A wind turbine operating data processing method and system

Abstract

The application relates to the technical field of data processing, in particular to a wind turbine operation data processing method and system, which comprises the following steps: obtaining original wind turbine operation data; performing segmented processing on the original wind turbine operation data to obtain a plurality of operation data blocks; performing vibration time domain feature extraction on the operation data blocks to obtain time domain feature information; performing temperature change feature extraction on the operation data blocks to obtain temperature feature information; and performing wind speed power correlation feature extraction on the operation data blocks to obtain wind speed power feature information. The application can significantly enhance the distinguishing and prediction ability of the equipment operation state by performing detailed time domain feature extraction, temperature change feature extraction and wind speed power correlation feature extraction on the wind turbine operation data, and combining a deep learning model to perform multi-source data fusion, and ensure the accuracy and efficiency of the prediction analysis results for different states.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a wind turbine operation data processing method and system. BACKGROUND

[0002] The wind turbine is a device for generating electricity by using wind energy. It converts the kinetic energy of wind into electrical energy. During operation, a large amount of high-dimensional time-series operation data generated by integrated sensors constitutes a digital mirror of the device state. Effective calculation and analysis of the data is the key to realizing predictive maintenance, improving operation efficiency and ensuring safety.

[0003] The traditional processing method usually relies on single-dimensional features or simple statistics, and only analyzes the time-domain or frequency-domain features of vibration data, which is difficult to capture the complex coupling relationship between different physical quantities, resulting in insufficient effective information density of the input of the subsequent analysis model, affecting the accuracy of data classification and identification. The same set of analysis models are used in the prior art to process the data, and complex deep diagnostic algorithms are executed, resulting in a waste of huge computing resources and processing delay. Using a simple model cannot perform in-depth and accurate analysis on data that truly exists as a precursor to fault risk or anomaly. Due to the lack of pre-identification of the data state, the processing strategy cannot be dynamically adjusted, and light trend prediction can be performed on healthy data. Medium complexity warning analysis is started on potential risk data. Only the most resource-intensive deep analysis model is called for confirmed abnormal data.

[0004] Therefore, it is urgent to propose a wind turbine operation data processing method that can perform multi-dimensional feature fusion and state classification on the data, and then dynamically call different analysis models according to the classification results to improve the accuracy of trend prediction, potential risk prediction and risk evolution prediction, thereby achieving dual improvement of the efficiency and accuracy of industrial time-series data processing. SUMMARY

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a wind turbine operation data processing method, comprising:

[0006] obtaining original wind turbine operation data; segmenting the original wind turbine operation data to obtain a plurality of operation data blocks;

[0007] extracting vibration time-domain features from the operation data blocks to obtain time-domain feature information, extracting temperature change features from the operation data blocks to obtain temperature feature information, and extracting wind speed-power correlation features from the operation data blocks to obtain wind speed-power feature information;

[0008] According to the time domain feature information, the temperature feature information and the wind speed power feature information, the original wind turbine operation data is classified into a normal operation state, a failure risk precursor state or an abnormal operation state;

[0009] In response to the original wind turbine operation data being in the normal operation state, the original wind turbine operation data is input into a basic analysis model to perform trend prediction and health degree assessment;

[0010] In response to the original wind turbine operation data being in the failure risk precursor state, the original wind turbine operation data is input into a warning analysis model to perform abnormal trend prediction and potential risk prediction;

[0011] In response to the original wind turbine operation data being in the abnormal operation state, the original wind turbine operation data is input into a deep analysis model to perform multi-source data fusion and risk evolution prediction.

[0012] Preferably, time domain vibration feature extraction is performed on the operation data block to obtain time domain feature information; temperature change feature extraction is performed on the operation data block to obtain temperature feature information; and wind speed power correlation feature extraction is performed on the operation data block to obtain wind speed power feature information, including:

[0013] Time domain statistics of vibration signals in the operation data block are calculated; wherein the time domain statistics include maximum value, minimum value, average value and root mean square value; and the time domain feature information is obtained according to the time domain statistics;

[0014] A temperature change curve of temperature monitoring data in the operation data block is calculated, and the temperature change rate and the temperature fluctuation range are determined according to the temperature change curve;

[0015] A linear regression coefficient of the temperature monitoring data and the operation time is calculated to obtain a temperature and operation time correlation coefficient; and the temperature feature information is obtained according to the temperature change rate, the temperature fluctuation range and the temperature and operation time correlation coefficient;

[0016] A Pearson correlation coefficient of wind speed collection data and power output data in the operation data block is calculated;

[0017] A wind speed mutation moment is identified, and a power response delay time corresponding to the wind speed mutation moment is calculated;

[0018] Power output data above the rated wind speed is counted to obtain a power saturation degree; and the wind speed power feature information is obtained according to the Pearson correlation coefficient, the power response delay time and the power saturation degree.

[0019] Preferably, according to the time domain feature information, the temperature feature information and the wind speed power feature information, the original wind turbine operation data is classified into a normal operation state, a fault risk precursor state or an abnormal operation state, comprising:

[0020] The time domain feature information, the temperature feature information and the wind speed power feature information are spliced to obtain a fusion operation feature information set;

[0021] The fusion operation feature information set is used to generate a similarity measure corresponding to the original wind turbine operation data, wherein the similarity measure is used to represent the similarity between each operation data block in the original wind turbine operation data;

[0022] The fusion operation feature information set is input into a pre-trained feature extraction network, and the feature extraction network outputs a state feature matrix;

[0023] According to the state feature matrix and the similarity measure, at least one same operation state set and an operation state summary corresponding to each same operation state set are determined in the original wind turbine operation data, wherein the state feature matrix is used to represent the state distribution of each operation data block in the original wind turbine operation data in the entire original wind turbine operation data;

[0024] The number of each same operation state set is counted, and according to the operation state summary of each same operation state set, the operation state type corresponding to each same operation state set is determined;

[0025] Based on the number of operation state types, the original wind turbine operation data is classified into a normal operation state, a fault risk precursor state or an abnormal operation state.

[0026] Preferably, the original wind turbine operation data is trend predicted and health degree evaluated to obtain a corresponding state evaluation vector, comprising:

[0027] The original wind turbine operation data is subjected to sliding window smoothing processing to obtain smoothed data, and the smoothed data is input into a long short-term memory network model to extract time series features;

[0028] The time series features and the original wind turbine operation data are superimposed, and after superimposition, full connection layer mapping is performed to obtain the state evaluation vector.

[0029] Preferably, the original wind turbine operation data is subjected to abnormal trend prediction and potential risk prediction to obtain corresponding early warning event data, comprising:

[0030] The original wind turbine operation data is subjected to wavelet transform to obtain a multi-scale feature map;

[0031] inputting the multi-scale feature map into a convolutional neural network model to extract local abnormal features;

[0032] performing similarity matching between the local abnormal features and a preset fault risk pattern library to obtain the early warning event data.

[0033] Preferably, the original wind turbine operation data is subjected to multi-source data fusion and risk evolution prediction to obtain a corresponding data processing package, including:

[0034] performing feature standardization processing on the original wind turbine operation data to obtain standardized feature data;

[0035] inputting the standardized feature data into a graph neural network model to construct a data correlation graph;

[0036] performing graph convolution operation on the data correlation graph to extract global fault risk features;

[0037] inputting the global fault risk features into a classifier to obtain the data processing package.

[0038] Preferably, the joint training method of the basic analysis model, the early warning analysis model and the deep analysis model includes:

[0039] a plurality of training data sets are obtained; wherein the training data set includes sample operation data of a wind turbine in different operating states and corresponding state labels;

[0040] input configuration processing is respectively performed on an initial basic analysis model, an initial early warning analysis model and an initial deep analysis model to obtain an input configuration strategy of each model; the input configuration strategy of any model is used to indicate the state label that needs to be processed by the model in the training process;

[0041] based on the input configuration strategy of each model, the sample operation data is input into the corresponding initial model for training to obtain an output result of each model;

[0042] determine the model loss information of each model according to the state label and the output result of each model; the model loss information of any model is fitted by a trend prediction loss, an abnormal trend prediction loss and a risk prediction loss, the trend prediction loss is used to represent the category deviation of the predicted operating state and the actual state, the abnormal trend prediction loss is used to represent the position deviation of the detected abnormal point and the actual abnormal point, and the risk prediction loss is used to represent the category deviation of the identified fault risk type and the actual fault risk type;

[0043] According to the model loss information of each model, the initial basic analysis model, the initial early warning analysis model and the initial deep analysis model are respectively optimized in parameters to obtain the basic analysis model, the early warning analysis model and the deep analysis model.

[0044] Preferably, according to the state label and the output result of each model, the model loss information of each model is determined, including:

[0045] A first weight ratio between the trend prediction loss and the abnormal trend prediction loss, and a second weight ratio between the trend prediction loss and the risk prediction loss are obtained;

[0046] The contribution degree of the trend prediction loss in the fitting process is adjusted by using the first weight ratio and the second weight ratio;

[0047] The contribution degree of the abnormal trend prediction loss and the risk prediction loss in the fitting process is adjusted by using a third weight ratio between the abnormal trend prediction loss and the risk prediction loss.

[0048] Preferably, the parameter optimization of the initial basic analysis model, the initial early warning analysis model and the initial deep analysis model includes:

[0049] In response to the trend prediction loss in the model loss information being greater than a first preset threshold, the parameters related to the basic operation feature extraction in the initial basic analysis model are preferentially adjusted;

[0050] In response to the abnormal trend prediction loss in the model loss information being greater than a second preset threshold, the parameters related to the abnormal trend in the initial early warning analysis model are preferentially adjusted;

[0051] In response to the risk prediction loss in the model loss information being greater than a third preset threshold, the parameters related to the risk type identification in the initial deep analysis model are preferentially adjusted.

[0052] A wind turbine operation data processing system suitable for the wind turbine operation data processing method described above, comprising:

[0053] A data block unit is configured to obtain original wind turbine operation data, and segment the original wind turbine operation data to obtain a plurality of operation data blocks;

[0054] A feature extraction unit is configured to extract time domain vibration features from the operation data blocks to obtain time domain feature information, extract temperature change features from the operation data blocks to obtain temperature feature information, and extract wind speed-power correlation features from the operation data blocks to obtain wind speed-power feature information;

[0055] The running classification unit is used for classifying the original wind turbine running data into a normal running state, a failure risk precursor state or an abnormal running state according to the time domain feature information, the temperature feature information and the wind speed power feature information.

[0056] The trend prediction unit is used for inputting the original wind turbine running data into a basic analysis model, performing trend prediction and health degree evaluation in response to the original wind turbine running data being in the normal running state.

[0057] The failure warning unit is used for inputting the original wind turbine running data into a warning analysis model, performing abnormal trend prediction and potential risk prediction in response to the original wind turbine running data being in the failure risk precursor state.

[0058] The failure diagnosis unit is used for inputting the original wind turbine running data into a deep analysis model, performing multi-source data fusion and risk evolution prediction in response to the original wind turbine running data being in the abnormal running state.

[0059] Compared with the prior art, the present application has the following beneficial effects:

[0060] (1) The present application can effectively learn and utilize the internal coupling relationship between different physical quantities by performing detailed time domain feature extraction, temperature change feature extraction and wind speed power correlation feature extraction on the wind turbine running data, and combining a deep learning model for multi-source data fusion. Moreover, the time domain feature, the temperature feature and the wind speed power feature are comprehensively analyzed, which significantly enhances the distinguishing and prediction ability of the equipment running state, thereby improving the accuracy of the running state classification and future trend prediction.

[0061] (2) The present application jointly trains a basic analysis model, a warning analysis model and a deep analysis model, and dynamically optimizes according to the loss information of each model, which avoids the calculation redundancy of a single complex model processing all data, realizes the on-demand allocation and optimized scheduling of computing resources, significantly improves the timeliness of large-scale industrial time series data processing tasks, and ensures the accuracy of the prediction results for different states. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 It is a step flow diagram of the overall method in an embodiment of the present application.

[0063] Figure 2 It is a system architecture diagram of the overall system in an embodiment of the present application.

[0064] In the figure: 1, data block unit; 2, feature extraction unit; 3, running classification unit; 4, trend prediction unit; 5, failure warning unit; 6, failure diagnosis unit. DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0066] Embodiment one, please refer to Figure 1 The present application provides a technical solution: a wind turbine operation data processing method, comprising:

[0067] S1, obtaining original wind turbine operation data; segmenting the original wind turbine operation data to obtain a plurality of operation data blocks;

[0068] S2, extracting vibration time domain features from the operation data blocks to obtain time domain feature information; extracting temperature change features from the operation data blocks to obtain temperature feature information; extracting wind speed-power correlation features from the operation data blocks to obtain wind speed-power feature information;

[0069] S3, classifying the original wind turbine operation data into a normal operation state, a fault risk precursor state or an abnormal operation state according to the time domain feature information, the temperature feature information and the wind speed-power feature information;

[0070] S4, in response to the original wind turbine operation data being in the normal operation state, inputting the original wind turbine operation data into a basic analysis model to perform trend prediction and health degree assessment;

[0071] S5, in response to the original wind turbine operation data being in the fault risk precursor state, inputting the original wind turbine operation data into a warning analysis model to perform abnormal trend prediction and potential risk prediction;

[0072] S6, in response to the original wind turbine operation data being in the abnormal operation state, inputting the original wind turbine operation data into a deep analysis model to perform multi-source data fusion and risk evolution prediction.

[0073] It should be noted that the original wind turbine operation data includes vibration signal data, temperature monitoring data, wind speed collection data and power output data; the time domain feature information includes mean value, variance, peak factor and kurtosis coefficient of the vibration signal; the temperature feature information includes temperature change rate, temperature fluctuation range and correlation coefficient of temperature and operation time; the wind speed-power feature information includes Pearson correlation coefficient of wind speed and power, power response delay time when wind speed suddenly changes and power saturation degree above rated wind speed.

[0074] In an optional embodiment, the time-domain feature information is obtained by extracting the time-domain features of the operation data block; the temperature feature information is obtained by extracting the temperature variation features of the operation data block; and the wind speed-power feature information is obtained by extracting the wind speed-power correlation features of the operation data block, including:

[0075] calculating the time-domain statistics of the vibration signal in the operation data block; wherein the time-domain statistics include maximum value, minimum value, average value and root mean square value; and obtaining the time-domain feature information according to the time-domain statistics;

[0076] calculating the temperature variation curve of the temperature monitoring data in the operation data block, determining the temperature variation rate and the temperature fluctuation range according to the temperature variation curve;

[0077] calculating the linear regression coefficient of the temperature monitoring data and the operation time, obtaining the correlation coefficient of the temperature and the operation time; and obtaining the temperature feature information according to the temperature variation rate, the temperature fluctuation range and the correlation coefficient of the temperature and the operation time;

[0078] calculating the Pearson correlation coefficient of the wind speed acquisition data and the power output data in the operation data block;

[0079] identifying the wind speed mutation time, and calculating the power response delay time corresponding to the wind speed mutation time;

[0080] counting the power output data above the rated wind speed to obtain the power saturation degree; and obtaining the wind speed-power feature information according to the Pearson correlation coefficient, the power response delay time and the power saturation degree.

[0081] It should be noted that the maximum value refers to the maximum amplitude in the vibration signal, representing the maximum intensity of vibration; the minimum value refers to the minimum amplitude in the vibration signal, representing the minimum intensity of vibration; the average value refers to the mean value of the vibration signal, reflecting the overall average level of the signal; the root mean square value is the value obtained by squaring, averaging and then taking the square root of the vibration signal, representing the effective vibration intensity of the signal; RMS value is often used to evaluate the energy or intensity of the signal; according to these time domain statistics, comprehensive characteristics about the vibration signal can be obtained, so as to evaluate whether there is abnormal vibration of the unit, which helps to judge the health status of the equipment; the temperature change curve is drawn according to the temperature monitoring data in the operation data block, and the curve of temperature change with time is drawn; the temperature change rate refers to the speed of temperature change per unit time; it can be obtained by calculating the slope of the temperature curve, and too fast or too slow rate may mean that the equipment has a problem or is being disturbed by the outside world; the temperature fluctuation range refers to the difference between the maximum and minimum values of temperature change in a period of time; a larger fluctuation range may indicate that the temperature control system of the equipment is unstable; by linear regression analysis of temperature monitoring data and operation time, the regression coefficient is obtained; this can reflect the trend and law of temperature change of the equipment during operation; combined with the temperature change rate, the fluctuation range and the linear regression coefficient, characteristic information about the temperature change of the equipment can be obtained, so as to help identify whether there is temperature anomaly or equipment overheating problem; the Pearson correlation coefficient is a statistical measure of the strength of the linear relationship between two variables; here, it is used to calculate the correlation between wind speed collection data and power output data; if the correlation between wind speed and power is strong, it means that the working efficiency of the wind turbine generator is highly related to the change of wind speed; when the wind speed changes sharply, it may cause the power output of the unit to change; by identifying the moment of wind speed mutation, the reaction of the unit in response to wind speed change can be observed; the power response delay time is the delay time of the change of the power output of the unit after the wind speed mutation occurs; a long delay time may indicate that the unit reacts slowly, which may indicate that the equipment has a slow response problem; the power saturation degree refers to the power output of the unit tending to saturation when the wind speed reaches a certain value (i.e. above the rated wind speed), and no longer increases significantly with the increase of wind speed; the power saturation degree describes the stability of the power output when the wind speed exceeds the rated value; combined with the Pearson correlation coefficient, the power response delay time and the power saturation degree, the working characteristics of the unit under different wind speed conditions can be extracted; for example, whether it can respond quickly to wind speed change, whether it is prone to power saturation, etc.

[0082] In an optional embodiment, according to the time domain feature information, the temperature feature information and the wind speed power feature information, the original wind turbine generator operation data is classified into a normal operation state, a fault risk precursor state or an abnormal operation state, comprising:

[0083] The time domain feature information, the temperature feature information and the wind speed power feature information are spliced to obtain a fusion operation feature information set;

[0084] A similarity measure corresponding to the original wind turbine operation data is generated based on the fusion operation feature information set, wherein the similarity measure is used to represent the similarity between each operation data block in the original wind turbine operation data;

[0085] The fusion operation feature information set is input into a pre-trained feature extraction network, and the feature extraction network outputs a state feature matrix;

[0086] At least one same-class operation state set and an operation state summary corresponding to each same-class operation state set are determined in the original wind turbine operation data according to the state feature matrix and the similarity measure; wherein the state feature matrix is used to represent the state distribution of each operation data block in the original wind turbine operation data in the entire original wind turbine operation data;

[0087] The number of each same-class operation state set is counted; and the operation state type corresponding to each same-class operation state set is determined according to the operation state summary of each same-class operation state set;

[0088] The original wind turbine operation data is classified into a normal operation state, a fault risk precursor state or an abnormal operation state based on the number of operation state types.

[0089] It should be noted that the three types of feature information are fused (spliced) to obtain a comprehensive "fusion running feature information set"; this set contains features extracted from multiple dimensions (vibration, temperature, wind speed power, etc.), which helps to comprehensively describe the running state of the wind turbine; by analyzing the fusion running feature information set, the similarity between each running data block is calculated; the similarity measure is used to measure the similarity between wind turbine data in different time periods or different running states; for example, Euclidean distance, cosine similarity, etc. Method can be used to measure the similarity of different data blocks; similarity measure helps to identify which data blocks exhibit similar features among multiple running data blocks of the wind turbine; this is very important for classifying and clustering different running states of the wind turbine; the fusion running feature information set is input into a pre-trained feature extraction network; this network is a deep learning model that can extract higher-order, abstract features from the fusion features after training; the role of the feature extraction network is to convert the input data into a state feature matrix, which represents the features of the wind turbine in different running states; the state feature matrix is a high-dimensional matrix, where each row represents the features of a data block, and the columns represent different feature dimensions; this matrix describes the state distribution of different data blocks during the entire running process of the wind turbine, helping to identify the running state corresponding to each data block; according to the state feature matrix and the similarity measure, at least one "same running state set" can be identified among all running data; these sets are composed of similar running data blocks, reflecting the similar state of the wind turbine in a specific time period; each same running state set will have a corresponding "running state summary", which is the representative feature or average feature of the set; for example, the average temperature, wind speed, vibration, etc. of a certain set can be used as the representative state of the set; count the number of all identified same running state sets to understand the distribution of each state set; by counting the number of these sets, it can be determined whether the unit is in a normal running state, a fault risk precursor state, or an abnormal running state.

[0090] In an optional embodiment, trend prediction and health assessment are performed on the original wind turbine running data to obtain a corresponding state assessment vector, including:

[0091] The original wind turbine running data is subjected to sliding window smoothing to obtain smoothed data;

[0092] The smoothed data is input into a long short-term memory network model to extract time series features;

[0093] The time series features are superimposed on the original wind turbine running data, and after superimposition, a full connection layer mapping is performed to obtain a state assessment vector.

[0094] It is worth noting that the sliding window is a common signal processing technique used to reduce noise and fluctuations in data, helping to extract smoother trends; it works by setting a window size, sliding over the original data, calculating the mean, weighted average or other statistics of data points within each window; in this way, short-term fluctuations can be smoothed out, retaining long-term trends in the data; smoothed data means that high-frequency noise and sudden fluctuations in the data have been removed, making it more stable for subsequent analysis; usually, this smoothing process helps to make the data more reliable and consistent when analyzing, avoiding the impact of short-term anomalies on the overall model training; Long Short-Term Memory (LSTM) LSTM (Long Short-Term Memory) is a special type of recurrent neural network (RNN) used to process and predict time series-based data; it can learn long-term dependencies in sequential data, effectively avoiding the gradient vanishing problem that traditional RNNs may encounter in long sequence processing; LSTM is very suitable for processing wind turbine operation data with time continuity; input the smoothed wind turbine operation data into the LSTM model, and the model will automatically learn and extract the time series features in the data; for example, LSTM can capture the dynamic changes of wind turbine operation in different time periods and find potential patterns and periodic changes; superimpose the time series features extracted by LSTM and the original wind turbine operation data, that is, combine these two pieces of information; usually, the features extracted by LSTM are implicit features learned through time series, which can reflect the trends and patterns of the data; combining these features with the original data can help the model better understand the specific operation of the wind turbine and enhance the model's expressive ability; the superposition operation not only retains the detailed information of the original data, but also incorporates the time series features extracted by LSTM, allowing the subsequent model to more comprehensively assess the operation status of the wind turbine; In deep neural networks, the fully connected layer (Fully Connected Layer) refers to the connection between each input neuron and output neuron; here, the superimposed data will be passed to the fully connected layer for processing; the fully connected layer is usually used to map the extracted features to a high-dimensional space, helping to further analyze the features of the data and produce the final prediction or assessment results; through the mapping of the fully connected layer, the model will produce an output based on the input data (including time series features and original data); the output can be a classification result (such as normal, fault risk, anomaly) or a continuous value (such as fault risk probability, health score, etc.).

[0095] In an optional embodiment, the original wind turbine operation data is subjected to abnormal trend prediction and potential risk prediction to obtain corresponding early warning event data, including:

[0096] Wavelet transform is performed on original wind turbine operation data to obtain a multi-scale feature map;

[0097] The multi-scale feature map is input into a convolutional neural network model to extract local abnormal features;

[0098] The local abnormal features are matched with a preset fault risk pattern library in terms of similarity to obtain early warning event data.

[0099] It should be noted that wavelet transform is an effective signal processing technology, which can extract the local characteristics of signals in different frequency and time domain through multi-scale decomposition; Unlike traditional Fourier transform, wavelet transform can provide local information of signals in time and frequency domain, which is very useful for analyzing non-stationary signals such as wind turbine operation data; Wavelet transform decomposes signals by changing the scale (or frequency) to obtain features of different scales; In the operation data of wind turbine, there may be fluctuations of different frequency and duration (such as short-term transient fault risk or long-term systemic problem), and wavelet transform can help extract features of different scales; The multi-scale feature map obtained is the result of collecting the features of signals in multiple scales, which can reflect the changes at different levels of data; Convolutional neural network (CNN) is a powerful deep learning model widely used in image processing, speech recognition and other tasks; In this scenario, CNN is used to process the multi-scale feature map obtained by wavelet transform; CNN can automatically extract useful features from input data through convolutional, pooling and fully connected layers, and further perform pattern recognition and classification; Through the convolution operation of CNN, the model can learn the local features in the data, which may include abnormal fluctuation patterns in the operation data of wind turbine; For example, there may be sudden abnormal fluctuations in the vibration signal of wind turbine or abnormal fluctuations in its operating parameters, and CNN can extract these local abnormal features from the multi-scale feature map; Through the filtering operation of convolutional layer, CNN can efficiently identify these local patterns different from normal operation state; The fault risk pattern library is a "database" containing different types of wind turbine fault risks, which contains the feature description of various possible fault risk patterns (such as mechanical fault risk, electrical fault risk, etc.); These fault risk patterns are usually obtained through historical data, expert knowledge or experimental research, which can help the system identify which patterns are related to a certain specific fault risk; After extracting local abnormal features, the model compares these features with the pre-set fault risk pattern library and calculates the similarity; This process is usually completed by calculating the distance between features (such as Euclidean distance, cosine similarity, etc.); If the extracted local abnormal features are highly similar to the features of a certain fault risk pattern in the library, it can be determined that the wind turbine may have the corresponding fault risk; The result of similarity matching can provide the basis for subsequent fault risk diagnosis; Through similarity matching, the system will generate early warning event data; If the detected local abnormal features have high similarity with the features of a certain fault risk pattern, the system will issue a warning signal to indicate that the wind turbine may have a specific fault risk.

[0100] In an optional embodiment, multi-source data fusion and risk evolution prediction are performed on the original wind turbine operation data to obtain corresponding data processing packages, including:

[0101] The original wind turbine operation data is subjected to feature standardization to obtain standardized feature data;

[0102] The standardized feature data is input into a graph neural network model to construct a data correlation graph;

[0103] The data correlation graph is subjected to graph convolution operation to extract global fault risk features;

[0104] The global fault risk features are input into a classifier to obtain a data processing package.

[0105] Note that the graph neural network (GNN) is a deep learning method specifically designed for processing graph-structured data; graph-structured data includes nodes (data points) and edges (relationships between nodes); in the fault risk diagnosis of wind turbines, the data collected by each sensor can be regarded as a node, and the relationship between different sensors (such as location adjacency, signal correlation, etc.) can be regarded as an edge to form a graph; by linking the standardized feature data with the relationship between sensors, we can construct a "data correlation graph"; in this graph, the nodes of the graph represent different sensors, and the edges represent the relationship between the sensors (such as the adjacent relationship of physical location or signal correlation); this graph structure can capture the spatial and signal correlation between sensors, so that the model can learn the mutual influence between different sensors through the structure of the graph; graph convolution (Graph Convolution) is a core operation in GNN, which is used for feature propagation and learning on graph data; traditional convolutional neural networks (CNN) operate on regular grid data, while graph convolution can work on unstructured graph data and propagate features through the connection relationship between nodes; through the graph convolution operation, the model can pass information and fuse the features of multiple nodes according to the correlation between sensors, and extract global fault risk features; these global features can include the overall operating state of the wind turbine, helping the model to identify abnormal patterns of the entire system, rather than just local fault risk information; through the graph convolution layer, the relationship between nodes is passed and updated layer by layer, so as to extract global fault risk features (such as the overall fault risk signal pattern of the system); the classifier is a model used for classification decision based on input features; common classifiers include support vector machines (SVM), decision trees, neural networks, etc.; here, the global fault risk features after graph convolution are passed to the classifier as input for fault risk type prediction; the output of the classifier is the data processing package of the wind turbine; by analyzing the global features extracted by the graph convolution network, the classifier can determine whether the wind turbine has a fault risk and determine the fault risk type (such as motor fault risk, transmission system fault risk, etc.); the final result will help maintenance personnel to timely discover the fault risk of the wind turbine and take appropriate maintenance measures, thereby improving the reliability and maintenance efficiency of the wind turbine.

[0106] In an optional embodiment, the joint training method of the basic analysis model, the early warning analysis model and the deep analysis model comprises:

[0107] A plurality of training data sets are obtained; wherein the training data set includes sample operating data of the wind turbine in different operating states and corresponding state labels;

[0108] The initial basic analysis model, the initial early warning analysis model and the initial deep analysis model are respectively subjected to input configuration processing to obtain an input configuration strategy of each model; and the input configuration strategy of any one model is used to indicate a state label to be processed by the model in a training process;

[0109] Based on the input configuration strategy of each model, sample operation data is input into the corresponding initial model for training to obtain an output result of each model;

[0110] According to the state label and the output result of each model, model loss information of each model is determined; the model loss information of any one model is fitted by a trend prediction loss, an abnormal trend prediction loss and a risk prediction loss, the trend prediction loss is used to represent a category deviation of a predicted operation state of the model and an actual state, the abnormal trend prediction loss is used to represent a position deviation of a detected abnormal point of the model and an actual abnormal point, and the risk prediction loss is used to represent a category deviation of a recognized fault risk type of the model and an actual fault risk type;

[0111] According to the model loss information of each model, the initial basic analysis model, the initial early warning analysis model and the initial deep analysis model are respectively subjected to parameter optimization to obtain a basic analysis model, an early warning analysis model and a deep analysis model.

[0112] It should be noted that the initial basic analysis model, the initial early warning analysis model, and the initial deep analysis model: these are the initial versions of the models, respectively for basic analysis, early warning analysis, and deep diagnosis; each model is designed for different needs of wind turbines, and the basic analysis model is usually used for simple analysis of the running state of the wind turbine; the early warning analysis model is used to predict whether the wind turbine is about to have a fault risk or anomaly, and to issue an early warning; the deep analysis model is used to more accurately diagnose the fault risk, analyze the specific fault risk type and location; for each model, determine which input data needs to be received by the model, and configure the training of the model based on these data; the input configuration strategy of each model indicates the state labels that the model needs to handle during training; for example, the basic analysis model may only care about two state labels, normal and fault risk, while the deep analysis model may need more detailed labels, such as different types of fault risk types; according to the input configuration strategy described above, the running data of the sample is input into the corresponding initial model for training; in this stage, the model will learn from the given data and labels, adjust its internal parameters, so that it can gradually fit the patterns and rules in the data; the loss function is an indicator that measures the difference between the model's predicted results and the true results; each model will calculate the loss value according to its output results and corresponding state labels; the trend prediction loss is used to measure the deviation between the model's predicted running state and the actual state; it focuses on whether the model accurately reflects the actual running state change in the time series; for example, can the model correctly predict the future running state of the equipment (such as from "normal" to "fault risk"); the abnormal trend prediction loss is used to measure the positional deviation between the model's detection of abnormal points and the actual abnormal points; if the model can correctly detect the abnormalities in the wind turbine (such as sudden changes in current, voltage, etc.), the abnormal trend prediction loss will be small; if the model fails to correctly identify the anomaly, the loss will increase; the risk prediction loss is used to measure the deviation between the model's identified fault risk type and the actual fault risk type; risk prediction is the core of this process, and the model needs to correctly identify different fault risk types (such as motor fault risk, transmission system fault risk, etc.); if the model incorrectly classifies the fault risk type, the loss will increase; based on the loss information of each model, adjust the parameters of the model through optimization algorithms (such as gradient descent) to make it perform better on the training set and reduce the loss value; through optimization, the model can improve accuracy, reduce errors, and better complete the task; adjust the basic analysis model through the state labels in the training data to better distinguish between normal and fault risk states; optimize the early warning analysis model through trend and anomaly information in the training data to predict potential fault risks of the wind turbine in advance; optimize the deep analysis model through risk prediction labels in the training data to accurately diagnose the fault risk type of the wind turbine;The basic analysis model, the early warning analysis model and the deep analysis model become more accurate versions after parameter optimization. The optimized models can perform more accurate analysis, early warning and fault risk diagnosis according to different operation data. Each optimized model has different tasks: the basic analysis model is responsible for the preliminary judgment of the operation state of the equipment, the early warning analysis model can discover potential fault risks in advance, and the deep analysis model can in-depth analyze and accurately diagnose the specific fault risk types.

[0113] In an optional embodiment, according to the state label and the output result of each model, model loss information of each model is determined, including:

[0114] A first weight ratio between the trend prediction loss and the abnormal trend prediction loss, and a second weight ratio between the trend prediction loss and the risk prediction loss are obtained;

[0115] The first weight ratio and the second weight ratio are used to adjust the contribution degree of the trend prediction loss in the fitting process;

[0116] A third weight ratio between the abnormal trend prediction loss and the risk prediction loss is used to adjust the contribution degree of the abnormal trend prediction loss and the risk prediction loss in the fitting process.

[0117] It should be noted that the first weight ratio is the ratio between the trend prediction loss and the abnormal trend prediction loss, which is used to indicate the relative importance of the two loss terms in model training; for example, if the first weight ratio is 2, the contribution of the trend prediction loss will be twice that of the abnormal trend prediction loss; this weight ratio is set by analyzing the relative importance of each task, the training needs of the model, data characteristics and other factors; the second weight ratio is the ratio between the trend prediction loss and the risk prediction loss, which determines the relative importance of the two in the total loss; for example, if the second weight ratio is 1.5, it means that the contribution of the trend prediction loss in training is 1.5 times that of the risk prediction loss; this ratio needs to be adjusted according to the training target and the focus of the task, for example, if diagnosing fault risk is more important than predicting trends, the weight of the risk prediction loss may be higher; based on the values of the first weight ratio and the second weight ratio, the weight of the trend prediction loss in the fitting process is adjusted; the fitting process is the process of model learning, in which the optimization algorithm tries to minimize the total loss function; according to the first weight ratio and the second weight ratio, the weight of the trend prediction loss may increase or decrease accordingly; for example, if the first weight ratio is larger (indicating that abnormal trend prediction is less important), the trend prediction loss will have a larger weight in the model optimization process, and the model will pay more attention to the accuracy of predicting trends; the third weight ratio represents the ratio between the abnormal trend prediction loss and the risk prediction loss, which is used to balance the contribution of the two tasks in the loss function; for example, if the third weight ratio is 2, it means that the weight of the abnormal trend prediction task is larger than that of the risk prediction task; adjusting the contribution of the two losses is to ensure the balance of the model in detecting abnormalities and identifying fault risks, and is set according to the final application scenario of the model; for example, if the accuracy of detecting abnormalities is more critical than identifying fault risk types, the weight of the abnormal trend prediction loss can be appropriately increased; in the training process, the final form of the loss function is to combine each loss term according to certain weights; for multi-task learning, the common practice is to give each loss term a weight, which reflects the importance of the tasks; according to the required task optimization target, the weights are gradually adjusted so that the model can correctly focus on the most important task during training; for example, if the model performs poorly in identifying fault risks, the weight of the risk prediction loss can be increased; if trend prediction is more important than abnormal trend prediction, the weight of the trend prediction loss is increased.

[0118] In an optional embodiment, the initial basic analysis model, the initial early warning analysis model and the initial deep analysis model are parameter optimized, including:

[0119] In response to the trend prediction loss in the model loss information being greater than a first preset threshold, the parameters related to the basic operation feature extraction in the initial basic analysis model are preferentially adjusted;

[0120] In response to the abnormal trend prediction loss in the model loss information being greater than a second preset threshold, parameters related to abnormal point detection in the initial early warning analysis model are preferentially adjusted;

[0121] In response to the risk prediction loss in the model loss information being greater than a third preset threshold, parameters related to fault risk type identification in the initial deep analysis model are preferentially adjusted.

[0122] It should be noted that the first preset threshold is a preset threshold, which is usually determined by experiment or experience; if the trend prediction loss exceeds this threshold, it indicates that the performance of the model in this respect is not up to standard; at this time, the parameters related to the basic running feature extraction are preferentially adjusted, such as the data preprocessing method, feature selection, and feature engineering method; by optimizing these features, the trend prediction ability of the model is improved, thereby reducing the loss; the second preset threshold is used to determine whether the abnormal trend prediction loss is too large; if the loss is greater than the second threshold, it indicates that the abnormal trend prediction function of the model is insufficient; abnormal point detection refers to identifying points in the data that do not conform to the normal mode; possible reasons include device fault risk, environmental abnormalities, etc.; when the abnormal trend prediction loss exceeds the preset threshold, the parameters related to abnormal point detection are preferentially adjusted, such as selecting a different abnormal trend prediction algorithm (such as a statistical method or a machine learning method), optimizing the sensitivity of abnormal trend prediction, adjusting the threshold of the abnormal trend prediction model, etc.; the risk prediction loss measures the performance of the model in identifying specific fault risk types; a larger loss means that the accuracy of the model in the risk prediction task is poor; the third preset threshold is a preset threshold used to determine whether the risk prediction task needs to be optimized; if the loss exceeds the third threshold, it indicates that the risk prediction effect is not ideal; fault risk type identification refers to identifying what type of fault risk the device has (e.g., battery fault risk, transmission system fault risk, etc.); each fault risk type has different symptoms, and the model needs to make accurate classification based on the input data; when the risk prediction loss is too large, the parameters related to fault risk type identification are preferentially adjusted, such as optimizing the structure of the classifier, adjusting the labels of the fault risk categories, using a more appropriate loss function, increasing the sample size of the fault risk data, etc.; this can improve the ability of the model to identify fault risk types, thereby reducing the loss.

[0123] Embodiment two, please refer to Figure 2 The present application provides a technical solution: a wind turbine operation data processing system, which is suitable for the wind turbine operation data processing method described above, comprising:

[0124] The data block unit 1 is used for obtaining original wind turbine operation data; the original wind turbine operation data is segmented and processed to obtain a plurality of operation data blocks;

[0125] The feature extraction unit 2 is configured to perform vibration time-domain feature extraction on the operation data block to obtain time-domain feature information, temperature change feature extraction on the operation data block to obtain temperature feature information, and wind speed-power correlation feature extraction on the operation data block to obtain wind speed-power feature information.

[0126] The operation classification unit 3 is configured to classify the original wind turbine operation data into a normal operation state, a fault risk precursor state or an abnormal operation state according to the time-domain feature information, the temperature feature information and the wind speed-power feature information.

[0127] The trend prediction unit 4 is configured to, in response to the original wind turbine operation data being in the normal operation state, input the original wind turbine operation data into a basic analysis model to perform trend prediction and health degree evaluation on the original wind turbine operation data, and obtain a corresponding state evaluation vector.

[0128] The fault early warning unit 5 is configured to, in response to the original wind turbine operation data being in the fault risk precursor state, input the original wind turbine operation data into a pre-warning analysis model to perform trend prediction and health degree evaluation.

[0129] The fault diagnosis unit 6 is configured to, in response to the original wind turbine operation data being in the abnormal operation state, input the original wind turbine operation data into a deep analysis model to perform multi-source data fusion and risk evolution prediction.

[0130] The embodiments of the present application are described in detail above in combination with the drawings, but the present application is not limited thereto, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present application.

Claims

1. A method for processing wind turbine operating data, characterized in that... ,include: Obtain raw wind turbine operating data; segment the raw wind turbine operating data to obtain multiple operating data blocks; Vibration time-domain features are extracted from the running data block to obtain time-domain feature information; temperature change features are extracted from the running data block to obtain temperature feature information; wind speed-power correlation features are extracted from the running data block to obtain wind speed-power feature information. Based on the time-domain feature information, the temperature feature information, and the wind speed and power feature information, the original wind turbine operating data is classified into normal operating state, fault risk precursor state, or abnormal operating state. In response to the original wind turbine operating data indicating a normal operating state, the original wind turbine operating data is input into the basic analysis model to perform trend prediction and health assessment. In response to the original wind turbine operating data being a precursor to the fault risk, the original wind turbine operating data is input into the early warning analysis model to perform abnormal trend prediction and potential risk prediction. In response to the abnormal operating state of the original wind turbine operating data, the original wind turbine operating data is input into the deep analysis model to perform multi-source data fusion and risk evolution prediction. Based on the time-domain feature information, the temperature feature information, and the wind speed and power feature information, the original wind turbine operating data is classified into normal operating state, fault risk precursor state, or abnormal operating state, including: The time-domain feature information, the temperature feature information, and the wind speed and power feature information are concatenated to obtain a fused operational feature information set. A similarity metric is generated based on the fused operational feature information set to represent the degree of similarity between the original wind turbine operational data. The fused running feature information set is input into a pre-trained feature extraction network, and the feature extraction network outputs a state feature matrix; Based on the state feature matrix and the similarity metric, at least one set of similar operating states and a summary of the operating states corresponding to each set of similar operating states are determined in the original wind turbine operating data; wherein, the state feature matrix is ​​used to represent the state distribution of each operating data block in the original wind turbine operating data in the entire original wind turbine operating data; Count the number of each type of running state set; determine the running state type corresponding to each type of running state set based on the running state summary of each type of running state set; Based on the number of the aforementioned operating status types, the original wind turbine operating data is classified into normal operating status, fault risk precursor status, or abnormal operating status.

2. The wind turbine operation data processing method according to claim 1, characterized in that... Vibration time-domain features are extracted from the running data block to obtain time-domain feature information; Temperature change features are extracted from the running data block to obtain temperature feature information; Wind speed and power correlation feature extraction is performed on the running data block to obtain wind speed and power feature information, including: Calculate the time-domain statistics of the vibration signal in the running data block; wherein, the time-domain statistics include the maximum value, minimum value, average value and root mean square value; obtain the time-domain feature information based on the time-domain statistics; Calculate the temperature change curve of the temperature monitoring data in the running data block, and determine the temperature change rate and temperature fluctuation range based on the temperature change curve; Calculate the linear regression coefficient between the temperature monitoring data and the running time to obtain the correlation coefficient between temperature and running time; obtain the temperature characteristic information based on the temperature change rate, the temperature fluctuation range, and the correlation coefficient between temperature and running time. Calculate the Pearson correlation coefficient between the wind speed acquisition data and the power output data in the aforementioned running data block; Identify the moment of sudden wind speed change and calculate the power response delay time corresponding to the moment of sudden wind speed change; statistically analyze the power output data above the rated wind speed to obtain the power saturation; obtain the wind speed power characteristic information based on the Pearson correlation coefficient, the power response delay time and the power saturation.

3. The wind turbine operation data processing method according to claim 2, characterized in that, The original wind turbine operating data is used for trend prediction and health assessment to obtain a corresponding state assessment vector, including: The original wind turbine operating data is smoothed using a sliding window process to obtain smoothed data; the smoothed data is then input into a long short-term memory network model to extract time series features. The time series features are superimposed with the original wind turbine operating data, and then mapped using a fully connected layer to obtain the state evaluation vector.

4. The wind turbine operation data processing method according to claim 3, characterized in that, The original wind turbine operating data is used to predict abnormal trends and potential risks, resulting in corresponding early warning event data, including: Wavelet transform is performed on the original wind turbine operating data to obtain a multi-scale feature map; The multi-scale feature map is input into a convolutional neural network model to extract local anomaly features; The local anomaly features are matched with a preset fault risk pattern library to obtain the early warning event data.

5. The wind turbine operation data processing method according to claim 4, characterized in that, The original wind turbine operating data is subjected to multi-source data fusion and risk evolution prediction to obtain a corresponding data processing package, including: The original wind turbine operating data is subjected to feature standardization processing to obtain standardized feature data; The standardized feature data is input into a graph neural network model to construct a data association graph; Perform graph convolution operation on the data association graph to extract global fault risk features; The global fault risk features are input into the classifier to obtain the data processing package.

6. The wind turbine operation data processing method according to claim 5, characterized in that, The joint training method for the basic analysis model, the early warning analysis model, and the deep analysis model includes: Obtain multiple sets of training datasets; wherein, the training datasets include sample operating data of wind turbines under different operating states and corresponding state labels; Input configuration processing is performed on the initial basic analysis model, the initial early warning analysis model, and the initial deep analysis model respectively to obtain the input configuration strategy for each model; the input configuration strategy of any model is used to indicate the state labels that the model needs to process during training. Based on the input configuration strategy of each model, the sample running data is input into the corresponding initial model for training, and the output results of each model are obtained. Based on the state label and output results of each model, the model loss information of each model is determined. The model loss information of any model is obtained by fitting the trend prediction loss, the abnormal trend prediction loss and the risk prediction loss. The trend prediction loss is used to represent the category deviation between the model's predicted running state and the actual state. The abnormal trend prediction loss is used to represent the position deviation between the abnormal points detected by the model and the actual abnormal points. The risk prediction loss is used to represent the category deviation between the fault risk type identified by the model and the actual fault risk type. Based on the model loss information of each model, the parameters of the initial basic analysis model, the initial early warning analysis model, and the initial deep analysis model are optimized respectively to obtain the basic analysis model, the early warning analysis model, and the deep analysis model.

7. A method for processing wind turbine operation data according to claim 6, characterized in that, Based on the state labels and output results of each model, determine the model loss information for each model, including: Obtain a first weighting ratio between the trend prediction loss and the abnormal trend prediction loss, and a second weighting ratio between the trend prediction loss and the risk prediction loss; The contribution of the trend prediction loss to the fitting process is adjusted using the first weight ratio and the second weight ratio. The contribution of the abnormal trend prediction loss and the risk prediction loss in the fitting process is adjusted by using the third weight ratio between them.

8. The wind turbine operation data processing method according to claim 7, characterized in that, The parameters of the initial basic analysis model, the initial early warning analysis model, and the initial deep analysis model are optimized, including: If the trend prediction loss in the model loss information is greater than a first preset threshold, then the parameters related to the extraction of basic operating features in the initial basic analysis model are adjusted first. If the abnormal trend prediction loss in the model loss information is greater than the second preset threshold, the parameters related to the abnormal trend in the initial early warning analysis model will be adjusted first. If the risk prediction loss in the model loss information is greater than a third preset threshold, the parameters related to risk type identification in the initial deep analysis model will be adjusted first.

9. A wind turbine operation data processing system, applicable to the wind turbine operation data processing method according to any one of claims 1-8, characterized in that, include: The data segmentation unit is used to acquire raw wind turbine operating data; The original wind turbine operating data is segmented to obtain multiple operating data blocks; The feature extraction unit is used to extract vibration time-domain features from the running data block to obtain time-domain feature information; Temperature change features are extracted from the running data block to obtain temperature feature information; Wind speed and power correlation features are extracted from the running data block to obtain wind speed and power feature information; The operation classification unit is used to classify the original wind turbine operation data into normal operation status, fault risk precursor status or abnormal operation status based on the time domain feature information, the temperature feature information and the wind speed and power feature information. The trend prediction unit is used to input the original wind turbine operating data into the basic analysis model in response to the original wind turbine operating data being in the normal operating state, and to perform trend prediction and health assessment. The fault early warning unit is used to respond to the original wind turbine operating data as a precursor to the fault risk state by inputting the original wind turbine operating data into the early warning analysis model and performing abnormal trend prediction and potential risk prediction. The fault diagnosis unit is used to respond to the abnormal operating state of the original wind turbine operating data by inputting the original wind turbine operating data into the deep analysis model and performing multi-source data fusion and risk evolution prediction.

Images