An icing blade aeroelastic flutter and whole machine mistuning identification method based on multi-dimensional dynamic feature engineering

Abstract

This invention discloses a method for identifying aeroelastic flutter of icing blades and overall turbine mistuning based on multidimensional dynamic feature engineering, belonging to the field of wind power technology. The method includes acquiring raw multivariate time series data and preprocessing it. Features are extracted from the preprocessed raw multivariate time series data from three dimensions: time domain, time-frequency domain wavelet energy, and control system domain, to obtain basic physical features. These basic physical features are then filtered based on the Gini index to obtain fault seed features. Interactive and adaptive features are constructed based on the fault seed features. The basic physical features, interactive features, and adaptive features are fused into a high-dimensional feature set, which is then input into a weighted ensemble learning classifier to identify the health, aeroelastic flutter, and overall turbine mistuning state of icing blades. This invention solves the problems of poor robustness and low accuracy in fault identification of traditional methods, providing a reliable guarantee for the safe and stable operation of wind turbines under icing conditions.

Description

Technical Field

[0001] This invention relates to the field of wind power technology, specifically to a method for identifying aeroelastic flutter of icing blades and overall turbine detuning based on multidimensional dynamic characteristic engineering. Background Technology

[0002] As wind power technology advances towards higher power output and longer, more flexible blades, the problem of blade icing in wind turbines operating in cold climates is becoming increasingly severe. Uneven icing alters the original aerodynamic shape and mass distribution of the blades, potentially inducing dangerous self-excited vibrations in the blade structure, known as aeroelastic flutter. Furthermore, it disrupts the overall mass and aerodynamic balance of the rotor, leading to turbine detuning and exacerbating fatigue loads on the drivetrain and tower. Both of these failure conditions seriously threaten the structural safety of the unit, reduce power generation efficiency, and shorten equipment lifespan.

[0003] While modern wind turbines can collect high-frequency operating data down to the millisecond level, existing data analysis methods still face challenges in handling complex, time-varying, and nonlinear operating conditions such as icing. Early fault characteristic signals such as aeroelastic flutter or overall turbine detuning are typically extremely weak. Traditional time-domain statistical analysis and frequency-domain analysis methods based on Fast Fourier Transform (FFT) are limited in their ability to process non-stationary and transient signals, making it difficult to effectively separate fault characteristics from the operating environment, resulting in low identification accuracy and a high false alarm rate.

[0004] More importantly, the features extracted by existing methods are usually static and lack adaptability to operating conditions. The vibration and control characteristics of wind turbines vary greatly under different wind speeds and power levels, making it difficult for static feature thresholds to be applicable to all operating conditions, resulting in insufficient robustness of diagnostic models. How to extract dynamic features that can sense the current operating conditions and perform self-adjustment and normalization is a problem that existing technologies have not yet solved. In addition, the data imbalance caused by the scarcity of fault samples in actual operation also seriously restricts the generalization ability and practical application effect of intelligent diagnostic models.

[0005] To address the aforementioned issues, there is an urgent need for a method for identifying aeroelastic flutter of icing blades and overall system mistuning based on multidimensional dynamic feature engineering. This method would solve the problems associated with traditional methods and improve the ability to identify aeroelastic flutter of icing blades and overall system mistuning. Summary of the Invention

[0006] The purpose of this invention is to provide a method for identifying aeroelastic flutter of icing blades and overall turbine detuning based on multidimensional dynamic feature engineering. This method solves the problems of poor robustness and low accuracy in fault identification of traditional methods, and provides a reliable guarantee for the safe and stable operation of wind turbines under icing conditions.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for identifying aeroelastic flutter of icing blades and overall system mistuning based on multidimensional dynamic feature engineering includes:

[0009] Step 1: Collect raw multivariate time series data and preprocess it;

[0010] Step 2: Extract features from the preprocessed original multivariate time series data from three dimensions: time domain, time-frequency domain wavelet energy, and control system domain, to obtain the basic physical features;

[0011] Step 3: Based on the Gini index, perform feature screening on the basic physical characteristics to obtain fault seed characteristics;

[0012] Step 4: Construct interactive features and adaptive features based on fault seed features;

[0013] Step 5: Fuse the basic physical features, interactive features, and adaptive features into a high-dimensional feature set. The high-dimensional feature set will be input into a weighted ensemble learning classifier to identify the health of icing blades, aeroelastic flutter, and overall detuning state.

[0014] Further, in step 1, raw multivariate time series data is collected and preprocessed, specifically as follows:

[0015] Simultaneously acquire millisecond-level operating data from the wind turbine monitoring and data acquisition system SCADA and high-frequency sensor data from the independent pitch controller IPC;

[0016] The operational data and high-frequency sensor data are aligned based on high-precision timestamps and divided into continuous time windows. Each data sample typically corresponds to a continuous operational data window with a duration of 1 to 10 minutes.

[0017] For missing and outlier values ​​within the window, a time-based interpolation algorithm and a pre- and post-value imputation method are used to process them, ultimately obtaining time series data samples.

[0018] Furthermore, the basic physical features extracted from the time domain dimension include RMS value, kurtosis, skewness, peak-to-peak value, root square amplitude, margin factor, impulse factor, and peak factor.

[0019] Furthermore, feature extraction is performed on the preprocessed original multivariate time series data from the wavelet energy dimension in the time-frequency domain, specifically as follows:

[0020] Continuous wavelet transform is used to perform time-frequency analysis on the original multivariate time series data. Wavelet coefficients are calculated and converted into energy time-frequency spectrum. The mean, standard deviation, kurtosis and maximum value of the energy time series are calculated within a preset frequency band and used as the basic physical characteristics of the energy dimension of the time-frequency domain wavelet.

[0021] Furthermore, the basic physical features extracted from the control system domain include blade standard deviation, blade inconsistency, mean tracking error, and maximum tracking error.

[0022] Furthermore, in step 3, feature filtering is performed on the basic physical characteristics based on the Gini index to obtain fault seed features, specifically:

[0023] Obtain the basic physical characteristics of all samples in the original multivariate time series data, construct a feature set, use the true state of the samples as labels, and train a random forest classifier based on the labels and feature set;

[0024] The average reduction in Gini impurity of each feature is extracted as the feature importance based on the trained random forest classifier, and fault seed features are selected by combining the fault mechanism.

[0025] Furthermore, the fault seed features include vibration kurtosis and tracking error mean, wherein the vibration kurtosis serves as a seed feature indicating aeroelastic flutter, and the tracking error mean serves as a seed feature indicating overall machine detuning.

[0026] Furthermore, in step 4, interactive features are constructed based on the fault seed features, specifically as follows:

[0027] The operating condition characteristics representing the real-time energy level of the unit are obtained and nonlinearly combined with the fault seed characteristics to construct the operating condition interaction characteristics.

[0028] Furthermore, in step 4, adaptive features are constructed based on the fault seed features, specifically as follows:

[0029] Using data from samples of multiple health states, a dynamic health threshold that can represent the health state of the local group is calculated for each seed feature. The seed feature value of the current sample and the dynamic health threshold are normalized to construct an adaptive feature.

[0030] In summary, the present invention has at least one of the following beneficial technical effects:

[0031] 1. High fault sensitivity: It creatively constructs an interactive feature of operating conditions, and nonlinearly combines the fault seed features selected through importance assessment with the operating condition features that represent the load. It can intelligently amplify the fault representation under high load conditions where faults are prone to occur, and effectively enhance the early weak fault signals.

[0032] 2. Strong feature robustness: By using the dynamic benchmark normalization method and establishing dynamic thresholds based on the wind turbine's own health data, the constructed features can adapt to the basic operating state of the unit, effectively overcoming feature drift caused by changes in operating conditions.

[0033] 3. High diagnostic accuracy and generalization ability: High-quality synthetic samples are created by scaling noise according to feature scale. Combined with the class weight balancing mechanism in model training and the hierarchical cross-validation strategy in evaluation, and finally adopting a performance-based dynamic weighted ensemble learning model, the classifier still has high accuracy and strong generalization ability when facing rare fault samples. From the four levels of data, model, evaluation to decision, it is ensured that the classifier still has high accuracy and strong generalization ability when facing rare fault samples.

[0034] 4. Strong systematic approach: A complete feature engineering framework is proposed, from multi-domain feature extraction and dynamic feature construction to state identification, providing a systematic solution for fault diagnosis problems under similar complex working conditions. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0036] Figure 2 This is a principle block diagram constructed based on the interactive features of the operating conditions;

[0037] Figure 3 This is a principle block diagram constructed based on dynamic benchmark normalization features;

[0038] Figure 4 This is a structural diagram of a weighted ensemble learning classification model. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0040] like Figure 1 As shown, this invention provides a method for identifying aeroelastic flutter of icing blades and overall engine detuning based on multidimensional dynamic feature engineering, including:

[0041] Step 1: Collect raw multivariate time series data and preprocess it;

[0042] Step 2: Extract features from the preprocessed original multivariate time series data from three dimensions: time domain, time-frequency domain wavelet energy, and control system domain, to obtain the basic physical features;

[0043] Step 3: Based on the Gini index, perform feature screening on the basic physical characteristics to obtain fault seed characteristics;

[0044] Step 4: Construct interactive features and adaptive features based on fault seed features;

[0045] Step 5: Fuse the basic physical features, interactive features, and adaptive features into a high-dimensional feature set. The high-dimensional feature set will be input into a weighted ensemble learning classifier to identify the health of icing blades, aeroelastic flutter, and overall detuning state.

[0046] The dataset processed by the method of this invention is a labeled dataset. ,in, This represents the original multivariate time series data of the i-th sample. N represents the total number of samples, and the data was collected from wind turbines under conditions where the blades were at risk of icing or had already iced.

[0047] In step 1, raw multivariate time series data is collected and preprocessed, specifically as follows:

[0048] The system simultaneously acquires millisecond-level operating data from the SCADA system and high-frequency sensor data from the IPC (Independent Pitch Controller). The data stream includes, but is not limited to, the active power of the turbine, blade vibration acceleration, and actual and commanded pitch angles of the three blades. The sampling frequency of the high-frequency data is preferably not less than 10Hz to ensure effective capture of high-frequency dynamic information such as flutter. The multi-source data is aligned based on a high-precision timestamp and divided into continuous time windows for analysis. Each data sample typically corresponds to a continuous operating data window with a duration of 1 to 10 minutes. Subsequently, a time-based interpolation algorithm and pre- and post-value filling method are used to process missing and outlier values ​​within the window to form high-quality time series data samples.

[0049] In step 2, feature extraction is performed on the preprocessed original multivariate time series data from three dimensions: time domain, time-frequency domain wavelet energy, and control system domain, to obtain the basic physical features, specifically:

[0050] 1. Feature extraction from preprocessed raw multivariate time series data in the time domain.

[0051] For the active power and blade vibration acceleration signals of the generator unit, a set of statistical characteristics that can comprehensively describe its dynamic behavior from multiple dimensions such as waveform distribution, energy level, and impact characteristics are calculated. Specifically, these characteristics include the following eight features:

[0052] The RMS (Root Mean Square) value is the most commonly used and classic indicator for measuring the overall energy or amplitude level of a signal. The more volatile the signal, the larger its RMS value. Its calculation formula is:

[0053] (1)

[0054] cliff ( Kurtosis (KK) measures the degree of peaks or spikes in the waveform distribution of a signal. A normal signal, such as Gaussian white noise, has a KK value of approximately 3. When impulsive components, such as early flutter, appear in the signal, the KK value increases significantly, making it a sensitive indicator for diagnosing impulsive faults. Its calculation formula is:

[0055] (2)

[0056] Skewness measures the asymmetry of a signal waveform about its mean. If the waveform is symmetrical, the skewness is zero. Certain unidirectional wear or asymmetrical faults can cause the signal waveform to be clipped or stretched in one direction, resulting in significant skewness. The formula for calculating skewness is:

[0057] (3)

[0058] Peak-to-peak value (P2P) measures the maximum fluctuation range or dynamic range of a signal over the entire time window, i.e., the maximum value of the signal. and minimum value The difference. The formula for calculating it is:

[0059] (4)

[0060] Root amplitude ( Similar to the RMS value, the square root (RRoot) is also a way to measure signal amplitude, but due to the square root operation, it is more sensitive to small fluctuations in the signal. Its calculation formula is:

[0061] (5)

[0062] The margin factor (MF) is a dimensionless indicator that measures the ratio of a signal's peak value to its root square amplitude. During impulsive faults, the peak value increases significantly, while the root square amplitude, which primarily reflects background fluctuations, remains relatively unchanged, causing the MF to rise sharply. This makes it a powerful indicator for diagnosing impulsive faults. Its calculation formula is:

[0063] (6)

[0064] The impulse factor (IF) measures the severity of signal spikes relative to the average daily fluctuation level; it is the ratio of the signal's peak value to its mean absolute value. The formula is:

[0065] (7)

[0066] Peak factor (CF), similar to impulse factor, measures impulse from an energy perspective; it is the ratio of the signal's peak value to its effective energy. Its calculation formula is:

[0067] (8)

[0068] In the above formula, Let N be the amplitude of the i-th sampling point in the time series signal, and N be the total length of the time series signal. This is the arithmetic mean of the signal sequence. Let x be the standard deviation of the signal sequence, and max(x) and min(x) be the maximum and minimum values ​​of the signal sequence. It represents the absolute value of the signal amplitude.

[0069] 2. Feature extraction from the preprocessed original multivariate time series data using wavelet energy in the time-frequency domain.

[0070] To capture non-stationary signals such as flutter, continuous wavelet transform (CWT) is used to perform time-frequency analysis on the vibration signals. Preferably, the complex Morlet wavelet is selected as the mother wavelet. Because of its good balance between time and frequency resolution, it is particularly suitable for accurately capturing the local characteristics of non-stationary, transient signals such as aeroelastic flutter. The formula for calculating the continuous wavelet transform of the vibration sequence x(t) is as follows:

[0071] (9)

[0072] Where 'a' represents the scale and 'b' represents the time shift. The result of the transformation is a set of wavelet coefficients. Each coefficient The magnitude of represents the intensity of the frequency component of signal x(t) at time point b, corresponding to scale a.

[0073] For further analysis, the wavelet coefficient matrix is ​​first transformed into an energy-time spectrum, that is, the energy at each time frequency point is calculated. The calculation formula is as follows:

[0074] (10)

[0075] Then, based on the wind turbine's dynamic characteristics, energy analysis is performed within three preset key frequency bands on this energy time-frequency spectrum. Preferably, within the 1P band, 3P band, and high-frequency flutter band, the energy time series of each band is obtained by averaging the energy values ​​of all frequencies within that band. The mean, standard deviation, kurtosis, and maximum value of this energy time series are then calculated as the final time-frequency domain characteristics. The 1P band range is set based on the fact that the rated operating speed range of large wind turbine units is typically 5-15 rpm; the 1P band can be set to 0.08Hz to 0.26Hz, the 3P band to 0.24Hz to 0.76Hz, and the high-frequency flutter band to 0.8Hz to 3Hz.

[0076] 3. Feature extraction from the preprocessed raw multivariate time series data from the perspective of the control system domain.

[0077] This section describes features used to quantify the performance of the pitch control system and the coordination of the three blade motions from different dimensions. This plays a crucial role in diagnosing overall engine mistuning caused by aerodynamic imbalances. Specifically, it includes the following features:

[0078] Blade standard deviation ( This feature calculates the standard deviation of the pitch angle time series of each of the three blades over the entire time window and takes their average. It is mainly used to measure the overall activity or intensity of the blade movement during pitch control. Under unstable conditions, blade movement will be more frequent, and the blade standard deviation will increase accordingly. The calculation formula is:

[0079] (11)

[0080] Among them, the standard deviation of the time series of the pitch angle of each blade The calculation method is as follows:

[0081] (12)

[0082] in, For the j-th blade at discrete time point t... The actual measured pitch angle value, Let j be the time series of the pitch angle of the j-th blade. The average time value is the average time value over the entire time window, where T is the total number of sampling points within that time window.

[0083] Blade inconsistency ( This feature is used to robustly quantify the degree of coordination among the three blades during motion. Increased inconsistency typically indicates differences in individual blade behavior caused by factors such as icing, control command delays, or actuator malfunctions. The calculation formula is:

[0084] (13)

[0085] Among them, the average pitch angle at each moment The calculation formula is:

[0086] (14)

[0087] Tracking error is a feature used to accurately characterize the precision of the pitch system's execution commands. When blade icing causes abnormal aerodynamic loads, these loads may exceed the actuator's capabilities, resulting in the actual pitch angle failing to accurately track the commanded value, thus increasing the tracking error. This invention statistically analyzes its mean and maximum values ​​to comprehensively evaluate this performance.

[0088] Mean tracking error ( The calculation formula is as follows:

[0089] (15)

[0090] Tracking error ( The formula for calculating the maximum value is:

[0091] (16)

[0092] in, For the j-th blade at discrete time point t... The actual measured pitch angle value, This refers to the pitch angle setpoint commanded by the pitch control system to all blades at a discrete time point t. To take the maximum value across all three blades and the entire time window.

[0093] In step 3, feature selection is performed on the basic physical characteristics based on the Gini index to obtain fault seed features, specifically:

[0094] To ensure the physical orientation and effectiveness of subsequent dynamic feature construction, the basic physical features extracted in step 2 need to be evaluated for importance, and the seed features most sensitive to the target fault need to be selected. This invention preferably uses a feature importance ranking algorithm based on the Gini index. Step 3, based on all the extracted basic physical features, selects fault seed features based on the Gini index as follows:

[0095] 1. Initial training and evaluation model: All basic physical features extracted for all training samples in step 2 are used as the feature set, and the true state of the samples (healthy, aeroelastic flutter, overall detuning) is used as the label. Using this feature set and labels, a random forest classifier is initially trained. The core purpose of this training is not to obtain a final diagnostic model, but to utilize the ability of random forest to automatically evaluate the contribution of features during training.

[0096] After initial training, the average reduction in Gini impurity for each basic feature in distinguishing between healthy, flutter, and system misalignment samples is extracted from the trained random forest model; this reduction is also called feature importance. For each decision tree in the forest, the greater the reduction in Gini impurity resulting from a node split, the stronger the discriminative power of the feature used for the split. For the i-th decision tree in the forest, the Gini index of node q... The calculation formula is:

[0097] (17)

[0098] Where C represents the number of categories. Let be the proportion of category c in the i-th tree node q.

[0099] Calculate the importance score of the j-th feature based on the Gini index. The principle is to calculate the average reduction in Gini impurity brought about by this feature across all decision trees. The calculation formula is:

[0100] (18)

[0101] In the formula, K is the total number of decision trees, and Q is the set of all nodes where feature j appears in the i-th tree. and The Gini index of the two new nodes after node q splits.

[0102] For all features The scores are sorted to obtain a feature importance ranking.

[0103] 2. Seed Feature Selection. To ensure the physical interpretability and fault-specificity of the seed features, this invention combines expert knowledge and fault mechanisms to select features with a strong physical correlation to the target fault phenomenon from the top of the ranking list as the final seed features. In this invention, vibration kurtosis is selected as the seed feature indicating aeroelastic flutter because the kurtosis index is extremely sensitive to impact vibration signals, can effectively capture the transient dynamic behavior of early flutter, has a high importance score, and a clear physical mechanism. The mean tracking error is selected as the seed feature indicating overall system detuning because it directly quantifies the decrease in blade control accuracy caused by blade aerodynamic imbalance, and is also an optimal result combining score and mechanism. At the same time, operating condition features that can represent the real-time load level of the unit, such as the effective power value, are selected for subsequent interactive feature construction.

[0104] In step 4, interactive features and adaptive features are constructed based on the fault seed features, specifically as follows:

[0105] Based on the fault seed features selected in step 3, advanced dynamic features that are robust to operating conditions and sensitive to faults are constructed from two dimensions, achieving a leap from static physical features to dynamic intelligent features. Specifically:

[0106] 1. Construction of Operating Condition Interaction Features. This feature is designed to address the problem that early, weak fault characteristics are easily drowned out by strong background noise during operation. For example... Figure 2 As shown, by using fault seed features Operating condition characteristics representing the real-time energy level of the unit By performing nonlinear combinations, a new working condition interaction feature is constructed. The core idea is to utilize the properties of the logarithmic function log(1+x). At low power levels, it has a relatively small impact on the characteristics, but in high-power fault-prone areas, it can nonlinearly and significantly amplify the fluctuations in fault characteristics, thereby improving the sensitivity of fault detection under critical operating conditions. The calculation formula is as follows:

[0107] (19)

[0108] in, Features of fault seeds selected from the leaderboard To characterize the operating conditions of the load, This is the final constructed working condition interaction feature.

[0109] In this invention, based on the results selected in step 3, vibration kurtosis and the mean tracking error are chosen as fault feature seeds. Select the effective power value as the operating characteristic of the load. A vibration kurtosis-power interaction index is constructed by selecting vibration kurtosis and RMS power values ​​to enhance sensitivity to aeroelastic flutter. The specific calculation formula is as follows:

[0110] (20)

[0111] The tracking error-power interaction index is constructed by selecting the mean tracking error and the effective power value to enhance sensitivity to overall system detuning. The specific calculation formula is as follows:

[0112] (twenty one)

[0113] 2. Adaptive Feature Construction. The purpose of constructing this feature is to characterize how many times the fault indication level of the current sample is compared to the health baseline. It eliminates drift caused by individual unit differences and changes in basic operating conditions, exhibiting stronger robustness. The adaptive characteristic is specifically reflected in two aspects: First, the dynamic health baseline is calculated using the health data of the monitored unit itself, rather than using a fixed value applicable to all units, thus adapting to individual differences between different units; second, the dynamic health baseline can be recalibrated as needed using the latest health data of the unit at different life stages, such as after major overhauls or seasonal changes, thereby adapting to performance drift caused by equipment aging or environmental changes. Figure 3 As shown, firstly, using sample data from multiple health states, a dynamic health threshold (T) representing the health state of the unit is calculated for each seed feature. The calculation method for this threshold originates from the 3σ criterion in statistical process control. This criterion, based on normal distribution theory, considers events in a stable system where a value deviates from its mean μ by more than three standard deviations σ as events with extremely low probability. The calculation formula is as follows:

[0114] (twenty two)

[0115] in, and , respectively, represent the mean and standard deviation of a certain seed feature across all healthy sample sets. T is the dynamic health threshold calculated from the healthy samples.

[0116] Then the current value of the seed feature of the sample to be diagnosed. Normalization is performed using the following formula to construct normalized features. :

[0117] (twenty three)

[0118] in, This represents the actual value of a seed feature of the current sample to be diagnosed. The final constructed dynamic benchmark normalized features are the adaptive features.

[0119] In this invention, an adaptive vibration kurtosis index is selected to construct the vibration kurtosis structure, and its calculation formula is as follows:

[0120] Adaptive vibration kurtosis index = (twenty four)

[0121] The adaptive tracking error exponent is constructed by selecting the tracking error, and its calculation formula is as follows:

[0122] Adaptive tracking error index = (25)

[0123] In step 5, the basic physical features, interaction features, and adaptive features are fused into a high-dimensional feature set. This high-dimensional feature set is then input into a weighted ensemble learning classifier to identify the health of icing blades, aeroelastic flutter, and overall system detuning. Specifically:

[0124] 1. Data Augmentation and Final Feature Set Construction. The basic features, interaction features, and adaptive features extracted and constructed in steps 2 and 4 are fused into a high-dimensional feature vector. For fault categories with small sample sizes in the training set, a noise injection method scaling by feature scale is used for data augmentation. Specifically, a Gaussian white noise matrix with a mean of 0 and a preset small standard deviation (e.g., 0.02) is generated; the standard deviation of each feature column in the minority class sample set is calculated, and the corresponding column in the noise matrix is ​​scaled using this standard deviation to obtain scaled noise that matches the feature scale; finally, the scaled noise is added to the sample features to form new synthetic samples, balancing the training dataset.

[0125] 2. High-dimensional feature standardization. To eliminate the adverse effects of differences in dimensions and numerical ranges between different features on model performance, especially support vector machines, standardization is required before inputting the final feature set into the model. Preferably, the Z-score standardization method is used, which transforms each feature dimension into a standard normal distribution with a mean of 0 and a standard deviation of 1 by subtracting the mean of the feature column and dividing by its standard deviation, thus ensuring that all features are on the same scale.

[0126] 3. Optimization, construction, and training of the weighted ensemble learning model. To systematically address the data imbalance problem and maximize model performance, this invention constructs a weighted ensemble model consisting of heterogeneous base learners, preferably a random forest classifier and a support vector machine classifier.

[0127] (1) Base Learner Hyperparameter Optimization: To ensure the model's generalization ability on imbalanced data, a hierarchical K-fold cross-validation method is used to optimize the model's hyperparameters. This method ensures that the proportions of healthy, flutter, and detuned samples in each fold remain consistent with the original dataset when partitioning the dataset. Specifically, through grid search and other methods, the optimal hyperparameter combinations are found for the two heterogeneous base learners, Random Forest (RF) and Support Vector Machine (SVC), respectively, within the framework of hierarchical K-fold cross-validation.

[0128] (2) Construction of weighted ensemble learning classifier: After the base learner is optimized, a weighted ensemble learning classifier based on a soft voting mechanism is finally constructed. In an exemplary configuration, the random forest classifier may contain 100 decision trees, and the support vector machine may use the radial basis function (RBF) as the kernel function.

[0129] (3) Final Model Training and Weight Allocation: During the final model training, the built-in class weight balancing mechanism of the base learners will be enabled to assign higher misclassification penalties to classes with fewer samples, thereby further mitigating the impact of sample imbalance. At the same time, the final voting weights of the ensemble learning classifiers are not fixed, but are dynamically allocated based on the performance of each base learner in cross-validation. The base learner with better performance will have a greater say in the final decision.

[0130] 4. Model Performance Evaluation. To comprehensively and objectively verify the diagnostic performance of the constructed model, the ensemble prediction results generated by the hierarchical K-fold cross-validation process described above are directly used for evaluation. The effectiveness of the model is quantified by calculating a set of recognized and comprehensive evaluation metrics, specifically including:

[0131] Confusion matrix: Used to visually examine the specific classification performance of the model in the three categories of health, aeroelastic flutter and whole machine detuning, and to analyze the possible confusion categories.

[0132] F1 score, precision, and recall: These comprehensively evaluate the model's overall classification ability when faced with imbalanced data, with particular attention to its precision and recall for minority class fault samples.

[0133] Overall accuracy: measures how well the model correctly classifies all samples.

[0134] 5. State Identification. When new sample data to be diagnosed is input, the final high-dimensional feature vector is first extracted according to the process of steps 1 to 4; then the feature vector is input into the trained final model, and the model will output the state category of the sample through a weighted soft voting mechanism, thereby achieving accurate identification of the health of icing blades, aeroelastic flutter or overall detuning state.

[0135] This invention provides an electronic device including a memory and a processor. The memory stores a computer program, and the processor runs the computer program to enable the electronic device to perform the method described above. The electronic device can be an edge computing device deployed on wind turbine blades, or a remote monitoring center server for receiving and processing data from the entire wind farm.

[0136] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0137] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0138] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0139] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0140] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0141] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A method for identifying aeroelastic flutter of icing blades and overall system mistuning based on multidimensional dynamic feature engineering, characterized in that, include: Step 1: Collect raw multivariate time series data and preprocess it; the raw multivariate time series data comes from the millisecond-level operating data of the wind turbine monitoring and data acquisition system SCADA and the high-frequency sensor data of the independent pitch controller IPC, which are acquired synchronously. Step 2: Extract features from the preprocessed original multivariate time series data from three dimensions: time domain, time-frequency domain wavelet energy, and control system domain, to obtain the basic physical features; The basic physical features extracted from the control system domain include blade standard deviation, blade inconsistency, mean tracking error, and maximum tracking error. Step 3: Based on the Gini index, perform feature screening on the basic physical characteristics to obtain fault seed characteristics; Step 4: Construct interactive features and adaptive features based on fault seed features; Interaction features are constructed based on fault seed features, specifically as follows: The operating condition characteristics representing the real-time energy level of the unit are obtained and nonlinearly combined with the fault seed characteristics to construct the operating condition interaction characteristics. Adaptive features are constructed based on fault seed features, specifically as follows: Using data from samples of multiple health states, a dynamic health threshold that can represent the health state of the local group is calculated for each seed feature. The seed feature value of the current sample and the dynamic health threshold are normalized to construct an adaptive feature. Step 5: Fuse the basic physical features, interactive features, and adaptive features into a high-dimensional feature set. The high-dimensional feature set will be input into a weighted ensemble learning classifier to identify the health of icing blades, aeroelastic flutter, and overall detuning state.

2. The method for identifying aeroelastic flutter of icing blades and overall system mistuning based on multidimensional dynamic feature engineering according to claim 1, characterized in that, In step 1, raw multivariate time series data is collected and preprocessed, specifically as follows: Simultaneously acquire millisecond-level operating data from the wind turbine monitoring and data acquisition system SCADA and high-frequency sensor data from the independent pitch controller IPC; The operational data and high-frequency sensor data are aligned based on high-precision timestamps and divided into continuous time windows. Each data sample typically corresponds to a continuous operational data window with a duration of 1 to 10 minutes. For missing and outlier values ​​within the window, a time-based interpolation algorithm and a pre- and post-value imputation method are used to process them, ultimately obtaining time series data samples.

3. The method for identifying aeroelastic flutter of icing blades and overall system mistuning based on multidimensional dynamic feature engineering according to claim 2, characterized in that, The basic physical features extracted from the time domain include RMS value, kurtosis, skewness, peak-to-peak value, root square magnitude, margin factor, impulse factor, and peak factor.

4. The method for identifying aeroelastic flutter of icing blades and overall system mistuning based on multidimensional dynamic feature engineering according to claim 3, characterized in that, Feature extraction is performed on the preprocessed original multivariate time series data from the wavelet energy dimension in the time-frequency domain, specifically as follows: Continuous wavelet transform is used to perform time-frequency analysis on the original multivariate time series data. Wavelet coefficients are calculated and converted into energy time-frequency spectrum. The mean, standard deviation, kurtosis and maximum value of the energy time series are calculated within a preset frequency band and used as the basic physical characteristics of the energy dimension of the time-frequency domain wavelet.

5. The method for identifying aeroelastic flutter of icing blades and overall system mistuning based on multidimensional dynamic feature engineering according to claim 4, characterized in that, In step 3, feature selection is performed on the basic physical characteristics based on the Gini index to obtain fault seed features, specifically: Obtain the basic physical characteristics of all samples in the original multivariate time series data, construct a feature set, use the true state of the samples as labels, and train a random forest classifier based on the labels and feature set; The average reduction in Gini impurity of each feature is extracted as the feature importance based on the trained random forest classifier, and fault seed features are selected by combining the fault mechanism.

6. The method for identifying aeroelastic flutter of icing blades and overall system mistuning based on multidimensional dynamic feature engineering according to claim 5, characterized in that, The fault seed features include vibration kurtosis and tracking error mean, wherein the vibration kurtosis serves as a seed feature indicating aeroelastic flutter, and the tracking error mean serves as a seed feature indicating overall machine detuning.

Images