Fan blade icing detection method and device and terminal equipment
By using SCADA data analysis and optimizing the SVM model with deep belief networks and genetic algorithms, high-order features are automatically extracted, solving the problems of low accuracy and insufficient generalization ability in wind turbine blade icing detection, and achieving efficient and easy-to-operate fault diagnosis.
Patent Information
- Application Number
- CN202411294282.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-14
- Publication Date
- 2024-11-15
AI Technical Summary
Existing methods for detecting icing on wind turbine blades suffer from insufficient generalization ability and low detection accuracy. Furthermore, they rely on specialized knowledge and are difficult to operate and maintain.
By collecting SCADA data, extracting features related to wind turbine blade icing, and combining deep belief networks and genetic algorithms to optimize the SVM model, high-order features are automatically extracted and the detection model is trained to form a wind turbine blade icing detection device and terminal equipment.
It significantly improves the accuracy and efficiency of wind turbine blade icing fault diagnosis, reduces reliance on specialized knowledge, and enhances the model's generalization ability and ease of operation.
Smart Images

Figure CN118959252A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of fan blade icing detection, and in particular to a fan blade icing detection method, device and terminal equipment. Background Art
[0002] Wind power generation is a renewable energy technology that converts wind energy into electrical energy. As a green energy source, wind power generation has been rapidly developed around the world, and with the advancement of technology and the reduction of costs, its application prospects are very broad. However, when wind turbines are operating in a low temperature and high humidity environment, the blades are prone to ice, which will reduce the operating efficiency of the fan, change the blade shape, and even induce blade breakage. Therefore, timely detection and elimination of blade icing failures is crucial to reduce or eliminate the impact of ice on the operating efficiency and equipment loss of the fan.
[0003] Existing methods for detecting ice on wind turbine blades mainly rely on direct observation or specific sensors for ice detection, which has problems such as blind observation areas, environmental sensitivity, high costs, and difficulty in maintenance. In addition, model technologies based on traditional artificial intelligence have shortcomings such as insufficient generalization ability, reliance on professional knowledge, and low detection accuracy. Summary of the invention
[0004] The purpose of the present invention is to provide a method, device, terminal equipment and computer-readable storage medium for detecting icing of fan blades in view of the deficiencies in the prior art, aiming to solve the problems of insufficient generalization ability and low detection accuracy of the existing methods for detecting icing of fan blades.
[0005] A first aspect of an embodiment of the present invention provides a method for detecting icing of a fan blade, comprising:
[0006] Collection steps: Collect SCADA data of the wind turbine to be tested;
[0007] Manual extraction steps: extracting the features related to wind turbine blade icing from SCADA data;
[0008] Automatic extraction step: Use deep belief network to automatically extract high-order features from SCADA data;
[0009] Optimization steps: Genetic algorithm is used to jointly optimize the feature quantity, high-order features and SVM model parameters;
[0010] Training steps: input the optimized feature quantity and high-order features into the SVM model training to form a fan blade icing detection model;
[0011] The automatic extraction step comprises:
[0012] Data standardization step: convert different dimensional data in SCADA data into the same standard range;
[0013] Construction steps: Build a deep belief network model for extracting data features layer by layer;
[0014] Fine-tuning step: add a classification layer on top of the deep belief network model;
[0015] High-order feature extraction step: The SCADA data after data standardization is input into the convolutional neural network to extract local features, and then passed to the trained deep belief network model to extract data features layer by layer to obtain high-order features. Specifically, when extracting features from raw data, adding a convolutional neural network to first perform local convolution on the raw data and then pass the processed features into the deep belief network significantly improves accuracy.
[0016] Furthermore, the construction step includes:
[0017] Design steps: Design the RBM layer. RBM is a two-layer neural network consisting of a visible layer and a hidden layer. The local feature data obtained by convolutional neural network processing of SCADA is input from the visible layer to determine the number of neurons in each layer of RBM, and the data features are obtained by training the hidden layer.
[0018] Unsupervised pre-training steps: Use the contrastive divergence algorithm to train the RBM layer layer by layer. The hidden layer output of the previous RBM layer is used as the input of the next RBM layer. Training is performed layer by layer, and the RBM layers are stacked to form a deep belief network model.
[0019] Furthermore, the training step includes:
[0020] Preprocessing step: taking the optimized feature quantity and high-order feature as the first data group, identifying and filling the missing values in the first data group, and standardizing the first data group so that the first data group falls within the range of [0, 1];
[0021] Selecting SVM model step: selecting an appropriate type of SVM model as basic model training according to the first data set, setting key parameters and optimizing;
[0022] Training the SVM model: using the preprocessed first data set and the optimized key parameters, fitting the training data through the SVM model to determine the number of support vectors and important parameters;
[0023] Evaluate the SVM model steps: Use test data to evaluate the model performance and adjust the model parameters based on the evaluation results to optimize the performance.
[0024] Furthermore, the training step further includes:
[0025] Inspection step: recording blade icing status data within a preset time period to form a second data set;
[0026] Visualization step: Draw an ice coverage growth curve based on the second data group, and roll it in a half-hour cycle to form an ice coverage trend graph.
[0027] Furthermore, the feature quantity includes:
[0028] The square of wind speed is expressed as Among them, dt represents a time increment, dM represents the change in the mass of ice on the surface of the object within the time dt, V represents the wind speed, and f(t) means the product of the three parameters of impact coefficient, adhesion coefficient and growth coefficient when the temperature changes with time;
[0029] Power residual, the power residual refers to the relative residual between the actual power and the power characteristic curve, and the power residual is expressed by the formula: Wherein, preal is the actual output power, ppre is the power predicted according to the power characteristic curve, and residual_p is the relative residual between the actual power and the power characteristic curve, referred to as the power residual.
[0030] Difference in ambient and cabin temperature;
[0031] The ratio of wind speed to fan power is expressed by the formula: Among them, wind_power is the ratio of wind speed to wind turbine power, windspeed is the ambient wind speed, and power is the grid-side active power.
[0032] Furthermore, the mass of ice dM on the surface of the object within the time dt is expressed by the formula dM=α1a2a3ρVSdt, where α1, α2, and α3 are the impact coefficient, adhesion coefficient, and growth coefficient, respectively. The impact coefficient represents the probability that a water droplet can successfully hit the surface of the object, the adhesion coefficient represents the probability that a water droplet can adhere to the surface of the object after hitting it, the growth coefficient represents the probability that a water droplet can exist after adhering to the surface of the object, ρ represents the water content in the air, V represents the wind speed, and S represents the surface area of the object.
[0033] Furthermore, the water content ρ is a constant, the adhesion coefficient α2 is 1, and the growth coefficient is a constant or 1, then α2α3ρS is determined to be a constant C, and the ice mass dM is expressed by the formula dM=Cα1Vdt. Then, the constants are merged into a constant, highlighting the variable part, so as to clearly understand and analyze the changing quantity.
[0034] A second aspect of an embodiment of the present application provides a device for detecting icing of fan blades, comprising:
[0035] Acquisition unit: used to collect SCADA data of the wind turbine to be tested;
[0036] Manual extraction unit: used to extract the characteristic quantities related to wind turbine blade icing from SCADA data;
[0037] Automatic extraction unit: used to automatically extract high-order features from SCADA data using a deep belief network;
[0038] Optimization unit: used to jointly optimize feature quantities, high-order features and SVM model parameters using genetic algorithms;
[0039] Training unit: used to input the optimized feature quantity and high-order features into SVM model training to form a fan blade icing detection model;
[0040] The automatic extraction unit comprises:
[0041] Data standardization unit: used to convert different dimensional data in SCADA data into the same standard range;
[0042] Construction unit: used to construct a deep belief network model for extracting data features layer by layer;
[0043] Fine-tuning unit: used to add a classification layer on top of the deep belief network model;
[0044] High-order feature extraction unit: It is used to input the SCADA data after data standardization into the convolutional neural network to extract local features, and then pass it to the trained deep belief network model to extract data features layer by layer to obtain high-order features.
[0045] A third aspect of an embodiment of the present application provides a terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first aspect when executing the computer program.
[0046] A fourth aspect of an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method described in the first aspect are implemented.
[0047] The beneficial effects of the present invention are:
[0048] Collect SCADA data of the wind turbine to be tested; extract feature quantities related to the icing of the wind turbine blades from the SCADA data; use a deep belief network to automatically extract high-order feature quantities from the SCADA data; use a genetic algorithm to jointly optimize the feature quantities, high-order features and SVM model parameters; input the optimized feature quantities and high-order features into the SVM model training to form a wind turbine blade icing detection model. The present invention significantly improves the accuracy and efficiency of wind turbine blade icing fault diagnosis by combining SCADA data analysis and artificial intelligence algorithms, especially SVM and genetic algorithms. In addition, a deep belief network is used to automatically extract high-order feature quantities from SCADA data, and it is input into the model training together with the remaining feature quantities, which further improves the accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 A schematic flow chart of a method for detecting icing of fan blades according to the present application;
[0050] Figure 2 A graph showing the relationship between wind turbine power and ambient wind speed is provided for this application;
[0051] Figure 3 A graph showing the relationship between time and ambient cabin temperature difference for this application;
[0052] Figure 4 A partial diagram of the relationship between wind speed and grid-side active power ratio is provided for this application;
[0053] Figure 5 A schematic diagram of a process for automatically extracting high-order features from SCADA data using a deep belief network is provided in this application;
[0054] Figure 6 A flow chart for establishing a SVM model for this application;
[0055] Figure 7 A flow chart of SVM model parameter optimization for this application;
[0056] Figure 8 A schematic diagram of a terminal device provided in one embodiment of the present application. DETAILED DESCRIPTION
[0057] The present invention is described in detail below in conjunction with the accompanying drawings. In the following description, for the purpose of illustration rather than limitation, specific details such as specific system structures, technologies, etc. are proposed so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits and methods are omitted to prevent unnecessary details from hindering the description of the present application.
[0058] Blade icing can reduce the efficiency of wind turbines, change blade shape, and even induce blade breakage. Therefore, timely detection and elimination of blade icing failures are crucial to reducing or eliminating the impact of icing on wind turbine operating efficiency and equipment loss. Currently, there are three main areas of research on wind turbine blade icing: (1) conducting experiments on airfoils in ice wind tunnels to study icing problems and aerodynamic characteristics under different operating conditions; (2) using numerical methods to simulate icing problems of two-dimensional airfoils of wind turbines under different operating conditions; and (3) icing detection and deicing research. Existing icing detection methods include direct observation, indirect monitoring, wind tunnel testing, and numerical simulation.
[0059] The simplest method of monitoring wind turbine blade icing is to observe the blades directly through optical devices. The literature installs a camera on the wind turbine and monitors blade icing by observing real-time images; designs an icing sensor based on the absorption and reflection of sound waves and other waves by icing; takes ambient temperature and humidity, wind turbine output power loss, etc. as research objects, and establishes a corresponding icing detection model based on their correlation with the degree of blade icing; and detects blade vibration to achieve blade icing status detection. However, direct observation method for monitoring wind turbine blade icing requires the installation of additional sensors on the blades, which increases costs and maintenance difficulties; in addition, the direct monitoring method may not cover the icing conditions in all areas, especially for large wind turbine blades, the layout and maintenance of the monitoring system will be more complicated and difficult.
[0060] Therefore, the Supervisory Control and Data Acquisition System (SCADA) data is used. The role of the SCADA system is as follows: the wind turbine SCADA system is responsible for collecting the operating status data of the wind turbine, which includes but is not limited to wind speed, wind direction, rotation speed, power, etc., providing important information for wind turbine status monitoring and fault diagnosis. A more accurate and real-time method for monitoring wind turbine blade icing is provided based on the deep learning model, avoiding the need to install sensors directly on the blades. That is, first, based on industry experience, data with a high correlation with wind turbine blade icing is selected from the SCADA monitoring data set as the model input to train the diagnostic model. Existing technical solutions: such as blade icing fault prediction based on BP neural network, which has relatively rough data selection and processing, resulting in insufficient prediction accuracy and a large demand for data; blade icing fault detection model based on long short-term memory network, but the long short-term memory network model (Long Short-Term Memory Model, LSTM) has problems such as insufficient accuracy; the support vector machine (SVM) method is a supervised learning model, which has stronger adaptability and better classification ability than other artificial intelligence methods when the number of learning samples is small.
[0061] Since traditional wind turbine blade icing detection methods may rely on direct observation or specific sensors, traditional direct detection technology may have observation blind spots or be sensitive to environmental conditions, resulting in incomplete or inaccurate detection. The existing model technology based on indirect monitoring methods with specific features may be effective under specific conditions, but may lack generalization capabilities and cannot adapt to changing environments and working conditions, resulting in the risk of misdiagnosis. In addition, the use of existing model technology is heavily dependent on the professional knowledge of researchers, and is difficult for non-professionals to operate and maintain.
[0062] Therefore, the embodiments of the present application provide a method, apparatus, terminal device and computer-readable storage medium for detecting icing of wind blades to solve the problems of insufficient generalization ability and low detection accuracy of existing methods for detecting icing of wind blades.
[0063] First, this application provides a method for detecting icing on fan blades. Figure 1 , Figure 1 A schematic flow chart of a method for detecting icing of a fan blade according to the present application is shown. Figure 1 As shown, the method for detecting icing of fan blades may include the following steps:
[0064] Collection steps: Collect SCADA data of the wind turbine to be tested;
[0065] Specifically, the wind turbine SCADA system collects a large amount of operating status data.
[0066] Manual extraction steps: extracting the features related to wind turbine blade icing from SCADA data;
[0067] Specifically, feature quantities closely related to the icing of wind turbine blades are extracted to provide more accurate input for fault diagnosis; in this step, the feature quantities include wind speed square, power residual, ambient and cabin temperature difference, and wind speed to wind turbine power ratio. The power residual refers to the relative residual between the actual power and the power characteristic curve, wherein the residual refers to the degree of deviation between the actual output power of the wind turbine and the power characteristic curve.
[0068] The specific content of the feature quantity is:
[0069] When the temperature of the environment around the unit blade is low enough and the amount of moisture in the environment reaches a certain level, water droplets under low temperature conditions collide with the blade surface along with the airflow and freeze on the blade surface. According to the Makkonen ice model of the object surface, under the condition of a given ambient temperature, the ice mass on the object surface per unit time can be expressed by formula (1):
[0070] dM=α1a2a3ρVSdt(1)
[0071] Where: α1, α2, α3 are the impact coefficient, adhesion coefficient and growth coefficient respectively.
[0072] The impact coefficient indicates the probability that a water droplet can successfully hit the surface of an object;
[0073] The adhesion coefficient indicates the probability that a water droplet can adhere to the surface of an object after hitting it;
[0074] The growth coefficient represents the probability that a water droplet can exist after it adheres to the surface of an object.
[0075] ρ represents the water content in the air, V represents the incoming flow velocity (wind speed), and S represents the surface area of the object, which is a constant.
[0076] Since the spatial variation of the fan blades is much smaller than the spatial variation of the air moisture content, the moisture content ρ can be considered a constant. The adhesion coefficient α2 is 1 in most cases. The growth coefficient α3 is 1 when rime ice forms. When rime ice forms, α3 is related to the moisture content, but the moisture content ρ can be considered a constant, so the growth coefficient α3 is also considered a constant. Then α2α3ρ can be considered a constant C, so the Makkonen icing model can be simplified as:
[0077] dM=Cα1Vdt(2)
[0078] By fitting the impact coefficient, we get α1∝f(V), and finally combine equation (2) to derive the ice mass relationship:
[0079]
[0080] When the temperature changes, the relationship between the ice mass is:
[0081]
[0082] The above formula shows that under given temperature conditions and when the air moisture content remains constant, the ice mass of the blade per unit time is proportional to V 1.1762 The characteristic quantity V that characterizes the icing of the fan blades can be obtained by 1.1762 , within the allowable error range, the square of the available wind speed V 2 replace.
[0083] Ice on wind turbine blades will cause power loss in the wind turbine. When the wind turbine is operating normally, the output power is stable and has a good correspondence with the ambient wind speed. Figure 2 shown.
[0084] The wind speed-power baseline fitted by the normal operation data is the power characteristic curve of the fan. After the blades are frozen, the actual operating power curve of the fan will deviate from the normal power characteristic curve, and the degree of deviation is equivalent to the severity of the blade icing. Obviously, the degree of deviation between the actual output power of the fan and the power characteristic curve, i.e. the residual, can be used to distinguish between blade icing and non-icing.
[0085] The relative residual between the actual power and the power characteristic curve, i.e. the relative deviation degree, is shown in formula (5):
[0086]
[0087] Where: p real is the actual output power,
[0088] p pre To predict the power based on the power characteristic curve,
[0089] residual_p is the relative residual between the actual power and the power characteristic curve, referred to as the power residual.
[0090] By analyzing the type of wind turbine operation data obtained, the square of wind speed and power residual were extracted as feature quantities for identifying icing on wind turbine blades.
[0091] Ice formation occurs when the temperature drops to a certain value. The ambient temperature is the horizontal axis, and the difference between the ambient temperature and the cabin temperature is the vertical axis. Figure 3 shown.
[0092] As can be seen from the figure, the icing data is concentrated above y=0, and the normal data is evenly distributed on both sides of y=0, which can easily distinguish between icing data and non-icing data.
[0093] With time as the horizontal axis and the ratio of wind speed to grid-side active power as the vertical axis, a graph is drawn, and the scattered data points are fitted into a curve. In order to make the data more concentrated and easier to observe, a constant is added to the wind speed and grid-side active power respectively, and the data with a denominator close to zero is amplified and then compared. Through multiple tests, it is found that the effect is better when the constant is 5, as shown in formula (6):
[0094]
[0095] Where: windspeed is the ambient wind speed, power is the grid-side active power.
[0096] like Figure 4 As shown, Figure 4 It is a local diagram of the ratio of wind speed to grid-side active power.
[0097] As can be seen from the figure, the ratio of wind speed to grid-side active power is basically distributed near y=0 during normal periods, and the distribution is relatively concentrated; while during the blade icing period, most of it is above y=0, and the curve changes greatly. These data features can better distinguish the blade icing state.
[0098] In addition, in order to improve the accuracy of the fault diagnosis model for wind turbine blade icing faults, multiple groups of high-level features of icing data automatically extracted from the SCADA system using a deep belief network (DBN) model and the above four groups of manually extracted features are used as inputs to the SVM to train the diagnostic model.
[0099] Automatic extraction step: Use deep belief network to automatically extract high-order features from SCADA data;
[0100] Specifically, the automatic extraction step includes:
[0101] Data standardization step: convert different dimensional data in SCADA data into the same standard range;
[0102] Construction steps: Build a deep belief network model for extracting data features layer by layer;
[0103] Specifically, the construction steps include:
[0104] Design steps: Design a restricted Boltzmann Machine (RBM) layer. RBM is a two-layer neural network consisting of a visible layer and a hidden layer. The raw data of SCADA is input into the visible layer to determine the number of neurons in each layer of RBM, and the data features are obtained by training the hidden layer.
[0105] Unsupervised pre-training steps: Use the contrastive divergence algorithm to train the RBM layer layer by layer. The hidden layer output of the previous RBM layer is used as the input of the next RBM layer. Training is performed layer by layer, and the RBM layers are stacked to form a deep belief network model.
[0106] Fine-tuning step: add a classification layer on top of the deep belief network model;
[0107] High-order feature extraction step: The SCADA data after data standardization is input into the convolutional neural network to extract local features, and then passed to the trained deep belief network model to extract data features layer by layer to obtain high-order features.
[0108] The process of automatically extracting high-order features from SCADA data using a deep belief network is as follows: Figure 5 As shown, from left to right, it is visible layer (input raw data) -> hidden layer 1 -> hidden layer 2 -> output layer (output features)
[0109] The specific process is as follows:
[0110] Data standardization: Convert different dimensional data in SCADA data into the same standard range to reduce the differences between features.
[0111] Structural design of deep belief network: Design the RBM layer. RBM is a two-layer neural network consisting of a visible layer and a hidden layer. It extracts its potential features by modeling the input data. Therefore, the purpose of this step is to build a network structure that can extract data features layer by layer.
[0112] The design process is as follows:
[0113] The raw data of SCADA is input from the visible layer to determine the number of neurons in each layer of RBM, and the data features are obtained by hidden layer training;
[0114] Unsupervised pre-training: Each layer of RBM is trained using the Contrastive Divergence (CD) algorithm, which adjusts the parameters of the RBM by minimizing the reconstruction error.
[0115] The hidden layer output of the previous RBM layer is used as the input of the next RBM layer. Training is performed layer by layer, and stacking RBMs forms a deep belief network DBN for complex feature extraction. Through this layer-by-layer progressive structure, DBN can capture high-level features of the data.
[0116] As an implementation method, based on optimizing the number of RBM neurons in each layer and the initialization of inter-layer weights, adaptive optimization algorithms (Adaptive Moment Estimation and Root Mean Square Propagation, Adam and RMSProp) are used to dynamically adjust the learning rate. Furthermore, these optimization algorithms automatically adjust the learning rate of the model by monitoring the gradient changes during the training process, thereby accelerating the convergence of the model, reducing training time, and improving the efficiency of feature extraction.
[0117] As another implementation, in order to further enhance the generalization ability of the model and prevent overfitting problems, the present invention introduces regularization techniques, including Dropout and L2 regularization methods. During the training process, randomly discarding some neuron connections (Dropout) and adding weight penalties (L2 regularization) can effectively prevent the model from overfitting the training data, thereby maintaining high performance in the test phase.
[0118] Fine-tune the DBN parameters to improve the classification ability of the model, that is, by adding a classification layer: add a classification layer (such as a softmax regression layer) on the top layer of DBN to classify the extracted features. In this way, labeled data is used for training, and the parameters of the entire network are adjusted through the back propagation algorithm. The model can be used to extract features and accurately classify the ice coverage status of the blades.
[0119] Advanced feature extraction: The preprocessed blade monitoring data is input into the Convolutional Neural Network (CNN), and the convolutional layer of CNN is used to extract the local features of SCADA data (such as the short-term trend of wind speed change, local anomalies of power residual, etc.). After the convolution operation, the pooling layer (such as maximum pooling or average pooling) is used to further reduce the dimension of the data and reduce the amount of calculation while retaining the key feature information. The superposition of multiple convolution and pooling layers enables CNN to extract more abstract and diverse local features layer by layer. The local feature vector extracted from CNN is passed as input to the deep belief network (DBN), and the data is forward propagated through each layer of RBM to extract features layer by layer. Finally, a high-dimensional feature representation is obtained from the output of the hidden layer of the top layer. These features contain deep information about blade icing failure. By fusing the local features of CNN and the deep features of DBN, the local detail information and global trend information required for wind turbine blade icing detection are captured at the same time. In the fusion process, feature concatenation or weighted fusion is used to merge the output feature vectors of CNN and DBN into a comprehensive feature set.
[0120] Optimization steps: Genetic algorithm is used to jointly optimize the feature quantity, high-order features and SVM model parameters;
[0121] According to the Makkonen icing model, it can be seen that there is a potential mathematical correlation between wind turbine blade icing and the SCADA data of wind turbine operating status; therefore, by screening the data that are closely related to wind turbine icing in the SCADA data during the wind turbine icing period, the four parameter groups of wind speed square (A), wind turbine power residual (B), ambient and cabin temperature difference (C), and wind speed to wind turbine power ratio (D) are analyzed, and the SVM model is trained with the advanced features of SCADA data automatically extracted by the deep belief network to form a fan blade icing detection algorithm model.
[0122] Training steps: The optimized feature quantities and high-order features are input into the SVM model training to form a fan blade icing detection model.
[0123] Specifically, the training step includes:
[0124] Preprocessing step: taking the optimized feature quantity and high-order feature as the first data group, identifying and filling the missing values in the first data group, and standardizing the first data group so that the first data group falls within the range of [0, 1];
[0125] Selecting SVM model step: selecting an appropriate type of SVM model as basic model training according to the first data set, setting key parameters and optimizing;
[0126] Training the SVM model: using the preprocessed first data set and the optimized key parameters, fitting the training data through the SVM model to determine the number of support vectors and important parameters;
[0127] Evaluate the SVM model steps: Use test data to evaluate the model performance and adjust the model parameters based on the evaluation results to optimize the performance.
[0128] Specifically, the training step further includes:
[0129] Inspection step: recording blade icing status data within a preset time period to form a second data set;
[0130] Visualization step: Draw an ice coverage growth curve based on the second data group, and roll it in a half-hour cycle to form an ice coverage trend graph.
[0131] The specific process is as follows:
[0132] Select four parameter groups (data group 1) of the data in the SCADA of any wind turbine under icing state for preprocessing;
[0133] Import required libraries; load dataset;
[0134] Data cleaning and handling missing values: First, identify missing values in the input data. For a small number of missing values, you can choose to delete or fill them with the mean or median; if there are more missing values, consider using the mode or a model based on other variables to predict the missing values. It is important to choose a method that matches the data characteristics and business needs to ensure that data cleaning does not introduce bias and affect the accuracy of the model.
[0135] Feature engineering: Generate new features or select important features. Use Deep Boltzmann Machine with Stochastic Pooling (DBNS) with Softmax to extract high-level features and manual features from SCADA system operation data and input them into SVM.
[0136] Data scaling or standardization: Scaling or standardization of wind turbine blade ice feature data is a key step to ensure model training efficiency and accuracy. Use minimum-maximum standardization (scaling the data to a range of 0 to 1) or Z-score standardization (making the data have a mean of 0 and a standard deviation of 1). This helps avoid the impact of data of different dimensions and magnitudes on model training and improves algorithm convergence speed and prediction performance.
[0137] Selecting an SVM model
[0138] Select the type of SVM: Gaussian kernel SVM, radial basis kernel SVM: When selecting the type of SVM model for wind turbine blade icing detection, the main consideration is the characteristics and distribution characteristics of the data. If the data features are linearly separable, that is, the data points can be clearly separated by a straight line or a hyperplane, a linear kernel SVM should be selected. However, for nonlinearly separable data, a nonlinear kernel function is needed to improve the generalization ability of the model. The Gaussian kernel (Radial Basis Function, RBF) is usually selected for its smoothness and flexibility, and is suitable for processing complex nonlinear relationships. The radial basis kernel SVM is suitable for scenarios with a large feature space or a large number of data points, and can handle more complex data distributions. When selecting, the computational complexity and prediction performance of the model should also be considered. The effects of different kernel functions should be evaluated through methods such as cross-validation to determine the most suitable SVM model type for the current data set and task. Considering the potential nonlinear correlation between the icing feature data and the SCADA data of the wind turbine operating status, the Gaussian kernel SVM should be selected as the basic model training.
[0139] Choosing appropriate parameters, such as penalty parameter C and kernel parameter gamma: The penalty parameter C and kernel function parameter gamma should be based on the balance between the generalization ability of the model and the degree of fit to the data. The penalty parameter C controls the model's tolerance to misclassification: a larger C value makes the model more sensitive to misclassification and may lead to overfitting; a smaller C value makes the model more lenient to misclassification and may lead to underfitting. The kernel function parameter gamma defines the influence range of a single training sample. For a Gaussian kernel, a smaller gamma value indicates a farther influence range, while a larger gamma value indicates the opposite. Usually, the best combination of these parameters is determined through grid search and cross-validation to find the best balance between controlling model complexity and improving prediction accuracy. This process needs to be adjusted according to the characteristics of the specific dataset and the performance of the model on the validation set.
[0140] Train the SVM model:
[0141] Fitting the SVM model using the training data: First, the SVM model distinguishes different categories by finding the optimal hyperplane in the data. Hyperplane is a core concept in the SVM model, especially when performing linear classification. In two-dimensional space, the hyperplane can be seen as a straight line that separates different categories. In higher-dimensional space, the hyperplane is an N-1-dimensional plane, where N is the number of features. The purpose of the hyperplane is to maximize the margin between categories, that is, to find the best decision boundary so that data points of different categories are separated as clearly as possible. In this process, not all training samples contribute to defining this hyperplane.
[0142] Determine the number of support vectors and important parameters: Support vectors are those data points located near (or on the edge of) the decision boundary, which are direct participants in the optimal hyperplane found by the SVM model. These vectors determine the location and direction of the hyperplane. The number of support vectors depends on the characteristics of the dataset and the complexity of the model.
[0143] The training steps are as follows: from the large amount of operating data collected in the wind turbine SCADA system, key feature quantities related to the height of wind turbine blade icing, such as squared wind speed, power residual, etc., are extracted through data analysis and feature engineering. Next, select an appropriate SVM type, such as Gaussian kernel SVM, which is suitable for processing complex nonlinear relationships. Then, set the key parameters of SVM, including the penalty parameter C and the kernel function parameter gamma, which can be optimized through grid search and cross-validation. Using the extracted feature quantities and optimized parameters, the training data is fit through the SVM model to determine the support vector and decision boundary. Finally, the model performance is evaluated through test data, and the parameters are adjusted to optimize the diagnostic accuracy and generalization ability of the model. The model establishment flow chart is shown below. Figure 6 shown.
[0144] Evaluate the SVM model:
[0145] Evaluating model performance using test data: The scheme for evaluating model performance using test data involves applying the SVM model to the test set, calculating the classification accuracy, and evaluating the diagnostic accuracy using the confusion matrix. The model is further optimized by adjusting the key parameters C and g of the SVM, as well as feature selection, to ensure its generalization ability and diagnostic accuracy.
[0146] Adjust model parameters to optimize performance: Take the key parameters C (penalty factor) and g (kernel function parameter) of SVM and feature selection as optimization targets. These parameters and feature selection are encoded into chromosomes, each of which represents a set of possible parameter and feature combinations. Then, we use these chromosomes (parameter combinations) to train the SVM model and evaluate the performance of the model through cross-validation, where performance is usually measured by classification accuracy. Performance evaluation can be imagined as an exam, where each set of parameter and feature combinations will receive a score. Next, we select excellent chromosomes based on these scores (fitness), "breed" (i.e., copy excellent parameter combinations), and introduce some mutations (i.e., small random adjustments) to produce a new generation of chromosomes. This process is like natural selection, and excellent genes (parameter combinations) are more likely to be retained. We repeat this process, optimizing generation after generation, until a certain number of iterations is reached, or the model performance is no longer significantly improved.
[0147] As an implementation method, in order to avoid premature convergence of genetic algorithms during the evolution process (i.e., reaching the local optimal solution too early), the exploration depth of the search space is limited. For this reason, an adaptive genetic algorithm (AGA) is introduced to dynamically adjust key parameters such as the crossover rate and mutation rate during the optimization process. That is, in AGA, the crossover rate and mutation rate will be adjusted in real time according to the distribution of population fitness. When the fitness of the population tends to converge, the crossover rate decreases and the mutation rate increases, which promotes the diversity of the population and prevents it from falling into the local optimum; when the fitness distribution of the population is more dispersed, the crossover rate increases and the mutation rate decreases to speed up the convergence. If it is detected that the population diversity decreases or the fitness growth stagnates, the algorithm will automatically adjust the parameters or introduce new candidate solutions (individuals) to re-diversify the population. Through this feedback mechanism, AGA maintains a high search efficiency and solution diversity throughout the evolution process, and enhances the robustness of parameter optimization.
[0148] Finally, the parameters and feature subsets corresponding to the optimal chromosome we obtain are the best SVM model parameters selected by the genetic algorithm. The model parameter optimization process is as follows: Figure 7 shown.
[0149] In this way, the genetic algorithm helps us search for the optimal solution in the parameter space, find the best C and g values, and the most appropriate feature subset Sopt, thereby improving the diagnostic accuracy of the SVM model.
[0150] The following are the steps to use the SVM training model to draw the ice growth curve:
[0151] The four parameter groups of the data in the SCADA of all wind turbines in the entire wind farm under ice-covered conditions are tested by the SVM model algorithm. Based on historical data and expert experience, thresholds that can represent different degrees of icing are set. If more than 90% of the data at the current moment show icing, we record it as 1, indicating strong icing; if more than 70% show icing, it is recorded as 0, indicating moderate icing; if more than 50%, it is recorded as -1, indicating slight icing.
[0152] Furthermore, the recorded values (with values {-1, 0, 1} for the freezing degree labels) at the current time, half an hour ago, 2 hours ago, and 4 hours ago are recorded respectively to form a new data group (data group 2);
[0153] Step 6: Draw an ice growth curve based on data set 2, and roll it over every half an hour to form an ice trend graph. The method is:
[0154] When data group 2 is [1,1,1,1] or [1,1,1,0], [1,1,0,1], [1,1,0,0], the recorded value increases by 1, indicating that the degree of icing increases;
[0155] When data group 2 is [0,1,1,1[or [1,0,1,1],[1,0,0,1],[1,0,1,0],[0,1,0,1],[0,1,1,0], the recorded value increases by 0, indicating that the degree of icing has not changed;
[0156] When data group 2 is [0,0,1,1],[0,0,0,0],[1,0,0,0],[0,1,0,0],[0,0,1,0],[0,0,0,1], the recorded value is subtracted by 1, indicating that the degree of icing has weakened;
[0157] It should be noted that the ice growth curve only expresses the ice growth trend and ice duration, and does not directly represent the ice thickness.
[0158] Preferably, as an implementation mode, based on the above-mentioned embodiment, the inventor creatively sets an algorithm formula based on a physical model after repeated experiments to more accurately predict the icing trend of the wind turbine blades, such as formula (7):
[0159] I=α(V 2 ρ'S)+β(P real -P pre )+γ(T env -T nacelle )+δ(wind_power) (7)
[0161] Where: I: icing index, used to evaluate the possibility of blade icing, α, β, γ, δ: weight coefficients, obtained through machine learning algorithm optimization, V: wind speed, ρ': air density, S: object surface area, i.e. blade surface area, P real : Actual power output, P pre : Predict power according to the power characteristic curve, T env : Ambient temperature, T nacelle : cabin temperature, wind_power: ratio of wind speed to wind turbine power. Formula (7) combines physical parameters and weight coefficients optimized by machine learning, which can more accurately predict the icing trend of wind turbine blades like weather forecasts, improve the accuracy and practicality of the detection model, and enhance the interpretability and credibility of the algorithm by introducing the physical model.
[0162] The present invention provides a more accurate, efficient and easy-to-operate diagnostic solution for the problem of wind turbine blade icing through big data analysis and artificial intelligence algorithms. The wind turbine SCADA system collects a large amount of operating status data, and selects the four parameter combinations of wind speed square (A), wind turbine power residual (B), ambient and cabin temperature difference (C), and wind speed to wind turbine power ratio (D) that are closely related to wind turbine icing for analysis. In addition, a deep belief network is used to automatically extract high-order features of the data from the station SCADA system data and the four groups of parameter features, which are jointly input into SVM training to form a wind turbine blade icing detection model.
[0163] The proposed solution aims to extract the feature quantities closely related to the icing of wind turbine blades through mechanism research and data exploration, provide more accurate input for fault diagnosis, and use genetic algorithms to jointly optimize the feature quantities and SVM parameters to improve the diagnostic accuracy and stability of the model, reduce the risk of misdiagnosis due to technical limitations, and provide a more reliable basis for wind turbine blade deicing decisions. In addition, through optimized feature selection and model training, the generalization ability of the diagnostic model is enhanced, enabling it to adapt to different environments and working conditions. At the same time, the use of automated data analysis processes and model diagnosis processes reduces dependence on professionals, making fault diagnosis more convenient and easy to operate. This is especially important for non-professionals.
[0164] In addition, the present invention enhances the generalization ability of the diagnostic model by optimizing feature selection and model training, so that it can adapt to different environments and working conditions. This is crucial to improving the stability and reliability of the model in practical applications. By reducing the risk of misdiagnosis, the present invention provides a more reliable basis for fan blade deicing decisions, which helps to reduce or eliminate the impact of icing on fan operating efficiency and equipment loss.
[0165] The present application provides a fan blade icing detection device comprising:
[0166] Acquisition unit: used to collect SCADA data of the wind turbine to be tested;
[0167] Manual extraction unit: used to extract the characteristic quantities related to wind turbine blade icing from SCADA data;
[0168] Automatic extraction unit: used to automatically extract high-order features from SCADA data using a deep belief network;
[0169] Optimization unit: used to jointly optimize feature quantities, high-order features and SVM model parameters using genetic algorithms;
[0170] Training unit: used to input the optimized feature quantity and high-order features into SVM model training to form a fan blade icing detection model;
[0171] The automatic extraction unit comprises:
[0172] Data standardization unit: used to convert different dimensional data in SCADA data into the same standard range;
[0173] Construction unit: used to construct a deep belief network model for extracting data features layer by layer;
[0174] Fine-tuning unit: used to add a classification layer on top of the deep belief network model;
[0175] High-order feature extraction unit: It is used to input the SCADA data after data standardization into the convolutional neural network to extract local features, and then pass it to the trained deep belief network model to extract data features layer by layer to obtain high-order features.
[0176] The present application provides a detection device for fan blade icing, which extracts feature quantities related to fan blade icing from SCADA data; uses a deep belief network to automatically extract high-order feature quantities from SCADA data; uses a genetic algorithm to jointly optimize feature quantities, high-order features and SVM model parameters; and inputs the optimized feature quantities and high-order features into SVM model training to form a fan blade icing detection model. The present invention significantly improves the accuracy and efficiency of fan blade icing fault diagnosis by combining SCADA data analysis and artificial intelligence algorithms, especially SVM and genetic algorithms. In addition, a deep belief network is used to automatically extract high-order feature quantities from SCADA data, and it is input into model training together with other feature quantities, which further improves the accuracy.
[0177] Figure 8 Schematic diagram of a terminal device provided by an embodiment of the present application. Figure 8As shown, a terminal device 100 of this embodiment includes: a processor 1000, a memory 1001, and a computer program 1002 stored in the memory 1001 and executable on the processor 1000, such as a method program for detecting icing of a fan blade. When the processor 1000 executes the computer program 1002, the steps in each of the above-mentioned methods for detecting icing of a fan blade are implemented, such as Figure 1 Alternatively, when the processor 1000 executes the computer program 1002, the functions of each unit in the above-mentioned device embodiments are realized.
[0178] Exemplarily, the computer program 1002 may be divided into one or more units, which are stored in the memory 1001 and executed by the processor 1000 to complete the present application. The one or more units may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program 1002 in the terminal device 100. For example, the computer program 1002 may be divided into the following specific functions of each unit:
[0179] Acquisition unit: used to collect SCADA data of the wind turbine to be tested;
[0180] Manual extraction unit: used to extract the characteristic quantities related to wind turbine blade icing from SCADA data;
[0181] Automatic extraction unit: used to automatically extract high-order features from SCADA data using a deep belief network;
[0182] Optimization unit: used to jointly optimize feature quantities, high-order features and SVM model parameters using genetic algorithms;
[0183] Training unit: used to input the optimized feature quantity and high-order features into SVM model training to form a fan blade icing detection model;
[0184] The automatic extraction unit comprises:
[0185] Data standardization unit: used to convert different dimensional data in SCADA data into the same standard range;
[0186] Construction unit: used to construct a deep belief network model for extracting data features layer by layer;
[0187] Fine-tuning unit: used to add a classification layer on top of the deep belief network model;
[0188] High-order feature extraction unit: used to input the SCADA data after data standardization into the convolutional neural network, and then pass it to the trained deep belief network model to extract data features layer by layer to obtain high-order features.
[0189] The terminal device includes but is not limited to a processor 1000 and a memory 1001. Those skilled in the art will appreciate that Figure 8 It is only an example of a terminal device 100 and does not constitute a limitation on the terminal device 100. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the terminal device may also include input and output devices, network access devices, buses, etc.
[0190] The processor 1000 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.
[0191] The memory 1001 may be an internal storage unit of the terminal device 100, such as a hard disk or memory of the terminal device 100. The memory 1001 may also be an external storage device of the terminal device 100, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the terminal device 100. Further, the memory 1001 may also include both the internal storage unit of the terminal device 100 and the external storage device. The memory 1001 is used to store the computer program and other programs and data required by the roaming control device. The memory 1001 may also be used to temporarily store data that has been output or is to be output.
[0192] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.
[0193] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.
[0194] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.
Claims
1. A method for detecting icing of fan blades, characterized in that: include: Collection steps: Collect SCADA data of the wind turbine to be tested; Manual extraction steps: extracting the features related to wind turbine blade icing from SCADA data; Automatic extraction step: Use deep belief network to automatically extract high-order features from SCADA data; Optimization steps: Genetic algorithm is used to jointly optimize the feature quantity, high-order features and SVM model parameters; Training steps: input the optimized feature quantity and high-order features into the SVM model training to form a fan blade icing detection model; The automatic extraction step comprises: Data standardization step: convert different dimensional data in SCADA data into the same standard range; Construction steps: Build a deep belief network model for extracting data features layer by layer; Fine-tuning step: add a classification layer on top of the deep belief network model; High-order feature extraction step: The SCADA data after data standardization is input into the convolutional neural network to extract local features, and then passed to the trained deep belief network model to extract data features layer by layer to obtain high-order features.
2. The method for detecting icing of fan blades according to claim 1, characterized in that: The construction steps include: Design steps: Design the RBM layer. RBM is a two-layer neural network consisting of a visible layer and a hidden layer. The local feature data obtained by convolutional neural network processing of SCADA is input from the visible layer to determine the number of neurons in each layer of RBM, and the data features are obtained by training the hidden layer. Unsupervised pre-training steps: Use the contrastive divergence algorithm to train the RBM layer layer by layer. The hidden layer output of the previous RBM layer is used as the input of the next RBM layer. Training is performed layer by layer, and the RBM layers are stacked to form a deep belief network model.
3. The method for detecting icing of fan blades according to claim 1, characterized in that: The training steps include: Preprocessing step: taking the optimized feature quantity and high-order feature as the first data group, identifying and filling the missing values in the first data group, and standardizing the first data group so that the first data group falls within the range of [0, 1]; Selecting SVM model step: selecting an appropriate type of SVM model as basic model training according to the first data set, setting key parameters and optimizing; Training the SVM model: using the preprocessed first data set and the optimized key parameters, fitting the training data through the SVM model to determine the number of support vectors and important parameters; Evaluate the SVM model steps: Use test data to evaluate the model performance and adjust the model parameters based on the evaluation results to optimize the performance.
4. The method for detecting icing of fan blades according to claim 1, characterized in that: The training step further includes: Inspection step: recording blade icing status data within a preset time period to form a second data set; Visualization step: Draw an ice coverage growth curve based on the second data group, and roll it in a half-hour cycle to form an ice coverage trend graph.
5. The method for detecting icing of fan blades according to claim 1, characterized in that: The feature quantities include: The square of wind speed is expressed as Among them, dt represents a time increment, dM represents the change in the mass of ice on the surface of the object within the time dt, V represents the wind speed, and f(t) means the product of the three parameters of impact coefficient, adhesion coefficient and growth coefficient when the temperature changes with time; Power residual, the power residual refers to the relative residual between the actual power and the power characteristic curve, and the power residual is expressed by the formula: Among them, p real is the actual output power, p pre To predict the power according to the power characteristic curve, residual_p is the relative residual between the actual power and the power characteristic curve, referred to as the power residual; Difference in ambient and cabin temperature; The ratio of wind speed to fan power is expressed by the formula: Among them, wind_power is the ratio of wind speed to wind turbine power, windspeed is the ambient wind speed, and power is the grid-side active power.
6. The method for detecting icing of fan blades according to claim 5, characterized in that: The mass of ice dM on the surface of an object within time dt is expressed by the formula dM=α1a2a3ρVSdt, where α1, α2, and α3 are the impact coefficient, adhesion coefficient, and growth coefficient, respectively. The impact coefficient represents the probability that a water droplet can successfully hit the surface of an object, the adhesion coefficient represents the probability that a water droplet can adhere to the surface of an object after hitting it, the growth coefficient represents the probability that a water droplet can exist after adhering to the surface of an object, ρ represents the water content in the air, V represents the wind speed, and S represents the surface area of the object.
7. The method for detecting icing of fan blades according to claim 6, characterized in that: The water content ρ is a constant, the adhesion coefficient α2 is 1, the growth coefficient is a constant or 1, then α2α3ρS is determined to be a constant C, and the ice mass dM is expressed by the formula dM=Cα1Vdt.
8. A device for detecting icing of fan blades, characterized in that: include: Acquisition unit: used to collect SCADA data of the wind turbine to be tested; Manual extraction unit: used to extract the characteristic quantities related to wind turbine blade icing from SCADA data; Automatic extraction unit: used to automatically extract high-order features from SCADA data using a deep belief network; Optimization unit: used to jointly optimize feature quantities, high-order features and SVM model parameters using genetic algorithms; Training unit: used to input the optimized feature quantity and high-order features into SVM model training to form a fan blade icing detection model; The automatic extraction unit comprises: Data standardization unit: used to convert different dimensional data in SCADA data into the same standard range; Construction unit: used to construct a deep belief network model for extracting data features layer by layer; Fine-tuning unit: used to add a classification layer on top of the deep belief network model; High-order feature extraction unit: It is used to input the SCADA data after data standardization into the convolutional neural network to extract local features, and then pass it to the trained deep belief network model to extract data features layer by layer to obtain high-order features.
9. A terminal device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method according to any one of claims 1 to 7 when executing the computer program.
10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.
