A fan blade icing detection method, system and electronic device based on mechanism feature fusion

Abstract

The application provides a fan blade icing detection method and system based on mechanism feature fusion and an electronic device, and relates to the field of wind power operation and condition monitoring. The SCADA system data set of a wind turbine generator is acquired, a new feature set with physical meaning is constructed based on the blade icing formation mechanism, the Makkonen icing empirical model and the effect of icing on the operation performance of the unit, all features are comprehensively correlated and sorted, the optimal feature group is selected, the blade icing time series statistical features are constructed, and the gradient boosting decision tree model based on multi-time series feature fusion is input, so that the fan blade icing detection is realized, the noise and model deviation are reduced, the data quality and training stability are improved, the accuracy and technicality of the icing early warning are improved, and the operation safety and power generation efficiency of the unit are improved.

Description

Technical Field

[0001] This invention relates to the field of wind power operation and maintenance and condition monitoring, and in particular to a method, system and electronic equipment for detecting icing on wind turbine blades based on mechanism feature fusion. Background Technology

[0002] Wind energy, as a clean and renewable energy source, is widely used in high-altitude, low-temperature, and high-humidity areas. Wind turbine blades are exposed to harsh environments for extended periods, making them highly susceptible to icing failures in winter. This can lead to deterioration of blade aerodynamic performance, reduced power generation efficiency, load imbalance, and in severe cases, blade breakage and turbine shutdown. Therefore, accurate detection and early warning of wind turbine blade icing are crucial for the safe operation and maintenance of wind farms and for improving power generation revenue.

[0003] However, existing research still has the following limitations: First, the formation of wind turbine icing essentially involves a series of complex energy dynamic processes. Traditional models often rely on transient characteristics, lacking sufficient early icing detection capabilities and failing to capture the gradual evolution of the icing state. Second, as black boxes, purely data-driven models ignore the intrinsic physical mechanisms of wind turbine blade icing and lack the ability to explore deeper causes of icing. Their generalization ability is limited under complex operating conditions, inevitably resulting in significant errors. Third, existing methods fail to fully utilize the physical mechanisms of icing to effectively constrain the features in the data-driven models, making it difficult to further improve model accuracy and robustness. Summary of the Invention

[0004] The main objective of this invention is to provide a method, system, and electronic device for detecting icing on wind turbine blades based on mechanistic feature fusion, thereby addressing the problems of insufficient early icing detection capability, insufficient feature representation capability, and low model accuracy and robustness.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for detecting icing on wind turbine blades based on mechanism feature fusion, comprising the following steps: S1: Obtain the data set of the wind turbine SCADA system and perform data preprocessing; S2: Obtain the preprocessed dataset and construct a new set of features with physical meaning based on the blade icing formation mechanism, the Makkonen icing empirical model, and the impact of icing on unit operating performance. The new set of features includes icing intensity mechanism features constructed based on the blade icing phase change process and the Makkonen icing mass growth model, performance deviation features constructed based on wind-power characteristics, temperature difference features constructed based on nacelle-environment thermal coupling, and balance features constructed based on three-blade balance characteristics. S3: The optimal feature group is obtained by comprehensively ranking all features based on Spearman correlation coefficient, chi-square test analysis and RFECV-LightGBM; S4: Constructing the time-series statistical characteristics of blade icing based on sliding window analysis; S5: Construct a gradient boosting decision tree model based on multi-time series feature fusion, and input the time series statistical features to be tested to realize wind turbine blade icing detection.

[0006] In a preferred embodiment, the data preprocessing includes: The data is labeled with its state, noisy data is removed using box plots, and missing values ​​are filled forward. The imbalance problem is addressed by combining threshold filtering for undersampling with time-preserving SMOTE oversampling. The specific process is as follows: S11: Obtain data from the wind turbine SCADA system; S12: Classify each data point according to the time periods of normal and icing states, and remove samples with unclear states; S13: Noisy data is identified and removed using box plots. Missing values ​​after removal are filled using a forward imputation method. The formulas for calculating the upper and lower limits are as follows: ; In the formula, k As control terms, Q1 and Q2 are the upper and lower quartiles, respectively; S14: Undersampling is performed using a threshold-based filtering method to remove unreasonable and shutdown data; and oversampling is performed using Smote, which preserves time series characteristics, to make the ratio of frozen samples to normal samples close to the preset value.

[0007] In the preferred embodiment, the SCADA dataset mentioned in step S1 contains 26 sets of continuous numerical features, including operating condition parameters, environmental parameters, and state parameters, with a sampling interval of 7 seconds. When denoising box plots, k=3 is used to identify extreme singularities. The preset ratio of normal samples to frozen samples after imbalance treatment is 1:1.

[0008] In the preferred scheme, the specific process of constructing a new feature set and filtering the original features in step S2 is as follows: S21: Analysis of the failure mechanism of wind turbine blade icing, identifying ambient temperature and wind speed as key variables; S22: Extracting ambient temperature based on Makkonen's icing empirical model. With the square of wind speed As an important mechanistic characteristic reflecting the amount of icing, the formula is: ; In the formula, For ambient wind speed; S23: Analysis of the impact of wind turbine blade icing failure on unit operating performance, including: The system constructs power residuals or wind speed to power ratios, constructs ambient temperature differences, and performs balance analysis on blade angle, blade speed, and pitch motor temperature parameters. Redundant features are eliminated, and new features are constructed.

[0009] In the preferred embodiment, step S23 specifically includes: Based on wind power characteristics, the formula for the power residual and wind speed-power ratio is as follows: ; WSP ; In the formula, This represents the actual output power of the fan at the same wind speed. This represents the theoretical output power of the fan at the same wind speed. WS For wind speed, P Power; The temperature difference characteristics are constructed based on the temperature difference between the environment and the cabin, and the formula is as follows: ; In the formula, ET For ambient temperature, IT Cabin temperature; Mean characteristics were constructed based on the three-blade balance analysis, including the average blade angle. 、 The average blade speed, average pitch motor temperature, average blade DC charging power supply temperature, and average blade charger DC current are expressed as follows: MPA= ; MPS= ; MPMT= ; MPNT= ; MPND= ; In the formula, For the angle of each wind turbine blade, For the speed of each wind turbine blade, Temperature of each pitch motor Temperature of the DC charging power supply for each blade The DC current of each blade charger; At the same time, redundant features such as single blade angle, single blade speed, single pitch motor temperature, and x / y direction acceleration are eliminated.

[0010] In the preferred scheme, step S3 provides statistical correlation evidence through Spearman correlation coefficient and chi-square test, completes model-driven feature importance verification through RFECV-LightGBM, and retains key features with high correlation through comprehensive ranking.

[0011] In the preferred embodiment, step S4 constructs time-series statistical features of leaf icing based on sliding window analysis. The sliding window is used to statistically analyze six time-series statistical features within a certain time window, namely the maximum value, minimum value, average value, standard deviation, median, and variance, which are used to characterize the leaf icing state, development, and severity.

[0012] In the preferred scheme, step S5 uses mechanism fusion features and time-series statistical features as inputs, and employs a gradient boosting decision tree to achieve binary classification detection of wind turbine blade icing, thereby obtaining the detection results.

[0013] Secondly, the present invention provides a wind turbine blade icing detection system based on mechanism feature fusion, characterized in that the system used in the aforementioned wind turbine blade icing detection method based on mechanism feature fusion comprises: The data acquisition and preprocessing module is used to acquire data from the wind turbine SCADA system and perform data tagging, noise processing, and imbalance data processing. The mechanism feature construction and screening module is used to acquire the preprocessed dataset and construct a new set of features with physical meaning based on the blade icing formation mechanism, the Makkonen icing empirical model, and the impact of icing on unit operating performance. The new feature set includes icing intensity mechanism features constructed based on the blade icing phase change process and the Makkonen icing mass growth model, performance deviation features constructed based on wind-power characteristics, temperature difference features constructed based on nacelle-environment thermal coupling, and equilibrium features constructed based on the three-blade equilibrium characteristics. The multi-criteria feature ranking module is used to perform comprehensive correlation ranking of features using three methods: Spearman correlation coefficient, chi-square test analysis and RFECV-LightGBM, and output the optimal feature set. The temporal feature extraction module is used to construct temporal statistical features of leaf icing based on sliding window analysis; The icing detection module is used to build and run a gradient boosting decision tree model based on multi-temporal feature fusion, and output the icing detection results of wind turbine blades.

[0014] Thirdly, the present invention provides an electronic device, including a memory and a processor; The memory is used to store computer programs; The processor is used to implement the wind turbine blade icing detection method based on mechanism feature fusion when executing the computer program.

[0015] This invention provides a wind turbine blade icing detection method based on mechanism feature fusion. By acquiring the SCADA system dataset of the wind turbine, and constructing a new set of features with physical meaning based on the blade icing formation mechanism, the Makkonen icing empirical model, and the impact of icing on the unit's operating performance, the method performs comprehensive correlation ranking of all features, selects the optimal feature group, constructs the blade icing time-series statistical features, and inputs them into a gradient boosting decision tree model based on multi-time-series feature fusion. This method realizes wind turbine blade icing detection, reduces noise and model bias, improves data quality and training stability, enhances the accuracy and technicality of icing early warning, and significantly improves the unit's operating safety and power generation efficiency.

[0016] This invention employs a feature construction and selection method that combines mechanistic analysis with data-driven approaches, enabling the model to possess both data learning capabilities and physical logic support, thus improving its robustness under complex and variable operating conditions. In the feature engineering stage, new features are innovatively constructed by combining the ice formation mechanism with empirical models, giving the data-driven model physical interpretability and enhancing its generalization ability. Furthermore, by extracting and fusing temporal dynamic features during the ice formation process, the model can more sensitively capture early ice formation and state evolution processes, improving early warning capabilities. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention; Figure 2 Scatter plots of ambient wind speed and power under normal and icing conditions; Figure 3 A scatter plot showing the angle relationships of each blade; Figure 4 A scatter plot showing the velocity relationships of each blade; Figure 5 A scatter plot showing the temperature relationships of each pitch motor; Figure 6 A scatter plot showing the temperature relationship of the DC charging power supply for each blade. Figure 7 A scatter plot showing the DC current relationship of each blade charger; Figure 8 A scatter plot of acceleration in the x-direction and acceleration in the y-direction; Figure 9 This is used to construct multi-temporal features of blade icing based on sliding window analysis. Detailed Implementation

[0018] Example 1 like Figure 1-9 As shown, a method for detecting icing on wind turbine blades based on mechanism feature fusion includes the following steps: S1: Obtain the data set of the wind turbine SCADA system and perform data preprocessing.

[0019] S2: Obtain the preprocessed dataset and construct a new set of features with physical meaning based on the blade icing formation mechanism, the Makkonen icing empirical model, and the impact of icing on unit operating performance. The new set of features includes icing intensity mechanism features constructed based on the blade icing phase change process and the Makkonen icing mass growth model, performance deviation features constructed based on wind-power characteristics, temperature difference features constructed based on nacelle-environment thermal coupling, and balance features constructed based on the three-blade balance characteristics.

[0020] S3: The optimal feature group is obtained by comprehensively ranking all features based on their correlation using Spearman correlation coefficient, chi-square test analysis, and RFECV-LightGBM.

[0021] S4: Constructing the time-series statistical features of leaf icing based on sliding window analysis.

[0022] S5: Construct a gradient boosting decision tree model based on multi-time series feature fusion, and input the time series statistical features to be tested to realize wind turbine blade icing detection.

[0023] This embodiment acquires the SCADA system dataset of wind turbine units and constructs a new set of features with physical meaning based on the blade icing formation mechanism, the Makkonen icing empirical model, and the impact of icing on unit operating performance. All features are then comprehensively ranked by correlation to obtain the optimal feature set. A time-series statistical feature of blade icing is constructed and input into a gradient boosting decision tree model based on multi-time-series feature fusion. This achieves wind turbine blade icing detection, reduces noise and model bias, improves data quality and training stability, enhances the accuracy and technicality of icing early warning, and improves unit operating safety and power generation efficiency.

[0024] In the preferred embodiment, data preprocessing includes: Data from the wind turbine SCADA system is acquired, the data is labeled with status, noisy data is removed using box plots, and missing values ​​are filled forward. Data imbalance is addressed by combining threshold filtering for undersampling with time-preserving SMOTE oversampling. The specific process is as follows: S11: Acquire SCADA system data for a wind turbine. The data covers SCADA operation data of a domestic wind turbine from November to December 2015, as well as the corresponding normal and fault state time periods. The SCADA sampling interval is approximately 7 seconds. The SCADA-monitored state variables involve various systems of the wind turbine operation. The dataset contains 26 sets of characteristic parameters closely related to the wind turbine blade icing problem. All of these characteristic variables are continuous numerical variables, covering multiple dimensions such as wind turbine operating parameters, environmental parameters, and state parameters.

[0025] Table 1 shows the physical meaning and units of the operating parameters. These parameters include the current engine speed, grid-side active power, wind turbine blade yaw angle, wind turbine blade speed during yaw, and the DC current of the chargers for each of the three blades of the three-bladed wind turbine. Figure 2 As shown, this visually demonstrates that the power output deviates significantly from the normal operating curve under the same wind speed in an icy state.

[0026] Table 1 Physical meaning and units of operating parameters

[0027] Table 2 shows the physical meaning and units of the environmental parameters. The environmental parameters include the ambient temperature at the location of the wind turbine blades, the real-time wind speed at the location of the wind turbine blades, the windward angle of the wind turbine blades, and the average wind direction angle over 25 seconds, which are obtained from wind speed and temperature sensors inside and outside the nacelle.

[0028] Table 2 Physical meaning and units of environmental parameters

[0029] like Figure 3 As shown, the data for normal and icing conditions highly overlap and exhibit a linear distribution at 45°, indicating that the correlation between the single blade angle and icing is extremely low, providing a basis for eliminating redundant features.

[0030] Table 3 shows the physical meaning and units of the environmental parameters. The state parameters include the temperature inside the nacelle, the temperature of the DC charging power supply for each of the three blades of the three-bladed wind turbine, the deflection angle and rotation speed of each of the three blades of the three-bladed wind turbine, the temperature of each of the three pitch generators, and the acceleration of the wind turbine vibration sensor in the x and y directions.

[0031] Table 3 Physical meaning and units of state parameters

[0032] S12: The SCADA system data for the wind turbine was labeled. The dataset centrally labeled the start and end times of the normal operation period and the icing period. Each data point could be categorized based on its normal operation and icing time periods. However, there was a segment of samples with unclear turbine status between the end of the normal operation period and the start of blade icing. These unclear samples were caused by computer downtime or human error and would interfere with model training; therefore, they were removed.

[0033] S13: Due to factors such as system malfunctions, sensor failures, and operator errors, the data acquisition equipment may contain noisy data. The box plot method is used to identify noisy data. The specific steps are as follows: Step 1: Sort the data values ​​of one of the features in ascending order, and sort the remaining training data in the same order.

[0034] Step 2: Use box plots to identify and remove noisy data. For the missing values ​​after removal, use forward imputation to calculate the upper quartile Q1, lower quartile Q2, median, and upper and lower limits Ucl and Lcl of this feature data. The formulas for calculating the upper and lower limits are as follows: ; In the formula, k As control terms, Q1 and Q2 are the upper and lower quartiles, respectively.

[0035] Based on empirical values, when k When the value is 3, the value of the feature data exceeds the upper and lower limits Ucl and Lcl, which are extreme singularities. In this embodiment... k Take 3.

[0036] Step 3: Remove samples containing extreme singularities in the feature data. Then continue to check the remaining features until all features have been checked.

[0037] Step 4: For the missing values ​​after removal, use the forward filling method to fill the missing values.

[0038] S14: Imbalanced data is processed. During actual operation, the wind turbine is in normal operation most of the time, and only in an icing state for a small period of time. The ratio of icing samples to normal samples is approximately 1:15, which is a typical sample class imbalance problem. Therefore, a method of undersampling based on threshold filtering and oversampling based on the time series characteristics preservation method (Smote) is used to process the ratio of normal samples to icing samples to a value close to the preset value of 1:1.

[0039] Undersampling based on threshold filtering involves removing unreasonable data from the wind turbine blades (characterized by low wind speed but high power output) and shutdown data (characterized by actual wind speed greater than 0 m / s but no power output). Oversampling based on the time-series characteristic preservation mechanism of Smote is as follows: Step 1, let the minority class sample set be (i=1, 2, ..., N) Identification using Euclidean distance Each sample of The nearest neighbor samples are labeled as , j {1, 2，… , K}; Step 2: Randomly select one from several nearest neighbor samples. Sample, and randomly generated random values ​​in [0, 1]. In order to obtain new samples : = + ; In the formula, represents the newly synthesized sample. ∈[0,1].

[0040] Step 3: Repeat step 2 in the same way to obtain... N A new sample: , new {1 2，… , N}

[0041] Step 4: Use KNN to find the nearest neighbor of the oversampled sample. Obtain the timestamp of the nearest neighbor and assign it to the oversampled sample.

[0042] Step 5: Sort the oversampled data according to the time column to ensure that the newly generated samples have reasonable time attributes, thereby maintaining the time series characteristics of the data. After processing, the ratio of frozen samples to normal samples in the final dataset is 1:1.

[0043] In this embodiment, a dual cleaning process of box plot denoising and temporal consistency verification is adopted, combined with threshold filtering undersampling and temporal preservation oversampling to balance the sample ratio, reduce the model bias caused by noise and class imbalance, and improve data quality and training stability.

[0044] In the preferred scheme, the specific process of constructing a new feature set and filtering the original features in step S2 is as follows: S21: Analysis of the Failure Mechanism of Wind Turbine Blades Icing: Icing on the surface of wind turbine blades is essentially a phase change process of water under low-temperature conditions. When the ambient temperature around the blades drops below freezing, supercooled droplets in the airflow come into contact with the blade surface, resulting in impact, stagnation, and energy exchange, which then solidifies and gradually accumulates into an ice layer. Ambient temperature is the most direct factor affecting wind turbine blade icing. When the ambient temperature around the unit is below zero, the blades come into contact with moist water vapor, and the moisture in the air condenses into ice and adheres to the surface of the unit's blades. The lower the temperature, the higher the degree of icing. Wind speed affects the convective heat transfer rate on the blade surface. The faster the wind speed, the higher the incoming velocity of liquid water droplets, resulting in more water impacting the blades and more severe icing. High wind speeds increase the impact velocity of water droplets on the blades and the convective heat transfer rate, thus promoting the formation of a thicker ice layer on the blades. Studies have shown that ambient temperature and wind speed are key variables affecting this process.

[0045] Factors affecting icing on wind turbine blades include, but are not limited to, the factors mentioned above. The causes of blade icing are complex, and traditional methods require comprehensive consideration of multiple factors and thorough analysis and evaluation. It is necessary to combine empirical models of blade icing to explore the causes of blade icing.

[0046] S22: Empirical Model Analysis of Wind Turbine Blade Icing: The Makkonen model was used to determine a simple functional relationship between observed parameters and icing mass, in order to extract key parameters characterizing the degree of icing. The icing process mainly consists of three stages: collision, capture, and freezing of supercooled water droplets. The expression for the growth rate of blade icing mass per unit time is: ; In the formula, Collision efficiency – the percentage of supercooled water droplets that can successfully impact the surface of a wind turbine blade; Sticking efficiency – the percentage of supercooled water droplets that adhere to the fan blades after successfully impacting them and do not fall off. Accretion efficiency is the probability that supercooled water droplets will remain frozen on the surface of the wind turbine blades without melting again. The moisture content in the air; Wind speed; This represents the effective cross-sectional area along the vector direction of the water droplet impact on the fan.

[0047] The relationship between icing amount and wind speed is as follows: ; In the formula, the symbol Indicates direct proportion. , , All are constants.

[0048] When using wind turbine monitoring data for modeling, utilize approximation It can fit the above model well; therefore, within the allowable error range, it can be used. The model is used as a substitute. It can be seen that under certain temperature and humidity conditions, water droplets remain uniform, and the amount of ice formation increases with increasing wind speed. Different temperature and humidity environments result in different degrees of ice formation; when the temperature... T and air moisture content When the amount of ice formation changes, it can be expressed as: .

[0049] In summary, based on the simplified analysis of the Makkonen icing empirical model, it was found that the ambient temperature... T Sum of wind speed squared It can serve as an important mechanistic characteristic reflecting the amount of ice formation.

[0050] The expression is: ; In the formula, For ambient wind speed.

[0051] S23: Analysis of the impact of wind turbine blade icing failure on unit operating performance, including: The system constructs power residuals or wind speed to power ratios, constructs ambient temperature differences, and performs balance analysis on blade angle, blade speed, and pitch motor temperature parameters. Redundant features are eliminated, and new features are constructed.

[0052] Blade icing can significantly affect the aerodynamic and strength characteristics of wind turbines. Wind turbine observation data records the operating status of the turbine, and analyzing the impact of blade icing on wind turbine performance helps to discover icing patterns and further extract the corresponding mechanistic characteristics.

[0053] (1) Output power As shown in the above formula for icing amount, under the same environment, the icing mass per unit time increases with increasing wind speed, indicating that wind speed is one of the important influencing factors of wind turbine blade icing. Due to the mechanism of wind turbines, the operating state of the turbine directly affects its power; therefore, it is necessary to analyze the relationship between wind speed and power, two important factors affecting wind turbine blade icing. The wind energy comprehensive utilization model reveals the relationship between wind speed and the active power of the wind turbine, as shown in the following equation: ; In the formula This refers to the windward area of ​​the wind turbine blades. This represents the actual wind energy utilization rate.

[0054] like Figure 2 As shown, wind speed and power The diagram shows a wind power characteristic curve relationship. Blue represents normal wind turbine operation data, while red represents wind turbine icing data. Under normal operating conditions, the wind turbine follows the wind power characteristic curve, and the power tends to stabilize after reaching a threshold. However, icing alters the surface roughness of the turbine blades and the aerodynamic shape of the blade airfoil, reducing the turbine's output power at the same wind speed. The actual output power deviates significantly from the normal output power fitting curve, and this deviation indicates the severity of icing.

[0055] Furthermore, the degree of deviation between the actual output power of the wind turbine and the fitted curve of the output power under normal operating conditions can be used to characterize the icing situation of the wind turbine blades, as shown in the following formula: ; In the formula, This represents the actual output power of the fan at the same wind speed. This represents the theoretical output power of the fan at the same wind speed. The relative residual between actual power and theoretical power is referred to as power residual.

[0056] Wind speed and power are highly correlated and important characteristics in the problem of wind turbine blade icing. WS and power P ratio WSP It can effectively characterize the icing state of leaves, as shown in the following formula: WSP ; In the formula, WS For wind speed, P Power.

[0057] (2) Ambient temperature difference Due to the unique environment in which wind turbines operate, ambient temperature can fluctuate significantly over time and depending on weather conditions. Icing affects the load distribution on the wind turbine blades, leading to abnormal equipment operation. Components within the nacelle, such as the main shaft, gearbox, and bearings, will generate heat, subsequently causing the cooling water to warm up. These changes will result in variations in the nacelle temperature.

[0058] The formula for ambient temperature difference is: ; In the formula, ET For ambient temperature, IT Cabin temperature; (3) Leaf balance analysis like Figures 3-7 As shown, to address the blade imbalance problem that may be caused by blade icing failure, a visualization analysis of the relevant characteristic parameters of the blades was performed. The data includes scatter plots of the relationships between blade angle, blade speed, blade charger DC current, blade DC charging power supply temperature, and blade charger DC current for all blades in both normal and icing states. It can be seen that the icing data for each blade's DC charging power supply temperature and each blade charger DC current largely overlaps with the normal data. The blade angle, blade speed, and pitch motor temperature values ​​of the three blades form a straight line with an inclination angle of 45°, indicating that the correlation between the blade angle, blade speed, pitch motor temperature value, and wind turbine blade icing is extremely low. In order to improve the performance and operating speed of the model, the blade angle, blade speed, and pitch motor temperature values ​​of blades 1, 2, and 3 were removed from the dataset, and their mean values ​​were used as new features.

[0059] The average angle of the wind turbine blades is denoted as MPA The average speed of the blade is denoted as MPS The average temperature of the pitch motor is denoted as MPMT The average temperature of the blade DC charging power supply is denoted as MPNT The average DC current of the blade charger is denoted as MPN ， The five new features are constructed in the following ways: MPA= ; MPS= ; MPMT= ; MPNT= ; MPND= ; In the formula, For the angle of each wind turbine blade, For the speed of each wind turbine blade, Temperature of each pitch motor Temperature of the DC charging power supply for each blade This represents the DC current of each blade charger.

[0060] like Figure 8As shown, the number of horizontal and vertical acceleration values ​​collected by the SCADA system is relatively small, which indirectly reflects that the sensitivity of the sensors collecting these two features is not very high. This results in the x-axis and y-axis acceleration values ​​of the wind turbine being collected as a single value within a certain range. Furthermore, there is significant overlap between the x-axis and y-axis acceleration values ​​in normal and icing data, so they cannot be used as distinguishing features for blade icing. Therefore, these two acceleration features should be removed from the dataset.

[0061] This embodiment integrates the icing phase change mechanism, the Makkonen empirical model, and the influence law on unit performance to construct features with strong physical meaning, such as wind speed square, power residual, wind speed-to-power ratio, ambient temperature difference, and blade balance mean. This significantly improves the ability to characterize the mechanism fusion features, enhances the sensitivity to icing conditions, and strengthens the interpretability of the model.

[0062] In the preferred scheme, step S3 provides statistical correlation evidence through Spearman correlation coefficient and chi-square test, completes model-driven feature importance verification through RFECV-LightGBM, and retains key features with high correlation through comprehensive ranking.

[0063] This embodiment combines the above three methods to more comprehensively evaluate the relationship between features and the icing state of wind turbine blades, retains key features with high correlation to form the optimal feature set for model input, and improves the completeness and accuracy of the data.

[0064] In the preferred solution, such as Figure 9 As shown, in step S4, the time-series statistical features of leaf icing are constructed based on sliding window analysis. The sliding window is used to statistically analyze the multi-time-series features within a certain time window, including six features: maximum value, minimum value, average value, standard deviation, median, and variance, which are used to characterize the leaf icing state, development, and severity.

[0065] By using a sliding window to statistically analyze six characteristics within a certain time window—maximum, minimum, average, standard deviation, median, and variance—it can be effectively used to characterize whether leaves are icing, and the development and severity of icing. In the formula... Represents time-series features. i Representing the i One sample, Representing the total sample size, the expression is: ; ; ; ; ; .

[0066] In this embodiment, by combining sliding window statistics, time-series trends, and variable-scale windows to construct multi-time-series features, the dynamic laws of the entire process of icing from initiation to development to aggravation are accurately depicted, improving the accuracy of identifying early weak icing signals, thereby enhancing the technicality and accuracy of icing early warning.

[0067] In the preferred embodiment, step S5 uses the mechanistic fusion features and temporal statistical features as inputs to build a gradient boosting decision tree model based on the fusion of multiple temporal features to realize the detection of wind turbine blade icing. That is, based on the construction of the temporal statistical features of blade icing by sliding window analysis, the gradient boosting decision tree model is used to realize the detection of wind turbine blade icing.

[0068] This embodiment addresses the challenge of a single model comprehensively characterizing the dynamic process of icing and the coupled effects of multiple mechanisms. It proposes a mechanism-data fusion feature construction method, combining icing formation mechanisms, empirical models, and performance impact analysis to construct physically meaningful temporal features. Furthermore, multi-criteria feature selection is used to filter key variables, enhancing the model's ability to characterize the icing process. The original features failed to fully reflect the physical mechanisms of icing, making it difficult for the model to learn effective discrimination patterns from the data. In the feature engineering stage, this embodiment combines icing formation mechanisms and empirical models to construct new features, giving the data-driven model physical interpretability and enhancing its generalization ability. Moreover, by extracting and fusing temporal dynamic features during the icing process, the model can more sensitively capture early icing and state evolution processes, improving early warning capabilities, enhancing generalization ability under complex weather and variable operating conditions, reducing false alarm rates, and improving detection accuracy.

[0069] Example 2 To further illustrate with reference to Example 1, a wind turbine blade icing detection system based on mechanism feature fusion is provided to implement the method in Example 1, comprising: Data acquisition and preprocessing module: used to acquire data from the wind turbine SCADA system, and perform data tagging, noise processing, and imbalance data processing; Mechanism Feature Construction and Screening Module: Used to analyze the blade icing failure mechanism, icing empirical model and the impact of failure on operating performance, construct a new feature set and remove redundant features; Multi-criteria feature ranking module: Used to perform comprehensive correlation ranking of features using three methods: Spearman correlation coefficient, chi-square test analysis and RFECV-LightGBM, and output the optimal feature set; Temporal feature extraction module: used to construct temporal statistical features of leaf icing based on sliding window analysis; Icing Detection Module: Used to build and run a gradient boosting decision tree model based on multi-temporal feature fusion, and output the icing detection results of wind turbine blades.

[0070] This embodiment provides a working process, working details and technical effects of a wind turbine blade icing detection method based on mechanism feature fusion, which can be referred to in Embodiment 1 and will not be repeated here.

[0071] Example 3 This application provides an electronic device, including a memory and a processor; Memory is used to store computer programs.

[0072] The processor is used to implement the wind turbine blade icing detection method based on mechanism feature fusion of Embodiment 1 when executing a computer program, wherein the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0073] For specific examples, the memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out memory (FIFO), and / or last-in-first-out memory (FILO), etc.; the processor may be limited to microprocessors of the STM32F105 series, ARM (Advanced RISC Machines), x86 architecture processors, or processors with integrated NPU (neural-network processing units).

[0074] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for detecting icing on wind turbine blades based on mechanistic feature fusion, characterized in that, Includes the following steps: S1: Obtain the data set of the wind turbine SCADA system and perform data preprocessing; S2: Obtain the preprocessed dataset and construct a new set of features with physical meaning based on the blade icing formation mechanism, the Makkonen icing empirical model, and the impact of icing on unit operating performance. The new set of features includes icing intensity mechanism features constructed based on the blade icing phase change process and the Makkonen icing mass growth model, performance deviation features constructed based on wind-power characteristics, temperature difference features constructed based on nacelle-environment thermal coupling, and balance features constructed based on three-blade balance characteristics. S3: The optimal feature group is obtained by comprehensively ranking all features based on Spearman correlation coefficient, chi-square test analysis and RFECV-LightGBM; S4: Constructing the time-series statistical characteristics of blade icing based on sliding window analysis; S5: Construct a gradient boosting decision tree model based on multi-time series feature fusion, and input the time series statistical features to be tested to realize wind turbine blade icing detection.

2. The wind turbine blade icing detection method based on mechanism feature fusion according to claim 1, characterized in that, The data preprocessing includes: The data is labeled with its state, noisy data is removed using box plots, and missing values ​​are filled forward. The imbalance problem is addressed by combining threshold filtering for undersampling with time-preserving SMOTE oversampling. The specific process is as follows: S11: Obtain data from the wind turbine SCADA system; S12: Classify each data point according to the time periods of normal and icing states, and remove samples with unclear states; S13: Noisy data is identified and removed using box plots. Missing values ​​after removal are filled using a forward imputation method. The formulas for calculating the upper and lower limits are as follows: ； In the formula, k As control terms, Q1 and Q2 are the upper and lower quartiles, respectively; S14: Undersampling is performed using a threshold-based filtering method to remove unreasonable and shutdown data; and oversampling is performed using Smote, which preserves time series characteristics, to make the ratio of frozen samples to normal samples close to the preset value.

3. The wind turbine blade icing detection method based on mechanism feature fusion according to claim 2, characterized in that, The SCADA dataset mentioned in step S1 contains 26 sets of continuous numerical features, including operating condition parameters, environmental parameters, and state parameters, with a sampling interval of 7 seconds. When denoising box plots, k=3 is used to identify extreme singularities. The preset ratio of normal samples to frozen samples after imbalance treatment is 1:

1.

4. The wind turbine blade icing detection method based on mechanism feature fusion according to claim 1, characterized in that, In step S2, the specific process of constructing a new feature set and filtering the original features is as follows: S21: Analysis of the failure mechanism of wind turbine blade icing, identifying ambient temperature and wind speed as key variables; S22: Extracting ambient temperature based on Makkonen's icing empirical model. With the square of wind speed As an important mechanistic characteristic reflecting the amount of icing, the formula is: ； In the formula, For ambient wind speed; S23: Analysis of the impact of wind turbine blade icing failure on unit operating performance, including: The system constructs power residuals or wind speed to power ratios, constructs ambient temperature differences, and performs balance analysis on blade angle, blade speed, and pitch motor temperature parameters. Redundant features are eliminated, and new features are constructed.

5. The wind turbine blade icing detection method based on mechanism feature fusion according to claim 4, characterized in that, Step S23 specifically includes: Based on wind power characteristics, the formula for the power residual and wind speed-power ratio is as follows: ； WSP ； In the formula, This represents the actual output power of the fan at the same wind speed. This represents the theoretical output power of the fan at the same wind speed. WS For wind speed, P Power; The temperature difference characteristics are constructed based on the temperature difference between the environment and the cabin, and the formula is as follows: ； In the formula, ET For ambient temperature, IT Cabin temperature; Mean characteristics were constructed based on the three-blade balance analysis, including the average blade angle. 、 The average blade speed, average pitch motor temperature, average blade DC charging power supply temperature, and average blade charger DC current are expressed as follows: MPA= ； MPS= ； MPMT= ； MPNT= ； MPND= ； In the formula, For the angle of each wind turbine blade, For the speed of each wind turbine blade, Temperature of each pitch motor Temperature of the DC charging power supply for each blade The DC current of each blade charger; At the same time, redundant features such as single blade angle, single blade speed, single pitch motor temperature, and x / y direction acceleration are eliminated.

6. The wind turbine blade icing detection method based on mechanism feature fusion according to claim 1, characterized in that, In step S3, statistical correlation is provided through Spearman correlation coefficient and chi-square test, and model-driven feature importance verification is completed through RFECV-LightGBM. The comprehensive ranking retains key features with high correlation.

7. The wind turbine blade icing detection method based on mechanism feature fusion according to claim 1, characterized in that, In step S4, the time-series statistical features of leaf icing are constructed based on sliding window analysis. The maximum, minimum, average, standard deviation, median, and variance of the features within a certain time window are statistically analyzed using the sliding window. These features are used to characterize the state, development, and severity of leaf icing.

8. The wind turbine blade icing detection method based on mechanism feature fusion according to claim 1, characterized in that, In step S5, the mechanism fusion features and time-series statistical features are used as inputs, and a gradient boosting decision tree is used to realize the binary classification detection of wind turbine blade icing and obtain the detection results.

9. A wind turbine blade icing detection system based on mechanism feature fusion, characterized in that, The system for implementing the method of any one of claims 1 to 8 comprises: The data acquisition and preprocessing module is used to acquire data from the wind turbine SCADA system and perform data tagging, noise processing, and imbalance data processing. The mechanism feature construction and screening module is used to acquire the preprocessed dataset and construct a new set of features with physical meaning based on the blade icing formation mechanism, the Makkonen icing empirical model, and the impact of icing on unit operating performance. The new feature set includes icing intensity mechanism features constructed based on the blade icing phase change process and the Makkonen icing mass growth model, performance deviation features constructed based on wind-power characteristics, temperature difference features constructed based on nacelle-environment thermal coupling, and equilibrium features constructed based on the three-blade equilibrium characteristics. The multi-criteria feature ranking module is used to perform comprehensive correlation ranking of features using three methods: Spearman correlation coefficient, chi-square test analysis and RFECV-LightGBM, and output the optimal feature set. The temporal feature extraction module is used to construct temporal statistical features of leaf icing based on sliding window analysis; The icing detection module is used to build and run a gradient boosting decision tree model based on multi-temporal feature fusion, and output the icing detection results of wind turbine blades.

10. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement, when executing the computer program, the wind turbine blade icing detection method based on mechanistic feature fusion as described in any one of claims 1 to 8.

Images