Wind power generation monitoring data acquisition clustering method, medium and system

Abstract

The invention provides a wind power generation monitoring data acquisition clustering method, medium and system, and belongs to the technical field of wind power generation monitoring data acquisition, and the method comprises the steps: setting an initial sampling frequency based on the operation parameters of a wind generating set, collecting and preprocessing the initial sampling frequency to obtain preprocessed data, and carrying out the component decomposition to obtain a stable component and a variable component; establishing an acquisition data quality evaluation matrix to calculate an acquisition contribution value, dividing the data into acquisition data clusters by adopting a multi-layer lightweight clustering method, performing multi-level feature extraction and fusion through a plurality of lightweight neural networks by adopting the multi-layer lightweight clustering method, determining an optimal sampling time window length based on the acquisition data clusters, and performing multi-level feature extraction and fusion based on the optimal sampling time window length. And the sampling interval is updated and applied to monitoring data acquisition of the wind generating set, so that the technical problem that differential sampling analysis of data with different importance degrees is difficult to realize due to the fact that acquisition and clustering of the wind power generation monitoring data are split in the prior art is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power generation monitoring data collection, and in particular, relates to a wind power generation monitoring data collection clustering method, medium and system. Background Art

[0002] As the proportion of wind power generation in the energy structure continues to increase, the importance of wind turbine operating status monitoring has become increasingly prominent. At present, wind turbine monitoring mainly includes two links: data acquisition and data analysis. Data acquisition is responsible for obtaining parameters such as voltage, current, speed, vibration, temperature, etc., and data analysis processes and analyzes the collected data through clustering and other methods. In the prior art, these two links are often separated and implemented independently. Data acquisition mainly adopts a fixed sampling frequency method, that is, various parameters are sampled at preset fixed time intervals. The setting of the sampling frequency is usually based on empirical values ​​or conservative estimates, and there is a lack of feedback adjustment mechanism for the data analysis results. This fixed sampling method is difficult to adapt to the monitoring needs of wind turbines under different operating conditions. When the unit is in a stable operating state, a large amount of redundant data will be generated, and important information may be missed when the state changes rapidly.

[0003] In terms of data analysis, existing technologies mainly use traditional algorithms such as K-means and hierarchical clustering to perform cluster analysis on the collected data to identify the operating mode and status characteristics of the unit. However, these clustering methods only process the collected historical data and cannot guide and optimize the data collection process. Even if it is found that some data clusters require higher frequency sampling or that the sampling of some data is redundant, it is impossible to adjust the sampling strategy in time.

[0004] This separation of data collection and analysis leads to the following problems: First, data collection cannot be adaptively adjusted according to the results of cluster analysis, resulting in a waste of sampling resources or loss of information. Second, cluster analysis cannot fully utilize the time series characteristics and parameter correlations in the data collection process, affecting the accuracy of the analysis results. Third, due to the lack of a feedback mechanism, it is difficult for the system to achieve differentiated sampling of data of different importance, reducing the efficiency of the monitoring system.

[0005] Especially in the monitoring of large wind farms, as the scale of data increases, the problem caused by the separation of data collection and analysis becomes more prominent. How to effectively feed back the cluster analysis results to the data collection process and realize the intelligent adjustment of the sampling strategy has become a technical problem that needs to be solved urgently in the field of wind power monitoring. Summary of the invention

[0006] In view of this, the present invention provides a wind power generation monitoring data collection and clustering method, medium and system, which can solve the technical problem that the existing technology often separates the collection and clustering of wind power generation monitoring data, making it difficult to achieve differentiated sampling and analysis of data of different importance.

[0007] The present invention is achieved in that:

[0008] The first aspect of the present invention provides a wind power generation monitoring data collection and clustering method, comprising: setting an initial sampling frequency based on the operating parameters of a wind turbine generator set to collect voltage data, current data, speed data, vibration data, and temperature data, preprocessing the voltage data, the current data, the speed data, the vibration data, and the temperature data to obtain preprocessed data, component decomposing the preprocessed data to obtain stable components and variable components, establishing a collection data quality assessment matrix to calculate a collection contribution value, and using a multi-layer lightweight clustering method to divide the data into collection data clusters. The multi-layer lightweight clustering method uses an electrical feature lightweight neural network, a mechanical feature lightweight neural network, an environmental feature lightweight neural network, a motor feature lightweight neural network, a mechanism feature lightweight neural network, and a cluster feature lightweight neural network to perform multi-level feature extraction and fusion, determine the optimal sampling time window length based on the collection data cluster, update the sampling interval and apply it to wind turbine generator set monitoring data collection.

[0009] The collected data includes voltage data, current data, rotation speed data, vibration data, and temperature data, and the preprocessing includes denoising and standardization.

[0010] Among them, the component decomposition obtains voltage stability component, voltage variation component, current stability component, current variation component, speed stability component, speed variation component, vibration stability component, vibration variation component, temperature stability component, and temperature variation component, and establishes wind turbine monitoring data collection evaluation indicators.

[0011] The acquisition data quality assessment matrix includes data integrity, data volatility and data correlation. The acquisition features are evaluated for importance based on the acquisition data quality assessment matrix to calculate the acquisition contribution value.

[0012] Among them, the multi-layer lightweight clustering method includes constructing an electrical feature lightweight neural network, a mechanical feature lightweight neural network, and an environmental feature lightweight neural network, and the electrical feature lightweight neural network inputs a voltage stability component, a voltage variation component, a current stability component, and a current variation component.

[0013] Among them, the mechanical feature lightweight neural network inputs the speed stability component, the speed change component, the vibration stability component, and the vibration change component, and the environmental feature lightweight neural network inputs the temperature stability component and the temperature change component.

[0014] Among them, the first network output of the electrical characteristic lightweight neural network and the second network output of the mechanical characteristic lightweight neural network are jointly input into the motor characteristic lightweight neural network, and the second network output of the mechanical characteristic lightweight neural network and the third network output of the environmental characteristic lightweight neural network are jointly input into the mechanism characteristic lightweight neural network.

[0015] Among them, the fourth network output of the motor feature lightweight neural network and the fifth network output of the mechanism feature lightweight neural network are jointly input into the cluster feature lightweight neural network, and the collected data cluster is determined according to the sixth network output of the cluster feature lightweight neural network.

[0016] On the basis of the above technical solution, a wind power generation monitoring data collection clustering method of the present invention can also be improved as follows:

[0017] Wherein, the multi-layer lightweight clustering method includes:

[0018] Constructing an electrical feature lightweight neural network, inputting the voltage stability component, the voltage variation component, the current stability component, and the current variation component into the electrical feature lightweight neural network; constructing a mechanical feature lightweight neural network, inputting the speed stability component, the speed variation component, the vibration stability component, and the vibration variation component into the mechanical feature lightweight neural network; constructing an environmental feature lightweight neural network, inputting the temperature stability component and the temperature variation component into the environmental feature lightweight neural network;

[0019] The first network output of the electrical characteristic lightweight neural network and the second network output of the mechanical characteristic lightweight neural network are jointly input into the motor characteristic lightweight neural network; the second network output of the mechanical characteristic lightweight neural network and the third network output of the environmental characteristic lightweight neural network are jointly input into the mechanism characteristic lightweight neural network;

[0020] The fourth network output of the motor feature lightweight neural network and the fifth network output of the mechanism feature lightweight neural network are jointly input into the clustering feature lightweight neural network;

[0021] The collected data cluster is determined according to the sixth network output of the clustering feature lightweight neural network.

[0022] Furthermore, the electrical characteristic lightweight neural network adopts a three-layer perceptron structure, specifically: the number of neurons in the input layer is 4, corresponding to the voltage stability component, the voltage variation component, the current stability component, and the current variation component; the number of neurons in the hidden layer is 8, and the ReLU activation function is adopted; the number of neurons in the output layer is 2, and the Sigmoid activation function is adopted.

[0023] Furthermore, the mechanical feature lightweight neural network adopts a three-layer perceptron structure, specifically: the number of neurons in the input layer is 4, corresponding to the speed stability component, the speed change component, the vibration stability component, and the vibration change component; the number of neurons in the hidden layer is 8, and the ReLU activation function is adopted; the number of neurons in the output layer is 2, and the Sigmoid activation function is adopted.

[0024] Furthermore, the environmental feature lightweight neural network adopts a three-layer perceptron structure, specifically: the number of neurons in the input layer is 2, corresponding to the temperature stability component and the temperature change component, the number of neurons in the hidden layer is 4, and the ReLU activation function is adopted; the number of neurons in the output layer is 2, and the Sigmoid activation function is adopted.

[0025] Furthermore, the step of jointly inputting the first network output of the electrical feature lightweight neural network and the second network output of the mechanical feature lightweight neural network into the motor feature lightweight neural network is specifically: splicing the first network output and the second network output into a 4-dimensional vector, and inputting it into the input layer of the motor feature lightweight neural network. The motor feature lightweight neural network adopts a two-layer structure, the number of hidden layer neurons is 6, the ReLU activation function is adopted, and the number of output layer neurons is 2.

[0026] Furthermore, the second network output of the mechanical feature lightweight neural network and the third network output of the environmental feature lightweight neural network are jointly input into the mechanism feature lightweight neural network step, specifically: the second network output and the third network output are spliced ​​into a 4-dimensional vector, and input into the input layer of the mechanism feature lightweight neural network. The mechanism feature lightweight neural network adopts a two-layer structure, the number of hidden layer neurons is 6, the ReLU activation function is used, and the number of output layer neurons is 2.

[0027] Furthermore, the fourth network output of the motor feature lightweight neural network and the fifth network output of the mechanism feature lightweight neural network are jointly input into the cluster feature lightweight neural network step, specifically: the fourth network output and the fifth network output are spliced ​​into a 4-dimensional vector, and input into the input layer of the cluster feature lightweight neural network, the cluster feature lightweight neural network adopts a three-layer structure, the number of neurons in the first hidden layer is 8, the number of neurons in the second hidden layer is 4, both use ReLU activation function, and the number of neurons in the output layer is the required number of clusters. The step of determining the data collection cluster according to the sixth network output of the cluster feature lightweight neural network is specifically: the sixth network output is subjected to Softmax normalization processing to obtain the probability distribution of each sample belonging to each cluster, and the samples are divided into the cluster category with the largest probability to form a data collection cluster.

[0028] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned wind power generation monitoring data collection and clustering method.

[0029] A third aspect of the present invention provides a wind power generation monitoring data collection and clustering system, which includes the above-mentioned computer-readable storage medium.

[0030] Compared with the prior art, the beneficial effects of a wind power generation monitoring data collection clustering method, medium and system provided by the present invention are:

[0031] First, the present invention breaks the limitation of the separation of collection and clustering in traditional methods and establishes a complete data quality evaluation system. By calculating the collection quality score and collection contribution value, an accurate evaluation of the importance of different monitoring parameters is achieved, providing a reliable basis for the adaptive adjustment of the sampling frequency. This sampling optimization mechanism based on clustering feedback ensures that the monitoring system can perform differentiated sampling according to the importance of the data.

[0032] Secondly, the present invention processes monitoring data by component decomposition, and extracts and fuses features through a multi-layer lightweight neural network. This processing method can not only effectively capture the essential characteristics of the data, but also achieve a deep integration of the acquisition process and cluster analysis. By establishing three basic networks of electrical characteristics, mechanical characteristics, and environmental characteristics, as well as two intermediate layer networks of motor characteristics and mechanism characteristics, the system can adaptively adjust the sampling strategies of different parameters.

[0033] It is particularly worth emphasizing that the present invention innovatively feeds back the clustering results to the sampling control process. By calculating the temporal correlation and spatial correlation of the data clusters and combining the acquisition contribution value to determine the optimal sampling time window length, the intelligent adjustment of the sampling frequency is achieved. The design of the sampling interval being inversely proportional to the acquisition contribution value ensures accurate sampling of important data clusters and moderate sampling of secondary data clusters.

[0034] In addition, the lightweight neural network structure used in the present invention has low computational complexity, and through the rational design of the number of network layers and neurons, real-time collaborative optimization of collection and clustering is achieved. This lightweight design enables the system to quickly respond to changes in clustering results and adjust the sampling strategy in a timely manner, which is particularly suitable for large-scale deployment and application in actual engineering.

[0035] In summary, the present invention solves the technical problem that the prior art often splits the collection and clustering of wind power generation monitoring data, making it difficult to achieve differentiated sampling and analysis of data of different importance. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A flow chart of the method provided by the present invention;

[0037] Figure 2 Schematic diagram of voltage data components after EMD decomposition in Example 2;

[0038] Figure 3 Schematic diagram of the temporal and spatial correlation variation trend of different data clusters in Example 2;

[0039] Figure 4 Schematic diagram of the optimal sampling time window length of each data cluster in Example 2;

[0040] Figure 5 This is a diagram showing the dynamic changes in the sampling frequency of each parameter under extreme weather conditions in Example 2. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0042] like Figure 1 FIG. 1 is a flow chart of a wind power generation monitoring data collection and clustering method provided by the first aspect of the present invention. The method comprises the following steps:

[0043] S01. Set the initial sampling frequency according to the operating parameters of the wind turbine generator set, and collect voltage data, current data, speed data, vibration data, and temperature data during the operation of the wind turbine generator set in real time;

[0044] S02, performing data preprocessing on the voltage data, current data, speed data, vibration data, and temperature data, including denoising and standardization, to obtain preprocessed data;

[0045] S03. Decompose the preprocessed data into components to obtain the voltage stable component and voltage variation component of the voltage data, the current stable component and current variation component of the current data, the speed stable component and speed variation component of the speed data, the vibration stable component and vibration variation component of the vibration data, and the temperature stable component and temperature variation component of the temperature data, and establish the wind turbine generator monitoring data collection evaluation index;

[0046] S04. Calculate the collection quality scores of voltage data, current data, speed data, vibration data, and temperature data according to the wind turbine generator monitoring data collection evaluation index, and establish a collection data quality evaluation matrix. The collection data quality evaluation matrix includes data integrity, data volatility, and data correlation;

[0047] S05. Based on the data collection quality assessment matrix, evaluate the importance of the collection characteristics of voltage data, current data, speed data, vibration data, and temperature data, and calculate the collection contribution values ​​of voltage data, current data, speed data, vibration data, and temperature data to the wind turbine operating status monitoring;

[0048] S06. Weighting the voltage data, current data, speed data, vibration data, and temperature data according to the collected contribution values, and using a multi-layer lightweight clustering method to divide the weighted voltage data, current data, speed data, vibration data, and temperature data into a plurality of collected data clusters;

[0049] S07. Based on the collected data quality assessment matrix, for each collected data cluster, calculate the time correlation and spatial correlation of the data within the collected data cluster;

[0050] S08, determining the optimal sampling time window length corresponding to the collected data cluster according to the time correlation and space correlation and the collection contribution value;

[0051] S09, adaptively adjusting the sampling frequency of the voltage data, current data, speed data, vibration data, and temperature data in the collected data cluster based on the optimal sampling time window length, and updating the sampling interval of the voltage data, current data, speed data, vibration data, and temperature data, wherein the sampling interval is inversely proportional to the collection contribution value;

[0052] S10, applying the sampling interval to the wind turbine generator set monitoring data collection to achieve cluster adaptive adjustment of the wind turbine generator set monitoring data collection.

[0053] The specific implementation methods of the above steps are described in detail below:

[0054] In step S01, the initialization process of wind turbine monitoring data collection first needs to determine the appropriate sampling frequency according to the actual operating conditions of the unit. During the operation of the wind turbine, wind speed and power output are two important operating parameters, and the changes in these parameters will directly affect the operating state of the unit. When the wind speed is in a relatively stable state, the operating parameters of the unit are also relatively stable. At this time, a relatively low sampling frequency can be used for data collection. However, when the wind speed changes drastically, the operating state of the unit will change rapidly. At this time, the sampling frequency needs to be increased to accurately capture these changes. Similarly, when the power output of the unit fluctuates greatly, the sampling frequency also needs to be increased accordingly. In this way, it can be ensured that sufficient data information can be obtained under different operating conditions. After determining the initial sampling frequency, the system will collect various types of data during the operation of the wind turbine in real time, including voltage data, current data, speed data, vibration data and temperature data. These data reflect different aspects of the unit operation process and are an important basis for state monitoring and fault diagnosis.

[0055] In step S02, in order to ensure the accuracy of subsequent analysis, the collected raw data needs to be preprocessed. The preprocessing process mainly includes two links: denoising and standardization. In the actual data collection process, due to the influence of various factors, the collected data often contains some noise signals. These noises may come from environmental interference, errors of the sensor itself, interference in the signal transmission process and other aspects. If these noises are not processed, the accuracy of subsequent analysis will be affected. The denoising process can adopt a variety of methods, such as the median filtering method can effectively remove the sudden impulse noise, and the wavelet denoising rule can effectively remove the background noise while retaining the important features of the signal. After completing the denoising process, the data needs to be standardized. This is because different types of data may have different dimensions and numerical ranges. Directly using the raw data for analysis may cause some parameters with larger values ​​to dominate the analysis process, while covering up the influence of other important parameters. Through standardization, all data can be unified within the same scale range, eliminating the influence of dimensions, making subsequent analysis more objective and accurate.

[0056] In step S03, component decomposition of the preprocessed data is a key processing step. The main purpose of this step is to decompose each type of data into stable components and variable components. This decomposition can better reflect the essential characteristics of the data. Taking the empirical mode decomposition method as an example, this method can adaptively decompose nonlinear and non-stationary signals into a series of components with different characteristics. In the decomposition process, it is first necessary to identify all local extreme points in the signal, and then construct the upper envelope and the lower envelope through these extreme points. The average value of the upper and lower envelopes constitutes an average envelope, and a new signal is obtained by subtracting the average envelope from the original signal. If the new signal meets the conditions of the intrinsic mode function, then it is an intrinsic mode function; if the conditions are not met, the above process needs to be repeated until an intrinsic mode function is obtained. In this way, a series of intrinsic mode functions and a residual signal can be obtained by stepwise decomposition. Among these decomposed components, the low-frequency component usually reflects the overall change trend of the signal and can be used as a stable component; while the high-frequency component reflects the local fluctuation characteristics of the signal and can be used as a variable component. Through this decomposition, the stable and variable components of voltage, current, speed, vibration and temperature data can be obtained respectively. These components provide an important basis for the subsequent establishment of monitoring data collection evaluation indicators.

[0057] In step S04, a comprehensive data quality evaluation system is established based on the stable components and variable components of various data obtained in the previous steps. This evaluation system mainly evaluates data quality from three dimensions: data integrity, data volatility, and data correlation. Data integrity mainly focuses on the relationship between the number of data points actually collected and the number of data points that should be collected theoretically. This indicator can reflect whether there is loss or interruption in the data collection process. In practical applications, data collection may be interrupted or lost due to various reasons, so the evaluation of data integrity is of great significance for ensuring data quality. Data volatility mainly focuses on the discreteness of data, and this indicator can reflect the stability of data. If the data volatility of a certain parameter is large, it means that the parameter may have undergone a large change during the monitoring period. This change may be related to the operating status of the equipment and needs special attention. Data correlation mainly focuses on the degree of association between data points. This indicator can reflect the continuity and reliability of data. There should be a certain correlation between adjacent data points. If the correlation is too low, it may indicate that there is an abnormality in the data collection process. The indicators of these three dimensions together constitute a data quality evaluation matrix, through which the quality status of each monitoring parameter data can be comprehensively evaluated.

[0058] In step S05, based on the data quality assessment matrix, the importance of each monitoring parameter is evaluated, and the contribution value of each parameter to the monitoring of the operating status of the wind turbine generator set is calculated. The core of this step is to comprehensively determine the importance of each parameter in the entire monitoring system by evaluating the performance of each parameter in different quality dimensions. In practical applications, different monitoring parameters have different importance in reflecting the operating status of the unit. For example, the changes in some parameters may directly reflect the key performance indicators of the unit, while other parameters may only be auxiliary monitoring indicators. Therefore, it is necessary to comprehensively evaluate the importance of each parameter based on the performance of each parameter in the three dimensions of data integrity, data volatility and data correlation, combined with the operating characteristics of the wind turbine generator set. This evaluation should not only consider the characteristics of the parameters themselves, but also the relationship between the parameters. For example, some parameters may have a strong correlation, and then it is necessary to appropriately reduce the importance weights of some of the parameters to avoid duplication of information. The acquisition contribution value calculated in this way can truly reflect the importance of each parameter in the monitoring process.

[0059] In step S06, it is a complex process to classify the weighted data using a multi-layer lightweight clustering method. This process first requires the construction of three basic lightweight neural networks, which are used to process electrical features, mechanical features, and environmental features respectively. The design of these three networks adopts a simple three-layer perceptron structure. By reasonably setting the number of neurons in the input layer, hidden layer, and output layer, the key information of various features can be effectively extracted. On this basis, the extraction of motor features and mechanism features is realized through two intermediate lightweight neural networks. Finally, through a clustering feature lightweight neural network, the previously extracted features are integrated to obtain the final clustering result. This multi-level feature extraction and fusion method can fully consider the correlation between different types of data, thereby obtaining a more reasonable data classification result. At the same time, due to the use of a lightweight network structure, the computational burden of the entire processing process is relatively small, which is suitable for application in practical engineering.

[0060] In step S07, for each data cluster obtained by multi-layer lightweight clustering, an in-depth correlation analysis needs to be performed. This analysis process mainly includes two aspects: time correlation analysis and spatial correlation analysis. Time correlation analysis focuses on the correlation degree of data points in the time series within the same data cluster. This correlation may be manifested as periodic changes, trend changes or other types of time dependencies of data. By analyzing these time correlations, the changing rules of data within the data cluster can be understood, providing a basis for subsequent sampling frequency adjustment. Spatial correlation analysis focuses on the correlation degree between different monitoring points within the same data cluster. Since a wind turbine is a complex electromechanical system, there may be complex correlations between monitoring data at different locations. By analyzing these spatial correlations, the mutual influence between various parts of the system can be better understood, which helps to optimize the data acquisition strategy.

[0061] In step S08, determining the optimal sampling time window length is a process that requires comprehensive consideration of multiple factors. This process needs to consider the time correlation, spatial correlation and collection contribution value of each parameter of the data cluster at the same time. When a data cluster shows a strong time correlation, it means that the data in the cluster has good time continuity. In this case, the length of the sampling time window can be appropriately increased and the sampling frequency can be reduced. On the contrary, if the time correlation is weak, it is necessary to shorten the sampling time window and increase the sampling frequency to better capture the changes in the data. Similarly, spatial correlation will also affect the selection of the optimal sampling time window. Strong spatial correlation indicates that there is a close relationship between different monitoring points. In this case, the sampling strategy can be optimized by collaboratively considering the data of multiple monitoring points. In addition, the collection contribution value of each parameter needs to be considered. For parameters with higher contribution values, a higher sampling frequency should be guaranteed to ensure that sufficient effective information is obtained.

[0062] In step S09, the sampling frequency of the monitoring parameters in each data cluster is adaptively adjusted according to the determined optimal sampling time window length. The core principle of this adjustment process is that the sampling frequency is inversely proportional to the acquisition contribution value. For parameters with higher acquisition contribution values, it means that these parameters are more important to the monitoring system, and a higher sampling frequency needs to be maintained to capture their detailed change information. For parameters with lower acquisition contribution values, the sampling frequency can be appropriately reduced, which can reduce the burden of data storage and processing without significantly affecting the overall performance of the monitoring system. This adaptive adjustment mechanism enables the system to optimize the efficiency of resource use while ensuring the monitoring effect. In addition, the adjustment of the sampling frequency also needs to take into account actual engineering constraints, including sensor performance limitations, data transmission bandwidth limitations, etc., to ensure that the adjusted sampling frequency is feasible.

[0063] In step S10, the adaptive sampling strategy determined previously is applied to the actual wind turbine monitoring data collection process. This is a dynamic process, and the system needs to continuously monitor the characteristic changes of each data cluster and adjust the sampling strategy in a timely manner according to these changes. During the actual operation process, the working state of the wind turbine may change, and these changes will be reflected in the characteristic changes of the monitoring data. The system needs to be able to identify these changes in a timely manner and adjust the sampling strategy accordingly. This dynamic adjustment mechanism ensures that the monitoring system can always maintain the optimal data collection effect and provide reliable data support for the status monitoring and fault diagnosis of wind turbines. At the same time, this clustering-based adaptive adjustment method also has good scalability, and new monitoring parameters can be added or evaluation indicators can be adjusted according to actual needs.

[0064] The calculation process involved in the present invention is described in detail below:

[0065] 1. The component decomposition calculation adopts the empirical mode decomposition method to decompose the signal x(t), which is specifically expressed as follows:

[0066]

[0067] In the formula, c i (t) is the eigenmode function; r n (t) is the residual term; n is the number of decomposition levels; t is the sampling time.

[0068] 2. The calculation process of the acquisition data quality assessment matrix Q is as follows:

[0069]

[0070] in:

[0071]

[0072] In the formula, v i is the sampling data value; N is the actual number of sampling points; N t is the theoretical number of sampling points; is the data average.

[0073] 3. The calculation process of the collection contribution value is as follows:

[0074]

[0075] In the formula, C i is the collection contribution value of the i-th type of data; α1, α2, α3 are weight coefficients, and satisfy q imax is the maximum value of the i-th column.

[0076] 4. Forward propagation calculation of electrical characteristics lightweight neural network:

[0077] h1=f(W1x+b1);

[0078] y1=g(W2h1+b2);

[0079] Where x is the input vector; W1, W2 are weight matrices; b1, b2 are bias vectors; f(·) is the ReLU activation function; g(·) is the Sigmoid activation function; h1 is the hidden layer output; y1 is the network output.

[0080] 5. The structure and calculation method of the mechanical feature network and the environmental feature network are the same, the difference lies in the different input dimensions.

[0081] 6. The feature fusion process is calculated as follows:

[0082] z = [y1; y2];

[0083] h f =f(W f z+b f );

[0084] y f =g(W o z+b o );

[0085] Where z is the concatenation vector; y1, y2 are the features to be fused; W f ,W o is the fusion layer weight; b f ,b o is the fusion layer bias; h f is the fusion hidden layer output; y f The fusion output.

[0086] 7. The final clustering process uses Softmax to calculate the category probability:

[0087]

[0088] In the formula, p k is the probability that the sample belongs to the kth class; y k is the output value of the network for the kth class; K is the total number of clustering categories.

[0089] In this plan:

[0090] 1. Component decomposition using EMD method can effectively process nonlinear and non-stationary signals;

[0091] 2. The quality assessment matrix comprehensively considers the three dimensions of completeness, volatility and relevance;

[0092] 3. The weight coefficient is introduced into the collection contribution value to achieve adaptive balance in different dimensions;

[0093] 4. The neural network structure adopts a lightweight design to reduce computational complexity;

[0094] 5. Feature fusion adopts a multi-level fusion strategy to retain feature information at different levels.

[0095] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned wind power generation monitoring data collection and clustering method.

[0096] A third aspect of the present invention provides a wind power generation monitoring data collection and clustering system, which includes the above-mentioned computer-readable storage medium.

[0097] The existing wind power generation monitoring data collection clustering methods mainly have problems such as fixed data collection strategy, limited feature extraction capability, and insufficient clustering algorithm performance. In terms of data collection, traditional methods generally use fixed sampling frequencies to acquire data, which cannot adaptively adjust the data characteristics of wind turbines under different operating conditions. Due to the significant differences in the changing characteristics of various parameters during the operation of wind turbines, it is difficult for fixed sampling methods to balance the relationship between data value and collection costs, which often leads to insufficient data collection under critical conditions and a large amount of redundant data under stable conditions. In terms of feature extraction, the existing technology often processes different types of monitoring data such as electrical features, mechanical features, and environmental features independently, ignoring the coupling relationship and physical association between them, resulting in the inability of the extracted features to fully reflect the system status. In terms of clustering algorithms, traditional methods such as K-means and hierarchical clustering have high requirements on data distribution, and the algorithm complexity increases significantly with the data scale, making it difficult to meet real-time processing requirements.

[0098] The wind power generation monitoring data acquisition clustering method proposed in the present invention achieves a good balance between data processing capability and computational efficiency through a multi-layer lightweight neural network. First, the lightweight network structure is used to reduce the complexity of the model. By streamlining the number of network layers and neurons, the computational load is significantly reduced while ensuring the feature extraction capability. This lightweight design enables the model to adapt to the edge computing environment and improves the practicality of the system. Secondly, the hierarchical feature extraction architecture built based on physical mechanisms fully considers the correlation between different types of data. The electrical feature network focuses on extracting characteristic patterns of electrical quantities such as voltage and current, the mechanical feature network focuses on capturing the dynamic characteristics of mechanical parameters such as speed and vibration, and the environmental feature network focuses on the influence of environmental factors such as temperature. This network design based on domain knowledge improves the interpretability and physical meaning of features.

[0099] At the feature fusion level, the present invention realizes multi-level feature fusion through the motor feature network and the mechanism feature network. The motor feature network combines electrical features with mechanical features, reflecting the coupling relationship between the electromagnetic characteristics and mechanical characteristics of the motor; the mechanism feature network integrates mechanical features with environmental features, reflecting the impact of environmental factors on the mechanical system. This progressive feature fusion strategy can gradually extract higher-level system features and provide more effective feature expression for subsequent clustering. The final clustering feature network realizes intelligent classification of monitoring data by comprehensively analyzing the features of each layer, providing a basis for optimizing the sampling strategy.

[0100] Another innovative point of the present invention is the use of an adaptive sampling mechanism driven by the collection contribution value. By establishing a data quality evaluation matrix that includes integrity, volatility, and correlation, the system can comprehensively evaluate the importance of different types of data. The collection contribution value calculated based on the evaluation results directly guides the dynamic adjustment of sampling parameters, thereby achieving optimal allocation of sampling resources. This data-driven sampling strategy enables the system to avoid the generation of redundant data while ensuring the collection of key information.

[0101] Compared with the existing technology, the lightweight neural network clustering method of the present invention has significant advantages in principle. The lightweight design reduces the complexity of the model, making the system more suitable for engineering practice; the hierarchical structure based on physical mechanism improves the interpretability of features; the multi-level feature fusion makes full use of the correlation between data; and the adaptive sampling mechanism realizes the intelligent allocation of collection resources. These innovations make the present invention have important theoretical value and application prospects in the field of wind power monitoring data collection.

[0102] In addition, the present invention improves the adaptability and maintainability of the system by optimizing the network structure and calculation process. The modular design makes the system easy to expand and upgrade, and the simplified network structure reduces the difficulty of debugging and maintenance. These features enable the present invention to better meet the needs of engineering practice and provide a new technical path for the intelligent upgrade of wind farm monitoring systems.

[0103] Specifically, the principle of the present invention is: first, at the data processing level, the present invention uses component decomposition to process monitoring data, decomposing complex time series signals into stable components and variable components. This decomposition method is based on signal processing theory and can effectively separate the basic characteristics and dynamic characteristics of the data. The stable component reflects the basic working state of the system and is suitable for a lower sampling frequency; while the variable component contains the dynamic characteristic information of the system and requires a higher sampling frequency to capture state changes.

[0104] Secondly, the present invention designs a multi-level lightweight neural network structure, including three basic feature networks and three fusion networks. This hierarchical network structure design follows the principle of "divide and conquer", and realizes the coordinated optimization of collection and clustering through layer-by-layer feature extraction and fusion. The basic feature network adopts a three-layer perceptron structure, provides nonlinear mapping capability through the ReLU activation function, and maps features to a fixed interval through the Sigmoid activation function.

[0105] At the feature fusion level, the motor feature network and the mechanism feature network establish the correlation between parameters by combining features at different levels. This feature fusion method is consistent with the physical working principle of wind turbines and can accurately reflect the degree of influence of different parameters on the system state. The final clustering feature network achieves deep fusion of features through a multi-layer structure and obtains reliable clustering results through Softmax normalization.

[0106] In terms of sampling optimization, the present invention designs an adaptive sampling mechanism based on clustering feedback based on information theory and control theory. By calculating the temporal correlation and spatial correlation of data clusters, the system can evaluate the importance and change law of different data. The design that the sampling interval is inversely proportional to the collection contribution value ensures the reasonable allocation of sampling resources. This design not only conforms to the Nyquist sampling theorem, but also can be optimized according to the actual importance of the data.

[0107] In general, the present invention realizes the coordinated optimization of collection and clustering through a multi-layer lightweight neural network, and its principle foundation is solid and its logic is rigorous. The system can adaptively adjust the sampling strategy according to the clustering results, and the clustering analysis can also make full use of the timing characteristics in the sampling process, forming a virtuous cycle of mutual promotion and common optimization between collection and clustering. This design idea not only solves the problem of separation between collection and clustering in traditional methods, but also provides a scalable and implementable engineering solution.

[0108] A specific embodiment 1 of the present invention is provided below. The specific implementation method of each step in this embodiment 1 is described in detail as follows: The specific implementation method of step S01 is to first determine the initial sampling frequency according to the operating parameters of the wind turbine generator set, such as wind speed, power output, etc. The purpose of this step is to ensure the real-time collection of important parameters during the operation of the unit. Specifically, different sampling frequencies will be set according to different operating states. For example, when the wind speed changes drastically or the power output fluctuates greatly, the sampling frequency will be appropriately increased to ensure that the time resolution of the data is within 0.5 seconds. At the same time, key parameter data such as voltage, current, speed, vibration and temperature during the operation of the unit will also be collected. The collection of these parameter data can provide basic data support for the subsequent monitoring and diagnosis of the operating status of the unit.

[0109] The specific implementation method of step S02 is to first pre-process the collected voltage, current, speed, vibration and temperature data. The pre-processing includes two steps: denoising and standardization. The purpose of denoising is to eliminate the noise components introduced by environmental interference, instrument accuracy and other factors, and retain the essential characteristics of the signal. The purpose of standardization is to unify the original parameter data of different dimensions into the same dimension, eliminate the influence of the dimension, and facilitate subsequent feature extraction and analysis. Denoising can use median filtering, in which the filter window length is set to 5; standardization uses z-score standardization, that is, the data is subtracted from the mean and divided by the standard deviation. The purpose of this step is to ensure that the subsequent data analysis is more accurate and reliable.

[0110] The specific implementation method of step S03 is to perform component decomposition on the pre-processed voltage, current, speed, vibration and temperature data. The purpose of component decomposition is to extract the stable component and the variable component in each parameter as the basic indicator for subsequent data evaluation. The specific component decomposition method is to use the empirical mode decomposition (EMD) algorithm. The EMD algorithm can adaptively decompose the nonlinear and non-stationary original signal into a series of intrinsic mode functions (IMFs), each IMF representing an oscillation component on a different time scale. Performing EMD decomposition on the voltage, current, speed, vibration and temperature data respectively can obtain their respective stable components (composed of low-frequency IMFs) and variable components (composed of high-frequency IMFs). The mathematical expression is as follows:

[0111]

[0112] Among them, c i (t) is the intrinsic mode function, r n (t) is the residual term, n is the number of decomposition levels, and t is the sampling time. These components reflect the stability and dynamic characteristics of the key parameters of the unit and provide an important basis for subsequent data quality evaluation.

[0113] The specific implementation of step S04 is to construct a 3×3 data quality assessment matrix Q based on the stable components and variable components of voltage, current, speed, vibration and temperature extracted in step S03. Each element of the matrix represents a certain quality characteristic of a certain parameter, including: data integrity (q 11 ), data volatility (q 12 ) and data correlation (q 13 ). The calculation formula is as follows:

[0114]

[0115] Among them, v i is the sampling data value, N is the actual number of sampling points, N t is the theoretical number of sampling points, is the data average. Data integrity reflects the ratio of the actual number of sampling points to the theoretical number of sampling points, and characterizes the completeness of the data; data volatility reflects the discreteness of the data, and characterizes the stability of the data; data correlation reflects the correlation between adjacent data points, and characterizes the temporal continuity of the data. By comprehensively considering these three dimensions, the quality status of each parameter data can be comprehensively evaluated.

[0116] The specific implementation method of step S05 is to calculate the contribution value C of each parameter data (voltage, current, speed, vibration, temperature) to the operation status monitoring of the wind turbine generator set according to the data quality evaluation matrix Q constructed in step S04. i This contribution value is obtained by weighted average of the three dimensions of the Q matrix. The weight coefficients α1, α2 and α3 can be adjusted according to the actual application requirements and meet Contribution value C i It reflects the relative importance of the parameter data in the overall monitoring system and provides a basis for subsequent adaptive sampling. The calculation formula is:

[0117]

[0118] Among them, q imax It represents the maximum value of the index in the i-th column. According to the actual application requirements, α1=0.4, α2=0.3, and α3=0.3 can be set.

[0119] The specific implementation method of step S06 is to first calculate the collection contribution value C of each parameter data obtained in step S05. i , weight the original data. Then a multi-level lightweight clustering method is used to divide the weighted voltage, current, speed, vibration and temperature data into several collection data clusters. This multi-level clustering method includes:

[0120] 1) Construct a lightweight neural network for electrical characteristics and input voltage stability component v s 、Voltage variation component v d , current stability component i s and the current variation component i d , a three-layer perceptron structure is adopted, the number of hidden layer neurons is 8, and the output is a 2-dimensional electrical feature vector y1. The forward propagation calculation is as follows:

[0121] h1=f(W1x+b1);

[0122] y1=g(W2h1+b2);

[0123] Among them, x is the input vector, W1, W2 are weight matrices, b1, b2 are bias vectors, f(·) is the ReLU activation function, g(·) is the Sigmoid activation function, and h1 is the hidden layer output.

[0124] 2) Construct a lightweight neural network based on mechanical characteristics and input the speed stability component n s , speed variation component n d , vibration stability component v b,s and vibration variation component v b,d ,The structure and calculation method are the same as above, and the output is a 2D mechanical characteristic vector y2.

[0125] 3) Construct a lightweight neural network of environmental features and input the temperature stability component T s and the temperature variation component T d ,The structure and calculation method are the same as above, and the output is a 2D environmental feature vector y3.

[0126] 4) Concatenate y1 and y2 and input them into the motor feature lightweight network, which adopts a two-layer structure with 6 hidden layer neurons and outputs a 2D motor feature vector y4.

[0127] 5) Concatenate y2 and y3 and input them into the lightweight network of mechanism characteristics. It adopts a two-layer structure with 6 hidden layer neurons and outputs a 2-dimensional mechanism feature vector y5.

[0128] 6) Concatenate y4 and y5 and input them into the clustering feature lightweight network with a three-layer structure. The number of neurons in the first hidden layer is 8, the number of neurons in the second hidden layer is 4, the number of neurons in the output layer is the required number of clusters K, and the clustering result y6 is output.

[0129] This multi-level feature fusion method can fully explore the intrinsic connections between different measurement parameters and extract more discriminative clustering features, thereby obtaining more accurate data collection cluster division results.

[0130] The specific implementation of step S07 is to calculate the time correlation and spatial correlation of the data within each cluster of collected data obtained in step S06. t reflects the temporal dependence and spatial correlation ρ between data points within the same cluster. s It reflects the degree of correlation between different sensor measurement points within the same cluster. These two indicators can measure the overall quality characteristics of the collected data cluster and provide a basis for the subsequent adaptive adjustment of the sampling frequency.

[0131] The specific implementation of step S08 is to calculate the time correlation ρ according to step S07. t and spatial correlation ρ s, combined with the collection contribution value C of each parameter data calculated in step S05 i , determine the optimal sampling time window length T corresponding to each data collection cluster win The specific method is, when ρ t and ρ s When it is higher, it means that the data has strong continuity and correlation, and the sampling frequency can be appropriately reduced, that is, T is increased. win On the contrary, it is necessary to increase the sampling frequency to capture more dynamic change information and reduce T win At the same time, collect the contribution value C i The higher the parameter, the higher the sampling frequency should be. win By comprehensively considering these factors, the optimal sampling time window length of each data cluster is finally determined, where T win The value range is [0.5,10] seconds.

[0132] The specific implementation of step S09 is to use the optimal sampling time window length T determined in step S08 as follows: win , the sampling frequency of voltage, current, speed, vibration and temperature data in each data collection cluster is adaptively adjusted. Specifically, the sampling frequency f s And the collection contribution value C i Inversely proportional, that is, the higher the contribution value of the parameter, the higher the sampling frequency; the lower the contribution value of the parameter, the lower the sampling frequency can be appropriately reduced. The calculation formula is:

[0133]

[0134] This adaptive adjustment mechanism can ensure the high temporal resolution acquisition of key parameter data while also taking into account the acquisition efficiency of secondary parameters, thereby optimizing resource utilization.

[0135] The specific implementation of step S10 is to apply the adaptive sampling frequency determined in step S09 to the actual wind turbine monitoring data collection process. During operation, the system will monitor the time correlation of each collected data cluster in real time. t and spatial correlation ρ s , and dynamically adjust the sampling frequency f accordingly s ,Ensure the high quality collection of key parameter data while taking into account the overall collection efficiency.,This adaptive clustering adjustment method can meet the data collection requirements in the complex ,operating environment of wind turbines and provide reliable basic data support for ,subsequent state monitoring and fault diagnosis.

[0136] In order to better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: A wind farm consists of 100 3MW wind turbines with a total installed capacity of 300MW. This wind farm is located in a remote area with a harsh surrounding environment. It is often affected by natural factors such as wind, low temperature, ice and snow, which poses a great challenge to the operation monitoring of the units. In order to ensure the safe and stable operation of the units, the operation and maintenance personnel urgently need an intelligent monitoring data collection method that can adapt to complex environments.

[0137] Based on the clustering method for wind power generation monitoring data collection proposed by the present invention, the operation and maintenance team of the wind farm decided to deploy this monitoring system throughout the entire farm after sufficient demonstration and practical verification. First, they determined the initial monitoring parameter sampling frequency based on the actual operating characteristics of the wind turbines. Due to the complex and changeable environment of the wind farm, there may be drastic changes in wind speed or large fluctuations in power output. In order to ensure high temporal resolution of key parameter data, the initial sampling frequency is set to 0.5 Hz, that is, 2 samples are collected per second. At the same time, the monitoring system will collect five key parameter data such as voltage, current, speed, vibration and temperature of 100 units in real time.

[0138] Next, the operation and maintenance personnel preprocessed the collected raw monitoring data. First, the median filter was used to remove noise, and the filter window length was set to 5. Then the z-score standardization method was used to unify the parameter data of different dimensions into the range of [-1,1] to eliminate the influence of dimension differences. After these two steps of preprocessing, the quality of the monitoring data was initially improved.

[0139] Next, the operation and maintenance personnel perform EMD component decomposition on the pre-processed monitoring data. Taking voltage data as an example, after EMD decomposition, the stable component v of the voltage can be obtained. s and the variable component v d Similarly, the other four parameter data were also decomposed by EMD, and the corresponding stable components and variable components were extracted, as shown in Table 1:

[0140] Table 1 Monitoring parameter components after EMD decomposition

[0141] parameter Stable Component Variable Component Voltage <![CDATA[v s ]]> <![CDATA[v d ]]> Current <![CDATA[i s ]]> <![CDATA[i d ]]> Speed <![CDATA[n s ]]> <![CDATA[n d ]]> vibration <![CDATA[v b,s ]]> <![CDATA[v b,d ]]> temperature <![CDATA[T s ]]> <![CDATA[T d ]]>

[0142] With these component data, the operation and maintenance personnel then constructed a 3×3 data quality assessment matrix Q. The specific calculation process is as follows:

[0143]

[0144] Among them, v i is the data value of the i-th sampling point, N is the actual number of sampling points, N t is the theoretical number of sampling points, is the average value of the data. Figure 2 As shown, the voltage data components after EMD decomposition are displayed, including the time domain waveforms of the original voltage signal, the stable component and the variable component.

[0145] Through calculation, the data quality assessment matrix Q of the 100 units in the wind farm is obtained, as shown in Table 2:

[0146] Table 2 Data quality assessment matrix Q;

[0147] parameter Data Integrity Data volatility Data correlation Voltage 0.98 0.12 0.85 Current 0.95 0.15 0.82 Speed 0.92 0.18 0.79 vibration 0.90 0.22 0.76 temperature 0.88 0.25 0.72

[0148] As can be seen from Table 2, the overall monitoring data quality of the wind farm is good and the data integrity is high, but there is still room for improvement in data volatility and correlation.

[0149] Based on this evaluation matrix, the operation and maintenance personnel then calculated the collection contribution value C of each parameter data. i , specifically as follows:

[0150]

[0151] Among them, α1=0.4, α2=0.3, α3=0.3. According to the calculation results, the collection contribution values ​​of 100 units are shown in Table 3:

[0152] Table 3 Collection contribution value C of each parameter data i ;

[0153]

[0154]

[0155] It can be seen from Table 3 that voltage and current data are more important in the entire monitoring system and have a larger collection contribution value, while vibration and temperature data are relatively less important. This lays the foundation for the subsequent adaptive sampling frequency adjustment.

[0156] Next, the operation and maintenance personnel used the multi-level lightweight neural network clustering method proposed in this invention to classify the monitoring data. Specifically, three lightweight neural networks for electrical characteristics, mechanical characteristics, and environmental characteristics were first constructed to extract feature vectors of different types of parameters. The input of the electrical characteristic network is v s 、v d 、i s and i d , a three-layer perceptron structure is used, the number of hidden layer neurons is 8, and the output is a 2D electrical feature vector y1. The input of the mechanical feature network is n s 、n d 、v b,s and vb,d , the structure and calculation method are the same as above, and the output is a 2D mechanical feature vector y2. The input of the environmental feature network is T s and T d ,It also adopts a three-layer perceptron structure, the number of hidden layer neurons is 4, and the output is a 2-dimensional environmental feature vector y3.

[0157] Then, y1 and y2 are concatenated and input into the motor feature network, which uses a two-layer structure and 6 hidden layer neurons, and outputs a 2D motor feature vector y4. y2 and y3 are concatenated and input into the mechanism feature network, which also uses a two-layer structure and 6 hidden layer neurons, and outputs a 2D mechanism feature vector y5. Finally, y4 and y5 are concatenated and input into the clustering feature network, which uses a three-layer structure, with 8 neurons in the first hidden layer, 4 neurons in the second hidden layer, and the number of neurons in the output layer equal to the required number of clusters K=10, and the clustering result y6 is output.

[0158] Based on this hierarchical feature extraction and fusion method, the monitoring data of the 100 units in the wind farm were finally divided into 10 collection data clusters. Figure 3 As shown, the changing trends of temporal correlation and spatial correlation of different data clusters are demonstrated.

[0159] Next, the operation and maintenance personnel calculated the time correlation ρ of the internal data for each data cluster. t and spatial correlation ρ s , as shown in Table 4:

[0160] Table 4 Correlation index of each collected data cluster

[0161] Cluster number <![CDATA[Time correlation ρ t > <![CDATA[Spatial correlation ρ s > 1 0.92 0.88 2 0.89 0.85 3 0.86 0.82 4 0.83 0.79 5 0.80 0.75 6 0.77 0.72 7 0.74 0.68 8 0.71 0.64 9 0.68 0.60 10 0.65 0.56

[0162] It can be seen from Table 4 that there are large differences in the temporal correlation and spatial correlation of different data clusters. Cluster 1 has the highest correlation, indicating that the data within the cluster has strong continuity and correlation; while cluster 10 has the lowest correlation, indicating that the data within the cluster has large discreteness and uncertainty.

[0163] Combined with the collection contribution value C of each parameter data obtained in step S05 i ,The operation and maintenance personnel determined the optimal sampling time window length T corresponding to each data collection cluster win , as shown in Table 5:

[0164] Table 5 The optimal sampling time window length T for each data collection cluster win ;

[0165]

[0166] It can be seen from Table 5 that for the voltage and current data with higher contribution values, the optimal sampling time window length is shorter, ranging from 0.5 to 1.0 seconds; while for the vibration and temperature data with lower contribution values, the optimal sampling time window length is longer, ranging from 3.0 to 6.0 seconds. Such an adaptive adjustment mechanism can not only ensure the high time resolution acquisition of key parameter data, but also take into account the acquisition efficiency of secondary parameters, thus optimizing the overall monitoring system performance. Figure 4 As shown, the optimal sampling time window length distribution of each data cluster is displayed in the form of a bar graph.

[0167] Finally, the wind farm operation and maintenance personnel apply the above-determined adaptive sampling frequency to the actual monitoring data collection process. During operation, the monitoring system will monitor the time correlation of each collected data cluster in real time. t and spatial correlation ρ s , and dynamically adjust the sampling frequency f according to the corresponding relationship between Table 4 and Table 5 s , the specific calculation formula is:

[0168]

[0169] Through this adaptive clustering adjustment method, the wind farm monitoring system can ensure high-quality collection of key parameter data under complex operating environments, while taking into account the optimization of overall collection efficiency, providing a reliable data basis for subsequent status monitoring and fault diagnosis.

[0170] Taking a 3MW wind turbine as an example, during operation under extreme weather conditions (wind speed reaches 25 meters per second, temperature drops to -20 degrees Celsius, and snow and ice cover is severe), the specific data collected by the monitoring system is as follows:

[0171] 1. The initial sampling frequency is set to 0.5 Hz, or 2 samples per second.

[0172] 2. After EMD component decomposition and data quality assessment, the monitoring parameter data of the unit is divided into cluster 3. According to Table 4 and Table 5, the time correlation of this cluster is t =0.86, spatial correlation ρ s =0.82, optimal sampling time window length T win =1.0 sec.

[0173] 3. Since the current data collection contribution value C in this cluster i =0.87 is the highest, so the sampling frequency f s =1 / T win = 1 Hz. The contribution of speed data collection is second, f s = 0.8 Hz; vibration and temperature data have lower contributions, fs =0.6 Hz.

[0174] 4. In the next 1 second, the monitoring system collected 2 voltage samples, 1 current sample, 1 speed sample, 1 vibration sample and 1 temperature sample respectively.

[0175] 5. As time went by, the operating status of the unit changed, and the monitoring system detected in real time that the time correlation and spatial correlation of each parameter decreased. Therefore, the system automatically adjusted the sampling frequency to 1.2 Hz for voltage and current, and 0.8 Hz for speed, vibration and temperature.

[0176] 6. Through continuous adaptive adjustment, the monitoring system ensures high-resolution acquisition of key parameter data while taking into account the acquisition efficiency of secondary parameters, fully reflecting the changes in the operating status of the unit under harsh environments, and providing a reliable basis for subsequent fault diagnosis. Figure 5 As shown, the dynamic changes of the sampling frequency of each parameter under extreme weather conditions are demonstrated.

[0177] It should be noted that the explanations of the variables involved in the present invention are shown in Table 6.

[0178] Table 6 Variable explanation table

[0179]

[0180] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A wind power generation monitoring data collection and clustering method, characterized in that: include: The voltage data, current data, rotational speed data, vibration data and temperature data are collected by setting the initial sampling frequency based on the operating parameters of the wind turbine generator set, the voltage data, current data, rotational speed data, vibration data and temperature data are preprocessed to obtain preprocessed data, the preprocessed data are component-decomposed to obtain stable components and variable components, a collection data quality assessment matrix is ​​established to calculate the collection contribution value, and a multi-layer lightweight clustering method is used to divide the data into collection data clusters. The multi-layer lightweight clustering method performs multi-layer feature extraction and fusion through electrical feature lightweight neural networks, mechanical feature lightweight neural networks, environmental feature lightweight neural networks, motor feature lightweight neural networks, mechanism feature lightweight neural networks and clustering feature lightweight neural networks, determines the optimal sampling time window length based on the collection data cluster, updates the sampling interval and applies it to the wind turbine generator set monitoring data collection.

2. The wind power generation monitoring data collection and clustering method according to claim 1 is characterized in that: The collected data includes voltage data, current data, rotation speed data, vibration data, and temperature data, and the preprocessing includes denoising and standardization.

3. The wind power generation monitoring data collection and clustering method according to claim 1 is characterized in that: The components are decomposed into voltage stability component, voltage variation component, current stability component, current variation component, speed stability component, speed variation component, vibration stability component, vibration variation component, temperature stability component, and temperature variation component, and wind turbine monitoring data collection evaluation indicators are established.

4. The wind power generation monitoring data collection and clustering method according to claim 1, characterized in that: The acquisition data quality assessment matrix includes data integrity, data volatility and data correlation. The acquisition features are evaluated for importance based on the acquisition data quality assessment matrix to calculate the acquisition contribution value.

5. The wind power generation monitoring data collection and clustering method according to claim 1, characterized in that: The multi-layer lightweight clustering method includes constructing an electrical feature lightweight neural network, a mechanical feature lightweight neural network, and an environmental feature lightweight neural network. The electrical feature lightweight neural network inputs a voltage stability component, a voltage variation component, a current stability component, and a current variation component.

6. The wind power generation monitoring data collection and clustering method according to claim 5, characterized in that: The mechanical characteristic lightweight neural network inputs a speed stability component, a speed variation component, a vibration stability component, and a vibration variation component; the environmental characteristic lightweight neural network inputs a temperature stability component and a temperature variation component.

7. The wind power generation monitoring data collection and clustering method according to claim 5, characterized in that: The first network output of the electrical feature lightweight neural network and the second network output of the mechanical feature lightweight neural network are jointly input into the motor feature lightweight neural network, and the second network output of the mechanical feature lightweight neural network and the third network output of the environmental feature lightweight neural network are jointly input into the mechanism feature lightweight neural network.

8. The wind power generation monitoring data collection and clustering method according to claim 7, characterized in that: The fourth network output of the motor feature lightweight neural network and the fifth network output of the mechanism feature lightweight neural network are jointly input into the cluster feature lightweight neural network, and the collected data cluster is determined according to the sixth network output of the cluster feature lightweight neural network.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed in a computer, they are used to execute the wind power generation monitoring data collection and clustering method according to any one of claims 1 to 8.

10. A wind power generation monitoring data collection and clustering system, characterized in that: A computer-readable storage medium comprising the computer-readable storage medium of claim 9.

Images