit is used to take the variable association graph and its corresponding monitoring variables as input to the graph autoencoder to obtain the reconstructed variable features; The anomaly score calculation unit is used to calculate the anomaly score between the features of the monitored variable and the reconstructed variable corresponding to the variable correlation diagram; The anomaly determination unit is used to compare the obtained anomaly score with the discrimination threshold, and to determine the anomaly of the gearbox of the wind turbine under test based on the comparison result.

[0079] Example 6 This embodiment also provides a computing device. The computing device includes a bus, a processor, a memory, and a communication interface. The processor, memory, and communication interface communicate with each other via the bus. The computing device can be a server or a terminal device. It should be understood that this application does not limit the number of processors and memory in the computing device.

[0080] A bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, a bus can include a path for transmitting information between various components of a computing device (e.g., memory, processor, communication interfaces).

[0081] The processor may include any one or more of the following: central processing unit (CPU), graphics processing unit (GPU), tensor processing unit (TPU), application specific integrated circuit (ASIC), field-programmable gate array (FPGA), microprocessor (MP), or digital signal processor (DSP).

[0082] Memory can include volatile memory, such as random access memory (RAM). Processors can also include non-volatile memory. volatile memory, such as read-only memory (ROM). ROM (memory only), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0083] The memory stores executable program code, which the processor executes to implement the functions of the aforementioned units, thereby achieving, for example, the method described in Embodiment 1. That is, the memory may store instructions for the methods and functions relating to the computing device in any of the above embodiments.

[0084] The communication interface uses transceiver modules such as, but not limited to, network interface cards and transceivers to enable communication between computing devices and other devices or communication networks.

[0085] Example 7 This embodiment also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, cause the processor to perform the methods and functions of the computing device involved in any of the above embodiments.

[0086] Generally, the various embodiments of this disclosure can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software, which can be executed by a controller, microprocessor, or other computing device. Although various aspects of the embodiments of this disclosure are shown and described as block diagrams, flowcharts, or represented using some other illustration, it should be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented as, as non-limiting examples, in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0087] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. An unsupervised fault diagnosis method for wind turbine gearboxes based on variable correlation graphs, characterized in that, Includes the following steps: Based on the real-time multi-source monitoring data of the gearbox of the wind turbine under test, a variable correlation diagram is constructed; The variable association graph and its corresponding monitoring variables are used as inputs to the graph autoencoder to obtain the reconstructed variable features; Calculate the anomaly score between the features of the monitored variables and the reconstructed variables corresponding to the variable correlation diagram; The obtained anomaly score is compared with the discrimination threshold, and the fault of the gearbox of the wind turbine under test is determined based on the comparison result.

2. The unsupervised fault diagnosis method for wind turbine gearboxes based on variable correlation graphs according to claim 1, characterized in that, Based on the real-time multi-source monitoring data of the gearbox of the wind turbine under test, a variable correlation diagram is constructed. The specific method is as follows: The real-time multi-source monitoring data includes multiple monitoring variables; Calculate the association strength among multiple monitored variables; Construct a variable association matrix based on the obtained association strength; Set association thresholds based on the variable association matrix; Based on the obtained association threshold, the variable association matrix is ​​sparsified to obtain the adjacency matrix; A variable association graph is constructed based on the obtained adjacency matrix.

3. The unsupervised fault diagnosis method for wind turbine gearboxes based on variable correlation graphs according to claim 1, characterized in that, The graph autoencoder includes an encoder and a decoder, wherein: The encoder expression is: in, For the first Layer node characteristics; For the first Layer training weight parameters; It is a non-linear activation function; The expression for the decoder is: in, This represents the final latent feature representation of the encoder output; Indicates the decoding function; This represents the reconstructed variable feature matrix.

4. The unsupervised fault diagnosis method for wind turbine gearboxes based on variable correlation graphs according to claim 1, characterized in that, The anomaly score between the monitored variable and the reconstructed variable features corresponding to the variable association diagram is calculated using the following formula: in, Indicates the first in real-time multi-source monitoring data The values ​​of the monitored variables; Indicates the corresponding monitoring variable The reconstructed value; Indicates the dimension of the variable.

5. The unsupervised fault diagnosis method for wind turbine gearboxes based on variable correlation graphs according to claim 1, characterized in that, The discrimination threshold is obtained according to the following formula: in, To determine the threshold; For the set of abnormal ratings; For quantile parameters; This is a quantile function.

6. The unsupervised fault diagnosis method for wind turbine gearboxes based on variable correlation graphs according to claim 1, characterized in that, The diagnostic method also includes: Based on the characteristics of the reconstructed variables, the fault location of the wind turbine gearbox is determined.

7. The unsupervised fault diagnosis method for wind turbine gearboxes based on variable correlation graphs according to claim 6, characterized in that, Based on the characteristics of reconstructed variables, the fault location of the wind turbine gearbox is located. The specific method is as follows: Based on the reconstructed variable characteristics, the local anomaly score value of each monitored variable is calculated separately; The monitoring variable corresponding to the maximum local anomaly score value is used as the anomaly monitoring variable; The fault location of the wind turbine gearbox is obtained based on anomaly monitoring variables.

8. An unsupervised fault diagnosis system for wind turbine gearboxes based on variable correlation graphs, characterized in that, include: The variable correlation graph construction unit is used to construct a variable correlation graph based on the real-time multi-source monitoring data of the gearbox of the wind turbine under test. The variable feature reconstruction unit is used to take the variable association graph and its corresponding monitoring variables as input to the graph autoencoder to obtain the reconstructed variable features; The anomaly score calculation unit is used to calculate the anomaly score between the features of the monitored variable and the reconstructed variable corresponding to the variable correlation diagram; The anomaly determination unit is used to compare the obtained anomaly score with the discrimination threshold, and to determine the anomaly of the gearbox of the wind turbine under test based on the comparison result.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing computer instructions that, when executed by the processor, cause the electronic device to perform the method of any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes computer-executable instructions that, when executed, implement the method of any one of claims 1 to 7.

Images