Wind turbine oil temperature regulation method and system based on sensor space-time fusion

Abstract

The application discloses a wind turbine oil temperature regulation method and system based on sensor space-time fusion, and belongs to the technical field of wind power equipment control. The existing regulation method does not consider the collaborative evolution relationship among multivariable sensors, which affects the regulation accuracy. The wind turbine oil temperature regulation method based on sensor space-time fusion comprises the following steps: first, obtaining sensor input data; second, obtaining sensor feature data; third, generating correlation data between sensors; fourth, generating sensor space-time feature data containing space features and time features, and using the same as input to predict the oil temperature and obtain an oil temperature prediction value; and fifth, generating a control strategy and outputting a control signal according to the control strategy. The application effectively fuses multiple types of sensor data in the bearing system of the wind turbine, can cope with complex changes of coexistence of slow drift and sudden transition, organically unifies prediction and control, and is more accurate in control.

Description

Technical Field

[0001] This invention relates to a method and system for controlling oil temperature in wind turbine generators based on sensor spatiotemporal fusion, belonging to the field of wind power equipment control technology. Background Technology

[0002] In wind power systems, generator bearing lubricating oil temperature is a crucial parameter for assessing equipment operational safety and energy efficiency. Abnormal oil temperature often indicates potential faults such as mechanical wear, lubrication degradation, or cooling system failure, and its changes exhibit a complex characteristic of both slow drift and sudden jumps. However, existing oil temperature monitoring and control technologies have the following main shortcomings: (1) Lagging monitoring methods and frequent false alarms: Currently, industrial sites mostly use fixed threshold alarms or rely on single-point temperature trend judgment. Once the oil temperature exceeds the set upper limit, a shutdown or load reduction command is triggered. Such methods ignore the co-evolution relationship between multiple variables, cannot capture early abnormal signs in time, and are sensitive to environmental disturbances, resulting in a high false alarm rate.

[0003] (2) Modeling methods ignore spatial coupling between sensors: Existing studies have attempted to introduce time series models such as long short-term memory networks to predict oil temperature. Although this can improve the accuracy to a certain extent, it usually only processes the temperature sequence itself and does not effectively integrate the spatial structure and dynamic coupling relationship between sensors such as oil temperature and speed. As a result, it is slow to respond to complex faults (such as slow change + sudden change) and has limited predictive ability.

[0004] (3) The prediction and control strategies are disconnected and the energy efficiency is not optimized: the existing prediction system only provides alarm signals, and the cooling pump and lubricating oil pump still operate at rated power or traditional PID logic. This may lead to "overcooling" and energy waste, and there is no flow regulation mechanism in the abnormal evolution process, which cannot work together to reduce the risk of failure and energy loss.

[0005] Due to the aforementioned technical problems, it is necessary to improve existing control technologies.

[0006] The information disclosed in this background section is only for understanding the background of the inventive concept, and therefore may include information that does not constitute prior art. Summary of the Invention

[0007] To address the aforementioned technical problems, one objective of this invention is to provide a wind turbine oil temperature control method based on sensor spatiotemporal fusion that effectively integrates multiple types of sensor data in the wind turbine bearing system, can cope with complex oil temperature changes involving both slow drift and sudden jumps, and organically unifies prediction and control. This method obtains the correlation data between sensors by extracting sensor feature data and integrating dynamic data correlation and physical coupling relationships. Based on the sensor feature data and correlation data, it obtains sensor spatiotemporal feature data, and uses this data as the basis for prediction, effectively improving prediction accuracy. By organically unifying the prediction results with the control strategy, the control becomes more precise and effective.

[0008] To address the aforementioned technical problems, another objective of this invention is to provide a wind turbine oil temperature control system based on sensor spatiotemporal fusion that effectively integrates multiple types of sensor data in the wind turbine bearing system, can cope with complex oil temperature changes involving both slow drift and sudden jumps, and organically unifies prediction and control. This system obtains correlation data between sensors by extracting sensor feature data and integrating dynamic data correlation and physical coupling relationships. Based on the sensor feature data and correlation data, it obtains sensor spatiotemporal feature data, and uses this data as the basis for prediction, effectively improving prediction accuracy. By organically unifying the prediction results with the control strategy, the control becomes more precise and effective.

[0009] To achieve one of the above objectives, the first technical solution of the present invention is as follows: A method for controlling oil temperature in wind turbine generators based on sensor spatiotemporal fusion includes the following steps: Step 1: Collect raw data from various types of sensors in the wind turbine bearing system in real time, and perform data cleaning to obtain sensor input data; the raw sensor data includes bearing lubricating oil temperature data; Step 2: Based on the sensor input data obtained in Step 1, set a time window, extract the data features within the time window, and obtain sensor feature data; the data features are used to reflect the sensor's real-time data, short-term dynamic change trends, and time-series statistical patterns. Step 3: Based on the sensor feature data obtained in Step 2, obtain the dynamic data correlation between sensors, and combine the physical coupling relationship between the sensors on site to generate correlation data between the sensors. Step 4: Based on the sensor feature data obtained in Step 2, and according to the correlation data, generate sensor spatiotemporal feature data containing spatial and temporal features, and use the sensor spatiotemporal feature data as input to predict the oil temperature at future times, thereby obtaining the oil temperature prediction value. Step 5: Based on the predicted oil temperature, generate a control strategy and output a control signal for controlling the oil temperature regulating device according to the control strategy.

[0010] As a preferred technical measure: Step 1: Real-time acquisition of raw data from various types of sensors in the wind turbine bearing system, followed by data cleaning. The method for obtaining sensor input data is as follows: Step 11: Sample the raw data of various types of sensors in the wind turbine bearing system at a fixed frequency and mark the time. The raw sensor data includes bearing lubricating oil temperature data, environmental data, and related equipment operation data. Step 12: Remove outliers from the data collected in Step 11; Step 13: Normalize the data obtained in Step 12 and fill in the missing values ​​to obtain the sensor input data.

[0011] As a preferred technical measure: In the second step, the method for extracting data features within the time window to obtain sensor feature data includes the following steps: Step 21: Set a sliding window for the sensor input data within the time window and slide the sliding window in fixed steps; Step 22: Calculate the multi-dimensional feature vector of the sensor input data in each sliding window. The multi-dimensional feature vector includes the normalized value of the real-time sensor data, the time difference between adjacent sensor times, and the mean, variance, and skewness of the sensor in the sliding window. Step 23: Based on all the multi-dimensional feature vectors obtained in Step 22, output the sensor feature matrix, which is the sensor feature data.

[0012] As a preferred technical measure: Step 3: Based on the sensor feature data obtained in Step 2, the dynamic data correlation between sensors is obtained, and the physical coupling relationship between the sensors in the field is comprehensively considered to generate the correlation data between the sensors. The method is as follows: Step 31: Construct a graph structure with each sensor as a vertex and the adjacency relationships between sensors as edges; Step 32: Assign static weights to the edges of the graph structure, wherein the static weights are determined based on the physical coupling relationships between the sensors in the field; Step 33: Based on sensor feature data, set up a dynamic calculation window and calculate the correlation index between each sensor within the dynamic calculation window to obtain dynamic weights; Step 34: Fuse the static weights and dynamic weights to obtain new edge weights, and update the edge weights of the graph structure; Step 35: Output the sensor adjacency matrix corresponding to the graph structure, i.e., the correlation data.

[0013] As a preferred technical measure: Step 4: Based on the sensor feature data obtained in Step 2, and according to the correlation data, the method for generating sensor spatiotemporal feature data containing spatial and temporal features is as follows: Step 41: Based on sensor feature data, for one of the sensors, calculate its attention score with each of its neighboring sensors according to the correlation data; Step 42: Process the attention scores of all neighboring sensors of this sensor to obtain the attention weights; Step 43: Change different attention to focus on different coupling modes between sensors, repeat steps 41 and 42 multiple times to obtain multiple attention weights; Step 44: Based on the multiple attention weights obtained in Step 43, the data features of neighboring sensors are weighted and aggregated onto the sensor to obtain the spatial aggregated feature value of the sensor; Step 45: Calculate the spatial aggregated feature values ​​of all sensors and output the sensor spatial feature matrix accordingly; Step 46: Based on sensor feature data, set the capture time for multiple layers with different granularities; Step 47: Capture the sensor time feature values ​​in each time layer and output the sensor time feature matrix accordingly; Step 48: Fuse the sensor spatial feature matrix obtained in step 45 and the sensor temporal feature matrix obtained in step 47 to obtain the sensor spatiotemporal feature matrix, that is, the sensor spatiotemporal feature data.

[0014] As a preferred technical measure: The fourth step also includes: using the spatiotemporal feature data of the sensor as input to predict the probability of anomalies at future times, and obtaining the predicted value of the anomaly probability.

[0015] As a preferred technical measure: The sixth step also includes: collecting feedback data from the oil temperature regulation device and optimizing the control strategy based on this feedback data.

[0016] As a preferred technical measure: Step 6: The method for collecting feedback data from the oil temperature regulation device and optimizing the control strategy based on this feedback data is as follows: Step 61: Collect feedback data from the oil temperature regulating device in real time and calculate the energy consumption index of the oil temperature regulating device; Step 62: Calculate the safety index based on the anomaly probability prediction value obtained in step 4; Step 63: Construct a reward function with energy consumption and safety as control objectives, and optimize the control strategy based on the energy consumption index obtained in Step 61 and the safety index obtained in Step 62.

[0017] As a preferred technical measure: In the fourth step, the prediction of oil temperature at future moments refers to the prediction of oil temperature at future moments through a neural network. This neural network is deployed at the edge of the wind turbine site, and its training process is deployed in the cloud. The cloud performs centralized modeling and training on the data uploaded by the edge and sends the training result parameters back to the edge. The training in the cloud includes: offline training, incremental training, and cross-scene transfer.

[0018] To achieve one of the above objectives, the second technical solution of the present invention is as follows: A wind turbine oil temperature control system based on sensor spatiotemporal fusion, which adopts the wind turbine oil temperature control method based on sensor spatiotemporal fusion described above in this invention; The oil temperature control system for this wind turbine includes the following modules: The data acquisition and cleaning module is used to acquire raw data from various types of sensors in the wind turbine bearing system in real time, and to clean the data to obtain sensor input data. The feature data extraction module is used to set a time window based on the sensor input data, extract the data features within the time window, and obtain sensor feature data. The correlation data generation module is used to obtain the dynamic data correlation between sensors based on sensor feature data, and to generate correlation data between sensors by integrating the physical coupling relationship between various sensors in the field. The spatiotemporal feature data generation and prediction module is used to generate sensor spatiotemporal feature data containing spatial and temporal features based on sensor feature data and the correlation data, and to predict the oil temperature at future times using the sensor spatiotemporal feature data as input to obtain the oil temperature prediction value. And a control module, used to generate a control strategy based on the oil temperature prediction value, and output a control signal for controlling the oil temperature regulating device according to the control strategy.

[0019] As a preferred technical measure: It also includes an optimization module, which is used to collect feedback data from the oil temperature regulating device and optimize the control strategy with reference to the feedback data.

[0020] As a preferred technical measure: It also includes the cloud, where the spatiotemporal feature data generation and prediction module uses a neural network to predict the oil temperature at future moments; the neural network is deployed at the edge of the wind turbine site, and its training process is deployed in the cloud, where the cloud performs centralized modeling and training on the data uploaded from the edge and sends the training result parameters to the edge; the training in the cloud includes: offline training, incremental training and cross-scene transfer.

[0021] Compared with existing technical solutions, the present invention has the following beneficial effects: This invention effectively integrates multiple types of sensor data in the wind turbine bearing system, overcoming the technical defects of relying on single-point temperature trend judgment. By integrating the physical coupling relationship and dynamic data correlation between multiple types of sensor data, more complete and real-time correlation data between various sensors can be obtained, constructing a multi-dimensional coupling relationship, thereby strengthening the collaborative evolution relationship between multiple variables. The dynamic data correlation can be dynamically updated and can quickly match the actual working conditions.

[0022] The data features extracted by this invention can reflect the real-time data, short-term dynamic change trends, and time-series statistical patterns of sensors, containing richer and more detailed temporal and spatial characteristics, effectively matching the complex on-site operating conditions of wind turbine bearing systems. The operating conditions during oil temperature control are extremely complex, with multiple slow and abrupt changes coexisting. These include slow drift conditions such as mechanical wear (e.g., bearing surface wear caused by long-term friction) and lubrication degradation (e.g., increased lubricating oil viscosity due to continuous long-term operation), as well as sudden transitions such as cooling system failures caused by malfunctions. The real-time data extracted by this invention can characterize the real-time state of the sensor, supporting the construction of spatial coupling relationships and real-time updates of dynamic data correlations; short-term dynamic change trends can reflect the short-term rate and direction of change, are particularly sensitive to "sudden transition anomalies," and can promptly reflect early abnormal signs; and time-series statistical patterns are more likely to capture long-term trends of slow drift. Both can not only assist and optimize dynamic data correlations but also serve as time features. The capture of data features and the fusion of physical coupling relationships and dynamic data correlations enable the control process to simultaneously perceive both gradual and sudden anomalies, significantly enhancing the system's ability to identify oil temperature anomalies under complex operating conditions and making it more sensitive to environmental disturbances. The spatiotemporal feature data of the sensors used as the basis for prediction in this invention not only considers the oil temperature sequence itself and effectively integrates the spatial structure and dynamic coupling relationship between oil temperature and other sensors, but also superimposes the features captured in time, making it more sensitive to complex faults (such as gradual changes + sudden changes), thereby achieving higher accuracy and earlier prediction, effectively dealing with complex oil temperature evolution processes. In addition, the prediction results are organically unified with the control strategy, overcoming the defect of the separation between prediction and control strategies in the prior art, making the control more precise and effective.

[0023] Furthermore, the present invention collects feedback data from the oil temperature regulating device and optimizes the control strategy by referring to the feedback data, thereby realizing a closed-loop control of the entire process of "sensing-prediction-regulation-feedback", resulting in better control effect.

[0024] Furthermore, this invention can use a neural network to predict the oil temperature at future moments; the neural network is deployed at the edge of the wind turbine site, and its training process is deployed in the cloud, which enables the edge to operate efficiently, with a single inference latency controlled within 50 milliseconds, meeting the application requirements of wind farms for real-time performance and low power consumption. Attached Figure Description

[0025] Figure 1 This is a flowchart of the control method described in the first specific embodiment of the present invention; Figure 2 A system diagram illustrating the control method described in the fourth specific embodiment of the present invention; Figure 3 This is a system structure diagram of a wind turbine oil temperature control system based on sensor spatiotemporal fusion, as described in this invention. Figure 4 This is a system structure diagram of another wind turbine oil temperature control system based on sensor spatiotemporal fusion described in this invention; Figure 5 This is a system structure diagram of another wind turbine oil temperature control system based on sensor spatiotemporal fusion described in this invention. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application. This invention covers any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined by the claims.

[0027] like Figure 1 As shown, the first specific embodiment of the control method of the present invention is as follows: A method for controlling oil temperature in wind turbine generators based on sensor spatiotemporal fusion includes the following steps: Step 1: Collect raw data from various types of sensors in the wind turbine bearing system in real time, and perform data cleaning to obtain sensor input data; the raw sensor data includes bearing lubricating oil temperature data; Step 2: Based on the sensor input data obtained in Step 1, set a time window, extract the data features within the time window, and obtain sensor feature data; the data features are used to reflect the sensor's real-time data, short-term dynamic change trends, and time-series statistical patterns. Step 3: Based on the sensor feature data obtained in Step 2, obtain the dynamic data correlation between sensors, and combine the physical coupling relationship between the sensors on site to generate correlation data between the sensors. Step 4: Based on the sensor feature data obtained in Step 2, and according to the correlation data, generate sensor spatiotemporal feature data containing spatial and temporal features, and use the sensor spatiotemporal feature data as input to predict the oil temperature at future times, thereby obtaining the oil temperature prediction value. Step 5: Based on the predicted oil temperature, a control strategy is generated, and a control signal for controlling the oil temperature regulating device is output according to the control strategy. The oil temperature regulating device can be a cooling pump, a proportional regulating valve, and an oil pump motor to achieve precise control of equipment heat dissipation and oil circulation.

[0028] A specific embodiment of the present invention for data acquisition and cleaning: Step 1: Real-time acquisition of raw data from various types of sensors in the wind turbine bearing system, followed by data cleaning. The method for obtaining sensor input data is as follows: Step 11: Sample the raw data of various types of sensors in the wind turbine bearing system at a fixed frequency and mark the time. The raw sensor data includes bearing lubricating oil temperature data, environmental data, and related equipment operation data. Among them, environmental data may include ambient temperature, and related equipment operation data may include instantaneous wind speed, wind turbine speed, winding temperature, generator front bearing temperature, and grid A-phase voltage. Step 12: Remove outliers from the data collected in Step 11; Step 13: Perform min-max normalization on the data obtained in Step 12 and fill in the missing values ​​to obtain the sensor input data.

[0029] A specific embodiment of the feature extraction method of the present invention: In the second step, the method for extracting data features within the time window to obtain sensor feature data includes the following steps: Step 21: Set a sliding window for the sensor input data within the time window and slide the sliding window in fixed steps; Step 22: Calculate the multi-dimensional feature vector of the sensor input data in each sliding window. The multi-dimensional feature vector includes the normalized value of the real-time sensor data, the time difference between adjacent sensor times, and the mean, variance, and skewness of the sensor in the sliding window. Step 23: Based on all the multi-dimensional feature vectors obtained in Step 22, output the sensor feature matrix, which is the sensor feature data.

[0030] Among them, the normalized value of the sensor's real-time data is used to reflect the sensor's real-time data; The time difference between adjacent moments of a sensor is used to reflect the short-term dynamic change trend of the sensor. The time difference directly describes the magnitude and direction of the change in sensor data between adjacent sampling moments. The mean, variance, and skewness of the sensor within the sliding window are used to reflect the time-series statistical regularity of the sensor. The mean reflects the average operating condition of the sensor within the sliding window and can be used as a benchmark to determine whether the current data deviates from the normal range; the variance reflects the fluctuation stability of the sensor data within the sliding window and can promptly capture "fluctuation-type anomalies"; the skewness reflects the symmetry of the sensor data distribution within the sliding window, quantifies the direction and degree of data distribution deviation, and can effectively identify "slow-change anomalies".

[0031] A specific embodiment of the present invention for generating correlated data: Step 3: Based on the sensor feature data obtained in Step 2, the dynamic data correlation between sensors is obtained, and the physical coupling relationship between the sensors in the field is comprehensively considered to generate the correlation data between the sensors. The method is as follows: Step 31: Construct a graph structure with each sensor as a vertex and the adjacency relationships between sensors as edges; Step 32: Assign static weights to the edges of the graph structure, wherein the static weights are determined based on the physical coupling relationships between the sensors in the field; Step 33: Based on sensor feature data, set up a dynamic calculation window and calculate the correlation index between each sensor within the dynamic calculation window to obtain dynamic weights; Step 34: Fuse the static weights and dynamic weights to obtain new edge weights, and update the edge weights of the graph structure; Step 35: Output the sensor adjacency matrix corresponding to the graph structure, i.e., the correlation data.

[0032] A specific embodiment of the present invention for generating spatiotemporal feature data of sensors: Step 4: Based on the sensor feature data obtained in Step 2, and according to the correlation data, the method for generating sensor spatiotemporal feature data containing spatial and temporal features is as follows: Step 41: Based on sensor feature data, for one of the sensors, calculate its attention score with each of its neighboring sensors according to the correlation data; Step 42: Process the attention scores of all neighboring sensors of this sensor to obtain the attention weights; Step 43: Change different attention to focus on different coupling modes between sensors, repeat steps 41 and 42 multiple times to obtain multiple attention weights; Step 44: Based on the multiple attention weights obtained in Step 43, the data features of neighboring sensors are weighted and aggregated onto the sensor to obtain the spatial aggregated feature value of the sensor; Step 45: Calculate the spatial aggregated feature values ​​of all sensors and output the sensor spatial feature matrix accordingly; Step 46: Based on sensor feature data, set the capture time for multiple layers with different granularities; Step 47: Capture the sensor time feature values ​​in each time layer and output the sensor time feature matrix accordingly; Step 48: Fuse the sensor spatial feature matrix obtained in step 45 and the sensor temporal feature matrix obtained in step 47 to obtain the sensor spatiotemporal feature matrix, that is, the sensor spatiotemporal feature data.

[0033] A second specific embodiment of the control method of the present invention: A wind turbine oil temperature control method based on sensor spatiotemporal fusion, in addition to the first specific embodiment, further includes in the fourth step: using sensor spatiotemporal feature data as input to predict the probability of anomalies at future times, and obtaining the predicted anomaly probability value.

[0034] The process also includes a sixth step: collecting feedback data from the oil temperature regulation device and optimizing the control strategy based on this feedback data. The feedback data may include the cooling pump frequency, valve opening, and actual flow rate.

[0035] A specific embodiment of the optimization method of the present invention: Step 6: The method for collecting feedback data from the oil temperature regulation device and optimizing the control strategy based on this feedback data is as follows: Step 61: Collect feedback data from the oil temperature regulating device in real time and calculate the energy consumption index of the oil temperature regulating device; Step 62: Calculate the safety index based on the anomaly probability prediction value obtained in step 4; Step 63: Construct a reward function with energy consumption and safety as control objectives, and optimize the control strategy based on the energy consumption index obtained in Step 61 and the safety index obtained in Step 62.

[0036] This embodiment describes a wind turbine oil temperature control method based on sensor spatiotemporal fusion, achieving closed-loop control throughout the entire process of "sensing-prediction-control-feedback," resulting in better control performance. By predicting the probability of anomalies, alarm signals can be provided in a timely manner. The control strategy is optimized using a reward function with energy consumption and safety as control objectives. The reward function penalizes high-risk and high-energy-consumption behaviors, which can synergistically reduce fault risk and energy loss, avoid overcooling, and improve system energy efficiency.

[0037] A third specific embodiment of the control method of the present invention: A wind turbine oil temperature control method based on sensor spatiotemporal fusion, in the fourth step of the first specific embodiment, the prediction of oil temperature at future moments refers to the prediction of oil temperature at future moments through a neural network; the neural network is deployed at the edge of the wind turbine site, and its training process is deployed in the cloud. The cloud performs centralized modeling and training on the data uploaded by the edge and sends the training result parameters back to the edge; the cloud training includes: offline training, incremental training, and cross-scene transfer.

[0038] This embodiment describes a wind turbine oil temperature control method based on sensor spatiotemporal fusion. The neural network is deployed at the edge of the wind turbine site, and its training process is deployed in the cloud, enabling efficient operation at the edge. The single inference latency is controlled within 50 milliseconds, meeting the application requirements of wind farms for real-time performance and low power consumption. In the cloud environment, a GPU cluster is used to centrally model and train the perception and inference data uploaded from the edge, improving the model's accuracy and generalization ability. During the incremental training phase, the latest data collected by the edge is automatically synchronized daily, and small-scale parameter updates are performed to maintain the model's continuous learning ability. Cross-scenario migration facilitates distributed deployment, upgrade maintenance, and system expansion in industrial applications.

[0039] like Figure 2 As shown, the fourth specific embodiment of the control method of the present invention is as follows: A method for controlling oil temperature in wind turbine generators based on sensor spatiotemporal fusion includes the following steps: Step 1: Real-time acquisition of raw data from various types of sensors in the wind turbine bearing system, followed by data cleaning to obtain sensor input data; the raw sensor data includes bearing lubricating oil temperature data, as detailed below: This step is located in the perception layer. The raw data from multiple sensor types includes eight categories of signals: wind turbine oil temperature, instantaneous wind speed, wind turbine speed, ambient temperature, winding temperature, generator front bearing temperature, and grid A-phase voltage. The data is collected once per minute, and a vector is generated every minute. Where N is the total number of sensors, each All include a timestamp.

[0040] Data cleaning includes outlier removal, normalization, and missing value imputation; specifically, it uses the improved Z-score method to remove outliers, i.e., when... When outliers are removed, MAD is the median absolute deviation. Then, each sensor sequence is subjected to min-max normalization to uniformly scale it to the [0,1] interval, resulting in normalized sensor data. The calculation formula is as follows:

[0041] Missing values ​​are filled using a strategy based on the duration of the missing value: linear interpolation is used for values ​​less than 5 minutes, and historical averages are used for values ​​greater than 5 minutes.

[0042] Step 2: Based on the sensor input data obtained in Step 1, set a time window, extract the data features within this time window, and obtain sensor feature data; the data features are used to reflect the sensor's real-time data, short-term dynamic change trends, and time-series statistical patterns, as follows: A time window T = 60 minutes is used to generate a 5-dimensional feature vector for each sensor within a sliding window that occurs once every minute. Including normalized values ​​and time differences Including statistical characteristics (mean, variance, skewness), the final output includes a normalized time series matrix. (T=60 minutes) and node feature matrix .

[0043] Step 3: Based on the sensor feature data obtained in Step 2, obtain the dynamic data correlation between sensors, and integrate the physical coupling relationship between the sensors in the field to generate correlation data between the sensors, as follows: This step is located at the edge, and the input is the normalized time series matrix output from the second step. Corresponding node feature matrix Then, a dynamic spatiotemporal graph structure is constructed to reflect the real-time correlation between multiple sensors. The set of vertices V, composed of sensor nodes, is used as the initial edge set, and the physical coupling relationships between different sensor data are set as the initial edge set. Initial edge weights This was determined based on prior engineering information. Subsequently, within a sliding time window (T=60 minutes), the Pearson correlation coefficient between node pairs was calculated. This is to capture the dynamic correlation of multi-source signals. The dynamic update formula for edge weights is:

[0044] After experimental optimization, the weight ratios were set to 0.6 and 0.4. Indicates time At any given moment, sensor nodes and The final weight of the edges between them; Represents sensor nodes and The initial physical connection weights between them are determined based on prior engineering knowledge. This indicates the sensor's performance over a past period of time (e.g., T=60 minutes). and The Pearson correlation coefficient of the data time series is used to quantify the statistical correlation between the two within the time window. To improve computational efficiency and robustness, edges with weights below 0.1 are pruned, ultimately generating a time-varying graph structure. The graph structure is further output as an adjacency matrix. That is, related data.

[0045] Step 4: Based on the sensor feature data obtained in Step 2, and according to the correlation data, generate sensor spatiotemporal feature data containing spatial and temporal features. Then, use this sensor spatiotemporal feature data as input to predict the oil temperature and anomaly probability at future times, obtaining the predicted oil temperature and anomaly probability values, as detailed below: This step relies on the spatiotemporal graph neural network STGNN deployed at the edge to perform real-time reasoning on the future operating state. Its network structure includes the following three main modules: spatial convolution module, temporal convolution module, and spatiotemporal fusion module. In the spatial convolution module, the multi-head graph attention mechanism is used to dynamically weight and aggregate neighbor node information. The processing flow is as follows: The input is the feature matrix H={ of all nodes in the graph. , ,…, }, where the feature vector of each node i is Where: N is the number of nodes, and F is the feature dimension; Each node feature is transformed using a shared linear transformation matrix. Perform mapping to obtain new node features : ,

[0046] The attention coefficient (single head) between nodes is calculated using the following formula for each pair of adjacent nodes. Calculate the unnormalized attention score :

[0047] in: This is a trainable attention weight vector; Let be the input feature vectors for nodes i and j, representing the state of each sensor node at the current time. is a trainable weight matrix; the weight matrix has dimensions F´×F; where F´ is the output dimension and F is the input dimension; it is used to map the input features to a higher-order feature space; This means concatenating the features of nodes i and j and using them as input, with a dimension of 2F'.

[0048] Based on the following formula, for all neighboring nodes of node i The attention scores are normalized to obtain the weights:

[0049] in The final normalized attention weight represents the degree of influence of node j on node i (the proportion among all neighbors). Let i be the set of neighboring nodes of node i.

[0050] The above steps are executed independently K times, generating K sets of attention weights and node feature representations. Each head output is:

[0051] in, It is the activation function ReLU, k=1,2,…,K.

[0052] The multi-head outputs are spliced ​​and aggregated according to the following formula:

[0053] in, Indicates feature concatenation operation; It is the representation of node i calculated by the k-th attention head; K is the number of attention heads; The outputs of all attention heads are concatenated to form a higher-dimensional feature vector.

[0054] The output of the spatial convolution module is the feature matrix. ={ , …, The data is then passed to the temporal convolution module for further spatiotemporal feature fusion.

[0055] The processing flow of the temporal convolution module is as follows: The input to the temporal convolution module is a matrix whose output feature dimension is C, the sampling time window is T, and the number of wind turbine nodes is N. Where: N is the number of nodes (number of wind turbine units); C is the current feature dimension of each node; T is the time step length; then temporal convolution is performed, including dilated convolution, nonlinear activation and regularization mechanisms; The dilated convolution uses dilation factors {1, 2, 4, 8}, corresponding to different temporal capture capabilities at various granularities across multiple layers. The nonlinear activation function ReLU is used to introduce nonlinearity and enhance the model's representational power. LayerNorm is employed for regularization to improve training stability, and Dropout controls overfitting.

[0056] The final output shape of the temporal convolution module :

[0057] Where C´ is the number of output channels; T´ is the size of the time dimension.

[0058] The spatiotemporal fusion module fuses spatial and temporal information through a spatiotemporal attention mechanism to generate a comprehensive representation, including attention weights. The calculation is as follows:

[0059] The spatiotemporal characteristics after fusion are After mapping through a fully connected layer, the predicted time points are output respectively. Oil temperature estimate (Where M is the number of temperature sensors) and the probability value of an anomaly occurring. ∈[0,1].

[0060] The above inference process runs efficiently on the edge computing unit, with a single inference latency controlled within 50 milliseconds, meeting the application requirements of wind farms for real-time performance and low power consumption.

[0061] Step 5: Based on the predicted oil temperature, generate a control strategy and output a control signal for controlling the oil temperature regulating device according to the control strategy; Step 6: Collect feedback data from the oil temperature regulation device and optimize the control strategy based on this feedback data, as follows: The control policy is generated by a reinforcement learning controller based on deep deterministic policy gradient (DDPG), with the predicted oil temperature as input. With anomaly probability The output is the key control variable: cooling pump frequency. Oil pump flow rate The control objective is optimized using the following reward function:

[0062] in: To determine the safe temperature threshold, after experimental optimization, the weighting coefficients were set to α=0.7 and β=0.3. Cooling pump frequency ; Oil pump flow rate ; and These are the equipment characteristic coefficients for the cooling pump and oil pump, respectively, and their values ​​are... 0.001 =0.2.

[0063] This step involves an edge-deployed execution control unit outputting control signals according to a strategy to drive a cooling pump (frequency range 20–60Hz), a proportional control valve (opening degree 0–100%), and an oil pump motor (flow rate 0–100L / min), achieving precise control over equipment heat dissipation and oil circulation. Simultaneously, the system collects real-time data on the operating status of the oil temperature regulation device (such as pump frequency, valve opening degree, and actual flow rate) and the latest round of sensor data, feeding this data back to the perception layer to update the dynamic graph structure and model input, thus achieving closed-loop evolution of the predictive and control system. The final output includes a set of control signals. and feedback data .

[0064] In this embodiment, an MQTT-based industrial message bus is used to connect the data flow paths of the perception layer, the edge inference end, and the execution control end, forming a complete data closed loop. The data flow path is: perception layer, edge end (graph construction and STGNN inference), execution layer (issuing control commands), and perception layer (acquiring feedback data). This closed loop automatically updates graph structure edge weights and adjusts STGNN parameters while providing data feedback, thereby achieving a self-learning mechanism and building a continuously evolving digital twin system. The final output includes real-time updated oil temperature prediction results, anomaly probability estimates, and energy-optimized control signals.

[0065] Furthermore, the training of the neural network mentioned above is conducted in the cloud. In the cloud environment, GPU clusters are used to centrally model and train the perception and inference data uploaded from the edge, improving the model's accuracy and generalization ability. The training process consists of three stages: offline training, incremental training, and cross-scene transfer.

[0066] First, during the offline training phase, the system comprehensively trains the Spatiotemporal Graph Neural Network (STGNN) based on historical labeled data (including oil temperature time series and anomaly labels). The optimization objective function is: ; Where: MSE represents the mean square error between the predicted value and the actual oil temperature; BCE is the probability of prediction anomalies. With actual label The binary cross-entropy between them; λ is the regularization coefficient, set to 0.1.

[0067] Secondly, during the incremental training phase, the system automatically synchronizes the latest data collected from the edge devices daily, performing small-scale parameter updates to maintain the model's continuous learning capability. This process typically takes less than 10 minutes, adapting to the pace of industrial deployment.

[0068] Finally, to achieve cross-scenario transfer of the model between different wind turbines, three strategies are proposed: first, graph structure fine-tuning, which updates only the edge weights in the graph. The first approach maintains the node feature mapping layer structure unchanged; the second approach introduces an adversarial domain adaptation mechanism, using a gradient inversion layer (GRL) to enable feature alignment on unlabeled new unit data. Its adversarial loss function is:

[0069]

[0070] It is the adversarial domain adaptation total loss, used to achieve consistency in feature distribution between the source and target domains. It is the prediction task loss, used to ensure the model's prediction accuracy and anomaly detection capability in the source domain. It is a domain classification loss, used to distinguish whether an input sample comes from the source domain or the target domain.

[0071] Wherein: γ is the tradeoff coefficient, set to 0.2; thirdly, an empirical replay strategy is adopted to retain key samples from the source domain in order to prevent catastrophic forgetting during the migration process.

[0072] The above training process outputs include: oil temperature prediction values ​​for the next 3 hours. Anomaly probability The updated model parameters are then deployed back to the edge for continuous iterative optimization.

[0073] This embodiment constructs a closed-loop control system around five stages: "data perception, graph modeling, intelligent reasoning, regulation and execution, and feedback evolution," achieving the following significant technical effects in bearing lubricating oil temperature anomaly prediction and energy efficiency control: First, the accuracy of oil temperature prediction is significantly improved. A dynamically constructed spatiotemporal graph structure is adopted, fusing the physical coupling relationships and temporal correlations between sensors. Through joint modeling using graph attention (GAT) and gated temporal convolution (Gated TCN), accurate modeling of the oil temperature evolution process is achieved. Compared to traditional models such as LSTM, the prediction accuracy is improved by more than 15%, and potential abnormal trends can be identified several hours in advance.

[0074] Second, it has a stronger ability to perceive complex anomalies; through the dynamic graph mechanism constructed by weighting static edges (physical connections) and dynamic edges (correlation coefficients), the model has the dual ability to perceive anomalies of gradual change (such as lubrication degradation) and sudden change (such as cooling failure), which significantly enhances the system's robustness in identifying oil temperature anomalies under complex working conditions.

[0075] Third, the control strategy combines intelligence and energy efficiency. Utilizing the Deep Deterministic Policy Gradient (DDPG) algorithm, continuous control commands are generated under the joint drive of predicted values ​​and anomaly probabilities, effectively achieving precise control of the cooling pump frequency and oil pump flow rate. The reward function simultaneously considers energy consumption and safety objectives, avoiding overcooling, improving system energy efficiency, and achieving energy savings of over 12%.

[0076] Fourth, the system has closed-loop evolution and self-learning capabilities; the system adopts an edge inference + cloud training architecture, with inference latency controlled within 50 milliseconds to meet the real-time control requirements of wind farms; at the same time, it realizes data closed-loop through MQTT bus, feeding the execution results back to the prediction module, supporting the online evolution and update of dynamic graph structure and model parameters, and has continuous learning capabilities.

[0077] Fifth, it is suitable for deployment in multiple scenarios and has good engineering adaptability; the system structure is clear, divided into a perception layer, an edge computing layer and a cloud training layer, and supports modular deployment and distributed expansion; the model has the ability to migrate across wind fields, and can quickly adapt to different models and operating conditions through graph structure fine-tuning and adversarial domain adaptation mechanisms to meet the needs of industrial-scale applications.

[0078] In summary, this invention has significant technical advantages in terms of prediction accuracy, anomaly detection capability, intelligent control effect, system real-time performance, and engineering deployability, and has broad application value.

[0079] like Figure 3 As shown, a specific embodiment of the control system of the present invention is as follows: A wind turbine oil temperature control system based on sensor spatiotemporal fusion, which adopts the wind turbine oil temperature control method based on sensor spatiotemporal fusion described in the first embodiment; The oil temperature control system for this wind turbine includes the following modules: The data acquisition and cleaning module is used to acquire raw data from various types of sensors in the wind turbine bearing system in real time, and to clean the data to obtain sensor input data. The feature data extraction module is used to set a time window based on the sensor input data, extract the data features within the time window, and obtain sensor feature data. The correlation data generation module is used to obtain the dynamic data correlation between sensors based on sensor feature data, and to generate correlation data between sensors by integrating the physical coupling relationship between various sensors in the field. The spatiotemporal feature data generation and prediction module is used to generate sensor spatiotemporal feature data containing spatial and temporal features based on sensor feature data and the correlation data, and to predict the oil temperature at future times using the sensor spatiotemporal feature data as input to obtain the oil temperature prediction value. And a control module, used to generate a control strategy based on the oil temperature prediction value, and output a control signal for controlling the oil temperature regulating device according to the control strategy.

[0080] like Figure 4 As shown, based on the wind turbine oil temperature control system described in this embodiment, an optimization module may also be included, which is used to collect feedback data from the oil temperature regulation device and optimize the control strategy with reference to the feedback data.

[0081] like Figure 5 As shown, based on the wind turbine oil temperature control system described in this embodiment, it may also include a cloud platform. The spatiotemporal feature data generation and prediction module predicts the oil temperature at future moments using a neural network. This neural network is deployed at the edge of the wind turbine site, and its training process is deployed in the cloud platform. The cloud platform performs centralized modeling and training on the data uploaded from the edge platform and sends the training result parameters back to the edge platform. The training in the cloud platform includes: offline training, incremental training, and cross-scenario transfer.

[0082] An embodiment of an electronic device applying the method of the present invention: An electronic device comprising: One or more processing units; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processing units, the one or more processing units implement the above-described method for controlling the oil temperature of wind turbine generators based on sensor spatiotemporal fusion.

[0083] The storage device can be internal memory, external memory, cache memory, or other special memory. The processing unit has signal processing capabilities and can be a general-purpose processor, a digital signal processor, an application-specific integrated circuit, an off-the-shelf programmable gate array, or other programmable logic device.

[0084] An embodiment of a medium using the method of the present invention: A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for controlling oil temperature in wind turbine generators based on sensor spatiotemporal fusion.

[0085] Computer-readable storage media refers to physical carriers capable of storing computer-recognizable data, instructions, or programs. These media must meet the core characteristic of being "readable by a computer" (i.e., the data exists in the form of electrical, magnetic, or optical signals and can be converted into binary information that a computer can process through appropriate devices). The physical carrier can be a magnetic storage medium, optical storage medium, semiconductor storage medium, or other storage media.

[0086] Finally, it should be noted that the above embodiments are merely specific implementations of the present invention and are not intended to limit it. The scope of protection of the present invention is not limited thereto and should be determined by the scope of the claims.

Claims

1. A method for controlling oil temperature in wind turbine generators based on sensor spatiotemporal fusion, characterized in that: Includes the following steps: Step 1: Collect raw data from various types of sensors in the wind turbine bearing system in real time, and perform data cleaning to obtain sensor input data; the raw sensor data includes bearing lubricating oil temperature data; Step 2: Based on the sensor input data obtained in Step 1, set a time window, extract the data features within the time window, and obtain sensor feature data; the data features are used to reflect the sensor's real-time data, short-term dynamic change trends, and time-series statistical patterns. Step 3: Based on the sensor feature data obtained in Step 2, obtain the dynamic data correlation between sensors, and combine the physical coupling relationship between the sensors on site to generate correlation data between the sensors. Step 4: Based on the sensor feature data obtained in Step 2, and according to the correlation data, generate sensor spatiotemporal feature data containing spatial and temporal features, and use the sensor spatiotemporal feature data as input to predict the oil temperature at future times, thereby obtaining the oil temperature prediction value. Step 5: Based on the predicted oil temperature, generate a control strategy and output a control signal for controlling the oil temperature regulating device according to the control strategy; The third step, based on the sensor feature data obtained in the second step, is to obtain the dynamic data correlation between sensors and, by integrating the physical coupling relationships between the sensors in the field, generate the correlation data between the sensors as follows: Step 31: Construct a graph structure with each sensor as a vertex and the adjacency relationships between sensors as edges; Step 32: Assign static weights to the edges of the graph structure, the static weights being determined based on the physical coupling relationships between the sensors in the field; Step 33: Based on the sensor feature data, set a dynamic calculation window and calculate the correlation index between the sensors within the dynamic calculation window to obtain dynamic weights; Step 34: Fuse the static weights and dynamic weights to obtain new edge weights and update the edge weights of the graph structure; Step 35: Output the sensor adjacency relationship matrix corresponding to the graph structure, i.e., the correlation data. The fourth step, based on the sensor feature data obtained in the second step and according to the correlation data, generates sensor spatiotemporal feature data containing spatial and temporal features as follows: Step 41: Based on the sensor feature data, for one sensor, calculate its attention score with each neighboring sensor according to the correlation data; Step 42: Process the attention scores of all neighboring sensors of the sensor to obtain attention weights; Step 43: Change different attention to focus on different coupling modes between sensors, repeat steps 41 and 42 multiple times to obtain multiple attention weights; Step 44: According to the multiple attention weights obtained in step 43, weighted aggregate the data features of neighboring sensors to the sensor to obtain the spatial aggregated feature value of the sensor; Step 45: Calculate the spatial aggregated feature values ​​of all sensors and output the sensor spatial feature matrix accordingly; Step 46: Based on the sensor feature data, set multiple layers of capture time with different granularities; Step 47: Capture the sensor temporal feature values ​​in each time layer and output the sensor temporal feature matrix accordingly; Step 48: Fuse the sensor spatial feature matrix obtained in step 45 and the sensor temporal feature matrix obtained in step 47 to obtain the sensor spatiotemporal feature matrix, i.e., the sensor spatiotemporal feature data. The fourth step involves using a spatiotemporal graph neural network (STGNN) deployed at the edge to perform real-time reasoning about future operating states. Its network structure includes three main modules: a spatial convolution module, a temporal convolution module, and a spatiotemporal fusion module. The spatial convolution module employs a multi-head graph attention mechanism for dynamically weighted aggregation of neighbor node information. The temporal convolution module takes a matrix as its input. Where N is the number of nodes; C is the current feature dimension of each node; T is the time step length, and temporal convolution is performed, which includes dilated convolution, nonlinear activation and regularization mechanisms; the spatiotemporal fusion module fuses spatial and temporal information through a spatiotemporal attention mechanism to generate a comprehensive representation.

2. The wind turbine oil temperature control method based on sensor spatiotemporal fusion as described in claim 1, characterized in that: Step 1: Real-time acquisition of raw data from various types of sensors in the wind turbine bearing system, followed by data cleaning. The method for obtaining sensor input data is as follows: Step 11: Sample the raw data of various types of sensors in the wind turbine bearing system at a fixed frequency and mark the time. The raw sensor data includes bearing lubricating oil temperature data, environmental data, and related equipment operation data. Step 12: Remove outliers from the data collected in Step 11; Step 13: Normalize the data obtained in Step 12 and fill in the missing values ​​to obtain the sensor input data.

3. The wind turbine oil temperature control method based on sensor spatiotemporal fusion as described in claim 1, characterized in that: In the second step, the method for extracting data features within the time window to obtain sensor feature data includes the following steps: Step 21: Set a sliding window for the sensor input data within the time window and slide the sliding window in fixed steps; Step 22: Calculate the multi-dimensional feature vector of the sensor input data in each sliding window. The multi-dimensional feature vector includes the normalized value of the real-time sensor data, the time difference between adjacent sensor times, and the mean, variance, and skewness of the sensor in the sliding window. Step 23: Based on all the multi-dimensional feature vectors obtained in Step 22, output the sensor feature matrix, which is the sensor feature data.

4. The wind turbine oil temperature control method based on sensor spatiotemporal fusion as described in claim 1, characterized in that: The fourth step also includes: using the spatiotemporal feature data of the sensor as input to predict the probability of anomalies at future times, and obtaining the predicted value of the anomaly probability.

5. The wind turbine oil temperature control method based on sensor spatiotemporal fusion as described in claim 4, characterized in that: The sixth step also includes: collecting feedback data from the oil temperature regulation device and optimizing the control strategy based on this feedback data.

6. The wind turbine oil temperature control method based on sensor spatiotemporal fusion as described in claim 5, characterized in that: Step 6: The method for collecting feedback data from the oil temperature regulation device and optimizing the control strategy based on this feedback data is as follows: Step 61: Collect feedback data from the oil temperature regulating device in real time and calculate the energy consumption index of the oil temperature regulating device; Step 62: Calculate the safety index based on the anomaly probability prediction value obtained in step 4; Step 63: Construct a reward function with energy consumption and safety as control objectives, and optimize the control strategy based on the energy consumption index obtained in Step 61 and the safety index obtained in Step 62.

7. A wind turbine oil temperature control method based on sensor spatiotemporal fusion as described in any one of claims 1-6, characterized in that: In the fourth step, the prediction of oil temperature at future moments refers to the prediction of oil temperature at future moments through a neural network. This neural network is deployed at the edge of the wind turbine site, and its training process is deployed in the cloud. The cloud performs centralized modeling and training on the data uploaded by the edge and sends the training result parameters back to the edge. The cloud-based training includes: offline training, incremental training, and cross-scenario transfer.

8. A wind turbine oil temperature control system based on sensor spatiotemporal fusion, characterized in that: It employs the wind turbine oil temperature control method based on sensor spatiotemporal fusion as described in any one of claims 1-7; The oil temperature control system for this wind turbine includes the following modules: The data acquisition and cleaning module is used to acquire raw data from various types of sensors in the wind turbine bearing system in real time, and to clean the data to obtain sensor input data. The feature data extraction module is used to set a time window based on the sensor input data, extract the data features within the time window, and obtain sensor feature data. The correlation data generation module is used to obtain the dynamic data correlation between sensors based on sensor feature data, and to generate correlation data between sensors by integrating the physical coupling relationship between various sensors in the field. The spatiotemporal feature data generation and prediction module is used to generate sensor spatiotemporal feature data containing spatial and temporal features based on sensor feature data and the correlation data, and to predict the oil temperature at future times using the sensor spatiotemporal feature data as input to obtain the oil temperature prediction value. And a control module, used to generate a control strategy based on the oil temperature prediction value, and output a control signal for controlling the oil temperature regulating device according to the control strategy.

Images