Wind turbine generator health degree analysis method fused with artificial intelligence algorithm

Abstract

The invention relates to the technical field of wind turbine generator monitoring, in particular to a wind turbine generator health degree analysis method fused with an artificial intelligence algorithm. The method comprises the following steps: collecting time sequence data, wake flow data and flow field simulation data of a wind turbine, and carrying out space-time alignment to generate a multi-modal data set; respectively extracting time sequence operation features and space correlation features through a time sequence convolutional network and a graph convolutional network, and fusing the time sequence operation features and the space correlation features into a unified feature vector; inputting a long short-term memory network and Gaussian process regression mixed model to evaluate the health state and predict the remaining service life of the key component; identifying a degradation stage based on a hidden semi-Markov model and generating a graded early warning signal; a preventive maintenance strategy is generated in combination with the deep Q network, and model parameters are dynamically updated through online learning. According to the method, through multi-modal data fusion and multi-algorithm collaborative modeling, spatial-temporal feature association is fully mined, and the comprehensiveness and accuracy of analysis are improved.

Description

Technical Field

[0001] This invention relates to the field of wind turbine monitoring technology, and more specifically, to a method for analyzing the health of wind turbines that integrates artificial intelligence algorithms. Background Technology

[0002] As the core equipment for new energy power generation, wind turbines operate in complex outdoor environments for extended periods. Their critical components (such as gearboxes and main shaft bearings) are susceptible to wind speed fluctuations, wake interference, and mechanical wear, leading to performance degradation and even shutdowns. Existing wind turbine health analysis methods largely rely on single-type data or traditional monitoring models, resulting in incomplete feature extraction and insufficient characterization of spatial correlations and temporal dependencies, making it difficult to meet the low-cost, high-power generation operation and maintenance requirements of wind farms. Therefore, this invention provides a wind turbine health analysis method that integrates artificial intelligence algorithms. Summary of the Invention

[0003] The purpose of this invention is to provide a wind turbine health analysis method that integrates artificial intelligence algorithms, in order to solve the problems of existing wind turbine health analysis methods mentioned in the background art, which mostly rely on single-type data or traditional monitoring models, resulting in incomplete feature extraction and insufficient characterization of spatial correlation and temporal dependence.

[0004] To achieve the above objectives, the present invention aims to provide a method for analyzing the health of wind turbine units by integrating artificial intelligence algorithms, comprising the following steps:

[0005] S1. Collect wind turbine time-series data, wake data, and flow field simulation data; perform spatial location registration and time series alignment to generate a multimodal dataset.

[0006] S2. Temporal convolutional networks are used to extract the operational features of wind turbine time-series data, and graph convolutional networks are used to extract the spatial correlation features of wake data and flow field simulation data. The operational features and spatial correlation features are then fused into a unified feature vector through an attention mechanism.

[0007] S3. Input the unified feature vector into a hybrid model based on long short-term memory network and Gaussian process regression to assess the current health status and predict the remaining service life of key components.

[0008] S4. Based on the health status assessment results, identify the degradation stage of the unit through a hidden semi-Markov model, and generate graded early warning signals according to preset thresholds.

[0009] S5. With the goal of minimizing operation and maintenance costs and maximizing power generation, a preventive maintenance strategy is generated by combining a deep Q network with early warning signals and the remaining service life.

[0010] S6. Based on the effectiveness of preventive maintenance strategies, the parameters of the hybrid model are dynamically updated through an online learning algorithm.

[0011] As a further improvement to this technical solution, in step S1, the specific steps involved in collecting wind turbine timing data, wake data, and flow field simulation data, and performing spatial position registration and time series alignment are as follows:

[0012] Collect wind turbine timing data from the SCADA system of the wind turbine, including wind speed, power, speed, vibration, temperature, and yaw angle;

[0013] The wake data between wind turbines is measured by lidar deployed in the wind farm, including wake velocity distribution, turbulence intensity, and wake deflection angle.

[0014] Flow field simulation data, including velocity field, pressure field, and vorticity field, are obtained through a computational fluid dynamics simulation platform.

[0015] A unified coordinate system for wind farms is established, and the physical positions of each wind turbine, sensor device, and simulation grid node are mapped to this coordinate system. Wake data and flow field simulation data are aligned to the same spatial grid through bilinear interpolation.

[0016] Using the timestamp of the SCADA system as a reference, the timing data of wind turbines, wake data and flow field simulation data are synchronized in time. The time interpolation method is used to unify data with different sampling frequencies to the same time resolution, forming a spatiotemporally aligned multimodal dataset.

[0017] As a further improvement to this technical solution, the specific steps involved in S2, which employs a temporal convolutional network to extract the operational features of wind turbine time-series data, are as follows:

[0018] Extract wind turbine time series data from multimodal datasets, including wind speed, power, generator speed, transmission chain vibration acceleration, bearing temperature, and yaw angle;

[0019] Constructing using causal convolution and dilated convolution A layered temporal convolutional network processes wind turbine time-series data. The input is a set of features at different time scales from each layer of the temporal convolutional network. These features are then concatenated along the channel dimension to form enhanced temporal features. ;

[0020] Enhanced temporal features Max pooling and average pooling are performed along the time dimension and then concatenated to form a global feature vector. This indicates the operational characteristics of the entire wind farm.

[0021] As a further improvement to this technical solution, the specific steps involved in S2, which employs graph convolutional networks to extract the spatial correlation features of wake data and flow field simulation data, are as follows:

[0022] Each turbine in the wind farm is considered a node in the diagram, totaling... Typhoon generators, constructing an undirected graph ,in For a set of nodes, Let it be the set of edges;

[0023] Based on the wind farm layout and prevailing wind direction, the adjacency relationships between nodes are defined, such as if the wind turbines... Located in the wind turbine Within the wake influence range, then establish an edge The adjacency matrix is ​​obtained. ;

[0024] For each wind turbine node The corresponding wake data and flow field simulation data are extracted from the multimodal dataset to form node feature vectors. The feature vectors of all nodes constitute the node feature matrix. The features of the nodal feature vector include wake velocity decay rate, turbulence intensity, wake deflection angle, local pressure gradient, and vorticity intensity.

[0025] Node feature matrix Adjacency Matrix In a two-layer graph convolutional network, each layer aggregates the features of neighboring nodes through an adjacency matrix, and outputs a spatial feature matrix. ;

[0026] Based on spatial feature matrix For each feature dimension, perform max pooling and average pooling on all nodes, and concatenate them into a global feature vector. This indicates the spatial correlation characteristics of the entire wind farm.

[0027] As a further improvement to this technical solution, the specific steps involved in S2, which fuse operational features and spatial correlation features into a unified feature vector through an attention mechanism, are as follows:

[0028] global feature vectors and global space feature vectors Concatenate them into a joint feature vector ;

[0029] Based on joint feature vectors Attention weights are calculated using a fully connected layer and a softmax function. ;

[0030] Based on attention weights For global feature vectors and global space feature vectors We perform weighted summation to obtain the weighted global feature vector. and weighted global space feature vector ;

[0031] The weighted global feature vector and weighted global space feature vector The features are concatenated and then dimensionality reduced and fused using a fully connected layer to obtain a unified feature vector. .

[0032] As a further improvement to this technical solution, the specific steps involved in constructing the hybrid model based on long short-term memory network and Gaussian process regression in step S3 are as follows:

[0033] Construct a two-layer long short-term memory network to unify the feature vector. As input, output the hidden state at the final time step. As an implicit representation of health status;

[0034] Establish a Gaussian process regression model and extract the hidden state at the final time step. As input features, they are used to predict the remaining service life of key components; where the Gaussian process assumes health indicators It follows a probability distribution defined by the mean function and the covariance function;

[0035] The hybrid model based on Long Short-Term Memory Network and Gaussian Process Regression is trained using historical running data to minimize the loss function, resulting in a well-trained hybrid model. The loss function consists of the mean squared error loss of the Long Short-Term Memory Network and the negative log-likelihood loss of the Gaussian Process.

[0036] As a further improvement to this technical solution, the specific steps involved in S3 for assessing the current health status and predicting the remaining service life of critical components are as follows:

[0037] The unified feature vector at the current time Input the trained hybrid model and output the hidden layer features through the Long Short-Term Memory network. The health score was obtained through linear transformation. ;

[0038] Hidden layer features Input a Gaussian process regression model to predict the mean value for each key component. and variance The probability distribution; among which, key components include the gearbox and spindle bearings;

[0039] Set the failure threshold as Iteratively predicts health indicators at future time steps until the mean is reached. Exceeding the failure threshold The elapsed time is the predicted remaining useful life. .

[0040] As a further improvement to this technical solution, the specific steps involved in S4, which use a hidden semi-Markov model to identify the degradation stage of the unit and generate graded early warning signals based on preset thresholds, are as follows:

[0041] Based on health scores, five latent states of unit degradation are defined, including healthy, minor degradation, moderate degradation, severe degradation, and impending failure.

[0042] Each wind turbine unit will be in time step Health score and remaining lifespan constitute the observation vector The observation sequence was obtained by arranging the observations in chronological order. ;

[0043] Assuming the duration of each degradation state follows a Gaussian distribution, the parameters of a hidden semi-Markov model are trained using historical data, and maximum likelihood estimation is used to optimize the observed sequence. The likelihood value is the largest; among them, the parameters of the hidden semi-Markov model include the distribution of the duration of each state, the state transition probability matrix, and the observation probability distribution;

[0044] The Viterbi algorithm is used to calculate the decoding state at each time step, resulting in the hidden state sequence. ;

[0045] Based on the current decoding state Generate a tiered early warning signal; if the decoded status is healthy... This indicates that the unit is operating normally and no warning needs to be issued.

[0046] If the decoding status is slightly degraded The system issued a blue alert, prompting operators to pay attention to the unit's trend and arrange routine inspections;

[0047] If the decoding status is moderately degraded The system issued a yellow alert, recommending preventative maintenance during the planned downtime.

[0048] If the decoding status is severely degraded The system issued an orange alert, requiring maintenance to be arranged as soon as possible to avoid unplanned downtime;

[0049] If the decoding status is about to expire The system issued a red alert and immediately shut down for maintenance to prevent malfunctions.

[0050] As a further improvement to this technical solution, in step S5, with the goal of minimizing operation and maintenance costs and maximizing power generation, a deep Q-network is used in conjunction with early warning signals and the remaining service life to generate a preventative maintenance strategy. The specific steps involved are as follows:

[0051] Using wind turbines and their operating environment as a reinforcement learning environment, a state vector is constructed that includes health score, remaining service life, graded early warning signals, current wind speed and real-time electricity price.

[0052] Four maintenance actions are set, including no maintenance, minor maintenance, medium maintenance, and major maintenance;

[0053] Based on state vectors and actions, a reward function is set that includes power generation revenue, maintenance costs, and fault penalties;

[0054] The state vector Input a deep Q-network, train it using historical data to obtain network parameters, the objective function is to minimize the temporal difference error, and the output is for each action. Values ​​are used to select the optimal action. The temporal difference error is calculated using the reward function.

[0055] Based on optimal action If the action is "no maintenance", continue monitoring; no operation is performed.

[0056] Otherwise, generate and output a preventative maintenance strategy that includes suggested maintenance time, maintenance type and content, estimated cost, and expected benefits.

[0057] As a further improvement to this technical solution, the specific steps involved in S6, which involve dynamically updating the parameters of the hybrid model through an online learning algorithm, are as follows:

[0058] Obtain recent wind turbine operating data, including wind speed, power, vibration, and temperature, and calculate a new unified feature vector. ;

[0059] The new unified feature vector Input the current hybrid model to obtain the predicted health score. ;

[0060] Calculate the true health score based on actual vibration and temperature data. Calculate its correlation with the predicted health score absolute error between ;

[0061] If absolute error If the error exceeds the threshold, an update is triggered.

[0062] Maintain a recent Each time a data window is updated, the latest set of data is added, the oldest set of data is removed, and the data in the window is used to perform a gradient descent update on the long short-term memory network in the hybrid model.

[0063] The length scale in Gaussian process regression is updated using exponentially weighted average smoothing. With signal variance ;

[0064] Record the update time, absolute error, and parameter adjustment amount. After updating the hybrid model, use the most recent data to verify it. If the absolute error decreases, the update is confirmed to be effective; otherwise, revert to the parameters of the previous version.

[0065] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0066] 1. In this wind turbine health analysis method that integrates artificial intelligence algorithms, multimodal data fusion and multi-algorithm collaborative modeling are used. It not only uses temporal convolutional networks to capture the long-term dependent features of turbine operation, but also uses graph convolutional networks to extract the spatial correlation features of wake and flow field. Combined with attention mechanism, it achieves accurate fusion of spatiotemporal features. Compared with the traditional analysis method of single data type and single algorithm, it solves the problems of one-sided health status assessment and high uncertainty in the prediction of the remaining life of key components, and greatly improves the comprehensiveness and accuracy of health analysis.

[0067] 2. In this wind turbine health analysis method that integrates artificial intelligence algorithms, the method achieves accurate identification and graded early warning of the unit degradation stage based on the hidden semi-Markov model. It combines a deep Q-network to generate a preventive maintenance strategy with the goal of minimizing operation and maintenance costs and maximizing power generation. The model parameters are dynamically updated through online learning. Compared with the traditional fixed threshold early warning and experience-based maintenance mode, this method solves the problems of delayed early warning, unreasonable maintenance timing leading to more unplanned downtime and high operation and maintenance costs. It realizes the optimization of the entire process from condition monitoring to intelligent operation and maintenance. Attached Figure Description

[0068] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0069] 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, and 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.

[0070] Example:

[0071] Please see Figure 1 As shown in the figure, this embodiment provides a method for analyzing the health of wind turbine units by integrating artificial intelligence algorithms, including the following steps:

[0072] S1. Collect wind turbine time-series data, wake data, and flow field simulation data; perform spatial location registration and time series alignment to generate a multimodal dataset.

[0073] In this embodiment, wind turbine timing data, including wind speed, power, speed, vibration, temperature, and yaw angle, are collected from the SCADA system of the wind turbine.

[0074] The wake data between wind turbines is measured by lidar deployed in the wind farm, including wake velocity distribution, turbulence intensity, and wake deflection angle.

[0075] Flow field simulation data, including velocity field, pressure field, and vorticity field, are obtained through a computational fluid dynamics simulation platform.

[0076] A unified coordinate system for wind farms is established, and the physical positions of each wind turbine, sensor device, and simulation grid node are mapped to this coordinate system. Wake data and flow field simulation data are aligned to the same spatial grid through bilinear interpolation.

[0077] Bilinear interpolation is a spatial interpolation method that estimates the value of a point based on the values ​​of four adjacent known points by performing two linear interpolations (first in the horizontal direction and then in the vertical direction).

[0078] Using the timestamp of the SCADA system as a reference, the timing data of wind turbines, wake data and flow field simulation data are synchronized in time. The time interpolation method is used to unify data with different sampling frequencies to the same time resolution, forming a spatiotemporally aligned multimodal dataset.

[0079] S2. Temporal convolutional networks are used to extract the operational features of wind turbine time-series data, and graph convolutional networks are used to extract the spatial correlation features of wake data and flow field simulation data. The operational features and spatial correlation features are then fused into a unified feature vector through an attention mechanism.

[0080] In this embodiment, wind turbine time series data is extracted from the multimodal dataset, including wind speed, power, generator speed, transmission chain vibration acceleration, bearing temperature, and yaw angle.

[0081] The number of sampling points for each wind turbine is set to be... , contains Several time-series variables; in practical applications, one can select... (That is, the past 24 hours, with a sampling interval of 1 minute) as the time window, Corresponding to the above six core operating parameters;

[0082] Constructing using causal convolution and dilated convolution Layers (e.g.) A temporal convolutional network is used to process the time-series data of wind turbines. The input is a set of features at different time scales from each layer of the temporal convolutional network. These features are then concatenated along the channel dimension to form enhanced temporal features. ;

[0083] Among them, causal convolution ensures the unidirectional dependency of time series modeling by constraining the convolution kernel to access only current and historical data, preventing the leakage of future information; dilated convolution inserts intervals into the standard convolution kernel, causing the receptive field to expand exponentially with the number of network layers, thereby capturing long-range temporal dependencies with fewer parameters; the calculation formula is as follows:

[0084]

[0085] in, For a moment The convolutional output features; For a moment The input vector; For the convolution kernel in the th Step weights For the kernel size (e.g.) ), The number of feature channels in the output after convolution (e.g.) The dilation factor of dilated convolution. It grows exponentially with the number of layers, that is, the first layer... Layer expansion factor ; For bias terms; ReLU As an activation function, it enhances the nonlinear expressive power;

[0086] The temporal convolutional network progressively increases the dimensionality of its output features at each layer (e.g., 32→64→128→256), adds residual connections after each layer to avoid the vanishing gradient problem, and finally extracts the feature vectors for each time step from the last layer. , The last layer has a feature dimension (e.g., 256). To fuse multi-scale temporal information, the features output from each layer are concatenated along the channel dimension to obtain enhanced temporal features. , For the first The feature dimensions of the layer ;

[0087] Enhanced temporal features Max pooling and average pooling are performed along the time dimension and then concatenated to form a global feature vector. This indicates the operational characteristics of the entire wind farm;

[0088] The formula is as follows:

[0089]

[0090] in, This indicates that the maximum value is taken along the time axis; This indicates that the average value is taken along the time axis; This represents a vector concatenation operation; The global feature vector is formed by compressing a variable-length time series into a fixed-length feature vector, which facilitates subsequent fusion with spatial features.

[0091] In this embodiment, each turbine in the wind farm is considered as a node in the diagram, totaling... Typhoon generators, constructing an undirected graph ,in For a set of nodes, Let it be the set of edges;

[0092] Based on the wind farm layout and prevailing wind direction (e.g., southerly or northerly), define the adjacency relationships between nodes, if the wind turbines Located in the wind turbine Within the wake influence range, then establish an edge The adjacency matrix is ​​obtained. ;

[0093] Adjacency matrix The definition is as follows:

[0094]

[0095] in, The adjacency matrix is ​​the first line, number Column elements, It is an adjacency matrix;

[0096] For each wind turbine node The corresponding wake data and flow field simulation data are extracted from the multimodal dataset to form node feature vectors. The feature vectors of all nodes constitute the node feature matrix. The features of the nodal feature vector include wake velocity decay rate, turbulence intensity, wake deflection angle, local pressure gradient, and vorticity intensity.

[0097] Node feature matrix Adjacency Matrix In a two-layer graph convolutional network, each layer aggregates the features of neighboring nodes through an adjacency matrix, and outputs a spatial feature matrix. ;

[0098] The formula for each layer of the graph convolutional network is as follows:

[0099]

[0100] in, To add self-connected adjacency matrices, It is the identity matrix; for The degree matrix, ; For the first The node feature matrix of a layered graph convolutional network. ; For the first The learnable weight matrix of a layered graph convolutional network; The ReLU activation function is used; the output dimension of the first layer of the graph convolutional network is set to 32, and the output dimension of the second layer of the graph convolutional network is set to 64, finally yielding the spatial feature matrix. ;

[0101] Based on spatial feature matrix For each feature dimension, perform max pooling and average pooling on all nodes, and concatenate them into a global feature vector. This indicates the spatial correlation characteristics of the entire wind farm;

[0102]

[0103] in, This represents the global spatial feature vector.

[0104] In this embodiment, the global feature vector and global space feature vectors Concatenate them into a joint feature vector ;

[0105]

[0106] in, Joint eigenvectors;

[0107] Based on joint feature vectors Attention weights are calculated using a fully connected layer and a softmax function. ;

[0108]

[0109] in, The weights of the fully connected layer, For biasing the fully connected layer; For attention weights, Represents the attention weights for temporal features. This represents the attention weights for spatial features, with the weight values ​​reflecting the importance of each feature in the fusion process; It is the softmax function;

[0110] Based on attention weights For global feature vectors and global space feature vectors We perform weighted summation to obtain the weighted global feature vector. and weighted global space feature vector ;

[0111]

[0112]

[0113] in, This is the weighted global feature vector. This is the weighted global space feature vector;

[0114] The weighted global feature vector and weighted global space feature vector The features are concatenated and then dimensionality reduced and fused using a fully connected layer to obtain a unified feature vector. ;

[0115]

[0116] in, For the fusion weights of the fully connected layer, For the fusion bias of the fully connected layer; To unify the feature vector, which simultaneously contains information on the wind turbine's operating status and the spatial environment, for subsequent health assessment and prediction; This is the ReLU activation function.

[0117] S3. Input the unified feature vector into a hybrid model based on long short-term memory network and Gaussian process regression to assess the current health status and predict the remaining service life of key components.

[0118] In this embodiment, a two-layer long short-term memory network is constructed to unify the feature vector. As input, output the hidden state at the final time step. As an implicit representation of health status;

[0119] The Long Short-Term Memory (LSTM) network has 128 memory units per layer. It processes the input vector sequentially, updating and passing the hidden state step-by-step. Finally, it extracts the hidden state at the last time step as the temporal feature representation of the healthy state, as shown in the following formula:

[0120]

[0121]

[0122] in, It is a long short-term memory network; express The hidden state at all times express The hidden state at any given moment; For the final time step; The hidden state at the final time step is used for subsequent health assessments;

[0123] Establish a Gaussian process regression model and extract the hidden state at the final time step. As input features, they are used to predict the remaining service life of key components (such as gearboxes and spindle bearings); where the Gaussian process assumes health indicators It follows a probability distribution defined by the mean function and the covariance function:

[0124]

[0125] in, For input feature samples, Indicates and The other input feature sample of the same dimension is all from , used to calculate covariance; As a health indicator; It is a mean function, and its value is set to 0; This represents a Gaussian process regression model; For the covariance function, the Matern 5 / 2 kernel function is selected, and its expression is:

[0126]

[0127] in, Indicates the Euclidean distance between features; The length scale parameter controls the smoothness of the function and is obtained through training on historical running data; The signal variance parameter controls the magnitude of output variation and is obtained through training with historical operating data; the Matern 5 / 2 kernel function can calculate similarity based on the distance between input features, thereby characterizing the smoothness and uncertainty of health status changes;

[0128] The hybrid model based on Long Short-Term Memory (LSTM) network and Gaussian process regression is trained using historical operating data (including samples from normal and degraded states) to minimize the loss function, resulting in a well-trained hybrid model. The loss function consists of the mean squared error loss from the LTM network and the negative log-likelihood loss from the Gaussian process.

[0129]

[0130] in, For the first The true health indicators of the operational data (such as vibration amplitude and temperature deviation). The first output of the Long Short-Term Memory Network Health status estimate of each running data point. For the marginal likelihood of a Gaussian process, The feature matrix of the training set; The value of the loss function. This is the mean square error loss term. This is the negative log-likelihood loss term.

[0131] In this embodiment, the unified feature vector at the current moment is... Input the trained hybrid model and output the hidden layer features through the Long Short-Term Memory network. The health score was obtained through linear transformation. ;

[0132] The scoring formula is as follows:

[0133]

[0134] in, The intermediate variable is obtained by linear transformation of the output of the Long Short-Term Memory network. These are the weighting parameters for the linear transformation. The bias parameters of this linear transformation are mapped to probability values ​​through the Sigmoid function; The health score is given, with 1 indicating perfect health and 0 indicating serious malfunction.

[0135] Hidden layer features Input a Gaussian process regression model to predict the mean value for each key component. and variance The probability distribution; among which, key components include the gearbox and spindle bearings;

[0136] Hidden layer features output by the Long Short-Term Memory network Inputting the data into a Gaussian process regression model yields the predicted distribution:

[0137]

[0138] in:

[0139]

[0140]

[0141] here, It is a test input With training input set The covariance vector between them It is the covariance matrix between the training input sets. It is the autocovariance of the test input; To predict the mean of the distribution, we represent the mean of the distribution. Most likely predicted value; The variance of the predicted distribution represents the degree of uncertainty in the prediction. For noise variance; The identity matrix has dimensions and covariance matrix. Consistent; Indicates that in the known training data (Input features) and Under the condition of (corresponding output), for a new test input Its output The predicted distribution; for gearboxes, vibration energy entropy is used as the input feature; for spindle bearings, the ratio of temperature gradient to peak value of vibration spectrum is used as the input feature;

[0142] Set the failure threshold as (e.g., gearbox vibration energy entropy > 0.8), predict health indicators for future time steps iteratively until the mean is reached. Exceeding the failure threshold The elapsed time is the predicted remaining useful life. ;

[0143] Simultaneously, output the 95% confidence interval:

[0144]

[0145] in, The failure time point corresponding to the mean prediction. This is the lower limit of the failure time point. This represents the upper limit of the failure time point. To predict the standard deviation of the distribution, This represents the remaining service life.

[0146] S4. Based on the health status assessment results, identify the degradation stage of the unit through a hidden semi-Markov model, and generate graded early warning signals according to preset thresholds.

[0147] In this embodiment, based on the health score, five hidden states of unit degradation are defined, including healthy, slightly degraded, moderately degraded, severely degraded, and about to fail.

[0148] For example:

[0149] For health, It is a slight degeneration. Moderate degradation, It is severely degraded; It is about to expire;

[0150] Each wind turbine unit will be in time step Health score and remaining lifespan constitute the observation vector The observation sequence was obtained by arranging the observations in chronological order. ;

[0151]

[0152] in, For a moment Health score, For a moment The remaining lifespan; For a moment The observation sequence; Indicates a time step (e.g., one observation point per day or week);

[0153]

[0154] in, For observation sequence;

[0155] Assuming the duration of each degradation state follows a Gaussian distribution, the parameters of a hidden semi-Markov model are trained using historical data, and maximum likelihood estimation is used to optimize the observed sequence. The likelihood value is the largest; among them, the parameters of the hidden semi-Markov model include the distribution of the duration of each state, the state transition probability matrix, and the observation probability distribution;

[0156] Specifically, the set of hidden states ;

[0157] Assuming each state Duration Follows a Gaussian distribution:

[0158]

[0159] in, For state Average duration of days, For state Standard deviation; For state The duration; Indicates a Gaussian distribution;

[0160] State transition probability matrix ,in Indicates from state Transferred to The probability (only allowing transitions to worse states, i.e.) );

[0161] Observational probability distribution Indicates the state The following observations The probability is assumed to follow a Gaussian distribution:

[0162]

[0163] in, For state The mean of the observed vectors, For state The covariance matrix of the observed vectors;

[0164] The Viterbi algorithm is used to calculate the decoding state at each time step, resulting in the hidden state sequence. ;

[0165]

[0166] in, It is a sequence of hidden states. For a moment The decoding state; the Viterbi algorithm is a dynamic programming algorithm used to find the most likely sequence of hidden states in a hidden Markov model, that is, the state path with the highest probability given the observation sequence;

[0167] Based on the current decoding state Generate a tiered early warning signal; if the decoded status is healthy... This indicates that the unit is operating normally and no warning needs to be issued.

[0168] If the decoding status is slightly degraded The system issued a blue alert, prompting operators to pay attention to the unit's trend and arrange routine inspections;

[0169] If the decoding status is moderately degraded The system issued a yellow alert, recommending preventative maintenance during the planned downtime.

[0170] If the decoding status is severely degraded The system issued an orange alert, requiring maintenance to be arranged as soon as possible to avoid unplanned downtime;

[0171] If the decoding status is about to expire The system issued a red alert and immediately shut down for maintenance to prevent malfunctions.

[0172] S5. With the goal of minimizing operation and maintenance costs and maximizing power generation, a preventive maintenance strategy is generated by combining a deep Q network with early warning signals and the remaining service life.

[0173] In this embodiment, the wind turbine and its operating environment are used as the reinforcement learning environment to construct a state vector that includes health score, remaining service life, graded early warning signals, current wind speed and real-time electricity price.

[0174]

[0175] in, The levels of the warning signals are: no warning, blue, yellow, orange, and red. Current wind speed (unit: m / s), sourced from SCADA system; The real-time electricity price (unit: yuan / kWh) is derived from the grid's price quote. It is a state vector;

[0176] Four maintenance actions are defined, including no maintenance, minor maintenance (such as lubrication and tightening), medium maintenance (such as component replacement), and major maintenance (such as overall gearbox overhaul).

[0177]

[0178] in, For action;

[0179] Based on state vectors and actions, a reward function is set that includes power generation revenue, maintenance costs, and fault penalties;

[0180]

[0181] in, ; The efficiency coefficient of a wind turbine (e.g., 0.45) reflects the efficiency of converting wind speed into electrical energy. For decision time intervals (e.g., 1 hour); The cost is set according to the type of operation (e.g., minor maintenance = 1000 yuan, medium maintenance = 5000 yuan, major maintenance = 20000 yuan). Health score Triggered at a certain time, set to a high cost (e.g., 50,000 yuan); For power generation revenue weighting coefficient, To maintain the cost weighting coefficient, The fault penalty weighting coefficient is used to balance objectives, such as... =1, =0.1, =10;

[0182] The state vector Input a deep Q-network, train it using historical data to obtain network parameters, the objective function is to minimize the temporal difference error, and the output is for each action. Values ​​are used to select the optimal action. The temporal difference error is calculated using the reward function.

[0183]

[0184] in, To select the appropriate action from all available options (such as no maintenance, minor maintenance, etc.) The action with the highest value; For each action value; The network parameters are obtained through training with historical data; Deep Q-network is a reinforcement learning algorithm that approximates the optimal action value function through deep neural networks and utilizes experience replay and target network stability training.

[0185] Based on optimal action If the action is "no maintenance", continue monitoring; no operation is performed.

[0186] Otherwise (for minor / medium / major maintenance), generate and output a preventative maintenance strategy that includes suggested maintenance time, maintenance type and content, estimated cost, and expected benefits;

[0187] Among them, it is recommended that the maintenance time be set in conjunction with wind speed forecasts and the off-peak electricity price period;

[0188] Maintenance types and contents are specified down to the components and operations;

[0189] Estimated costs are as follows: minor maintenance = 1,000 yuan, medium maintenance = 5,000 yuan, major maintenance = 20,000 yuan;

[0190] The expected return is the anticipated losses from the failures that are to be avoided and the restored power generation.

[0191] S6. Based on the effectiveness of preventive maintenance strategies, the parameters of the hybrid model are dynamically updated through an online learning algorithm;

[0192] In this embodiment, wind turbine operation data for a recent period (e.g., the past 24 hours) is obtained from the SCADA system, including wind speed, power, vibration, and temperature, and a new unified feature vector is calculated. (The method is the same as steps S2-S3);

[0193] The new unified feature vector Input the current hybrid model to obtain the predicted health score. ;

[0194] Calculate the true health score based on actual vibration and temperature data. Calculate its correlation with the predicted health score absolute error between ;

[0195]

[0196] in, This is the absolute error; For a true health score, To predict health scores;

[0197] If absolute error If the error exceeds the threshold (e.g., 0.15), an update is triggered.

[0198] Maintain a recent Group data window (e.g.) (Groups, each containing 24 hours of data), with the latest group of data (including uniform feature vectors) added each time an update is performed. and real health score Remove the oldest set of data, and use this window of data to perform a gradient descent update on the Long Short-Term Memory network in the hybrid model, with the learning rate set to [value missing]. ;

[0199] The update formula is expressed as:

[0200]

[0201] in, To update the weight parameters of the previous Long Short-Term Memory network; To update the weight parameters of the Long Short-Term Memory network;

[0202] The length scale in Gaussian process regression is updated using exponentially weighted average smoothing. With signal variance ;

[0203]

[0204]

[0205] in, The length scale before the update. For the updated length scale, The length scale is re-estimated from the latest sliding window data (such as the most recent 100 sets of data); The signal variance before the update. For the updated signal variance, The signal variance is re-estimated from the latest sliding window data;

[0206] Record the update time, absolute error, and parameter adjustment amount. After updating the hybrid model, use the data from the last 3 days to verify. If the absolute error decreases, the update is confirmed to be effective; otherwise, revert to the parameters of the previous version.

[0207] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing the health of wind turbine units by integrating artificial intelligence algorithms, characterized in that: Includes the following steps: S1. Collect wind turbine time-series data, wake data, and flow field simulation data; perform spatial location registration and time series alignment to generate a multimodal dataset. S2. Temporal convolutional networks are used to extract the operational features of wind turbine time-series data, and graph convolutional networks are used to extract the spatial correlation features of wake data and flow field simulation data. The operational features and spatial correlation features are then fused into a unified feature vector through an attention mechanism. S3. Input the unified feature vector into a hybrid model based on long short-term memory network and Gaussian process regression to assess the current health status and predict the remaining service life of key components. S4. Based on the health status assessment results, identify the degradation stage of the unit through a hidden semi-Markov model, and generate graded early warning signals according to preset thresholds. S5. With the goal of minimizing operation and maintenance costs and maximizing power generation, a preventive maintenance strategy is generated by combining a deep Q network with early warning signals and the remaining service life. S6. Based on the effectiveness of preventive maintenance strategies, the parameters of the hybrid model are dynamically updated through an online learning algorithm.

2. The method for analyzing the health of wind turbine units by integrating artificial intelligence algorithms according to claim 1, characterized in that: In step S1, the specific steps involved in collecting wind turbine time-series data, wake data, and flow field simulation data, and performing spatial location registration and time-series alignment are as follows: Collect wind turbine timing data from the SCADA system of the wind turbine, including wind speed, power, speed, vibration, temperature, and yaw angle; The wake data between wind turbines is measured by lidar deployed in the wind farm, including wake velocity distribution, turbulence intensity, and wake deflection angle. Flow field simulation data, including velocity field, pressure field, and vorticity field, are obtained through a computational fluid dynamics simulation platform. A unified coordinate system for wind farms is established, and the physical positions of each wind turbine, sensor device, and simulation grid node are mapped to this coordinate system. Wake data and flow field simulation data are aligned to the same spatial grid through bilinear interpolation. Using the timestamp of the SCADA system as a reference, the timing data of wind turbines, wake data and flow field simulation data are synchronized in time. The time interpolation method is used to unify data with different sampling frequencies to the same time resolution, forming a spatiotemporally aligned multimodal dataset.

3. The method for analyzing the health of wind turbine units by integrating artificial intelligence algorithms according to claim 1, characterized in that: In step S2, the specific steps involved in extracting the operational features of wind turbine time-series data using a temporal convolutional network are as follows: Extract wind turbine time series data from multimodal datasets, including wind speed, power, generator speed, transmission chain vibration acceleration, bearing temperature, and yaw angle; Constructing using causal convolution and dilated convolution A layered temporal convolutional network processes wind turbine time-series data. The input is a set of features at different time scales from each layer of the temporal convolutional network. These features are then concatenated along the channel dimension to form enhanced temporal features. ; Enhanced temporal features Max pooling and average pooling are performed along the time dimension and then concatenated to form a global feature vector. This indicates the operational characteristics of the entire wind farm.

4. The wind turbine health analysis method integrating artificial intelligence algorithms according to claim 3, characterized in that: In step S2, the specific steps involved in extracting the spatial correlation features of wake data and flow field simulation data using graph convolutional networks are as follows: Each turbine in the wind farm is considered a node in the diagram, totaling... Typhoon generators, constructing an undirected graph ,in For a set of nodes, Let it be the set of edges; Based on the wind farm layout and prevailing wind direction, the adjacency relationships between nodes are defined, such as if the wind turbines... Located in the wind turbine Within the wake influence range, then establish an edge The adjacency matrix is ​​obtained. ; For each wind turbine node The corresponding wake data and flow field simulation data are extracted from the multimodal dataset to form node feature vectors. The feature vectors of all nodes constitute the node feature matrix. The features of the nodal feature vector include wake velocity decay rate, turbulence intensity, wake deflection angle, local pressure gradient, and vorticity intensity. Node feature matrix Adjacency Matrix In a two-layer graph convolutional network, each layer aggregates the features of neighboring nodes through an adjacency matrix, and outputs a spatial feature matrix. ; Based on spatial feature matrix For each feature dimension, perform max pooling and average pooling on all nodes, and concatenate them into a global feature vector. This indicates the spatial correlation characteristics of the entire wind farm.

5. The wind turbine health analysis method integrating artificial intelligence algorithms according to claim 4, characterized in that: In step S2, the specific steps involved in fusing operational features and spatial correlation features into a unified feature vector through an attention mechanism are as follows: global feature vectors and global space feature vectors Concatenate them into a joint feature vector ; Based on joint feature vectors Attention weights are calculated using a fully connected layer and a softmax function. ; Based on attention weights For global feature vectors and global space feature vectors We perform weighted summation to obtain the weighted global feature vector. and weighted global space feature vector ; The weighted global feature vector and weighted global space feature vector The features are concatenated and then dimensionality reduced and fused using a fully connected layer to obtain a unified feature vector. .

6. The method for analyzing the health of wind turbine units by integrating artificial intelligence algorithms according to claim 1, characterized in that: In S3, the specific steps involved in constructing the hybrid model based on long short-term memory network and Gaussian process regression are as follows: Construct a two-layer long short-term memory network to unify the feature vector. As input, output the hidden state at the final time step. As an implicit representation of health status; Establish a Gaussian process regression model and extract the hidden state at the final time step. As input features, they are used to predict the remaining service life of key components; where the Gaussian process assumes health indicators It follows a probability distribution defined by the mean function and the covariance function; The hybrid model based on Long Short-Term Memory Network and Gaussian Process Regression is trained using historical running data to minimize the loss function, resulting in a well-trained hybrid model. The loss function consists of the mean squared error loss of the Long Short-Term Memory Network and the negative log-likelihood loss of the Gaussian Process.

7. The method for analyzing the health of wind turbine units by integrating artificial intelligence algorithms according to claim 6, characterized in that: In step S3, the specific steps involved in assessing the current health status and predicting the remaining service life of critical components are as follows: The unified feature vector at the current time Input the trained hybrid model and output the hidden layer features through the Long Short-Term Memory network. The health score was obtained through linear transformation. ; Hidden layer features Input a Gaussian process regression model to predict the mean value for each key component. and variance The probability distribution; among which, key components include the gearbox and spindle bearings; Set the failure threshold as Iteratively predicts health indicators at future time steps until the mean is reached. Exceeding the failure threshold The elapsed time is the predicted remaining useful life. .

8. The method for analyzing the health of wind turbine units by integrating artificial intelligence algorithms according to claim 1, characterized in that: In step S4, the specific steps involved in identifying the degradation stage of the unit using a hidden semi-Markov model and generating graded early warning signals based on preset thresholds are as follows: Based on health scores, five latent states of unit degradation are defined, including healthy, minor degradation, moderate degradation, severe degradation, and impending failure. Each wind turbine unit will be in time step Health score and remaining lifespan constitute the observation vector The observation sequence was obtained by arranging the observations in chronological order. ; Assuming the duration of each degradation state follows a Gaussian distribution, the parameters of a hidden semi-Markov model are trained using historical data, and maximum likelihood estimation is used to optimize the observed sequence. The likelihood value is the largest; among them, the parameters of the hidden semi-Markov model include the distribution of the duration of each state, the state transition probability matrix, and the observation probability distribution; The Viterbi algorithm is used to calculate the decoding state at each time step, resulting in the hidden state sequence. ; Based on the current decoding state Generate a tiered early warning signal; if the decoded status is healthy... This indicates that the unit is operating normally and no warning needs to be issued. If the decoding status is slightly degraded The system issued a blue alert, reminding operators to pay attention to the unit's trend and arrange routine inspections; If the decoding status is moderately degraded The system issued a yellow alert, recommending preventative maintenance during the planned downtime. If the decoding status is severely degraded The system issued an orange alert, requiring maintenance to be arranged as soon as possible to avoid unplanned downtime; If the decoding status is about to expire The system issued a red alert and immediately shut down for maintenance to prevent malfunctions.

9. The method for analyzing the health of wind turbine units by integrating artificial intelligence algorithms according to claim 1, characterized in that: In step S5, with the goal of minimizing operation and maintenance costs and maximizing power generation, the specific steps involved in generating a preventative maintenance strategy using a deep Q-network combined with early warning signals and the remaining service life are as follows: Using wind turbines and their operating environment as a reinforcement learning environment, a state vector is constructed that includes health score, remaining service life, graded early warning signals, current wind speed and real-time electricity price. Four maintenance actions are set, including no maintenance, minor maintenance, medium maintenance, and major maintenance; Based on state vectors and actions, a reward function is set that includes power generation revenue, maintenance costs, and fault penalties; The state vector Input a deep Q-network, train it using historical data to obtain network parameters, the objective function is to minimize the temporal difference error, and the output is for each action. Values ​​are used to select the optimal action. The temporal difference error is calculated using the reward function. Based on optimal action If the action is "no maintenance", continue monitoring; no operation is performed. Otherwise, generate and output a preventative maintenance strategy that includes suggested maintenance time, maintenance type and content, estimated cost, and expected benefits.

10. The method for analyzing the health of wind turbine units by integrating artificial intelligence algorithms according to claim 1, characterized in that: In step S6, the specific steps involved in dynamically updating the parameters of the hybrid model through an online learning algorithm are as follows: Obtain recent wind turbine operating data, including wind speed, power, vibration, and temperature, and calculate a new unified feature vector. ; The new unified feature vector Input the current hybrid model to obtain the predicted health score. ; Calculate the true health score based on actual vibration and temperature data. Calculate its correlation with the predicted health score absolute error between ; If absolute error If the error exceeds the threshold, an update is triggered. Maintain a recent Each time a data window is updated, the latest set of data is added, the oldest set of data is removed, and the data in the window is used to perform a gradient descent update on the long short-term memory network in the hybrid model. The length scale in Gaussian process regression is updated using exponentially weighted average smoothing. With signal variance ; Record the update time, absolute error, and parameter adjustment amount. After updating the hybrid model, use the most recent data to verify it. If the absolute error decreases, the update is confirmed to be effective; otherwise, revert to the parameters of the previous version.

Images