Fan fault risk quantification early warning method and system based on class-incremental prototype graph

Abstract

This application relates to a method and system for quantitative early warning of wind turbine fault risks based on incremental prototype graphs, belonging to the field of wind turbine fault diagnosis and operation and maintenance. The method includes: extracting wind turbine monitoring time-series features and historical operation and maintenance text features, and fusing them to obtain fused features; constructing a prototype graph based on the fused features; determining the wind turbine status category of the time-series monitoring data of the wind turbine to be evaluated based on the prototype graph; calculating the risk coefficient when the wind turbine status category is a fault category, and determining the early warning level based on the risk coefficient; and updating the prototype graph when the time-series monitoring data of the wind turbine to be evaluated meets the triggering conditions. This application models the fuzzy relationship between data and fault modes by constructing a prototype graph, realizing accurate and intelligent early warning under dynamic wind turbine fault risks, and is applicable to the predictive and preventive monitoring and maintenance of wind turbine equipment status in wind farms.

Description

Technical Field

[0001] This application relates to the field of wind turbine fault diagnosis and operation and maintenance, specifically to a method and system for quantifying and warning of wind turbine fault risks based on incremental prototype maps. Background Technology

[0002] Wind turbines are often deployed in complex outdoor environments, subjected to long-term conditions such as variable loads, strong vibrations, and drastic temperature and humidity fluctuations. Core components like gearboxes, generators, and blades are prone to performance degradation and failures. The level of operation and maintenance management in a wind farm directly determines the equipment's operational stability and power generation efficiency. Industry statistics show that unplanned outages due to sudden failures account for a high percentage of wind farm shutdowns, causing significant economic losses and severely impacting the continuity and stability of power supply. Therefore, accurate identification, quantitative assessment, and efficient handling of wind turbine failure risks have become a key requirement for the full lifecycle operation and maintenance management of wind farms.

[0003] Most existing intelligent fault diagnosis methods are based on deep learning technology, which constructs complex neural network models to learn features from collected equipment operating data to achieve fault early warning. However, the fault early warning function achieved by this method is mainly reflected in ideal laboratory environments. Its practical application effectiveness in real industrial environments is usually greatly reduced due to the dual limitations of data and models, easily leading to serious consequences such as false alarms, missed alarms, and misreporting. In summary, existing methods mainly suffer from the following key application bottlenecks: (1) Insufficient utility of fault warning labels. Historical SCADA (Supervisory Control And Data Acquisition) alarm data is collected to construct a training set for training the warning model. However, SCADA system alarm data is secondary data processed by the wind turbine manufacturer. An alarm signal is only generated when the fault has caused the equipment to shut down completely and reached the preset alarm threshold. The utility of fault warning labels is extremely low. Existing methods directly use this type of data to train the model, which results in the model learning the characteristic patterns after the fault occurs. It cannot truly capture the weak signs in the early stage of fault evolution and does not meet the preemptive requirement of fault warning. In essence, it is still a post-event alarm rather than a pre-event warning.

[0004] (2) Most existing intelligent fault diagnosis methods are static models and can only process static data. During long-term continuous operation of wind turbines, new types of faults that have not been seen in the training phase will constantly emerge as the equipment ages and the operating conditions change. However, most existing intelligent fault diagnosis methods are static models, and their knowledge boundaries are fixed within the distribution range of the training data. They can only process known static fault data that has been trained, and lack the ability to learn online and dynamically expand. They cannot adaptively learn new types of faults that constantly emerge during the operation of wind turbines. They are prone to false alarms and false alarms when dealing with new faults, and require frequent manual retraining and parameter adjustment, which increases the operation and maintenance costs and makes it difficult to cope with dynamic and changing actual industrial fault scenarios. Summary of the Invention

[0005] To overcome at least one deficiency in the prior art, this application provides a method and system for quantifying and warning of wind turbine failure risk based on incremental prototype maps.

[0006] Firstly, a method for quantitative early warning of wind turbine failure risk based on incremental prototype graphs is provided, including: Historical time-series monitoring data of wind turbines is acquired. The historical time-series monitoring data of wind turbines includes multiple time-series samples. Features are extracted from each time-series sample to obtain the time-series features corresponding to each time-series sample. Each time-series sample corresponds to a wind turbine status category. Feature extraction is performed on the text data of historical operation and maintenance reports of wind turbines to obtain text features; The temporal features and textual features corresponding to each time series sample are fused to obtain the fused features corresponding to each time series sample; Based on the fusion features, the class prototype of each wind turbine state category is determined; a class prototype graph is constructed with class prototypes as nodes, and the edge weight between any two class prototypes in the class prototype graph is determined based on the distance between the two class prototypes and the similarity of the category semantic description. Feature extraction is performed on the time-series monitoring data of the wind turbine to be evaluated to obtain the time-series features to be evaluated; the time-series features to be evaluated and the text features are fused to obtain the fused features of the time-series monitoring data of the wind turbine to be evaluated. Calculate the Euclidean distance between the fusion features of the time-series monitoring data of the wind turbine to be evaluated and each class prototype in the class prototype map, determine the class prototype corresponding to the minimum value of the Euclidean distance, and use the wind turbine state category of the determined class prototype as the wind turbine state category of the time-series monitoring data of the wind turbine to be evaluated. If the wind turbine status category of the wind turbine time-series monitoring data to be evaluated is not the normal category, then the risk coefficient is calculated, and the warning level is determined based on the risk coefficient.

[0007] In one embodiment, feature extraction is performed on each time series sample to obtain the time series features corresponding to each time series sample, including: The time series samples are D-dimensional data, and the time series samples are divided into multiple Patch data blocks; Perform 1D convolutional encoding on the D-dimensional data of each Patch data block to obtain local feature vectors; A multi-head attention mechanism is used to process local feature vectors to capture the correlation of features in different dimensions, and a global temporal feature vector for each patch data block is output through a fully connected layer. The global temporal feature vectors of all Patch data blocks are concatenated to form the temporal features corresponding to the temporal samples.

[0008] In one embodiment, feature extraction is performed on the text data of historical wind turbine operation and maintenance reports to obtain text features, including: Preprocess the text data of the historical operation and maintenance report of the wind turbine to obtain the preprocessed text data; The pre-trained language model is used to extract features from the pre-processed text data to obtain the global semantic feature vector of the text. The TF-IDF algorithm was used to extract keyword features from the preprocessed text data; The text features are obtained by concatenating the global semantic feature vector of the text and the keyword features.

[0009] In one embodiment, the temporal features and textual features corresponding to each time series sample are fused to obtain the fused features corresponding to each time series sample, including: Z-score standardization is applied to both temporal features and text features to obtain standardized temporal features and standardized text features. A linear projection layer is used to map the standardized temporal features and standardized text features to the same dimension, resulting in mapped temporal features and mapped text features; By concatenating the mapped temporal features and the mapped text features, the concatenated features are obtained. The spliced ​​features are processed through an attention mechanism to obtain the fusion weights; The mapped temporal features and mapped text features are weighted and fused according to the fusion weights to obtain the fused features.

[0010] In one embodiment, the class prototype for each wind turbine state category is determined based on the fusion characteristics, calculated using the following formula:

[0011] in, Fan status category class prototype, Fan status category The number of time series samples under the current conditions. Fan status category Time series samples below The corresponding fusion features.

[0012] In one embodiment, the risk coefficient is calculated using the following method: The deviation distance between the fused features of the time-series monitoring data of the wind turbine to be evaluated and the class prototype of the normal category is calculated using the following formula:

[0013] in, For the deviation distance, To evaluate the fusion characteristics of the wind turbine time-series monitoring data, The class prototype for the normal category. It is an L2 norm; The deviation distance is converted into an initial risk factor using the following formula:

[0014] in, The initial risk coefficient, As the risk coefficient weight; The initial risk coefficient is corrected to obtain the corrected risk coefficient using the following formula:

[0015] in, This is the corrected risk coefficient. The Euclidean distance between the fused features of the wind turbine time-series monitoring data to be evaluated and the class prototype of the corresponding wind turbine state category is given. This represents the maximum deviation distance for the corresponding wind turbine status category.

[0016] In one embodiment, determining the wind turbine failure level based on a risk factor includes: If the risk coefficient This is a Level 3 warning; This is the warning threshold; like ≤ Risk coefficient It is a Level II warning; Emergency threshold; If the risk coefficient is ≥ This is a Level 1 warning.

[0017] In one embodiment, the method further includes: If the minimum value of the Euclidean distance is greater than the set threshold, the class prototype graph is updated. The update process includes: updating the original class prototype, generating a new class prototype, and updating the edge weights in the class prototype graph.

[0018] In one embodiment, updating the original class prototype uses the following formula:

[0019] in, The result of updating the original class prototype. For the original class prototype, The original class prototype weight, The fusion characteristics of the time-series monitoring data of the wind turbine to be evaluated; Generate new class prototypes, including: The fusion characteristics of the wind turbine time-series monitoring data to be evaluated As the initial value for the new class prototype, the prototype cross-entropy loss function is used to optimize the new class prototype, resulting in the final new class prototype. The prototype cross-entropy loss function is:

[0020] in, The prototype cross-entropy loss function is... This is the distance weighting coefficient. The Euclidean distance between the new class prototype and the original class prototype of the wind turbine state category c is given, where N is the number of the original class prototypes.

[0021] Secondly, a wind turbine fault risk quantification and early warning system based on incremental prototype graphs is provided, including: The time-series feature extraction module is used to acquire historical time-series monitoring data of wind turbines. The historical time-series monitoring data of wind turbines includes multiple time-series samples. Feature extraction is performed on each time-series sample to obtain the time-series features corresponding to each time-series sample; each time-series sample corresponds to a wind turbine status category. The text feature extraction module is used to extract features from the text data of historical wind turbine operation and maintenance reports to obtain text features. The feature fusion module is used to fuse the temporal features and textual features corresponding to each time series sample to obtain the fused features corresponding to each time series sample. The class prototype graph construction module is used to determine the class prototype of each wind turbine state category based on the fused features; the class prototype graph is constructed with the class prototype as nodes, and the edge weight between any two class prototypes in the class prototype graph is determined based on the distance between the two class prototypes and the similarity of the category semantic description. The module for obtaining the feature to be evaluated is used to extract features from the time-series monitoring data of the wind turbine to be evaluated, and obtain the time-series features to be evaluated; the time-series features to be evaluated and the text features are fused to obtain the fused features of the time-series monitoring data of the wind turbine to be evaluated. The wind turbine status category determination module is used to calculate the Euclidean distance between the fusion features of the time-series monitoring data of the wind turbine to be evaluated and each class prototype in the class prototype map, determine the class prototype corresponding to the minimum value of the Euclidean distance, and use the wind turbine status category of the determined class prototype as the wind turbine status category of the time-series monitoring data of the wind turbine to be evaluated. The early warning module is used to calculate the risk coefficient and determine the early warning level based on the risk coefficient if the wind turbine status category in the time-series monitoring data to be evaluated is not the normal category.

[0022] Compared with the prior art, this application has the following beneficial effects: This application models the fuzzy relationship between data and fault modes by constructing a prototype graph, realizing accurate and intelligent early warning under dynamic wind turbine fault risk, which is applicable to wind farm wind turbine equipment status prediction, preventive monitoring and maintenance. Attached Figure Description

[0023] This application can be better understood by referring to the description given below in conjunction with the accompanying drawings, which, together with the detailed description below, are incorporated in and form part of this specification. In the drawings: Figure 1 A flowchart of a wind turbine fault risk quantification and early warning method based on incremental prototype graphs is shown. Figure 2 The diagram illustrates the principle of adaptive incremental update of class prototype graph. Detailed Implementation

[0024] Exemplary embodiments of the present application will be described below with reference to the accompanying drawings. For clarity and brevity, not all features of the actual embodiments are described in the specification. However, it should be understood that many embodiment-specific decisions can be made in the development of any such actual embodiment to achieve the developer’s specific objectives, and these decisions may vary as the embodiments differ.

[0025] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the device structure closely related to the solution of this application is shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0026] It should be understood that this application is not limited to the described embodiments by virtue of the following description with reference to the accompanying drawings. In this document, embodiments may be combined with each other, features may be substituted or borrowed between different embodiments, and one or more features may be omitted in one embodiment, where feasible.

[0027] This application provides a method for quantitative early warning of wind turbine failure risk based on incremental prototype graphs. Figure 1A flowchart of a wind turbine fault risk quantification and early warning method based on incremental prototype mapping is shown. (See [link]) Figure 1 The method mainly includes the following steps: Step S1: Obtain historical time-series monitoring data of the wind turbine. The historical time-series monitoring data of the wind turbine includes multiple time-series samples. Extract features from each time-series sample to obtain the time-series features corresponding to each time-series sample. Each time-series sample corresponds to a wind turbine status category.

[0028] Here, the historical time-series monitoring data of wind turbines is multi-dimensional data. Based on high-precision timestamps, the historical time-series monitoring data of wind turbines in different time periods are indexed and matched, and missing values ​​are repaired using linear interpolation. The original dimensions are unified into the target dimensions through a dimension mapping matrix, and invalid dimensions are eliminated.

[0029] Then, outlier handling is performed: the mean of the data for each dimension is calculated. with standard deviation If data points satisfy If the value is not found, it is considered an outlier and replaced with the median of that dimension. This represents the d-th dimension.

[0030] Specifically, feature extraction is performed on each time series sample to obtain the time series features corresponding to each time series sample, including: First, the historical time-series monitoring data of the wind turbine is divided into multiple time-series samples according to time. The time-series samples are D-dimensional data, and the time-series samples are further divided into multiple Patch data blocks. Here, multiple patch data blocks are obtained based on the patch encoding method, including: setting the patch length. With sliding step size Calculate the overlap rate , No. The index range of each patch is This yields the Patch data block.

[0031] Then, 1D convolutional encoding is performed on the D-dimensional data of each Patch data block to obtain local feature vectors; Then, a multi-head attention mechanism is used to process the local feature vectors to capture the correlation of features in different dimensions, and the global temporal feature vector of each patch data block is output through a fully connected layer.

[0032] Then, the global temporal feature vectors of all Patch data blocks are concatenated to form the temporal features corresponding to the temporal samples.

[0033] Step S2: Extract features from the text data of the historical operation and maintenance report of the wind turbine to obtain text features.

[0034] Specifically, firstly, the text data of the historical operation and maintenance report of the wind turbine is preprocessed to obtain preprocessed text data; here, the preprocessing includes: removing special characters and redundant information.

[0035] Then, a pre-trained language model is used to extract features from the preprocessed text data to obtain the global semantic feature vector of the text. Then, the TF-IDF algorithm is used to extract keyword features from the preprocessed text data; Then, the global semantic feature vector of the text and the keyword features are concatenated to obtain the text features.

[0036] Step S3: The temporal features and textual features corresponding to each time series sample are fused to obtain the fused features corresponding to each time series sample.

[0037] Here, temporal features and textual features are fused to obtain complete temporal-semantic features suitable for wind power fault assessment tasks.

[0038] Step S4: Based on the fusion features, determine the class prototype for each wind turbine state category; construct a class prototype graph with class prototypes as nodes, and determine the edge weight between any two class prototypes in the class prototype graph based on the distance between the two class prototypes and the similarity of the category semantic description.

[0039] Here, the wind turbine status categories include normal, minor fault, and severe fault. Each wind turbine status category corresponds to multiple time series samples. The class prototype of each wind turbine status category is determined using the following formula:

[0040] in, Fan status category class prototype, Fan status category The number of time series samples under the current conditions. Fan status category Time series samples below The corresponding fusion features.

[0041] The edge weight between any two class prototypes in the class prototype graph can be calculated using the following formula:

[0042] in, The fan status category is The edge weight between the two class prototypes, The distance between the prototypes of the two classes. The similarity of the category semantic descriptions of two class prototypes. Fan status category semantic description, Fan status category Semantic description.

[0043] Step S5: Extract features from the time-series monitoring data of the wind turbine to be evaluated to obtain the time-series features to be evaluated; fuse the time-series features to be evaluated with the text features to obtain the fused features of the time-series monitoring data of the wind turbine to be evaluated.

[0044] Step S6: Calculate the Euclidean distance between the fusion features of the wind turbine time-series monitoring data to be evaluated and each class prototype in the class prototype map, determine the class prototype corresponding to the minimum value of the Euclidean distance, and use the wind turbine state category of the determined class prototype as the wind turbine state category of the wind turbine time-series monitoring data to be evaluated.

[0045] Step S7: If the wind turbine status category of the wind turbine time-series monitoring data to be evaluated is not the normal category, calculate the risk coefficient and determine the warning level based on the risk coefficient.

[0046] Here, if the wind turbine status category of the wind turbine time-series monitoring data to be evaluated is the normal category, then no warning is required.

[0047] In this embodiment, a prototype-based graph is constructed to model the fuzzy relationship between data and fault modes, thereby achieving precise and intelligent early warning under dynamic wind turbine fault risks. This is applicable to the predictive, preventive monitoring and maintenance of wind turbine equipment status in wind farms.

[0048] In one embodiment, step S3 involves fusing the temporal features and textual features corresponding to each time series sample to obtain the fused features corresponding to each time series sample, including: First, Z-score standardization is performed on the temporal features and text features respectively to obtain standardized temporal features and standardized text features;

[0049] in, For standardized features, These are time-series features or text features. The characteristic mean, The characteristic standard deviation; Then, a linear projection layer is used to map the standardized temporal features and standardized text features to the same dimension, resulting in mapped temporal features and mapped text features.

[0050]

[0051] in, For the mapped time series features, For the mapped text features, For standardized time series features, For standardized text features, , Let be the projection weight matrix. , This is a bias term.

[0052] Then, the mapped temporal features and the mapped text features are concatenated to obtain the concatenated features. ;

[0053] Then, splice the features. The fusion weights are calculated using an attention mechanism. ;

[0054] in, For global average pooling, These are learnable parameters.

[0055] Then, based on the fusion weights The mapped temporal features and mapped text features are weighted and fused to obtain the fused features. .

[0056]

[0057] In one embodiment, step S7, the method for calculating the risk coefficient includes: First, the deviation distance between the fused features of the time-series monitoring data of the wind turbine to be evaluated and the class prototype of the normal category is calculated using the following formula:

[0058] in, For the deviation distance, To evaluate the fusion characteristics of the wind turbine time-series monitoring data, The class prototype for the normal category. It is an L2 norm; Then, the deviation distance is converted into an initial risk coefficient using the following formula:

[0059] in, The initial risk coefficient, The risk coefficient weights are obtained by calibration using historical failure data. Then, the initial risk coefficient is corrected to obtain the corrected risk coefficient, using the following formula:

[0060] in, This is the corrected risk coefficient. The Euclidean distance between the fused features of the wind turbine time-series monitoring data to be evaluated and the class prototype of the corresponding wind turbine state category is given. This represents the maximum deviation distance for the corresponding wind turbine status category; it is an empirical value.

[0061] Specifically, step S7, determining the wind turbine failure level based on the risk factor, includes: If the risk coefficient This is a Level 3 warning; The warning threshold is the set value. like ≤ Risk coefficient It is a Level II warning; This is the emergency threshold, which is a set value. If the risk coefficient is ≥ This is a Level 1 warning.

[0062] In one embodiment, the wind turbine fault risk quantification and early warning method based on incremental prototype mapping further includes: If the minimum value of the Euclidean distance is greater than the set threshold, the class prototype graph is updated. The update process includes: updating the original class prototype, generating a new class prototype, and updating the edge weights in the class prototype graph. Figure 2 The diagram illustrates the principle of adaptive incremental update of class prototype graph.

[0063] Specifically, the existing class prototype is updated using the following formula:

[0064] in, The result of updating the original class prototype. For the original class prototype, The original class prototype weight, The fusion characteristics of the time-series monitoring data of the wind turbine to be evaluated; Generate new class prototypes, including: The fusion characteristics of the wind turbine time-series monitoring data to be evaluated As the initial value for the new class prototype, the prototype cross-entropy loss function is used to optimize the new class prototype, encouraging the incremental class prototype to maintain its distinctiveness from the existing class prototype, while optimizing the internal cohesion of the incremental class prototype, resulting in the final new class prototype. The prototype cross-entropy loss function is:

[0065] in, The prototype cross-entropy loss function is... This is the distance weighting coefficient. The Euclidean distance between the new class prototype and the original class prototype of the wind turbine state category c is given, where N is the number of the original class prototypes.

[0066] Update edge weights in the class prototype graph: Based on the updated class prototype graph, recalculate the distance and semantic similarity between all categories, and update the edge weights of the class prototype graph.

[0067] Employing the same inventive concept as the wind turbine fault risk quantification and early warning method based on incremental prototype graphs, this embodiment also provides a corresponding wind turbine fault risk quantification and early warning system based on incremental prototype graphs, including: The time-series feature extraction module is used to acquire historical time-series monitoring data of wind turbines. The historical time-series monitoring data of wind turbines includes multiple time-series samples. Feature extraction is performed on each time-series sample to obtain the time-series features corresponding to each time-series sample; each time-series sample corresponds to a wind turbine status category. The text feature extraction module is used to extract features from the text data of historical wind turbine operation and maintenance reports to obtain text features. The feature fusion module is used to fuse the temporal features and textual features corresponding to each time series sample to obtain the fused features corresponding to each time series sample. The class prototype graph construction module is used to determine the class prototype of each wind turbine state category based on the fused features; the class prototype graph is constructed with the class prototype as nodes, and the edge weight between any two class prototypes in the class prototype graph is determined based on the distance between the two class prototypes and the similarity of the category semantic description. The module for obtaining the feature to be evaluated is used to extract features from the time-series monitoring data of the wind turbine to be evaluated, and obtain the time-series features to be evaluated; the time-series features to be evaluated and the text features are fused to obtain the fused features of the time-series monitoring data of the wind turbine to be evaluated. The wind turbine status category determination module is used to calculate the Euclidean distance between the fusion features of the time-series monitoring data of the wind turbine to be evaluated and each class prototype in the class prototype map, determine the class prototype corresponding to the minimum value of the Euclidean distance, and use the wind turbine status category of the determined class prototype as the wind turbine status category of the time-series monitoring data of the wind turbine to be evaluated. The early warning module is used to calculate the risk coefficient and determine the early warning level based on the risk coefficient if the wind turbine status category in the time-series monitoring data to be evaluated is not the normal category.

[0068] The wind turbine fault risk quantification and early warning system based on incremental prototype graphs in this embodiment has the same inventive concept as the wind turbine fault risk quantification and early warning method based on incremental prototype graphs described above. Therefore, the specific implementation of this system can be found in the embodiment section of the wind turbine fault risk quantification and early warning method based on incremental prototype graphs described above, and its technical effects correspond to the technical effects of the above method, so it will not be repeated here.

[0069] In summary, this application solves the problems of low efficiency of fault warning labels, difficulty in modeling fault-data fuzzy relationships, low accuracy and delayed warning timing, and poor dynamic adaptability of existing technologies. It achieves accurate and intelligent early warning under dynamic wind turbine fault risk and is applicable to the predictive, preventive monitoring and maintenance of wind turbine equipment status in wind farms.

[0070] The above descriptions are merely various embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for quantitative early warning of wind turbine failure risk based on incremental prototype graphs, characterized in that, include: Historical time-series monitoring data of wind turbines is acquired. The historical time-series monitoring data of wind turbines includes multiple time-series samples. Feature extraction is performed on each time-series sample to obtain the time-series features corresponding to each time-series sample. Each time-series sample corresponds to a wind turbine status category. Feature extraction is performed on the text data of historical operation and maintenance reports of wind turbines to obtain text features; The temporal features and text features corresponding to each time series sample are fused to obtain the fused features corresponding to each time series sample. Based on the fusion features, determine the class prototype for each wind turbine state category; A class prototype graph is constructed using class prototypes as nodes. The edge weight between any two class prototypes in the class prototype graph is determined based on the distance between the two class prototypes and the similarity of the category semantic descriptions. Feature extraction is performed on the time-series monitoring data of the wind turbine to be evaluated to obtain the time-series features to be evaluated; the time-series features to be evaluated and the text features are fused to obtain the fused features of the time-series monitoring data of the wind turbine to be evaluated. Calculate the Euclidean distance between the fusion features of the wind turbine time-series monitoring data to be evaluated and each class prototype in the class prototype map, determine the class prototype corresponding to the minimum value of the Euclidean distance, and take the wind turbine state category of the determined class prototype as the wind turbine state category of the wind turbine time-series monitoring data to be evaluated. If the wind turbine status category of the wind turbine time-series monitoring data to be evaluated is not the normal category, then a risk coefficient is calculated, and the warning level is determined based on the risk coefficient.

2. The method as described in claim 1, characterized in that, in, Feature extraction is performed on each time series sample to obtain the time series features corresponding to each time series sample, including: The time series samples are D-dimensional data, and the time series samples are divided into multiple Patch data blocks; Perform 1D convolutional encoding on the D-dimensional data of each Patch data block to obtain local feature vectors; A multi-head attention mechanism is used to process local feature vectors to capture the correlation of features in different dimensions, and a global temporal feature vector for each patch data block is output through a fully connected layer. The global temporal feature vectors of all Patch data blocks are concatenated to form the temporal features corresponding to the temporal samples.

3. The method as described in claim 1, characterized in that, in, Feature extraction was performed on the text data of historical wind turbine operation and maintenance reports to obtain text features, including: Preprocess the text data of the historical operation and maintenance report of the wind turbine to obtain the preprocessed text data; The preprocessed text data is used to extract features using a pre-trained language model to obtain a global semantic feature vector of the text. The TF-IDF algorithm is used to extract keyword features from the preprocessed text data; The text global semantic feature vector and the keyword features are concatenated to obtain the text features.

4. The method as described in claim 1, characterized in that, in, The temporal features and text features corresponding to each time series sample are fused to obtain the fused features corresponding to each time series sample, including: Z-score standardization is applied to both temporal features and text features to obtain standardized temporal features and standardized text features. A linear projection layer is used to map the standardized temporal features and the standardized text features to the same dimension, resulting in mapped temporal features and mapped text features; By concatenating the mapped temporal features and the mapped text features, the concatenated features are obtained. The splicing features are calculated using an attention mechanism to obtain the fusion weights; The mapped temporal features and the mapped text features are weighted and fused according to the fusion weights to obtain fused features.

5. The method as described in claim 1, characterized in that, in, Based on the fusion features, the class prototype for each wind turbine state category is determined using the following formula: in, Fan status category class prototype, Fan status category The number of time series samples under the current conditions. Fan status category Time series samples below The corresponding fusion features.

6. The method as described in claim 1, characterized in that, The method for calculating the risk coefficient includes: The deviation distance between the fused features of the time-series monitoring data of the wind turbine to be evaluated and the class prototype of the normal category is calculated using the following formula: in, For the deviation distance, To evaluate the fusion characteristics of the wind turbine time-series monitoring data, The class prototype for the normal category. It is an L2 norm; The deviation distance is converted into an initial risk coefficient using the following formula: in, The initial risk coefficient, As the risk coefficient weight; The initial risk coefficient is corrected to obtain the corrected risk coefficient using the following formula: in, This is the corrected risk coefficient. The Euclidean distance between the fused features of the wind turbine time-series monitoring data to be evaluated and the class prototype of the corresponding wind turbine state category is given. This represents the maximum deviation distance for the corresponding wind turbine status category.

7. The method as described in claim 1, characterized in that, in, The wind turbine failure level is determined based on the aforementioned risk coefficient, including: If the risk coefficient This is a Level 3 warning; This is the warning threshold; like ≤ Risk coefficient It is a Level II warning; Emergency threshold; If the risk coefficient is ≥ This is a Level 1 warning.

8. The method as described in claim 1, characterized in that, The method further includes: If the minimum value of the Euclidean distance is greater than the set threshold, the class prototype graph is updated. The update process includes: updating the original class prototype, generating a new class prototype, and updating the edge weights in the class prototype graph.

9. The method as described in claim 8, characterized in that, in, To update an existing class prototype, use the following formula: in, The result of updating the original class prototype. For the original class prototype, For the original class prototype weight, The fusion characteristics of the time-series monitoring data of the wind turbine to be evaluated; Generate new class prototypes, including: The fusion characteristics of the wind turbine time-series monitoring data to be evaluated As the initial value for the new class prototype, the prototype cross-entropy loss function is used to optimize the new class prototype to obtain the final new class prototype. The prototype cross-entropy loss function is: in, The prototype cross-entropy loss function is... This is the distance weighting coefficient. The Euclidean distance between the new class prototype and the original class prototype of the wind turbine state category c is given, where N is the number of the original class prototypes.

10. A wind turbine fault risk quantification and early warning system based on incremental prototype graphs, characterized in that, include: The time-series feature extraction module is used to acquire historical time-series monitoring data of wind turbines. The historical time-series monitoring data of wind turbines includes multiple time-series samples. Feature extraction is performed on each time-series sample to obtain the time-series features corresponding to each time-series sample. Each time-series sample corresponds to a wind turbine status category. The text feature extraction module is used to extract features from the text data of historical wind turbine operation and maintenance reports to obtain text features. The feature fusion module is used to fuse the temporal features and text features corresponding to each time series sample to obtain the fused features corresponding to each time series sample. A class prototype graph construction module is used to determine the class prototype of each wind turbine state category based on the fused features; A class prototype graph is constructed using class prototypes as nodes. The edge weight between any two class prototypes in the class prototype graph is determined based on the distance between the two class prototypes and the similarity of the category semantic descriptions. The module for obtaining the feature to be evaluated is used to extract features from the time-series monitoring data of the wind turbine to be evaluated, thereby obtaining the time-series features to be evaluated; the time-series features to be evaluated and the text features are fused to obtain the fused features of the time-series monitoring data of the wind turbine to be evaluated. The wind turbine status category determination module is used to calculate the Euclidean distance between the fusion features of the time-series monitoring data of the wind turbine to be evaluated and each class prototype in the class prototype map, determine the class prototype corresponding to the minimum value of the Euclidean distance, and use the wind turbine status category of the determined class prototype as the wind turbine status category of the time-series monitoring data of the wind turbine to be evaluated. The early warning module is used to calculate a risk coefficient and determine the early warning level based on the risk coefficient if the wind turbine status category of the wind turbine time-series monitoring data to be evaluated is not a normal category.

Images