Wind power intelligent operation and maintenance management method and system based on deep learning

Abstract

The invention discloses a wind power intelligent operation and maintenance management method and system based on deep learning. The method comprises the following steps: acquiring a graph structure data set; based on the graph structure data set, modeling a space-time dependency relationship in the fan group and between the fan and the environment by using a space-time graph neural network combined with a multi-head attention mechanism and a GCN graph convolutional network to obtain space-time features; performing feature alignment and fusion through a cross-modal attention mechanism to obtain fusion features; inputting the fusion features into a GRU gating circulation network, outputting residual effective life probability distribution of the wind turbine generator components and dynamic risk levels of multiple types of faults in future 72 hours, and obtaining a prediction result; and based on a deep reinforcement learning model, taking the prediction result as input, taking multi-objective optimization as rewards, and generating a dynamic maintenance decision scheme. The operation and maintenance cost is reduced, the unit availability rate is improved, and a large-scale wind power cluster operation and maintenance scene is adapted.

Description

Technical Field

[0001] This invention relates to the field of wind power equipment operation and maintenance management technology, and in particular to a wind power intelligent operation and maintenance management method and system based on deep learning. Background Technology

[0002] With the large-scale development of the wind power industry, wind turbines are mostly deployed in remote and complex environments. Their operating status is significantly affected by environmental factors such as wind speed and temperature. Failures of key components can easily lead to downtime losses and safety risks, greatly increasing the difficulty of operation and maintenance management. Traditional wind power operation and maintenance mainly relies on periodic inspections and reactive repairs, which are prone to over-maintenance or under-maintenance, increasing operation and maintenance costs. Reactive repairs are slow to respond, and the escalation of faults can easily trigger cascading damage. Existing intelligent operation and maintenance technologies mostly focus on single types of data and do not fully explore the spatiotemporal correlation and complementary value of multimodal data in wind turbine clusters, resulting in low fault prediction accuracy and weak early warning capabilities. Maintenance decisions rely heavily on manual judgment, making it impossible to achieve dynamic optimization based on real-time status and difficult to adapt to the refined operation and maintenance needs of large-scale wind power clusters. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a wind power intelligent operation and maintenance management method and system based on deep learning.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned deep learning-based intelligent operation and maintenance management method for wind power: The operation data of wind turbine units is collected, and after preprocessing the operation, a graph structure dataset is constructed with wind turbines as nodes and spatiotemporal associations as edges. Based on the graph structure dataset, a spatiotemporal graph neural network combining multi-head attention mechanism and GCN graph convolutional network is used to model the spatiotemporal dependencies within the wind turbine group and between the wind turbines and the environment, thereby obtaining spatiotemporal features; We use 1DCNN to extract deep temporal features from graph structure datasets, use an improved lightweight visual Transformer to extract image defect features, and use a cross-modal attention mechanism to align and fuse features to obtain fused features. The fused features are input into a GRU gated recurrent network, which outputs the probability distribution of the remaining effective life of wind turbine components and the dynamic risk level of multiple types of faults in the next 72 hours, thus obtaining the prediction results. Based on a deep reinforcement learning model, using the prediction results as input and multi-objective optimization as reward, a dynamic maintenance decision scheme is generated.

[0005] Furthermore, in the aforementioned deep learning-based intelligent operation and maintenance management method for wind power, the step of collecting wind turbine operating data, preprocessing the operation data, and constructing a graph structure dataset with wind turbines as nodes and spatiotemporal relationships as edges includes: Collect wind turbine operating data, including at least time-series operating data, vibration spectrum data, and infrared and visible light image data; The 3σ criterion is used to identify outliers in time series data and vibration data. Missing values ​​are filled in by interpolation. Image data is denoised to remove blurry, overexposed, or severely occluded invalid images, resulting in cleaned data. Z-score standardization was performed on time-series data and vibration spectrum data to eliminate differences in the magnitude of different parameters; image data was normalized to the [0,1] interval, and grayscale was performed after unifying the image size to obtain standard data.

[0006] Furthermore, in the aforementioned deep learning-based intelligent operation and maintenance management method for wind power, the step of collecting wind turbine operating data, preprocessing the operation data, and constructing a graph structure dataset with wind turbines as nodes and spatiotemporal relationships as edges includes: Based on the timestamp, the time-series data and vibration spectrum data of the same wind turbine are synchronized and aligned to obtain preprocessed data; Each wind turbine is constructed as a node, and the node feature vector is composed of the statistical characteristics of preprocessed time series data, vibration spectrum data, and the inherent properties of the corresponding wind turbine. Based on the spatiotemporal correlation characteristics, the connection relationship and weight of the edges are defined. The temporal edge is established between nodes of the same wind turbine at different times, and the weight is determined by the similarity of the operating status at adjacent times. The spatial edge is established between adjacent wind turbines, and the weight is calculated by comprehensively considering the correlation of wind speed and wind direction between wind turbines and geographical distance.

[0007] Furthermore, in the aforementioned deep learning-based intelligent operation and maintenance management method for wind power, the spatiotemporal graph neural network modeling the spatiotemporal dependencies within the wind turbine cluster and between the wind turbines and the environment, based on the graph structure dataset and utilizing a multi-head attention mechanism combined with a GCN graph convolutional network, yields spatiotemporal features, including: By using a multi-head attention mechanism to calculate the dependency weights between features at different times for the temporal feature sequence of each wind turbine node, the attention features in the time dimension are obtained. Based on the spatial edge weights in the graph structure, a spatial attention matrix is ​​constructed. The spatial interaction dependencies between different wind turbine nodes are learned through a multi-head attention mechanism to obtain the attention features in the spatial dimension. Attention features in the temporal and spatial dimensions are used as input to the GCN graph convolutional network. The neighborhood features of each node are aggregated through graph convolution operations, and the output is a spatiotemporal feature that integrates spatiotemporal dependencies.

[0008] Furthermore, in the aforementioned deep learning-based intelligent operation and maintenance management method for wind power, the method utilizes 1DCNN to extract deep temporal features from graph structure datasets, employs an improved lightweight visual Transformer to extract image defect features, and performs feature alignment and fusion through a cross-modal attention mechanism to obtain fused features, including: Calculate the similarity matrix between the deep temporal feature vector and the image defect feature vector, and dynamically adjust the contribution of the two modal features through similarity weights to obtain the aligned temporal features and image defect features; The aligned temporal features and image defect features are weighted and summed, and then nonlinearly transformed through a fully connected layer to output a fused feature vector.

[0009] Furthermore, in the aforementioned deep learning-based intelligent operation and maintenance management method for wind power, the fused features are input into a GRU gated recurrent network, and the remaining effective life probability distribution of wind turbine components and the dynamic risk levels of multiple types of faults within the next 72 hours are output to obtain prediction results, including: A GRU gated recurrent network model is constructed, and the gating mechanism is used to capture the long-short-term dependencies in the fused features. The first branch of the output layer uses the Softmax activation function to map the GRU output features to the probability distribution of the remaining effective lifetime. The second branch of the output layer predicts the dynamic risk level of each type of fault in the next 72 hours. The Sigmoid activation function is used in combination with weighted summation to divide the risk level into 5 levels. At the same time, the risk level change curve of each hour is output to obtain the prediction result.

[0010] Furthermore, in the aforementioned deep learning-based intelligent operation and maintenance management method for wind power, the generation of dynamic maintenance decision schemes based on a deep reinforcement learning model, using the prediction results as input and multi-objective optimization as reward, includes: Defined as the operating environment of wind turbine units, including the current operating status of the wind turbine, wind farm environmental conditions, maintenance resource allocation, and maintenance cost constraints; Based on the DQN deep Q-network as the core of the agent, the maintenance decision strategy is learned through neural network. The input is the environmental state vector and the output is the Q value of different maintenance actions.

[0011] Furthermore, in the deep learning-based intelligent operation and maintenance management system for wind power, the system includes the following modules: The wind power data acquisition module is used to collect the operating data of wind turbine units, and after preprocessing the operation, it constructs a graph structure dataset with wind turbines as nodes and spatiotemporal associations as edges. The spatiotemporal feature acquisition module is used to model the spatiotemporal dependencies within the wind turbine group and between the wind turbines and the environment based on the graph structure dataset by using a spatiotemporal graph neural network that combines a multi-head attention mechanism with a GCN graph convolutional network, and to obtain spatiotemporal features. The feature calculation and fusion module is used to extract deep temporal features from graph structure datasets using 1DCNN, extract image defect features using an improved lightweight visual Transformer, and perform feature alignment and fusion through a cross-modal attention mechanism to obtain fused features. The unit status prediction module is used to input the fused features into the GRU gated loop network, output the probability distribution of the remaining effective life of wind turbine components and the dynamic risk level of multiple types of faults in the next 72 hours, and obtain the prediction results. The dynamic maintenance decision module is used to generate dynamic maintenance decision schemes based on a deep reinforcement learning model, taking the prediction results as input and multi-objective optimization as reward.

[0012] Furthermore, in the deep learning-based intelligent operation and maintenance management system for wind power, the unit status prediction module includes the following sub-modules: A submodule is constructed to build a GRU gated recurrent network model. The gating mechanism is used to capture the long-short-term dependencies in the fused features. The first branch of the output layer uses the Softmax activation function to map the GRU output features to the probability distribution of the remaining effective lifetime. The summation submodule is used to predict the dynamic risk level of each type of fault in the second branch of the output layer within the next 72 hours. It uses the Sigmoid activation function combined with weighted summation to divide the risk level into 5 levels and outputs the risk level change curve for each hour to obtain the prediction result.

[0013] Furthermore, in the deep learning-based intelligent operation and maintenance management system for wind power, the dynamic maintenance decision module includes the following sub-modules: Define the submodule, which is used to define the wind turbine operating environment, including the current operating status of the wind turbine, wind farm environmental conditions, maintenance resource configuration and maintenance cost constraints; The output submodule is used to learn maintenance decision strategies through neural networks based on the DQN deep Q network as the core of the agent. The input is the environmental state vector and the output is the Q value of different maintenance actions.

[0014] Its beneficial effects lie in the fact that by integrating source data and graph structure modeling, it can achieve in-depth mining of spatiotemporal correlation and multi-dimensional information of wind turbine clusters, solving the problem of data fragmentation in traditional technologies and laying the foundation for accurate prediction. The synergistic application of spatiotemporal graph neural networks and cross-modal attention mechanisms effectively improves the extraction accuracy of spatiotemporal features and multimodal defect features, significantly reducing the prediction error of the remaining effective lifetime probability distribution and improving the accuracy of fault risk level prediction in the next 72 hours, thus achieving accurate early warning of faults. The combination of GRU model and deep reinforcement learning transforms the prediction results into dynamic maintenance decisions, taking into account multiple objectives of maintenance cost, downtime and safety risk optimization, reducing blind maintenance, and is expected to reduce operation and maintenance costs, improve unit availability, and adapt to large-scale wind power cluster operation and maintenance scenarios. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of the wind power intelligent operation and maintenance management method based on deep learning in this invention. Figure 2 This is a schematic diagram of the second embodiment of the wind power intelligent operation and maintenance management method based on deep learning in this invention. Figure 3 This is a schematic diagram of the first embodiment of the wind power intelligent operation and maintenance management system based on deep learning in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, a deep learning-based intelligent operation and maintenance management method for wind power includes the following steps: Step 101: Collect the operating data of the wind turbine units, and after preprocessing the operation, construct a graph structure dataset with wind turbines as nodes and spatiotemporal associations as edges; Specifically, in this embodiment, the wind turbine's operating data is collected, including at least time-series operating data, vibration spectrum data, and infrared and visible light image data. The 3σ criterion is used to identify outliers in the time-series and vibration data, and missing values ​​are filled in using interpolation. The image data is then denoised to remove blurry, overexposed, or severely obscured invalid images, resulting in clean data. The time-series and vibration spectrum data are Z-score standardized to eliminate differences in the magnitude of different parameters. The image data is normalized to the [0,1] interval, and after unifying the image size, it is converted to grayscale to obtain standard data. Based on timestamps, the time-series data and vibration spectrum data of the same wind turbine are synchronized and aligned to obtain preprocessed data. Each wind turbine is constructed as a node, and the node feature vector is composed of the statistical characteristics of the preprocessed time-series data and vibration spectrum data, as well as the inherent attributes of the corresponding wind turbine. The connection relationship and weight of the edges are defined based on the spatiotemporal correlation characteristics. The time edge is established between nodes of the same wind turbine at different times, and the weight is determined by the similarity of the operating status at adjacent times. The spatial edge is established between adjacent wind turbines, and the weight is calculated by comprehensively considering the correlation of wind speed and wind direction between wind turbines and geographical distance.

[0020] The time-series operation data is collected through the unit's built-in SCADA system and sensor network, which collects key parameters such as generator speed, nacelle wind speed, hub azimuth angle, blade angle, gearbox oil temperature, hydraulic system pressure, converter output power, and grid voltage / current. The sampling frequency is set to 1Hz to ensure that the dynamic operating status of the unit is captured. At the same time, the collection timestamp, corresponding wind turbine number, wind farm area and other related information are recorded to lay the foundation for subsequent spatiotemporal analysis.

[0021] Vibration spectrum data is collected by installing piezoelectric accelerometers on the surfaces of critical components prone to failure, such as gearboxes, spindles, generators, and bearings. Vibration signals in three dimensions, X / Y / Z axes, are collected at a sampling frequency of 20kHz. The time-domain vibration signals are converted into frequency-domain spectrum data through FFT (Fast Fourier Transform) to obtain characteristic information such as vibration amplitude and phase in different frequency bands, accurately capturing early signs of component failure.

[0022] Infrared and visible light image data were acquired using a combination of drone inspections and fixed-position monitoring. Infrared images focused on critical heat-generating components such as gearbox bearings, generator windings, and control cabinets. Acquisition was conducted under stable ambient temperatures, avoiding rain, snow, and periods of strong sunlight, with a resolution of at least 1920×1080. Parameters such as the location of the corresponding component, shooting distance, and ambient temperature were recorded. Visible light images focused on blade surfaces, the external structure of the nacelle, and tower connections, clearly revealing surface defects such as cracks, corrosion, and dirt accumulation. Information such as lighting conditions and shooting angle was recorded simultaneously.

[0023] Data cleaning involves using the 3σ criterion to identify outliers in time-series and vibration data. Interpolation methods are used: linear interpolation is applied to time-series data, and spline interpolation is applied to fill in missing values ​​for vibration data. Image data is denoised: median filtering is applied to infrared images, and Gaussian filtering is applied to visible light images. Invalid images that are blurry, overexposed, or severely occluded are removed.

[0024] Data standardization / normalization: Time series data and vibration spectrum data are Z-score standardized to eliminate differences in the magnitude of different parameters; Image data are normalized to the [0,1] interval, and the image size is unified. Infrared and visible light images are adjusted to 640×480 pixels and grayscale is performed. Infrared images or RGB channels are standardized, and visible light images are also normalized.

[0025] Data alignment is performed by synchronizing the time-series data and vibration spectrum data of the same wind turbine based on the timestamp; the wind turbine operating parameters at the corresponding acquisition time are matched to the image data to establish an image operating status association mapping.

[0026] Using wind turbine units as the core nodes, a spatiotemporal correlation graph structure is constructed. The specific steps include node definition, where each wind turbine is constructed as a node. The node feature vector is composed of the statistical characteristics of the preprocessed time series data and vibration spectrum data, including mean, variance, peak value, kurtosis, and the inherent attributes of the corresponding wind turbine, such as model, installed capacity, years of operation, and geographical location.

[0027] Edge construction is based on spatiotemporal correlation characteristics to define the connection relationships and weights of edges. Temporal edges establish connections between nodes of the same wind turbine at different times, and the weight is determined by the similarity of the operating states at adjacent times; spatial edges establish connections between adjacent wind turbines, and the weight is calculated by comprehensively considering the correlation of wind speed and direction between wind turbines and geographical distance; environmental correlation edges establish connections between wind farm environmental monitoring nodes and surrounding wind turbine nodes, and the weight is determined by the significance of the impact of environmental parameters on wind turbine operation.

[0028] The dataset is partitioned into training, validation, and test sets in a 7:2:1 ratio to ensure consistency in the distribution of stroke components, operating conditions, and environmental conditions across the sets.

[0029] Step 102: Based on the graph structure dataset, a spatiotemporal graph neural network combining multi-head attention mechanism and GCN graph convolutional network is used to model the spatiotemporal dependencies within the wind turbine cluster and between the wind turbines and the environment, obtaining spatiotemporal features. Specifically, in this embodiment, the multi-head attention mechanism is used to calculate the dependency weights between features at different times for the temporal feature sequence of each wind turbine node, obtaining the attention features in the time dimension. Based on the spatial edge weights in the graph structure, a spatial attention matrix is ​​constructed, and the spatial interaction dependencies between different wind turbine nodes are learned through the multi-head attention mechanism, obtaining the attention features in the spatial dimension. The attention features in the time dimension and the attention features in the spatial dimension are used as input to the GCN graph convolutional network, and the neighborhood features of each node are aggregated through graph convolution operations to output spatiotemporal features that fuse spatiotemporal dependencies.

[0030] The model architecture is designed to construct a spatiotemporal graph neural network model that integrates a multi-head attention mechanism and a GCN graph convolutional network. The overall architecture consists of four core modules: an input layer, a spatiotemporal attention layer, a GCN convolutional layer, and a feature fusion layer, which can accurately capture the spatiotemporal dependencies within the wind turbine group and between the wind turbine and the environment.

[0031] Temporal attention computation employs a multi-head self-attention mechanism to calculate the dependency weights between features at different times for the temporal feature sequence of each wind turbine node. The temporal feature sequence is divided into multiple subsequences, and each subsequence learns temporal dependencies through an independent attention head. The outputs of each attention head are then concatenated and linearly transformed to obtain the attention features in the time dimension.

[0032] Spatial attention computation is performed by constructing a spatial attention matrix based on the spatial edge weights in the graph structure. A multi-head attention mechanism is then used to learn the spatial interaction dependencies between different wind turbine nodes. For each node, its feature similarity with neighboring nodes is calculated, and the contribution of neighboring nodes is dynamically adjusted based on the edge weights to achieve spatial-dimensional attention feature extraction.

[0033] The features output by the spatiotemporal attention mechanism are used as input to the GCN. The neighborhood features of each node are aggregated through graph convolution operations. Based on the adjacency matrix of the graph structure, the features of each node are weighted and summed, with the weights being the spatiotemporal correlation weights of the edges, to achieve the initial aggregation of neighborhood information. The ReLU activation function is used to introduce a nonlinear transformation to enhance the feature representation capability of the model. 23 GCN convolutional layers are set to gradually deepen feature extraction, and finally output a spatiotemporal feature vector that integrates spatiotemporal dependencies.

[0034] Using the graph structure data of the training set as input, the mean squared error loss function (MSE) is adopted. The deviation between the actual operating state of the wind turbine and the predicted state of the model is used as the optimization objective. The Adam optimizer is used for model training, and the learning rate is initially set to 0.001. Through an early stopping strategy, training is stopped if the loss on the validation set does not decrease for 10 consecutive epochs to avoid overfitting.

[0035] Step 103: Use 1DCNN to extract deep temporal features from the graph structure dataset, use an improved lightweight visual Transformer to extract image defect features, and use a cross-modal attention mechanism to align and fuse features to obtain fused features. Specifically, in this embodiment, the similarity matrix between the deep temporal feature vector and the image defect feature vector is calculated. The contribution of the two modal features is dynamically adjusted by the similarity weight to obtain the aligned temporal features and image defect features. The aligned temporal features and image defect features are weighted and summed, and then nonlinearly transformed through a fully connected layer to output the fused feature vector.

[0036] For the temporal features in graph-structured data, a 1DCNN model is constructed using preprocessed temporal operation data and vibration spectrum data to extract deep temporal features. The model structure is as follows: Input layer: Temporal feature vector dimension 128 → Convolutional layer 1 (64 kernels, kernel size 3, stride 1, ReLU activation function) → Max pooling layer 1 (kernel size 2, stride 2) → Convolutional layer 2 (128 kernels, kernel size 3, stride 1, ReLU activation function) → Max pooling layer 2 (kernel size 2, stride 2) → Fully connected layer, output dimension 256, finally obtaining the deep temporal feature vector. Training and optimization: The 1DCNN model is trained solely using temporal data, employing cross-entropy loss as the loss function. A temporal feature classification task, such as wind turbine operation condition classification, is used as the auxiliary objective. The optimizer is SGD with a learning rate of 0.005, and training continues until model convergence.

[0037] For infrared and visible light image data, the standard visual Transformer (ViT) is improved with lightweight design to extract image defect features. The model is improved by reducing the number of layers in the Transformer encoder from 12 to 6 and the number of attention heads from 12 to 8. Depthwise separable convolutions are used to replace some fully connected layers, reducing the total number of model parameters to 1 / 3 of the original model. The image block size is increased in the input layer to improve feature extraction efficiency.

[0038] The feature extraction process involves inputting the preprocessed image into the improved ViT model, and through image segmentation, positional encoding, multi-head self-attention calculation, and encoder feature aggregation, the final output is an image defect feature vector with a dimension of 256.

[0039] For training and optimization, the model is trained using image datasets labeled with defect types such as cracks, corrosion, and abnormal heating. The loss function is cross-entropy loss, with defect classification as the objective. The optimizer is AdamW, and the learning rate is 0.0001. Data augmentation techniques, such as image rotation, flipping, and random cropping, are used to improve the model's generalization ability.

[0040] A cross-modal attention mechanism is constructed to achieve the alignment and fusion of deep temporal features and image defect features. Feature alignment is performed by calculating the similarity matrix between the deep temporal feature vector and the image defect feature vector. Dot product attention is used for calculation, and the contribution of the two modal features is dynamically adjusted through similarity weights so that the fused features can accurately associate the corresponding relationship between the defects in the running image.

[0041] Feature fusion involves weighted summation of aligned temporal features and image defect features, followed by nonlinear transformation through a fully connected layer, outputting a fused feature vector with a dimension of 512. The fused features include both dynamic temporal information of unit operation and appearance and thermal defect information of key components.

[0042] Step 104: Input the fused features into the GRU gated recurrent network, output the probability distribution of the remaining effective life of the wind turbine components and the dynamic risk level of multiple types of faults in the next 72 hours, and obtain the prediction results; Specifically, in this embodiment, a GRU gated recurrent network model is constructed. The gating mechanism is used to capture the long-term and short-term dependencies in the fused features. The first branch of the output layer uses the Softmax activation function to map the GRU output features to the probability distribution of the remaining effective lifetime. The second branch of the output layer predicts the dynamic risk level of each type of fault in the next 72 hours. The Sigmoid activation function is used in combination with weighted summation to divide the risk level into 5 levels. At the same time, the risk level change curve of each hour is output to obtain the prediction result.

[0043] A GRU gated recurrent network model is constructed. Utilizing its gating mechanism, the resetting and updating gates capture the long-short-term dependencies in the fused features, achieving accurate prediction of remaining lifetime and fault risk level. The model structure is as follows: Input layer: fused feature vector dimension 512 → GRU layer 1, 256 hidden units, dropout rate 0.2 → GRU layer 2, 128 hidden units, dropout rate 0.2 → fully connected layer 1; Output layer: dimension 64, activation function ReLU → output layer, divided into two branches: remaining lifetime probability distribution branch and fault risk level branch.

[0044] In the first branch of the output layer, the Softmax activation function is used to map the GRU output features to a probability distribution of the remaining effective lifespan. Specifically, the remaining effective lifespan is divided into multiple continuous intervals, and the model outputs the probability value of each interval, clearly presenting the uncertainty range of the remaining lifespan of the unit components through the probability distribution.

[0045] In the second branch of the output layer, for common wind turbine fault types such as gearbox faults, bearing faults, blade faults, generator faults, and hydraulic system faults, the dynamic risk level of each type of fault is predicted for the next 72 hours. A sigmoid activation function combined with weighted summation is used to classify the risk level into 5 levels: Level 1, no risk; Level 2, low risk; Level 3, medium risk; Level 4, high risk; and Level 5, extremely high risk. Simultaneously, hourly risk level change curves are output to achieve dynamic risk monitoring.

[0046] Using fused features as input, a joint loss function is employed: cross-entropy loss + KL divergence loss. Cross-entropy loss is used for fault risk level classification, and KL divergence loss is used for fitting the remaining lifetime probability distribution. The model is trained using the Adam optimizer with a learning rate of 0.001. The model's prediction accuracy is validated using a test set, requiring a remaining lifetime prediction error ≤10% and a fault risk level prediction accuracy ≥90%.

[0047] Step 105: Based on a deep reinforcement learning model, using the prediction results as input and multi-objective optimization as reward, generate a dynamic maintenance decision scheme.

[0048] Specifically, in this embodiment, the wind turbine operating environment is defined as including the current operating status of the wind turbine, wind farm environmental conditions, maintenance resource configuration, and maintenance cost constraints. Based on the DQN deep Q network as the core of the agent, the maintenance decision strategy is learned through the neural network. The input is the environmental state vector, and the output is the Q value of different maintenance actions.

[0049] A deep reinforcement learning framework for constructing an agent's environmental reward function is proposed. The environment is defined as the wind turbine's operating environment, which includes the wind turbine's current operating status, the prediction results from step 4, real-time operating data, wind farm environmental conditions, maintenance resource allocation, maintenance personnel, spare parts inventory, maintenance equipment, maintenance cost constraints, and other elements.

[0050] The intelligent agent employs a Deep Q-Network (DQN) as its core, learning maintenance decision-making strategies through neural networks. The input is an environmental state vector, and the output is the Q-value and priority of different maintenance actions. The action space defines a set of maintenance actions, including no maintenance, routine inspection, partial repair, for low / medium risk faults, complete replacement, for high / extremely high risk faults, and emergency shutdown for repair. Each action corresponds to a specific execution process and resource requirements.

[0051] The reward function is designed around multi-objective optimization, comprehensively considering factors such as maintenance cost, downtime, failure loss, and safety risk. The specific formula is: Reward Value = α × (1 / Maintenance Cost / Maximum Allowable Cost) + β × (1 / Downtime / Maximum Allowable Downtime) + γ × (1 / Failure Loss / Historical Maximum Failure Loss) + δ × Safety Risk Coefficient. Here, α, β, γ, and δ are weighting coefficients adjusted according to the wind farm's operational priority, and their sum is 1. The safety risk coefficient is determined by the failure risk level, with 1 for level 1 and 0 for level 5. The reward function guides the agent to learn the optimal maintenance strategy that minimizes cost, efficiency, and risk.

[0052] A DQN agent is trained using an experience replay mechanism and a target network update strategy. Experience replay stores the agent's interactions with the environment, including state, action, reward, and next-state experience data, in an experience pool. Randomly sampled batches of data are used to train the model, improving training stability. Target network updates involve setting up a target network and an evaluation network, updating the target network parameters at regular intervals to avoid Q-value estimation bias. Exploration and exploitation employ an ε-greedy strategy to balance exploration, trying new actions and exploiting existing ones, selecting the action with the highest current Q-value. ε is initially set to 0.9 and gradually decays to 0.1 with each training step.

[0053] The agent takes the prediction results from step 4 and the real-time environmental status as input, outputs the optimal maintenance action, and generates a dynamic maintenance decision plan that includes maintenance time, maintenance content, required resources, and expected results. Simultaneously, it monitors the wind turbine's operating status and changes in fault risk in real time. If the risk level exceeds a preset threshold or the operating status becomes abnormal, the agent dynamically adjusts the maintenance plan to ensure the timeliness and accuracy of maintenance decisions.

[0054] Its beneficial effects lie in the fact that by integrating source data and graph structure modeling, it can achieve in-depth mining of spatiotemporal correlation and multi-dimensional information of wind turbine clusters, solving the problem of data fragmentation in traditional technologies and laying the foundation for accurate prediction. The synergistic application of spatiotemporal graph neural networks and cross-modal attention mechanisms effectively improves the extraction accuracy of spatiotemporal features and multimodal defect features, significantly reducing the prediction error of the remaining effective lifetime probability distribution and improving the accuracy of fault risk level prediction in the next 72 hours, thus achieving accurate early warning of faults. The combination of GRU model and deep reinforcement learning transforms the prediction results into dynamic maintenance decisions, taking into account multiple objectives of maintenance cost, downtime and safety risk optimization, reducing blind maintenance, and is expected to reduce operation and maintenance costs, improve unit availability, and adapt to large-scale wind power cluster operation and maintenance scenarios.

[0055] Please see Figure 2 In the deep learning-based intelligent operation and maintenance management method for wind power, the process of collecting wind turbine operation data, preprocessing the operation data, and constructing a graph structure dataset with wind turbines as nodes and spatiotemporal relationships as edges includes the following steps: Step 201: Collect wind turbine operating data, including at least time-series operating data, vibration spectrum data, and infrared and visible light image data; Step 202: Use the 3σ criterion to identify outliers in time series data and vibration data, fill in missing values ​​by interpolation, denoise the image data, and remove invalid images that are blurry, overexposed, or severely occluded to obtain cleaned data. Step 203: Perform Z-score standardization on the time series data and vibration spectrum data to eliminate differences in the magnitude of different parameters; normalize the image data to the [0,1] interval, unify the image size, and then perform grayscale conversion to obtain standard data.

[0056] The above describes embodiments of the deep learning-based intelligent operation and maintenance management method for wind power according to the present invention. Please refer to [link / reference]. Figure 3 In the deep learning-based intelligent operation and maintenance management system for wind power, the system includes the following modules: The wind power data acquisition module is used to collect the operating data of wind turbine units, and after preprocessing the operation, it constructs a graph structure dataset with wind turbines as nodes and spatiotemporal associations as edges. The spatiotemporal feature acquisition module is used to model the spatiotemporal dependencies within the wind turbine cluster and between the wind turbines and the environment based on the graph structure dataset, using a spatiotemporal graph neural network that combines multi-head attention mechanism with GCN graph convolutional network, and obtain spatiotemporal features. The feature calculation and fusion module is used to extract deep temporal features from graph structure datasets using 1DCNN, extract image defect features using an improved lightweight visual Transformer, and perform feature alignment and fusion through a cross-modal attention mechanism to obtain fused features. The unit status prediction module is used to input the fused features into the GRU gated cyclic network, output the probability distribution of the remaining effective life of the wind turbine components and the dynamic risk level of multiple types of faults in the next 72 hours, and obtain the prediction results. The dynamic maintenance decision module is used to generate dynamic maintenance decision schemes based on deep reinforcement learning models, taking prediction results as input and multi-objective optimization as reward.

[0057] 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 technical solutions and their equivalents.

Claims

1. A wind power intelligent operation and maintenance management method based on deep learning, characterized in that, The deep learning-based intelligent operation and maintenance management method for wind power includes the following steps: The operation data of wind turbine units is collected, and after preprocessing the operation, a graph structure dataset is constructed with wind turbines as nodes and spatiotemporal associations as edges. Based on the graph structure dataset, a spatiotemporal graph neural network combining multi-head attention mechanism and GCN graph convolutional network is used to model the spatiotemporal dependencies within the wind turbine group and between the wind turbines and the environment, thereby obtaining spatiotemporal features; We use 1DCNN to extract deep temporal features from graph structure datasets, use an improved lightweight visual Transformer to extract image defect features, and use a cross-modal attention mechanism to align and fuse features to obtain fused features. The fused features are input into a GRU gated recurrent network, which outputs the probability distribution of the remaining effective life of wind turbine components and the dynamic risk level of multiple types of faults in the next 72 hours, thus obtaining the prediction results. Based on a deep reinforcement learning model, using the prediction results as input and multi-objective optimization as reward, a dynamic maintenance decision scheme is generated.

2. The wind power intelligent operation and maintenance management method based on deep learning as described in claim 1, characterized in that, The collected wind turbine operating data is preprocessed and then used to construct a graph-structured dataset with wind turbines as nodes and spatiotemporal relationships as edges, including: Collect wind turbine operating data, including at least time-series operating data, vibration spectrum data, and infrared and visible light image data; The 3σ criterion is used to identify outliers in time series data and vibration data. Missing values ​​are filled in by interpolation. Image data is denoised to remove blurry, overexposed, or severely occluded invalid images, resulting in cleaned data. Z-score standardization was performed on time-series data and vibration spectrum data to eliminate differences in the magnitude of different parameters; image data was normalized to the [0,1] interval, and grayscale was performed after unifying the image size to obtain standard data.

3. The wind power intelligent operation and maintenance management method based on deep learning as described in claim 1, characterized in that, The collected wind turbine operating data is preprocessed and then used to construct a graph-structured dataset with wind turbines as nodes and spatiotemporal relationships as edges, including: Based on the timestamp, the time-series data and vibration spectrum data of the same wind turbine are synchronized and aligned to obtain preprocessed data; Each wind turbine is constructed as a node, and the node feature vector is composed of the statistical characteristics of preprocessed time series data, vibration spectrum data, and the inherent properties of the corresponding wind turbine. Based on the spatiotemporal correlation characteristics, the connection relationship and weight of the edges are defined. The temporal edge is established between nodes of the same wind turbine at different times, and the weight is determined by the similarity of the operating status at adjacent times. The spatial edge is established between adjacent wind turbines, and the weight is calculated by comprehensively considering the correlation of wind speed and wind direction between wind turbines and geographical distance.

4. The wind power intelligent operation and maintenance management method based on deep learning as described in claim 1, characterized in that, Based on the graph structure dataset, a spatiotemporal graph neural network combining multi-head attention mechanism and GCN graph convolutional network is used to model the spatiotemporal dependencies within the wind turbine cluster and between the wind turbines and the environment, obtaining spatiotemporal features, including: By using a multi-head attention mechanism to calculate the dependency weights between features at different times for the temporal feature sequence of each wind turbine node, the attention features in the time dimension are obtained. Based on the spatial edge weights in the graph structure, a spatial attention matrix is ​​constructed. The spatial interaction dependencies between different wind turbine nodes are learned through a multi-head attention mechanism to obtain the attention features in the spatial dimension. Attention features in the temporal and spatial dimensions are used as input to the GCN graph convolutional network. The neighborhood features of each node are aggregated through graph convolution operations, and the output is a spatiotemporal feature that integrates spatiotemporal dependencies.

5. The wind power intelligent operation and maintenance management method based on deep learning as described in claim 1, characterized in that, The method utilizes 1DCNN to extract deep temporal features from graph structure datasets, employs an improved lightweight visual Transformer to extract image defect features, and performs feature alignment and fusion through a cross-modal attention mechanism to obtain fused features, including: Calculate the similarity matrix between the deep temporal feature vector and the image defect feature vector, and dynamically adjust the contribution of the two modal features through similarity weights to obtain the aligned temporal features and image defect features; The aligned temporal features and image defect features are weighted and summed, and then nonlinearly transformed through a fully connected layer to output a fused feature vector.

6. The wind power intelligent operation and maintenance management method based on deep learning as described in claim 1, characterized in that, The fused features are input into a GRU-gated recurrent network, which outputs the probability distribution of the remaining effective lifespan of wind turbine components and the dynamic risk levels of various types of faults within the next 72 hours, yielding prediction results, including: A GRU gated recurrent network model is constructed, and the gating mechanism is used to capture the long-short-term dependencies in the fused features. The first branch of the output layer uses the Softmax activation function to map the GRU output features to the probability distribution of the remaining effective lifetime. The second branch of the output layer predicts the dynamic risk level of each type of fault in the next 72 hours. The Sigmoid activation function is used in combination with weighted summation to divide the risk level into 5 levels. At the same time, the risk level change curve of each hour is output to obtain the prediction result.

7. The wind power intelligent operation and maintenance management method based on deep learning as described in claim 1, characterized in that, The method based on a deep reinforcement learning model, using the prediction results as input and multi-objective optimization as reward, generates a dynamic maintenance decision scheme, including: Defined as the operating environment of wind turbine units, including the current operating status of the wind turbine, wind farm environmental conditions, maintenance resource allocation, and maintenance cost constraints; Based on the DQN deep Q-network as the core of the agent, the maintenance decision strategy is learned through neural network. The input is the environmental state vector and the output is the Q value of different maintenance actions.

8. A wind power intelligent operation and maintenance management system based on deep learning, characterized in that, The deep learning-based intelligent operation and maintenance management system for wind power includes the following modules: The wind power data acquisition module is used to collect the operating data of wind turbine units, and after preprocessing the operation, it constructs a graph structure dataset with wind turbines as nodes and spatiotemporal associations as edges. The spatiotemporal feature acquisition module is used to model the spatiotemporal dependencies within the wind turbine group and between the wind turbines and the environment based on the graph structure dataset by using a spatiotemporal graph neural network that combines a multi-head attention mechanism with a GCN graph convolutional network, and to obtain spatiotemporal features. The feature calculation and fusion module is used to extract deep temporal features from graph structure datasets using 1DCNN, extract image defect features using an improved lightweight visual Transformer, and perform feature alignment and fusion through a cross-modal attention mechanism to obtain fused features. The unit status prediction module is used to input the fused features into the GRU gated loop network, output the probability distribution of the remaining effective life of wind turbine components and the dynamic risk level of multiple types of faults in the next 72 hours, and obtain the prediction results. The dynamic maintenance decision module is used to generate dynamic maintenance decision schemes based on a deep reinforcement learning model, taking the prediction results as input and multi-objective optimization as reward.

9. The wind power intelligent operation and maintenance management system based on deep learning as described in claim 8, characterized in that, The unit status prediction module includes the following sub-modules: A submodule is constructed to build a GRU gated recurrent network model. The gating mechanism is used to capture the long-short-term dependencies in the fused features. The first branch of the output layer uses the Softmax activation function to map the GRU output features to the probability distribution of the remaining effective lifetime. The summation submodule is used to predict the dynamic risk level of each type of fault in the second branch of the output layer within the next 72 hours. It uses the Sigmoid activation function combined with weighted summation to divide the risk level into 5 levels and outputs the risk level change curve for each hour to obtain the prediction result.

10. The wind power intelligent operation and maintenance management system based on deep learning as described in claim 8, characterized in that, The dynamic maintenance decision module includes the following sub-modules: Define the submodule, which is used to define the wind turbine operating environment, including the current operating status of the wind turbine, wind farm environmental conditions, maintenance resource configuration and maintenance cost constraints; The output submodule is used to learn maintenance decision strategies through neural networks based on the DQN deep Q network as the core of the agent. The input is the environmental state vector and the output is the Q value of different maintenance actions.

Images