A fan blade icing intelligent detection method and system

Abstract

The application discloses a kind of fan blade icing intelligent detection method and system, comprising: obtaining original data, and obtaining standardized feature vector by missing value interpolation and normalization;By automatic encoder, low-dimensional latent feature is extracted, and neural dynamics state is generated by inputting physical notification neural network;Based on power deviation, physical residual is constructed and constraint strength is dynamically adjusted using meta-learning;Neural dynamics state is linearly transformed into query, key and value vector, physical residual and geodesic distance correction attention weight are introduced, and refined feature vector is generated;A comprehensive loss function weighted by data fitting loss, physical residual constraint loss and icing differential operator constraint loss is constructed;Define the fisher information metric of network parameter, and construct natural gradient update rule on riemannian manifold to optimize parameters;Output fan icing probability.The application has the advantages of high detection accuracy, strong robustness, good physical consistency and interpretable decision, etc.

Description

Technical Field

[0001] This invention belongs to the field of wind power equipment condition monitoring technology, and in particular relates to an intelligent detection method and system for wind turbine blade icing. Background Technology

[0002] Wind turbine blades are prone to icing in low-temperature and humid environments, leading to decreased power generation efficiency, accelerated equipment wear, and even safety risks. Traditional blade icing detection methods include sensor-based active detection, model-based prediction, and passive de-icing measures based on materials science. However, existing technologies suffer from the following problems: they are highly dependent on domain experience and additional sensors, resulting in high costs and poor scalability; data-driven methods are often limited by imbalanced samples and high-dimensional SCADA data, making it difficult to capture long-term dependent features; traditional CNN or LSTM architectures cannot efficiently process large-scale multivariate time series data, resulting in insufficient accuracy and interpretability.

[0003] Therefore, there is an urgent need for a high-precision, real-time deployable method for detecting icing on wind turbine blades that can handle high-dimensional unbalanced data. Summary of the Invention

[0004] This invention provides a method and system for intelligent detection of icing on wind turbine blades based on a Riemann adaptive multi-scale spatiotemporal physical information neural network, aiming to solve the technical problems of low detection accuracy and poor robustness in the prior art.

[0005] In a first aspect, the present invention provides an intelligent detection method for wind turbine blade icing, comprising:

[0006] Based on the raw data of wind turbine operation obtained by the SCADA system, the raw data is imputed and normalized to obtain a standardized feature vector.

[0007] The standardized feature vector is input into an autoencoder to extract a low-dimensional latent feature vector, and the low-dimensional latent feature vector is input into a physical information neural network. The spatiotemporal evolution features caused by leaf icing are extracted using a multi-scale spectral sensing kernel to generate a neurodynamic state.

[0008] Based on the aforementioned neurodynamic state, the power deviation is calculated using aerodynamic equations, a physical residual function is constructed, and a meta-learning mechanism is used to dynamically adjust the physical constraint strength to obtain the physical constraint coefficient.

[0009] The neurodynamic state is transformed linearly to generate query vector, key vector and value vector respectively. The dot product of query vector and key vector is calculated. The physical residual function and manifold geodesic distance are introduced to correct the dot product result to obtain the physically corrected attention weight. The attention weight is then used to weight and aggregate the value vector to generate a refined feature vector.

[0010] A comprehensive loss function is constructed, which is obtained by weighted summation of data fitting loss, physical residual constraint loss and icing differential operator constraint loss, wherein the weight of physical residual constraint loss is the physical constraint coefficient;

[0011] Define the Fisher information metric for the network parameters in the physical informant neural network, construct a natural gradient update rule on the Riemannian manifold, and use the natural gradient update rule to optimize the network parameters to minimize the comprehensive loss function, thereby obtaining the optimized network parameters;

[0012] Using the optimized network parameters, the output layer of the physical information neural network is driven to map the refined feature vector to the probability of wind turbine icing.

[0013] Secondly, the present invention provides an intelligent detection system for wind turbine blade icing, comprising:

[0014] The acquisition module is configured to acquire raw data of wind turbine operation based on the SCADA system, and perform missing value interpolation and normalization on the raw data to obtain a standardized feature vector.

[0015] The extraction module is configured to input the standardized feature vector into an autoencoder to extract a low-dimensional latent feature vector, and input the low-dimensional latent feature vector into a physical information neural network to extract the spatiotemporal evolution features caused by leaf icing using a multi-scale spectral sensing kernel, thereby generating a neurodynamic state.

[0016] The first construction module is configured to calculate the power deviation based on the neural dynamics state and the aerodynamic equation, construct the physical residual function, and use a meta-learning mechanism to dynamically adjust the physical constraint strength to obtain the physical constraint coefficient.

[0017] The correction module is configured to generate query vector, key vector and value vector respectively by linear transformation of the neurodynamic state, calculate the dot product of query vector and key vector, and introduce the physical residual function and manifold geodesic distance to correct the dot product result to obtain physically corrected attention weights, and use the attention weights to perform weighted aggregation on the value vector to generate refined feature vectors.

[0018] The second construction module is configured to construct a comprehensive loss function, which is obtained by weighted summation of data fitting loss, physical residual constraint loss and icing differential operator constraint loss, wherein the weight of physical residual constraint loss is the physical constraint coefficient;

[0019] The optimization module is configured to define the Fisher information metric for the network parameters in the physical informative neural network, construct a natural gradient update rule on the Riemannian manifold, and optimize the network parameters using the natural gradient update rule to minimize the comprehensive loss function, thereby obtaining the optimized network parameters.

[0020] The mapping module is configured to use optimized network parameters to drive the output layer of the physical information neural network to map the refined feature vector to the wind turbine icing probability.

[0021] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the intelligent detection method for wind turbine blade icing according to any embodiment of the present invention.

[0022] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the steps of the intelligent detection method for wind turbine blade icing according to any embodiment of the present invention.

[0023] The intelligent detection method and system for wind turbine blade icing in this application extracts deep nonlinear features through an autoencoder and combines it with a multi-scale spectral sensing kernel to capture the spatiotemporal evolution of blade icing, effectively improving the accuracy and sensitivity of icing state identification and reducing false alarms and missed alarms. It introduces physical residual functions and Riemannian manifold natural gradient optimization to eliminate feature distribution shifts under complex operating conditions. Attention weights are corrected based on geodesic distance to effectively filter measurement noise in the original data, ensuring stable convergence of the model under extreme data. Aerodynamic power deviation constraints and icing differential operators are embedded in the loss function, making the model output conform to the physical evolution of blade icing, avoiding purely data-driven non-physical interpretations, and improving the reliability of prediction results. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart of an intelligent detection method for wind turbine blade icing provided in an embodiment of the present invention;

[0026] Figure 2 This is a structural block diagram of an intelligent detection system for wind turbine blade icing provided in an embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Please see Figure 1 The diagram shows a flowchart of an intelligent detection method for wind turbine blade icing according to this application.

[0030] like Figure 1 As shown, the intelligent detection method for wind turbine blade icing specifically includes the following steps:

[0031] Step S101: Obtain the raw data of wind turbine operation based on the SCADA system, and perform missing value interpolation and normalization on the raw data to obtain a standardized feature vector.

[0032] In this step, raw data is collected using a SCADA (Supervisory Control And Data Acquisition) system, with a sampling interval of 10 minutes. The raw data includes wind speed, wind direction, ambient temperature, relative humidity, air pressure, impeller speed, blade pitch angle, main shaft torque, wind power, generator speed, theoretical power, power deviation, blade speed deviation, humidity saturation index, and icing status label.

[0033] Missing values ​​are imputed using a sliding window mean, with a window radius of 3. If the proportion of missing values ​​within the window exceeds 50%, the historical mean of that feature is used for imputation. Then, Z-score standardization is applied to make each feature have a mean of 0 and a standard deviation of 1.

[0034] Step S102: Input the standardized feature vector into the autoencoder to extract the low-dimensional latent feature vector, and input the low-dimensional latent feature vector into the physical information neural network. Use the multi-scale spectral sensing kernel to extract the spatiotemporal evolution features caused by leaf icing and generate the neurodynamic state.

[0035] In this step, the Physically Informed Neural Network (PINN) is essentially a continuous-time neural manifold dynamics network, consisting of an initial mapping layer and a multi-scale spectral sensing kernel extraction layer.

[0036] The standardized feature vector Input auto encoder, the auto encoder is composed of The network consists of a fully connected layer, with each layer sequentially connected to a batch normalization layer and a ReLU activation layer, as expressed in the following expression:

[0037] ,

[0038] ,

[0039] ,

[0040] In the formula, This is the output of the fully connected layer l. Let be the weight matrix of the l-th fully connected layer of the autoencoder. The output feature vector of the (l-1)th layer, For the bias term of the fully connected layer l of the autoencoder, For batch normalization function, To modify the activation function of the linear unit;

[0041] go through After processing by a fully connected network, the low-dimensional latent feature vector is obtained as the output. , This is the output feature vector of the autoencoder.

[0042] It should be noted that the low-dimensional latent feature vector Mapped to the initial state through a fully connected layer. ,in, For the mapping weight matrix, For mapping bias terms;

[0043] By convolving the continuous function space states from past time points using a multi-scale spectral sensing kernel, the current neural dynamic state is generated, expressed as:

[0044] ,

[0045] In the formula, Let be the neurodynamic state at time t. It is a non-linear activation function. It is a learnable multi-scale spectral sensing kernel. For a continuous function space at historical moments state, It is a linear weight matrix. Let be the state of the continuous function space at the current time t.

[0046] Step S103: Based on the neural dynamics state, the power deviation is calculated by combining the aerodynamic equations, a physical residual function is constructed, and a meta-learning mechanism is used to dynamically adjust the physical constraint strength to obtain the physical constraint coefficient.

[0047] In this step, the predicted power deviation is calculated based on the neurodynamic state and combined with the aerodynamic power equation. The expression is as follows:

[0048] ,

[0049] In the formula, Theoretical power, The actual power calculated based on the neurodynamic state;

[0050] Construct the physical residual function, with the following expression:

[0051] ,

[0052] In the formula, For physical residual functions, The measured power deviation of the SCADA system;

[0053] Based on the physical residual function, a meta-learning mechanism is used to dynamically adjust the physical constraint coefficients, as expressed in the following expression:

[0054] ,

[0055] In the formula, For physical constraint coefficients, These are the initial values ​​for the physical constraint coefficients. For smooth activation function, The meta-learning rate, For physical residual loss Partial derivatives with respect to the physical constraint coefficients.

[0056] Step S104: The neural dynamic state is transformed linearly to generate a query vector, a key vector, and a value vector, respectively. The dot product of the query vector and the key vector is calculated, and the physical residual function and the manifold geodesic distance are introduced to correct the dot product result to obtain the physically corrected attention weight. The attention weight is then used to perform weighted aggregation on the value vector to generate a refined feature vector.

[0057] In this step, the neurodynamic state is transformed into a query vector through three independent linear transformations. Key vector Sum value vector ,in, The neurodynamic state at time point i. For the neurodynamic state at time point j, , , All are learnable linear transformation weight matrices;

[0058] The corrected attention score is calculated using the following expression:

[0059] ,

[0060] In the formula, The adjusted attention score between the i-th and j-th time points. Scaling factor Let be the physical residual value at time point i. , All are adjustment coefficients. For Riemannian manifolds and Geodesic distance between them;

[0061] The physically corrected attention weights are calculated using the following expression:

[0062] ,

[0063] In the formula, The attention weights are physically corrected. The adjusted attention score between the i-th time point and the k-th time point. It is an exponential function;

[0064] The value vector is weighted and aggregated using the attention weights to generate a refined feature vector, expressed as follows:

[0065] ,

[0066] In the formula, This is the refined feature vector generated at the i-th time point after attention-weighted aggregation.

[0067] Step S105: Construct a comprehensive loss function, which is obtained by weighted summation of data fitting loss, physical residual constraint loss and icing differential operator constraint loss, wherein the weight of physical residual constraint loss is the physical constraint coefficient.

[0068] In this step, the expression for the comprehensive loss function is:

[0069] ,

[0070] ,

[0071] ,

[0072] ,

[0073] ,

[0074] In the formula, For the comprehensive loss function, , , All of these are preset static hyperparameters. For dynamic adaptive data fitting loss, The physical constraint coefficients are for the t-th iteration and change dynamically as the training progresses. This is the residual loss due to aerodynamic power deviation. For the dynamic consistency loss of the icing differential operator, The number of iterations for optimizing the current manifold parameters. The physical constraint coefficients for the (k+1)th iteration are... Let be the physical constraint coefficients for the k-th iteration. For learning rate, The sigmoid activation function is used. This represents the partial derivative of the aerodynamic power deviation residual loss with respect to the physical constraint coefficient. This represents the total number of samples in the current training batch. This is the sample imbalance adjustment coefficient, used to balance the baseline weights of frozen and unfrozen samples. To inform the output layer of the neural network about the predicted icing probability of the i-th sample, To focus parameters, This is the label for the true freezing state of the i-th sample. The predicted power deviation value is derived by combining neurodynamic state with Bates' law and blade element theory of wind turbine blades. Let i be the neurodynamic state vector of the i-th sample at the current time. This represents the actual monitored power deviation value. It is a nonlinear icing differential operator. Let be the first-order forward derivative of the neurodynamic state with respect to time. A physical evolution kinetic function characterizing the mass accumulation and thermodynamic phase transition process of ice layer on the blade surface. , , These are the key meteorological driving variables in the input of the i-th sample, corresponding to ambient temperature, relative humidity, and wind speed, respectively.

[0075] Step S106: Define the Fisher information metric for the network parameters in the physical informative neural network, construct a natural gradient update rule on the Riemannian manifold, and optimize the network parameters using the natural gradient update rule to minimize the comprehensive loss function, thereby obtaining the optimized network parameters.

[0076] In this step, let the network parameters be... Network parameters The Fisher information metric is defined as:

[0077] ,

[0078] In the formula, For network parameters The Fisher information metric matrix, Let p be the mathematical expectation of the probability distribution. Given input data x and network parameters Under the condition, output state The conditional probability distribution, Log-conditional probability of network parameters gradient, It is the transpose symbol;

[0079] On a Riemannian manifold, a natural gradient update rule is constructed, expressed as follows:

[0080] ,

[0081] In the formula, The parameter value at the k-th iteration. The parameter value at the (k+1)th iteration. For learning rate, This is the inverse of the Fisher information metric matrix. The comprehensive loss function is applied to the network parameters at step k. The gradient;

[0082] Iteratively execute the update rules until convergence to obtain the optimized network parameters. .

[0083] Step S107: Using the optimized network parameters, drive the output layer in the physical information neural network to map the refined feature vector to the wind turbine icing probability.

[0084] In this step, the optimized network parameters are used. Configure the weight matrix of the output layer and bias terms ;

[0085] The refined feature vector The input and output layers, after linear transformation, are mapped to icing probabilities using the Sigmoid activation function, expressed as:

[0086] ,

[0087] In the formula, This represents the probability of wind turbine icing. This is the Sigmoid activation function.

[0088] In summary, the method of this application acquires raw SCADA data, obtains standardized feature vectors through missing value imputation and normalization; extracts low-dimensional latent features through an autoencoder, inputs them into a physical informative neural network to generate a neurodynamic state; constructs physical residuals based on power bias and uses meta-learning to dynamically adjust constraint strength; linearly transforms the neurodynamic state into query, key, and value vectors, introduces physical residuals and geodesic distance to correct attention weights, and generates refined feature vectors; constructs a comprehensive loss function weighted by data fitting loss, physical residual constraint loss, and icing differential operator constraint loss; defines a Fisher information metric for network parameters, constructs a natural gradient update rule on a Riemannian manifold to optimize parameters; and outputs the wind turbine icing probability. This method has advantages such as high detection accuracy, strong robustness, good physical consistency, and interpretable decision-making.

[0089] Please see Figure 2 The diagram shows a structural block diagram of an intelligent detection system for wind turbine blade icing according to this application.

[0090] like Figure 2 As shown, the intelligent detection system for wind turbine blade icing 200 includes an acquisition module 210, an extraction module 220, a first construction module 230, a correction module 240, a second construction module 250, an optimization module 260, and a mapping module 270.

[0091] The acquisition module 210 is configured to acquire raw data of wind turbine operation based on the SCADA system, perform missing value imputation and normalization on the raw data to obtain a standardized feature vector; the extraction module 220 is configured to input the standardized feature vector into an autoencoder to extract a low-dimensional latent feature vector, and input the low-dimensional latent feature vector into a physical feedback neural network to extract the spatiotemporal evolution features caused by blade icing using a multi-scale spectral sensing kernel to generate a neurodynamic state; the first construction module 230 is configured to calculate the power deviation based on the neurodynamic state and aerodynamic equations, construct a physical residual function, and dynamically adjust the physical constraint strength using a meta-learning mechanism to obtain physical constraint coefficients; the correction module 240 is configured to generate a query vector, a key vector, and a value vector from the neurodynamic state through linear transformation, calculate the dot product of the query vector and the key vector, and introduce the physical residual. The dot product result is corrected using the function and manifold geodesic distance to obtain physically corrected attention weights. These attention weights are then used to weighted aggregate the value vectors to generate refined feature vectors. A second construction module 250 is configured to construct a comprehensive loss function, which is obtained by weighted summation of data fitting loss, physical residual constraint loss, and icing differential operator constraint loss, where the weight of the physical residual constraint loss is the physical constraint coefficient. An optimization module 260 is configured to define a Fisher information metric for the network parameters in the physical informative neural network, construct a natural gradient update rule on the Riemannian manifold, and optimize the network parameters using the natural gradient update rule to minimize the comprehensive loss function, resulting in optimized network parameters. A mapping module 270 is configured to use the optimized network parameters to drive the output layer of the physical informative neural network to map the refined feature vectors to the wind turbine icing probability.

[0092] It should be understood that Figure 2 The modules and references described in the document Figure 1 The steps described in the text correspond to those in the method described above. Therefore, the operations, features, and corresponding technical effects described above also apply to the method described in the text. Figure 2 The various modules in the document will not be described in detail here.

[0093] In other embodiments, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the intelligent detection method for wind turbine blade icing in any of the above method embodiments.

[0094] In one embodiment, the computer-readable storage medium of the present invention stores computer-executable instructions, which are configured as follows:

[0095] Based on the raw data of wind turbine operation obtained by the SCADA system, the raw data is imputed and normalized to obtain a standardized feature vector.

[0096] The standardized feature vector is input into an autoencoder to extract a low-dimensional latent feature vector, and the low-dimensional latent feature vector is input into a physical information neural network. The spatiotemporal evolution features caused by leaf icing are extracted using a multi-scale spectral sensing kernel to generate a neurodynamic state.

[0097] Based on the aforementioned neurodynamic state, the power deviation is calculated using aerodynamic equations, a physical residual function is constructed, and a meta-learning mechanism is used to dynamically adjust the physical constraint strength to obtain the physical constraint coefficient.

[0098] The neurodynamic state is transformed linearly to generate query vector, key vector and value vector respectively. The dot product of query vector and key vector is calculated. The physical residual function and manifold geodesic distance are introduced to correct the dot product result to obtain the physically corrected attention weight. The attention weight is then used to weight and aggregate the value vector to generate a refined feature vector.

[0099] A comprehensive loss function is constructed, which is obtained by weighted summation of data fitting loss, physical residual constraint loss and icing differential operator constraint loss, wherein the weight of physical residual constraint loss is the physical constraint coefficient;

[0100] Define the Fisher information metric for the network parameters in the physical informant neural network, construct a natural gradient update rule on the Riemannian manifold, and use the natural gradient update rule to optimize the network parameters to minimize the comprehensive loss function, thereby obtaining the optimized network parameters;

[0101] Using the optimized network parameters, the output layer of the physical information neural network is driven to map the refined feature vector to the probability of wind turbine icing.

[0102] Computer-readable storage media may include a stored program area and a stored data area, wherein the stored program area may store an operating system and an application program required for at least one function; the stored data area may store data created based on the use of the intelligent wind turbine blade icing detection system, etc. Furthermore, the computer-readable storage medium may include high-speed random access memory, and may also include memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include memory remotely configured relative to a processor, which can be connected to the intelligent wind turbine blade icing detection system via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0103] Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiment of the present invention, such as... Figure 3 As shown, the device includes a processor 310 and a memory 320. The electronic device may also include an input device 330 and an output device 340. The processor 310, memory 320, input device 330, and output device 340 can be connected via a bus or other means. Figure 3 Taking a bus connection as an example, the memory 320 is the computer-readable storage medium described above. The processor 310 executes various server functions and data processing by running non-volatile software programs, instructions, and modules stored in the memory 320, thereby implementing the intelligent wind turbine blade icing detection method described in the above embodiment. The input device 330 can receive input digital or character information and generate key signal inputs related to user settings and function control of the intelligent wind turbine blade icing detection system. The output device 340 may include a display screen or other display device.

[0104] The aforementioned electronic device can execute the method provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in the embodiments of the present invention.

[0105] In one implementation, the aforementioned electronic device is applied to a smart detection system for wind turbine blade icing, serving as a client, and includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:

[0106] Based on the raw data of wind turbine operation obtained by the SCADA system, the raw data is imputed and normalized to obtain a standardized feature vector.

[0107] The standardized feature vector is input into an autoencoder to extract a low-dimensional latent feature vector, and the low-dimensional latent feature vector is input into a physical information neural network. The spatiotemporal evolution features caused by leaf icing are extracted using a multi-scale spectral sensing kernel to generate a neurodynamic state.

[0108] Based on the aforementioned neurodynamic state, the power deviation is calculated using aerodynamic equations, a physical residual function is constructed, and a meta-learning mechanism is used to dynamically adjust the physical constraint strength to obtain the physical constraint coefficient.

[0109] The neurodynamic state is transformed linearly to generate query vector, key vector and value vector respectively. The dot product of query vector and key vector is calculated. The physical residual function and manifold geodesic distance are introduced to correct the dot product result to obtain the physically corrected attention weight. The attention weight is then used to weight and aggregate the value vector to generate a refined feature vector.

[0110] A comprehensive loss function is constructed, which is obtained by weighted summation of data fitting loss, physical residual constraint loss and icing differential operator constraint loss, wherein the weight of physical residual constraint loss is the physical constraint coefficient;

[0111] Define the Fisher information metric for the network parameters in the physical informant neural network, construct a natural gradient update rule on the Riemannian manifold, and use the natural gradient update rule to optimize the network parameters to minimize the comprehensive loss function, thereby obtaining the optimized network parameters;

[0112] Using the optimized network parameters, the output layer of the physical information neural network is driven to map the refined feature vector to the probability of wind turbine icing.

[0113] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart detection method for icing on wind turbine blades, characterized in that, include: Based on the raw data of wind turbine operation obtained by the SCADA system, the raw data is imputed for missing values ​​and normalized to obtain a standardized feature vector. The standardized feature vector is input into an autoencoder to extract a low-dimensional latent feature vector, and the low-dimensional latent feature vector is input into a physical information neural network. The spatiotemporal evolution features caused by leaf icing are extracted using a multi-scale spectral sensing kernel to generate a neurodynamic state. Based on the aforementioned neurodynamic state, the power deviation is calculated using aerodynamic equations, a physical residual function is constructed, and a meta-learning mechanism is used to dynamically adjust the physical constraint strength to obtain the physical constraint coefficient. The neurodynamic state is transformed linearly to generate query vector, key vector and value vector respectively. The dot product of query vector and key vector is calculated. The physical residual function and manifold geodesic distance are introduced to correct the dot product result to obtain the physically corrected attention weight. The attention weight is then used to weight and aggregate the value vector to generate a refined feature vector. A comprehensive loss function is constructed, which is obtained by weighted summation of data fitting loss, physical residual constraint loss and icing differential operator constraint loss, wherein the weight of physical residual constraint loss is the physical constraint coefficient; Define the Fisher information metric for the network parameters in the physical informative neural network, construct a natural gradient update rule on the Riemannian manifold, and use the natural gradient update rule to optimize the network parameters to minimize the comprehensive loss function, thereby obtaining the optimized network parameters; Using the optimized network parameters, the output layer of the physical information neural network is driven to map the refined feature vector to the wind turbine icing probability.

2. The intelligent detection method for wind turbine blade icing according to claim 1, characterized in that, The step of inputting the standardized feature vector into the autoencoder to extract the low-dimensional latent feature vector includes: The standardized feature vector Input auto encoder, the auto encoder is composed of The network consists of a fully connected layer, with each layer sequentially connected to a batch normalization layer and a ReLU activation layer, as expressed in the following expression: ， ， ， In the formula, This is the output of the fully connected layer l. Let be the weight matrix of the l-th fully connected layer of the autoencoder. The output feature vector of the (l-1)th layer, For the bias term of the fully connected layer l of the autoencoder, For batch normalization function, To modify the activation function of the linear unit; go through After processing by a fully connected network, the low-dimensional latent feature vector is obtained as the output. , This is the output feature vector of the autoencoder.

3. The intelligent detection method for wind turbine blade icing according to claim 1, characterized in that, The step of inputting the low-dimensional latent feature vector into the physical information neural network, extracting the spatiotemporal evolution features caused by leaf icing using a multi-scale spectral sensing kernel, and generating a neural dynamic state includes: low-dimensional latent feature vectors Mapped to the initial state through a fully connected layer. ,in, For the mapping weight matrix, For mapping bias terms; By convolving the continuous function space states from past time points using a multi-scale spectral sensing kernel, the current neural dynamic state is generated, expressed as: ， In the formula, Let be the neurodynamic state at time t. It is a non-linear activation function. It is a learnable multi-scale spectral sensing kernel. For a continuous function space at historical moments state, It is a linear weight matrix. Let be the state of the continuous function space at the current time t.

4. The intelligent detection method for wind turbine blade icing according to claim 1, characterized in that, Based on the aforementioned neurodynamic state, the power deviation is calculated using aerodynamic equations, a physical residual function is constructed, and a meta-learning mechanism is employed to dynamically adjust the physical constraint strength, resulting in physical constraint coefficients including: Based on the neurodynamic state and combined with the aerodynamic power equation, the predicted power deviation is calculated as follows: ， In the formula, Theoretical power, The actual power calculated based on the neurodynamic state; Construct the physical residual function, with the following expression: ， In the formula, For physical residual functions, The measured power deviation collected by the SCADA system; Based on the physical residual function, a meta-learning mechanism is used to dynamically adjust the physical constraint coefficients, as expressed in the following expression: ， In the formula, For physical constraint coefficients, These are the initial values ​​for the physical constraint coefficients. For smooth activation function, The meta-learning rate, For physical residual loss Partial derivatives with respect to the physical constraint coefficients.

5. The intelligent detection method for wind turbine blade icing according to claim 1, characterized in that, The process involves generating query vectors, key vectors, and value vectors from the neurodynamic state through linear transformation, calculating the dot product of the query vector and the key vector, and correcting the dot product result using the physical residual function and manifold geodesic distance to obtain physically corrected attention weights. These attention weights are then used to weight and aggregate the value vectors to generate refined feature vectors. The neurodynamic state is used to generate query vectors through three independent linear transformations. Key vector Sum value vector ,in, The neurodynamic state at time point i. For the neurodynamic state at time point j, , , All are learnable linear transformation weight matrices; The corrected attention score is calculated using the following expression: ， In the formula, The adjusted attention score between the i-th and j-th time points. Scaling factor Let be the physical residual value at time point i. , All are adjustment coefficients. For Riemannian manifolds and Geodesic distance between them; The physically corrected attention weights are calculated using the following expression: ， In the formula, The attention weights are physically corrected. The adjusted attention score between the i-th time point and the k-th time point. It is an exponential function; The value vector is weighted and aggregated using the attention weights to generate a refined feature vector, expressed as follows: ， In the formula, This is the refined feature vector generated at the i-th time point after attention-weighted aggregation.

6. The intelligent detection method for wind turbine blade icing according to claim 1, characterized in that, The expression for the comprehensive loss function is: ， ， ， ， ， In the formula, For the comprehensive loss function, , , All of these are preset static hyperparameters. For dynamic adaptive data fitting loss, The physical constraint coefficients are for the t-th iteration and change dynamically as the training progresses. This is the residual loss due to aerodynamic power deviation. For the dynamic consistency loss of the icing differential operator, The number of iterations for optimizing the current manifold parameters. The physical constraint coefficients for the (k+1)th iteration are... Let be the physical constraint coefficients for the k-th iteration. For learning rate, It is a sigmoid smooth activation function. This represents the partial derivative of the aerodynamic power deviation residual loss with respect to the physical constraint coefficient. This represents the total number of samples in the current training batch. This is the sample imbalance adjustment coefficient, used to balance the baseline weights of frozen and unfrozen samples. To inform the output layer of the neural network about the predicted icing probability of the i-th sample, To focus parameters, This is the label for the true freezing state of the i-th sample. The predicted power deviation value is derived by combining neurodynamic state with Bates' law and blade element theory of wind turbine blades. Let i be the neurodynamic state vector of the i-th sample at the current time. This represents the actual monitored power deviation value. It is a nonlinear icing differential operator. Let be the first-order forward derivative of the neurodynamic state with respect to time. A physical evolution kinetic function characterizing the mass accumulation and thermodynamic phase transition process of ice layer on the blade surface. , , These are the key meteorological driving variables in the input of the i-th sample, corresponding to ambient temperature, relative humidity, and wind speed, respectively.

7. The intelligent detection method for wind turbine blade icing according to claim 1, characterized in that, The definition of the Fisher information metric for the network parameters in the physical information neural network, the construction of a natural gradient update rule on the Riemannian manifold, and the optimization of the network parameters using the natural gradient update rule to minimize the comprehensive loss function, yielding optimized network parameters including: Let the network parameters be Network parameters The Fisher information metric is defined as: ， In the formula, For network parameters The Fisher information metric matrix, Let p be the mathematical expectation of the probability distribution. Given input data x and network parameters Under the condition, output state The conditional probability distribution, Log-conditional probability of network parameters gradient, It is the transpose symbol; On a Riemannian manifold, a natural gradient update rule is constructed, expressed as follows: ， In the formula, The parameter value at the k-th iteration. The parameter value at the (k+1)th iteration. For learning rate, This is the inverse of the Fisher information metric matrix. The comprehensive loss function is applied to the network parameters at step k. The gradient; Iteratively execute the update rules until convergence to obtain the optimized network parameters. .

8. The intelligent detection method for wind turbine blade icing according to claim 1, characterized in that, The step of using optimized network parameters to drive the output layer of the physical information neural network to map the refined feature vector to the wind turbine icing probability includes: Using optimized network parameters Configure the weight matrix of the output layer and bias terms ; The refined feature vector The input and output layers, after linear transformation, are mapped to icing probabilities using the Sigmoid activation function, expressed as: ， In the formula, This represents the probability of wind turbine icing. This is the Sigmoid activation function.

9. A smart detection system for icing on wind turbine blades, characterized in that, include: The acquisition module is configured to acquire raw data of wind turbine operation based on the SCADA system, and perform missing value interpolation and normalization on the raw data to obtain a standardized feature vector. The extraction module is configured to input the standardized feature vector into an autoencoder to extract a low-dimensional latent feature vector, and input the low-dimensional latent feature vector into a physical information neural network to extract the spatiotemporal evolution features caused by leaf icing using a multi-scale spectral sensing kernel, thereby generating a neurodynamic state. The first construction module is configured to calculate the power deviation based on the neural dynamic state and the aerodynamic equation, construct the physical residual function, and dynamically adjust the physical constraint strength using a meta-learning mechanism to obtain the physical constraint coefficient. The correction module is configured to generate query vector, key vector and value vector respectively by linear transformation of the neurodynamic state, calculate the dot product of query vector and key vector, and introduce the physical residual function and manifold geodesic distance to correct the dot product result to obtain physically corrected attention weights, and use the attention weights to perform weighted aggregation on the value vector to generate refined feature vectors. The second construction module is configured to construct a comprehensive loss function, which is obtained by weighted summation of data fitting loss, physical residual constraint loss and icing differential operator constraint loss, wherein the weight of physical residual constraint loss is the physical constraint coefficient; The optimization module is configured to define the Fisher information metric for the network parameters in the physical informative neural network, construct a natural gradient update rule on the Riemannian manifold, and optimize the network parameters using the natural gradient update rule to minimize the comprehensive loss function, thereby obtaining the optimized network parameters. The mapping module is configured to use optimized network parameters to drive the output layer of the physical information neural network to map the refined feature vector to the wind turbine icing probability.

Images