Offshore wind turbine monitoring data missing intelligent filling method based on multi-head gated residual network

Abstract

The application discloses an offshore wind turbine monitoring data missing intelligent filling method based on a multi-head gated residual network, relates to the technical field of artificial intelligence, and comprises the following steps: a multi-head gated residual network model is constructed; the multi-head gated residual network model is trained by using a model input data set, model parameters are optimized, and the model is converged until a complete training missing data filling model is obtained. Through the construction of three parallel feature extraction heads, including a space-time feature extraction head, a working condition mode extraction head and a missing mode extraction head, the time sequence dependence, the spatial correlation, the working condition coupling characteristics and the missing distribution mode can be captured from the offshore wind turbine monitoring data respectively and cooperatively, the deep fusion and feature extraction of the multi-source heterogeneous information under the complex operation environment are realized, and thus the accuracy of the missing value filling and the context understanding ability are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and particularly relates to a method for intelligent filling of missing offshore wind turbine monitoring data based on a multi-head gated residual network. BACKGROUND

[0002] With the development of large-scale and deep-sea offshore wind power, the single capacity of offshore wind turbine generators continues to increase. The operation state monitoring and data analysis of the offshore wind turbine generators are crucial for ensuring the safety of the generators, optimizing the operation and maintenance strategies, and improving the power generation efficiency. However, due to the harsh offshore environment and complex working conditions, data missing often occurs in the data acquisition and transmission process of the wind turbine monitoring system due to sensor failure, communication interruption, extreme weather and other factors. These missing values destroy the time sequence integrity of the data, seriously restricting the reliability of advanced applications such as fault diagnosis, performance evaluation and life prediction based on data, and becoming one of the key bottlenecks restricting the improvement of intelligent operation and maintenance level of offshore wind power.

[0003] At present, the filling technology for time series data missing has developed from traditional statistical methods to various schemes including machine learning. Traditional methods such as mean filling, forward / backward filling and linear interpolation are simple and easy to implement, but they ignore the inherent time sequence dependence and complex nonlinear correlation between multiple variables, and have limited filling accuracy on strong dynamic and multi-coupled system data such as offshore wind turbines. The methods based on statistical models (such as ARIMA and Kalman filter) and shallow machine learning (such as support vector regression and random forest) have improved the modeling ability to some extent, but their feature expression ability and adaptability to long-term dependence and complex missing patterns are still insufficient. In recent years, deep learning models, especially recurrent neural networks and variational autoencoders, have been introduced into this field due to their strong sequence modeling ability, showing better potential.

[0004] However, the existing deep learning methods still face significant challenges when applied to the filling of multi-variable monitoring data of offshore wind turbines. First, wind turbine data contains rich spatio-temporal features, dynamic patterns that change with operating conditions, and specific missing distribution patterns, and a single model architecture cannot capture these heterogeneous information simultaneously. Second, most models implicitly handle missing data, failing to explicitly model the impact of missing patterns on context understanding, resulting in performance degradation in complex scenarios such as continuous missing and random large block missing. Third, simply applying a complete data trained model to a real missing environment has the problem of inconsistency between training and inference scenarios, and lacks an effective mechanism to check the physical consistency and engineering reasonableness of the filling results.

[0005] Therefore, it is necessary to invent a method for intelligent filling of missing offshore wind turbine monitoring data based on a multi-head gated residual network to solve the above problems. SUMMARY

[0006] The offshore wind turbine monitoring data missing intelligent filling method based on the multi-head gated residual network aims to solve the problems in the background art.

[0007] To achieve the above-mentioned purpose, the offshore wind turbine monitoring data missing intelligent filling method based on the multi-head gated residual network comprises the following steps:

[0008] S1, obtaining a historical sequence of offshore wind turbine multivariate operating parameters containing missing values, and performing standardization preprocessing on the historical sequence to obtain a standardized sequence, so as to construct a model input data set;

[0009] S2, constructing a multi-head gated residual network model; the model comprises three feature extraction heads arranged in parallel, a gated fusion module and a residual prediction module; the three feature extraction heads comprise a space-time feature extraction head, a working condition mode extraction head and a missing mode extraction head;

[0010] S3, training the multi-head gated residual network model by using the model input data set, optimizing the model parameters, until the model converges, and obtaining a trained missing data filling model;

[0011] S4, inputting a real-time operating parameter sequence of the offshore wind turbine containing missing values to be filled into the trained missing data filling model, and outputting a complete operating parameter sequence of the wind turbine in which the missing values have been filled by the model;

[0012] S5, performing a rationality check on the complete operating parameter sequence of the wind turbine in which the missing values have been filled; if the check passes, outputting the filling result as the final filling data; if the check fails, outputting an abnormality check alarm or calling a backup filling result.

[0013] The technical effects and advantages of the present application are as follows:

[0014] 1. The present application can capture time series dependence, spatial correlation, working condition coupling characteristics and missing distribution mode from offshore wind turbine monitoring data respectively and cooperatively by constructing three feature extraction heads arranged in parallel, including a space-time feature extraction head, a working condition mode extraction head and a missing mode extraction head, realizing deep fusion and feature extraction of multi-source heterogeneous information in a complex operating environment, and thus significantly improving the accuracy of missing value filling and context understanding ability;

[0015] 2. The present application introduces a gated fusion module, dynamically calculates and allocates adaptive weights of outputs of each feature extraction head according to real-time context of input sequences by using a learnable weight generation network, realizes intelligent weighted fusion of different feature contributions, and this mechanism enables the model to flexibly cope with different missing scenarios and operating conditions, enhancing the adaptability and robustness of the model.

[0016] 3、The application converts the learning goal of the model from directly predicting the complete sequence to predicting the residual (correction amount) relative to an initial reference sequence by adopting a residual prediction module, which reduces the difficulty of direct regression of the model, makes the training process more stable and efficient, and effectively combines the preliminary linear estimation and nonlinear deep features by adding the reference sequence, thereby further improving the accuracy of the final filling result;

[0017] 4、The application trains the model by designing a loss function composed of a reconstruction loss and a regularization loss, which not only drives the model to accurately fit the known data, but also effectively prevents the model from overfitting through the regularization term and encourages the gating fusion module to balance the use of each feature source, thereby promoting the generalization ability and feature utilization efficiency of the model;

[0018] 5、The application adds a rationality checking step based on physical rules and statistical laws at the output end of the model to automatically check the filled sequence for multiple indicators such as extreme value, continuity and multivariate correlation consistency, which provides reliable quality control and safety valve for the filling result, and can trigger an alarm or call a backup plan when the result is abnormal, thereby greatly guaranteeing the physical rationality and system safety of the filling result in subsequent engineering applications. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The method flowchart of the application.

[0020] Figure 2 The standardized preprocessing flowchart of the application.

[0021] Figure 3 The multi-head gating residual network model structure diagram of the application.

[0022] Figure 4 The gating fusion module flowchart of the application.

[0023] Figure 5 The rationality checking flowchart of the application. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0025] The application provides a method for filling missing data in a time series as follows Figure 1The illustrated offshore wind turbine monitoring data missing intelligent filling method based on the multi-head gated residual network specifically comprises the following steps:

[0026] S1, obtain the offshore wind turbine multivariate operating parameter history sequence containing missing values, and perform standardization preprocessing on the history sequence to obtain a standardized sequence to construct a model input data set;

[0027] Further, in the above technical solution, the standardization preprocessing is specifically: for each operating parameter variable, the mean and standard deviation are calculated based on the non-missing values in its history data, and the Z-Score standardization method is used for processing; at the same time, a binary missing mask matrix is generated to identify the missing state of each data point in the history sequence.

[0028] It should be known that the "obtaining the offshore wind turbine multivariate operating parameter history sequence containing missing values" specifically comprises: extracting the multivariate time series data within a set time window from the monitoring and data acquisition system or the historical database of the offshore wind turbine generator. The multivariate operating parameters at least include: wind speed, wind turbine power, generator speed, pitch angle, gearbox oil temperature, and cabin vibration amplitude. The sequence is arranged in a fixed sampling interval in the time dimension to form a matrix like X∈R^(T×N), wherein T is the total number of time steps, and N is the total number of operating parameter variables. The missing values in the sequence are caused by data acquisition failure or transmission interruption, and are represented by a specific identifier in the initial data.

[0029] Reference Figure 2 , the "standardization preprocessing" is executed in the following sub-steps in specific implementation:

[0030] S11, missing value identification and sequence construction: read the original history sequence, uniformly replace the identifiers representing missing values in it with the numerical value NaN, and ensure that the sequence is strictly aligned according to the time stamp to obtain an initial data matrix X raw .

[0031] S12, mask matrix generation: synchronously create a binary mask matrix M∈{0,1}^(T×N)with the same dimension as X raw . For each data point x raw in X {t，n} , if x {t，n} is NaN, then m {t，n} =0 in the corresponding position of the mask matrix, indicating that the point is missing; if x {t，n} is a valid value, then m {t，n} =1, indicating that the point exists.

[0032] Where t is the index of time step, e.g. t = 1 represents the first sampling time in the sequence, t = 2 represents the second sampling time, and so on, which is used to locate the position of data in time dimension; n is the index of operating parameter variable, e.g. n = 1 can represent "wind speed", n = 2 can represent "wind turbine power", and so on, which is used to locate the position of data in variable dimension.

[0033] S13, statistical quantity calculation: for the nth variable, collect X raw all m {t，n} =1 positions corresponding valid values, calculate the mean μ n and the standard deviation σ n of the valid values of the variable.

[0034] S14, Z-Score standardization: using the calculated μ n and σ n , standardize each valid data point in the initial data matrix X raw to obtain the standardized sequence matrix X norm . The calculation formula for any position (t, n) is as follows:

[0035] If m {t，n} =1 (valid data point): x norm，{t，n} = (x raw，{t，n} - μ n ) / σ n ;

[0036] If m {t，n} =0 (missing data point): x norm，{t，n} still remains as NaN or a preset placeholder (such as 0), but its missing state has been completely captured by the mask matrix M;

[0037] Where x raw，{t，n} is the original historical sequence matrix X raw , the original data value at the tth time step and the nth variable position, which is a valid monitoring value or a missing NaN.

[0038] x norm，{t，n} is the standardized sequence matrix X norm , the standardized data value at the tth time step and the nth variable position.

[0039] The "model input data set" is composed of the standardized sequence Xnorm and its corresponding binary missing mask matrix M, which is used as the input of the model in the subsequent steps.

[0040] Reference Figure 3S2, a multi-head gated residual network model is constructed; the model comprises three characteristic extraction heads arranged in parallel, a gated fusion module, and a residual prediction module;

[0041] Further, in the above technical solution, the three characteristic extraction heads arranged in parallel comprise:

[0042] a space-time characteristic extraction head for capturing the dependence of the fan operation parameters in the time and space dimensions;

[0043] a working condition mode extraction head for learning the steady-state and transient correlation mode between parameters under different fan operation conditions;

[0044] a missing mode extraction head for explicitly modeling the distribution mode of the missing values and the influence of the missing values on the context information.

[0045] It should be noted that the three characteristic extraction heads take the standardized sequence X norm and the binary missing mask matrix M obtained in step S1 as common basic inputs. Each characteristic extraction head independently operates and aims to deconstruct the information related to missing value prediction from different angles, and finally outputs characteristic vectors of the same dimension. Specifically:

[0046] the space-time characteristic extraction head is implemented as a subnetwork stacked by a time series convolution network and a self-attention mechanism in sequence, and the input of the subnetwork is the standardized sequence X norm . First, X norm is input into a one-dimensional time series convolution network layer, which includes multiple convolution kernels with a length of k (for example, k=3), and by performing causal convolution in the time dimension, local time pattern features of each variable sequence are extracted. Subsequently, the output of the time series convolution layer is fed into a multi-head self-attention layer. In this layer, each time step of the sequence is regarded as a feature vector, and by calculating attention scores, the model can adaptively capture global long-term dependencies across the entire time window and simultaneously model the spatial correlation between different variables (channels). The final output of the space-time characteristic extraction head is the aggregated representation of the feature matrix after the transformation of the self-attention layer in the feature dimension, denoted as vector H st .

[0047] The working condition mode extraction head aims to deeply integrate the prior knowledge of the fan operation into the feature representation. In specific implementation, an additional working condition label sequence L is required, which is in one-to-one correspondence with the time sequence X norm in the time dimension, and each label l ta working state (e.g., "shutdown", "startup", "rated generation", "de-rated operation", "fault", etc.) of the wind turbine at time step t. First, the discrete working condition label is embedded to map it into a dense working condition feature vector E t . Then, at each time step t, the normalized data vector x norm，t is concatenated with the corresponding working condition embedding vector E t to form an augmented input vector [x norm，t ; E t ]. The augmented input of the whole time series is fed into a multi-layer perceptron (MLP) which contains at least one hidden layer with a non-linear activation function to learn the complex non-linear coupling relationship between working conditions and operating parameters. The output of the working condition pattern extraction head is a fixed-dimensional vector H op obtained by pooling (e.g., global average pooling) the time series features of the output of the MLP.

[0048] The missing pattern extraction head explicitly uses the missing mask M to guide feature learning to distinguish the different contexts of data missing and normal observation. Specifically, the normalized sequence X norm is concatenated with the binary missing mask matrix M in the feature dimension to form a two-channel input [X norm ; M]. The concatenated tensor is fed into a network layer based on the gated recurrent unit. The gating mechanism (update gate and reset gate) of the GRU unit enables it to effectively learn and remember the dependencies in the sequence. In this design, the mask M directly informs the model of the reliability of the data at each position as an additional input signal, so that the GRU can learn how to adjust its dependence on historical information and update strategy under different patterns of continuous missing, intermittent missing, etc., and thus extract context features sensitive to missing. The output of the missing pattern extraction head is a feature vector H mp obtained by taking the hidden state at the last time step of the GRU or aggregating the hidden states of all time steps.

[0049] Further, in the above technical solution, the spatio-temporal feature extraction head is composed of a time series convolution network layer and an attention mechanism layer in series; the time series convolution network layer is used to extract local time features, and the attention mechanism layer is used to model global time dependencies and spatial correlations between different operating parameter variables.

[0050] The working condition pattern extraction head takes the concatenation of the encoded working condition label vector and the normalized sequence as input, and extracts features deeply coupled with a specific working condition through a multi-layer perceptron.

[0051] The missing pattern extraction head extracts the context features in the missing pattern by generating a missing mask matrix with the same dimension as the input sequence, and inputting the mask matrix and the normalized sequence into a network layer based on a gated recurrent unit to learn the context features in the missing pattern.

[0052] It should be noted that the specific network structure of each feature extraction head and the data flow are as follows:

[0053] Specific implementation of the spatio-temporal feature extraction head:

[0054] The temporal convolution network layer adopts one-dimensional causal convolution to ensure that the output at time t only depends on the input at time t and before time t, which meets the characteristics of time series prediction. The number of convolution kernels is set to C1, and the length of the convolution kernel is K (for example, K=3). The input of the layer is the normalized sequence X norm ∈R^(T×N), after convolution and nonlinear activation function (such as ReLU), the local feature tensor F conv ∈R^(T×C1)is output.

[0055] The attention mechanism layer adopts a multi-head self-attention mechanism. First, F conv is split into h heads and calculated in parallel by three different linear projection layers to generate query Q, key K, and value V. The number of attention heads is set to h, and the C1-dimensional features are divided into h heads for parallel attention calculation. The attention function is a scaled dot-product attention. For each head i, the output is:

[0056] head i =Attention(Q i ,K i ,V i )=softmax((Q i *K i ^T) / sqrt(d k ))*V i ,

[0057] Where d k is the dimension of the key vector.

[0058] The outputs of all heads are concatenated in the feature dimension, and then fused by a linear projection layer W o to obtain the output F attn of the attention layer ∈R^(T×C1). This process models the association between time steps (global temporal dependency) and different convolution channels representing feature patterns (spatial correlation).

[0059] Output generation: apply global average pooling to F attn in the time dimension T to obtain a C1-dimensional feature vector H st as the final output of the extraction head.

[0060] Implementation of working condition mode extraction head:

[0061] Working condition label encoding: there are P predefined working conditions (e.g., stop = 0, start = 1, rated operation = 2, …). First, the working condition label sequence L in integer form is one-hot encoded to obtain a T x P matrix. Then, the one-hot vector is mapped to a dense working condition embedding vector E ∈ R^ (T x D e ) through a learnable embedding layer, where D e is the embedding dimension.

[0062] Data concatenation: at each time step t, the normalized data vector x norm，t ∈ R^N is concatenated with the corresponding working condition embedding vector e t ∈ R^ (D e ) to obtain the enhanced vector c oncatt =[x norm，t ; e t ] ∈ R^ (N + D e ). The entire sequence forms a tensor F concat ∈ R^ (T x (N + D e )).

[0063] Multi-layer perception processing: input F concat into an MLP containing at least one hidden layer. The MLP can be defined as:

[0064] MLP (F concat ) = W2 * σ (W1 * F concat + b1) + b2,

[0065] where W1, b1, W2, b2 are learnable parameters, and σ is a nonlinear activation function (such as ReLU or GELU). The output dimension of the MLP is the same as the time step length T, but the feature dimension is transformed to D mlp .

[0066] Output generation: perform global average pooling on the time series features of the MLP output in the time dimension T to obtain a D mlp dimensional feature vector H op as the final output of the extraction head.

[0067] Implementation of missing mode extraction head:

[0068] Input preparation: directly use the binary missing mask matrix M ∈ R^ (T x N) generated in step S1. Concatenate the normalized sequence X norm and the mask matrix M in the feature dimension (i.e., variable dimension N) to form the joint input F missing =[X norm; M] ∈ R^(T×2N). Here, concatenation means that for each original variable, the model receives both its standardized value (if available) and a flag indicating whether it is missing.

[0069] GRU network processing: The joint input F missing is fed into a gated recurrent unit layer in a time step order. The GRU layer can contain a single layer or multiple layers with a hidden state dimension D gru . The GRU unit computes the current hidden state h missingt at each time step t from the current input f {t-1} and the previous hidden state h t through mechanisms of update gate z t and reset gate r t . This process enables it to learn dynamic contextual dependencies affected by the missing data pattern in the sequence.

[0070] Output generation: The hidden state h T at the last time step of the GRU layer is taken as the summary feature of the sequence, i.e., a D gru -dimensional feature vector H mp is obtained as the final output of the extraction head. Alternatively, a pooling operation can be performed on the hidden states of all time steps to generate H mp .

[0071] Further, in the above technical solution, with reference Figure 4 to the gated fusion module, the outputs of the three feature extraction heads are dynamically weighted and integrated; the following steps are specifically performed:

[0072] A21, receiving output vectors from each of the feature extraction heads;

[0073] A22, calculating an adaptive weight vector for each output vector of the feature extraction head according to the current input sequence context through a learnable gated weight generation network;

[0074] A23, performing weighted summation of each output vector and its corresponding adaptive weight vector to generate a fused comprehensive feature vector.

[0075] It should be noted that the specific workflow and technical details of the gated fusion module are as follows:

[0076] A21, input reception: the gated fusion module receives output vectors from K feature extraction heads. According to the foregoing embodiments, K = 3, which are the spatiotemporal feature vector H st ∈ R^D, the operating mode feature vector H op ∈ R^D, and the missing pattern feature vector H mpX∈R^D. Here we assume that the outputs of each feature extractor head have been unified to the same dimension D, e.g., D = 128, by linear projection.

[0077] A22, Adaptive weight computation: This is the core of the gated fusion. The "learnable gating weight generation network" is a light-weight neural network, whose design goal is to evaluate the relative importance of the information provided by each feature extractor head based on the overall context of the current input sample. The specific implementation is:

[0078] Context information construction: In order to capture the global context, the standardized sequence X norm is flattened or aggregated in some form (e.g., average pooling along the time and feature dimensions, respectively) to form a global context vector C. At the same time, it is also possible to consider the statistical properties (such as the missing rate) of the binary missing mask matrix M as an additional input. A preferred implementation is to fuse or select the key representatives of the inputs of the K feature extractor heads as the input of the gating network, for example, using some aggregated representation of X norm .

[0079] Gating network forward propagation: The global context vector C is input into the gating weight generation network. This network consists of a multi-layer perceptron, whose last layer is a linear layer, outputting a K-dimensional raw weight score vector g ∈ R^K. For example, g = W g *σ(W c *C + b c ) + b g , where W c , b c , W g , b g are learnable parameters, and σ is an activation function.

[0080] Weight normalization: In order to obtain reasonable, additive weights, the raw weight score g is normalized by applying the softmax function to obtain the final adaptive weight vector α ∈ R^K, and Σα i = 1. That is:

[0081] α = softmax(g) = [exp(g1) / Σexp(g j ), exp(g2) / Σexp(g j ),..., exp(g K ) / Σexp(g j )],

[0082] where α iAdaptive weight scalar corresponding to the i-th feature extraction head. This enables the model to dynamically decide whether to rely more on spatiotemporal features, operating condition features, or missing pattern features according to the current input (e.g., specific missing pattern, operating condition stage).

[0083] A23, Feature Weighted Fusion: The output vector H i Adaptive weight scalar a corresponding to it i Weighted sum to generate the fused comprehensive feature vector H fused ∈R^D. The calculation formula is:

[0084] H fused =Σ(α i *H i ), where i=1, 2,..., K.

[0085] This step realizes the multi-perspective dynamic integration of information. For example, when the input sequence is severely missing, the gating network allocates higher weights to the missing pattern extraction head; when the fan is in a stable rated operating condition, more reliance is placed on the operating condition pattern extraction head and the spatiotemporal feature extraction head.

[0086] The output of the gating fusion module, i.e., the comprehensive feature vector H fused , will serve as the input to the subsequent residual prediction module.

[0087] Further, in the above technical solution, the residual prediction module takes the comprehensive feature vector as input and performs the following steps:

[0088] B21, Map the comprehensive feature vector through a fully connected layer to preliminarily generate the filling prediction value of the missing value;

[0089] B22, Perform residual addition of the filling prediction value and the initial reference sequence to generate the final filling output;

[0090] B23, Perform inverse normalization processing on the final filling output to restore it to the original physical dimension.

[0091] Further, in the above technical solution, the initial reference sequence is generated by inputting the historical sequence obtained in S1 to a linear projection layer, and the output of the linear projection layer is the initial reference sequence.

[0092] It should be noted that the specific implementation details of the residual prediction module and the initial reference sequence generation are as follows:

[0093] Regarding the generation of the initial reference sequence:

[0094] The initial baseline sequence is the starting point of the residual learning, which is a preliminary estimate of the original input sequence. The specific generation method is: input the original historical sequence containing missing values obtained in step S1 (or its form before standardization, denoted as X raw ) into an independent linear projection layer. The linear projection layer is composed of a weight matrix W base ∈R^(N×N)and a bias vector b base ∈R^N. For each time step t of the input sequence, the original data vector x raw，t ∈R^N (the missing position can be temporarily replaced by 0 or other neutral values), calculate:

[0095] b t =W base *x rawt +b base ,

[0096] After calculating all time steps t=1,..., T, the initial baseline sequence B ∈ R^(T×N)is obtained. This operation aims to map the original data to a feature space more suitable as a prediction baseline through a simple linear transformation. In the model training process, W base and b base are optimized together as learnable parameters.

[0097] Detailed steps of the residual prediction module:

[0098] B21, preliminary filling prediction value generation:

[0099] Take the comprehensive feature vector H fused ∈R^D output by the gating fusion module as input. First, map it to a vector matching the target output dimension through a fully connected layer (linear layer). The weight matrix of the fully connected layer is W pred1 ∈R^(D×(T*N)),and the bias is b pred1 ∈R^(T*N). Calculate the flattened prediction vector:

[0100] P flat =W pred1 *H fused +b pred1 .

[0101] Then, reshape P flat to the same time step T and variable number N as the input sequence to obtain the preliminary filling prediction value P ∈ R^(T×N), where P represents the "residual" or "increment" relative to the initial baseline sequence B that the model considers.

[0102] B22, residual addition to generate the final filling output:

[0103] The filling prediction value (residual) P obtained in the previous step is added element by element to the initial reference sequence B to generate the filling output X in the final standardized space of the module filled，norm ∈R^(T×N)。The calculation formula is:

[0104] X filled，norm =B+P,

[0105] This is the core embodiment of the residual learning idea: the model does not directly predict the complete sequence, but predicts a correction amount P which is easy to learn, and adds it to the basic linear estimate B to obtain a more accurate final result. X filled，norm The originally missing positions in the sequence are filled with the calculated values.

[0106] B23、De-standardization processing:

[0107] In order to convert the filling result back to a value with actual physical meaning, the X filled，norm needs to be de-standardized (inverse Z-Score transformation). The mean μ n and standard deviation σ n calculated and saved for each variable n in step S1 are used to calculate each element x filled，norm in X filled，norm，{t，n} as follows:

[0108] X filled，raw，{t，n} =x filled，norm，{t，n} *σ n +μ n ,

[0109] After performing this operation on all elements, the complete filled sequence X filled，raw ∈R^(T×N) is obtained, which is restored to the original physical dimension. This sequence is the final filling result of the model for the input missing sequence.

[0110] S3, train the multi-head gated residual network model using the model input data set, optimize the model parameters until the model converges, and obtain a trained missing data filling model;

[0111] Further, in the above technical solution, the loss function L used to train the multi-head gated residual network model is the sum of the reconstruction loss and the regularization loss, specifically represented as:

[0112] L=L rec +λ*L reg ,

[0113] Where L rec is the reconstruction loss term, L reg is the regularization loss term, and λ is the regularization coefficient.

[0114] It is necessary to know that each term of the loss function L is specifically defined, and the entire training process needs to be carried out under a specific optimization strategy, and the specific implementation details are as follows:

[0115] Specific calculation of reconstruction loss L rec :

[0116] The reconstruction loss is used to measure the difference between the model's imputed values and the true observed values. Since the training data itself contains missing values, the loss should only be calculated on known true observed locations. In practice, a masked mean squared error is used as the reconstruction loss function.

[0117] Calculation object: Let the model output the imputed sequence after inverse standardization for a batch of training data X filled，raw , the corresponding true complete sequence (pre-completed or simulated generated in the training set) is X true , and the binary missing mask matrix is M (where 1 represents an observed point and 0 represents an original missing point).

[0118] Calculation formula:

[0119] L rec = (1 / sum(M))*∑ {t，n} [M {t，n} * (X filled，raw，{t，n} -X true{t，n} )^2],

[0120] Where sum(M) represents the total number of all valid observed points in the current batch, and ∑ {t，n} represents the sum of all time steps t and all variables n. This calculation ensures that the loss function only penalizes the model's prediction error on known data points, thereby learning the ability to fill in based on valid context.

[0121] Specific composition of regularization loss L reg :

[0122] The regularization loss term L reg aims to prevent the model from overfitting and guide it to learn more robust or more interpretable features. In practice, L reg is the sum of multiple regularization terms, mainly including:

[0123] L2 weight decay: penalize the square sum of all trainable parameters in the model (such as convolution kernel weights, fully connected layer weights, biases, etc., but not limited to). This term is represented as ||θ||^2, where θ represents all model parameters.

[0124] Gated weight entropy regularization: To encourage the model to fully utilize the three feature extraction heads and avoid the gated fusion module always favoring a single feature head, entropy regularization can be added to the adaptive weight vector α calculated in step A22. For an input sample, its gated weight entropy is defined as:

[0125] H(α) = -Σ(α) i *log(α) i Minimizing negative entropy (i.e., -H(α)) can promote a more uniform weight distribution.

[0126] Therefore, L reg One specific implementation method can be defined as:

[0127] L reg =β*||θ||^2+γ*[-H(α)]

[0128] Here, β and γ are hyperparameters that control the strength of the two regularization terms, and can be considered together with λ or set separately.

[0129] The regularization coefficient λ is a preset hyperparameter used to balance the importance of reconstruction loss and regularization loss. In practice, a suitable value of λ can be selected by performing cross-validation on the validation set, for example, by performing a grid search from the set {0.001, 0.01, 0.1}.

[0130] The model is trained using a gradient descent-based optimization algorithm, such as the Adam optimizer. The training process includes forward propagation to calculate the loss, backpropagation to calculate the gradient, and updating the model parameters using the optimizer. Hyperparameters such as the number of training epochs, batch size, and learning rate are set. The model is considered to have converged and training stops when the reconstruction loss on the independent validation set no longer decreases significantly or reaches the preset number of epochs.

[0131] S4. Input the sequence of real-time operating parameters of offshore wind turbines containing missing values ​​into the fully trained missing data filling model, and the model outputs a complete sequence of wind turbine operating parameters with missing values ​​filled.

[0132] In practice, this step is performed as follows:

[0133] S41. Real-time Sequence Acquisition and Alignment: Acquire the real-time multivariate operating parameter sequence within a set time window from the offshore wind turbine monitoring system. The variable composition and sampling interval of the sequence are consistent with the historical sequence in step S1, including missing values ​​caused by data acquisition or transmission interruptions. This sequence is denoted as X. real-time，raw .

[0134] S42. Real-time sequence normalization and mask generation: Using the mean μ of each variable calculated based on historical non-missing values ​​saved in step S1. n and standard deviation σ n , for X real-time，raw Each valid data point in the sequence is Z-score standardized to obtain the standardized real-time sequence X. real-time，norm At the same time, generate a [something] with X. real-time，norm Binary missing mask matrix M of the same dimension real-time , is used to identify its real-time missing status (1 indicates valid, 0 indicates missing).

[0135] S43, Model Loading and Input Construction: Load the multi-head gated residual network model, which has been trained and has fixed parameters in step S3, from the storage medium. X real-time，norm With M real-time Together, they constitute the model input dataset. If the model requires operating condition labels as input, the corresponding operating condition label sequence L for the time window is simultaneously acquired or inferred. real-time And perform embedding encoding.

[0136] S44. Forward Inference of the Model: The model constructed in step S43 is input and fed into the fully trained multi-head gated residual network model. The model sequentially performs forward propagation calculations through its three parallel feature extraction heads (spatiotemporal feature extraction head, working condition mode extraction head, and missing mode extraction head), the gated fusion module, and the residual prediction module.

[0137] S45. Output Generation and Denormalization: The residual prediction module outputs the complete sequence X in the normalized space. filled，norm Then, using the μ saved in step S1 n and σ n , for X filled，norm Perform destandardization, the calculation formula is x filled，raw，t，n =x filled，norm，t，n ×σ n +μ n , where x filled，raw，t，n x represents the final filled value at the nth time step and the position of the running parameter variable in the completed sequence after filling, under the original physical dimensions; filled，norm，t，n Within the standardized space, this refers to the standardized value at the nth operating parameter variable position at the t-th time step in the sequence directly output by the model residual prediction module; the final wind turbine operating parameter sequence X, restored to its original physical dimensions and with missing values ​​filled, is obtained through this calculation formula. filled，raw This sequence is the output of step S4 and will be sent to step S5 for validity verification.

[0138] S5, performing rationality verification on the complete fan operation parameter sequence with missing values filled; if the verification passes, outputting the filling result as the final filling data; if the verification fails, outputting a verification exception alarm or calling a backup filling result.

[0139] To verify the effectiveness of the method, a plurality of experiments were performed on the actual monitoring data set of the offshore wind turbine, and the performance of the method and the conventional filling method under different missing rates was compared. The experimental data is shown in the following table:

[0140]

[0141] Table explanation:

[0142] The experimental data is derived from the actual monitoring data of a certain offshore wind farm for 6 months, including wind speed, power, speed and other variables.

[0143] The missing rate is generated by random masking, and each experiment is repeated 10 times to take the average value.

[0144] The mean square error (MSE) and the mean absolute error (MAE) are calculated after standardization, and the smaller the value, the more accurate the filling.

[0145] The rationality verification pass rate refers to the proportion of filling results that meet the physical and statistical rules in S5.

[0146] Further, in the above technical solution, the reference Figure 5 , the rationality verification specifically includes the following steps:

[0147] S51, based on physical rules and statistical laws, a plurality of verification indexes are calculated for the filled sequence; the plurality of verification indexes include sequence extreme rationality index, data point interval change continuity index and multi-variable physical correlation consistency index;

[0148] S52, comparing the plurality of calculated verification indexes with the corresponding preset threshold range based on historical normal data statistics;

[0149] S53, if all verification indexes are within their corresponding preset threshold range, it is determined that the verification passes; if any verification index exceeds its corresponding preset threshold range, it is determined that the verification fails.

[0150] It should be noted that the specific implementation of the rationality verification relies on a pre-established quantifiable verification rule library and threshold set, and the detailed technical details are as follows:

[0151] The specific calculation method and threshold setting of the verification index:

[0152] S51-1, Sequence extreme rationality index calculation and threshold:

[0153] Calculation: For each operating parameter variable (e.g. wind speed v, power P, rotation speed ω) in the filled sequence X filled，raw , calculate its minimum value min var and maximum value max var in the time window to be checked, respectively.

[0154] Threshold: Based on long-term historical normal data, statistics the absolute physical range of each variable under all normal operating conditions. For example, the threshold range of wind speed v can be set as [v minhist , v maxhist ], where v minhist is 0 m / s (no wind), and v maxhist is the cut-out wind speed of the fan (e.g. 25 m / s). The threshold range of power P is [0, rated power].

[0155] Comparison: Check whether min var is greater than or equal to v minhist , and whether max var is less than or equal to v maxhist . If so, this item of the variable passes.

[0156] S51-2, Change continuity between data points index calculation and threshold:

[0157] Calculation: For each operating parameter variable in the filled sequence X filled，raw , calculate the first-order difference (change) sequence Δx t = x {t+1} - x t of its adjacent time step data. Then, calculate the absolute value maximum max abs，diff or standard deviation std diff of the difference sequence as the index of change severity.

[0158] Threshold: Based on historical normal data, statistics the absolute value distribution of the difference between adjacent time steps during the smooth running period (excluding transient processes such as start-up and shutdown). The threshold can be set as a certain high percentile (e.g. 99th percentile) d 99 of the historical difference absolute value, or "mean + 3 times standard deviation".

[0159] Comparison: Check whether max abs，diff is less than or equal to d 99 . If so, this item of the variable passes, and this check aims to identify the mutation point that does not conform to the physical inertia law caused by incorrect filling.

[0160] S51-3, Multivariate physical correlation consistency index calculation and threshold:

[0161] Calculation: Based on the aerodynamics and control system principles of the wind turbine, check the correlation between key variables, most typically the power-wind speed characteristic curve.

[0162] For the filled wind speed v filled and power P filled , according to the theoretical power curve of the wind turbine or the empirical relationship fitted based on a large amount of historical normal data, calculate the corresponding theoretical power estimate P filled at v the .

[0163] Calculate the relative deviation of the filled power and the theoretical power at the actual wind speed: δ = |P filled -P the | / (rated power) (or other normalization methods).

[0164] Calculate the average or maximum value of δ in the entire time window as the consistency indicator c index .

[0165] Threshold: Based on historical normal data, statistics the distribution of the relative deviation between the actual power and the theoretical power estimate. The threshold can be set to a certain high percentile (such as the 95th percentile) of the deviation distribution c 95 .

[0166] Comparison: Check if c index is less than or equal to c 95 . If it meets, this variable pair passes this indicator. Other correlations can also be checked, such as the correlation between the pitch angle and the power in a specific wind speed segment, etc.

[0167] S52, threshold comparison and S53, specific implementation logic of verification decision:

[0168] Formation of the preset threshold range: Before model deployment, the sample data labeled as normal and without missing in the training set need to be used to calculate all the above indicator values offline, and set reasonable threshold ranges for each indicator variable through statistical analysis (such as calculating the mean and standard deviation, determining the percentile). These thresholds are stored as verification configuration files.

[0169] Real-time verification process: for each filled sequence X filled，raw output by the model:

[0170] First step, calculate all the pre-defined verification indicator values according to the above method.

[0171] Second step, compare each indicator value with the corresponding preset threshold range in the configuration file one by one.

[0172] Third step, if all the index values are within their respective threshold range, it is determined that "check pass", X filled，raw as the final filling data output to subsequent applications.

[0173] Fourth step, if any index value exceeds its threshold range, it is determined that "check fail". At this time, the system executes the preset exception handling strategy, for example:

[0174] Output check exception alarm: record alarm log and send notification to upstream system, prompting that the filling result this time is low in reliability and needs manual review.

[0175] Call backup filling result: trigger a conservative backup filling strategy, for example, use the historical same period mean of the variable, the value at the previous time point or the simple linear interpolation result as the alternative output, to ensure the basic operation of the system when the model may fail.

[0176] Through the above specific and quantifiable check steps, it is ensured that the model filling result not only meets the data-driven rules, but also satisfies the basic physical constraints and engineering common sense, significantly improving the reliability and safety of the filling system.

[0177] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and not for limiting the present application, although the present application is described in detail with reference to the foregoing embodiments, for those skilled in the art, it still can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. A method for intelligent filling of missing monitoring data of offshore wind turbines based on a multi-head gated residual network, characterized in that, Specifically comprising the following steps: S1, obtaining a historical sequence of offshore wind turbine multi-variable operating parameters containing missing values, and performing standardization preprocessing on the historical sequence to obtain a standardized sequence, so as to construct a model input data set; S2, constructing a multi-head gated residual network model; the model comprises a plurality of feature extraction heads, a gated fusion module and a residual prediction module arranged in parallel; the plurality of feature extraction heads at least comprise a space-time feature extraction head, an operating condition mode extraction head and a missing mode extraction head; The space-time feature extraction head is used for capturing the dependence of wind turbine operating parameters in time and space dimensions; The operating condition mode extraction head is used for learning the steady-state and transient correlation mode between parameters under different wind turbine operating conditions; The missing mode extraction head is used for explicitly modeling the distribution mode of data missing values and its influence on context information; The space-time feature extraction head is composed of a time series convolution network layer and an attention mechanism layer in series; the time series convolution network layer is used for extracting local time features, and the attention mechanism layer is used for modeling global time dependence and spatial correlation between different operating parameter variables; The operating condition mode extraction head inputs the encoded operating condition label vector and the spliced standardized sequence, and extracts a feature representation deeply coupled with the different wind turbine operating conditions through a multi-layer perception; The missing mode extraction head generates a missing mask matrix with the same dimension as the input sequence, and inputs the mask matrix and the standardized sequence into a network layer based on a gated recurrent unit, so as to learn the context features under the missing mode; S3, training the multi-head gated residual network model using the model input data set, optimizing the model parameters until the model converges, and obtaining a trained missing data filling model; When training the multi-head gated residual network model, the loss function L used is the sum of the reconstruction loss and the regularization loss, which is specifically represented as: L = L rec + λ * L reg , wherein L rec is a reconstruction loss term, L reg is a regularization loss term, and λ is a regularization coefficient. S4, inputting the real-time operating parameter sequence of offshore wind turbine containing missing values to be filled into the trained missing data filling model, and outputting a complete wind turbine operating parameter sequence with missing values filled by the model; S5, performing a reasonableness check on the complete wind turbine operating parameter sequence with missing values filled; if the check passes, outputting the complete wind turbine operating parameter sequence with missing values filled as the final filling data; if the check fails, outputting a check exception alarm or calling a backup filling result.

2. The offshore wind turbine monitoring data missing intelligent filling method based on the multi-head gated residual network of claim 1, wherein, The standardization preprocessing is specifically: for each operating parameter variable, calculating the mean and standard deviation based on the non-missing values in its historical data, and processing by Z-Score standardization method; at the same time, a binary missing mask matrix is generated to identify the missing state of each data point in the historical sequence.

3. The offshore wind turbine monitoring data missing intelligent filling method based on the multi-head gated residual network of claim 1, wherein, The gated fusion module is used for dynamically weighting and integrating the outputs of the plurality of feature extraction heads; specifically performing the following steps: A21, receiving output vectors from each feature extraction head; A22. Calculate the adaptive weight vector of each output vector corresponding to the feature extraction head according to the current input sequence context through a learnable gating weight generation network; A23. Weighted sum each output vector and its corresponding adaptive weight vector to generate the fused comprehensive feature vector.

4. The offshore wind turbine monitoring data missing intelligent filling method based on the multi-head gated residual network of claim 3, characterized in that, The residual prediction module takes the comprehensive feature vector as input and performs the following steps: B21. Map the comprehensive feature vector through a fully connected layer to preliminarily generate the filling prediction value of the missing value; B22. Perform residual addition between the filling prediction value and the initial reference sequence to generate the final filling output; B23. Perform inverse normalization processing on the final filling output to restore it to the original physical dimension.

5. The offshore wind turbine monitoring data missing intelligent filling method based on the multi-head gated residual network of claim 4, wherein, The initial reference sequence is generated by inputting the historical sequence obtained in S1 into a linear projection layer, and the output of the linear projection layer is the initial reference sequence.

6. The offshore wind turbine monitoring data missing intelligent filling method based on the multi-head gated residual network of claim 1, wherein, The reasonableness verification specifically includes the following steps: S51. Calculate multiple verification indexes based on physical rules and statistical laws for the complete fan operation parameter sequence with missing values filled; the multiple verification indexes include sequence extreme value reasonableness index, data point change continuity index, and multivariate physical correlation consistency index; S52. Compare the calculated multiple verification indexes with the corresponding preset threshold range based on historical normal data statistics; S53. If all verification indexes are within their corresponding preset threshold range, it is determined that the verification is passed; if any verification index exceeds its corresponding preset threshold range, it is determined that the verification is not passed.

Images