Wind power generation equipment state monitoring method and system combining Internet of Things and machine learning

Abstract

The invention provides a wind power generation equipment state monitoring method and system combining the Internet of Things and machine learning, and relates to the technical field of wind power generation equipment monitoring. A real-time operation data set containing physical state data and environment influence data is collected, then a time sequence association network of the real-time operation data set, an operation scene and a component operation state is established, and then a hierarchical machine learning reasoning model is called to carry out collaborative reasoning on the real-time operation data set and the time sequence association network; and generating a real-time state evaluation result of each component, generating a component operation regulation and control scheme according to the real-time state evaluation result of each component, transmitting the real-time state evaluation result of each component and the component operation regulation and control scheme to the equipment control terminal, and updating a time sequence association network and reasoning model parameters in combination with a feedback regulation and control result. The equipment state can be accurately monitored, the equipment operation is effectively regulated and controlled, and the operation efficiency and reliability are improved.

Description

Technical Field

[0001] The invention relates to the field of wind power equipment monitoring technology, and more specifically, to a method and system for monitoring the condition of wind power equipment that combines the Internet of Things and machine learning. Background Technology

[0002] In the field of wind power generation, the stable operation of wind power equipment is crucial for ensuring the reliability and efficiency of power supply. With the continuous expansion of wind power equipment scale and increasing complexity, accurate monitoring and effective control of the status of its various components has become an urgent problem to be solved.

[0003] Currently, traditional methods for monitoring the condition of wind power equipment mainly rely on periodic manual inspections and simple sensor data collection. Manual inspections are not only time-consuming and labor-intensive, but also struggle to detect potential problems in equipment components in real time, making it difficult to take timely and effective preventative measures. Simple sensor data collection typically only monitors certain specific parameters of equipment components, lacking a comprehensive consideration of the component's operating environment. Because the physical state and environmental influences of various components in wind power equipment change under different operating scenarios, single-parameter monitoring cannot comprehensively and accurately reflect the actual operating status of the components. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for monitoring the condition of wind power generation equipment by combining the Internet of Things and machine learning, the method comprising:

[0005] Based on the dynamic configuration of IoT sensing nodes for the operation scenarios of various components of wind power generation equipment, a set of real-time operation data is collected. The operation scenarios include the physical action mode and environmental interaction mode of the components during operation, and the set of real-time operation data includes the physical state data and environmental impact data generated by the components under different operation scenarios.

[0006] A time-series correlation network is established between the real-time operating data set and the operating scenarios and component operating states. The time-series correlation network is used to characterize the component operating states corresponding to the changes of the real-time operating data set over time under different operating scenarios.

[0007] A hierarchical machine learning inference model is invoked to perform collaborative inference processing on the real-time running data set and time series association network to generate real-time status evaluation results for each component of the wind power generation equipment.

[0008] A component operation control plan is generated based on the real-time status assessment results. The component operation control plan includes component operation specifications for different operating states.

[0009] The real-time status assessment results and component operation control schemes are transmitted to the equipment control terminal. Based on the component control results fed back by the equipment control terminal, the inference parameters of the time-series correlation network and hierarchical machine learning inference model are updated.

[0010] Furthermore, embodiments of the present invention also provide a wind power generation equipment condition monitoring system combining the Internet of Things and machine learning, comprising:

[0011] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the above-described method for monitoring the condition of wind power generation equipment combining the Internet of Things and machine learning by executing the machine-executable instructions.

[0012] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a wind power generation equipment condition monitoring system combining Internet of Things and machine learning reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the wind power generation equipment condition monitoring system combining Internet of Things and machine learning to execute the above-described wind power generation equipment condition monitoring method combining Internet of Things and machine learning.

[0013] Based on the above, by dynamically configuring IoT sensing nodes according to the operating scenarios of various components of wind power generation equipment, a real-time operating data set is collected. This set includes physical state data and environmental impact data generated by components under different operating scenarios, taking into account the physical action mode of components during operation and their interaction with the environment. Then, a time-series correlation network between the real-time operating data set and the operating scenarios and component operating states is established. The intrinsic relationship between data and component states is analyzed in depth from the dimensions of time and correlation. It can present the component operating states corresponding to the changes of the real-time operating data set over time under different operating scenarios. A hierarchical machine learning inference model is called to perform collaborative inference processing on the real-time operating data set and the time-series correlation network, giving full play to the advantages of machine learning in data processing and pattern recognition, and generating accurate and detailed real-time status assessment results for various components of wind power generation equipment. The component operation control scheme generated based on the real-time status assessment results can formulate corresponding operating procedures for components in different operating states. The real-time status assessment results and component operation control schemes are transmitted to the equipment control terminal. The inference parameters of the time-series correlation network and hierarchical machine learning inference model are updated in combination with the feedback component control results. This can continuously optimize the performance and accuracy of the model, realize continuous and accurate monitoring and effective control of the status of wind power generation equipment, improve the operating efficiency and reliability of wind power generation equipment, and reduce the equipment failure rate and maintenance costs. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the execution flow of the wind power generation equipment status monitoring method combining the Internet of Things and machine learning provided in an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of exemplary hardware and software components of a wind power equipment condition monitoring system that combines the Internet of Things and machine learning, provided in an embodiment of the present invention. Detailed Implementation

[0016] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a wind power generation equipment condition monitoring method combining the Internet of Things (IoT) and machine learning, provided in one embodiment of the present invention. The following is a detailed description of this wind power generation equipment condition monitoring method combining the Internet of Things (IoT) and machine learning.

[0017] Step S110: Dynamically configure IoT sensing nodes based on the operating scenarios of each component of the wind power generation equipment, and collect real-time operating data sets. The operating scenarios include the physical action mode and environmental interaction mode of the component during operation, and the real-time operating data sets include the physical state data and environmental impact data generated by the component under different operating scenarios.

[0018] In this embodiment, the blade component of a wind turbine is used as an application example throughout the text. This blade component involves various operating scenarios during operation, requiring dynamic configuration of IoT sensing nodes and collection of relevant data based on these scenarios. The blade component is a key component of a wind turbine for capturing wind energy, and its operating status directly affects the power generation efficiency and safety of the entire device.

[0019] Step S111: Analyze the physical interaction mode of each component of the wind power generation equipment during operation. The physical interaction mode includes the transmission contact mode between components, the force distribution mode, and the motion trajectory characteristics.

[0020] The blade assembly and hub are connected by bolts to achieve transmission contact. This transmission contact method allows the hub to drive the blade to rotate. During rotation, the bolted connection will bear significant torque and shear force. The force distribution of the blade during operation is relatively complex. From the blade root to the blade tip, the centrifugal force and aerodynamic load vary due to different distances from the center of rotation. The blade root usually experiences greater force, while the blade tip experiences relatively less force but is subject to a larger aerodynamic bending moment. In terms of motion trajectory characteristics, the blade moves in a circular motion around the center of the hub, and its motion trajectory is circular. Under different wind speeds, the rotational speed of the blade will change, resulting in changes in the linear velocity of each point on the motion trajectory.

[0021] Step S112: Analyze the environmental interaction mode of each component during operation. The environmental interaction mode includes the contact angle between the component and the airflow, the heat exchange path with surrounding components, and the interaction mode with environmental impurities.

[0022] The contact angle between the blades and the airflow changes with wind speed and direction. When the wind direction shifts, the blades adjust their angle of attack accordingly to maximize wind energy capture. During operation, the blades exchange heat with the surrounding air, and the heat generated by internal electrical components (such as lightning protection devices) also exchanges with the external environment through the blade walls. Additionally, there is heat conduction at the blade-hub connection. The interaction with environmental impurities primarily involves the erosion of the blade surface by dust, salt, and other contaminants. In some coastal areas, salt spray can also corrode the blade materials, affecting their lifespan.

[0023] Step S113: Integrate the physical action mode and the environmental interaction mode, divide the corresponding operation scenarios of each component, and determine the operation characteristics of the component in each operation scenario.

[0024] For example, when the wind speed is within a certain range and the wind direction is stable, the blades are in normal power generation operation. At this time, the blade's transmission contact method is stable, the force distribution is relatively uniform, the motion trajectory is regular, and the contact angle with the airflow is at its optimal windward position. Heat exchange is mainly through air convection, and the effect of environmental impurities is relatively weak. When encountering strong winds, the blades will adjust to a feathering position to reduce the wind-receiving area. At this time, the blade's transmission contact method changes, the force at the bolt connections increases, the rotational speed of the motion trajectory decreases, the contact angle with the airflow is smaller, and the heat exchange with surrounding components may increase due to the increased wind speed. Simultaneously, the impact force of environmental impurities also increases, thus constituting a strong wind feathering operation scenario. The blade's operating characteristics differ under each operating scenario. The operating characteristics of the normal power generation scenario are high power generation efficiency, stable physical state, and environmental interaction; the operating characteristics of the strong wind feathering operation scenario are low rotational speed, high force, and strong environmental interaction.

[0025] Step S114: Select a matching IoT sensing node type based on the physical action characteristics of the component in each operating scenario. The IoT sensing node type must match the component's transmission contact method, force distribution pattern, and motion trajectory characteristics.

[0026] Under normal power generation operation, the blade's transmission contact is stable, with uniform force distribution and a regular motion trajectory. For the transmission contact method, a vibration sensor capable of monitoring the tightness of bolt connections is selected. This sensor type matches the bolt transmission contact method between the blade and the hub, allowing real-time monitoring of vibration at the connection points to determine if loosening is present. For the force distribution pattern, strain gauge sensors are selected, which can be fitted to different positions on the blade surface to match the force distribution pattern from the blade root to the blade tip, accurately collecting force data at different locations. Regarding motion trajectory characteristics, a gyroscope sensor is used. This sensor can be mounted on the blade to capture the motion trajectory characteristics of the blade's circular motion, including rotation angle and angular velocity data. In strong wind feathering operation scenarios, due to increased force at the bolt connections, in addition to the aforementioned vibration sensor, a pressure sensor can be added for more precise monitoring of the force. The type of pressure sensor must be compatible with the larger force distribution pattern under these conditions.

[0027] Step S115: Determine the installation location of the IoT sensing node based on the environmental interaction characteristics of the components in each operating scenario.

[0028] In normal power generation operation, the blade's contact angle with the airflow is at its optimal windward position. To accurately monitor this contact angle, the wind direction sensor is installed at the center of the blade's leading edge. This orientation allows for direct sensing of the airflow direction, thus accurately acquiring contact angle data. The heat exchange paths with surrounding components primarily involve heat exchange between the blade and air, and between the blade and the hub. Temperature sensors should be installed on the blade surface near the hub and near internal electrical components to monitor temperature changes along different heat exchange paths. Regarding the interaction with environmental impurities, considering that impurities mainly corrode the blade surface, surface corrosion sensors are installed on the windward and leeward sides of the blade. The windward side is where impurities directly impact, while the leeward side may accumulate impurities due to eddies. These installation orientations effectively monitor the effects of environmental impurities. In strong wind feathering operation, the contact angle between the blade and the airflow is smaller. In this case, the wind direction sensor's installation orientation does not need to be changed, but it must still be ensured that it accurately captures the wind direction even with the blade in feathering position.

[0029] Step S116: Deploy the selected type of IoT sensing node at the determined installation location, and set the data collection interval of the sensing node according to the frequency of change of the operating scenario.

[0030] For the blade components, vibration sensors, strain gauge sensors, gyroscope sensors, wind direction sensors, temperature sensors, and surface corrosion sensors are deployed at the determined installation locations under both normal power generation and strong wind feathering operation scenarios. The frequency of changes in operating scenarios depends on local meteorological conditions. For example, in a certain region, normal power generation operation may last for several days, while strong wind feathering operation occurs less frequently, perhaps only a few times a year. For normal power generation operation scenarios, due to their longer duration and relatively stable state, the data acquisition interval is set to 5 minutes to ensure data accuracy while avoiding data redundancy. Although strong wind feathering operation scenarios occur less frequently, once they do, the blade state changes rapidly. Therefore, the data acquisition interval for this scenario is set to 1 minute to capture changes in blade state more promptly. During deployment, it is necessary to ensure that the sensors are securely installed to prevent loosening due to blade rotation, and to implement protective measures for the sensors to cope with environmental impacts under different operating scenarios.

[0031] Step S117: Activate the IoT sensing node to collect physical state data generated by the component in different operating scenarios during the corresponding time periods. The physical state data includes vibration change data, force intensity data, temperature change data, and movement speed data of the component during transmission.

[0032] During normal power generation operation, vibration sensors activate, collecting vibration data at the connection between the blade and hub bolts. This data reflects the stability of the connection; an abnormally large increase in vibration amplitude may indicate bolt loosening. Strain gauge sensors collect stress data at different locations on the blade. Strain gauges at the blade root monitor significant tensile and bending stresses, while those at the blade tip primarily collect stress data generated by centrifugal force. Temperature sensors collect temperature data, including temperature changes caused by heat exchange between the blade surface and the air, as well as temperature changes generated by the operation of internal electrical components. Gyroscope and velocity sensors work together to collect blade velocity data, including rotational and linear speeds. During strong wind and feathering operation, all sensors continue to operate. The vibration data collected by the vibration sensors may differ due to increased stress; strain gauge sensors will collect greater stress data than in normal operation, while the velocity data will show a decrease in blade rotational speed.

[0033] Step S118: Collect environmental impact data of the environment in which the component is located during different operating scenarios and time periods. The environmental impact data includes airflow speed change data, airflow direction change data, ambient temperature stability data, ambient humidity change data, and ambient impurity concentration data.

[0034] During normal power generation operation, wind direction and wind speed sensors installed on the leading edge of the blades collect data on changes in airflow velocity and direction, monitoring wind speed and direction changes in real time. Ambient temperature stability data is collected by an ambient temperature sensor installed at the bottom of the tower. This location is relatively stable and reflects the temperature of the environment surrounding the entire wind turbine, exhibiting a degree of stability and not significantly affected by blade rotation. Ambient humidity change data is collected by a humidity sensor, which can be installed on the hub near the blades to monitor changes in air humidity. Ambient impurity concentration data is collected by an impurity concentration sensor, installed at an appropriate location around the blades, to monitor the concentration of impurities such as dust and salt in the air. During strong wind and feathering operation, airflow velocity data will show a significant increase in wind speed, airflow direction change data may show unstable wind direction, and ambient impurity concentration data may increase due to ground impurities being stirred up by strong winds.

[0035] Step S119: Organize the collected physical state data and environmental impact data according to the operation scenario, and associate the physical state data and environmental impact data of the same time period under the same operation scenario to form a real-time operation data set.

[0036] Physical state data, such as vibration changes, stress intensity, temperature changes, and velocity, collected during normal power generation operation periods, are grouped into one category. Environmental impact data, including airflow velocity changes, airflow direction changes, ambient temperature stability, ambient humidity changes, and environmental impurity concentrations, are also categorized accordingly. Then, the physical state data and environmental impact data for the same time period within the same normal power generation operation scenario are correlated. For example, blade vibration change data for a given minute is correlated with airflow velocity and direction changes for that minute, forming a subset of real-time operational data for that time period within that operational scenario. Similarly, data from strong wind feathering operation scenarios are categorized and correlated to form corresponding subsets of real-time operational data. Finally, all real-time operational data subsets from all operational scenarios are integrated to form a set of real-time operational data for the blade components.

[0037] Step S1110: Arrange the real-time running data set in order of collection time and label the running scenario identifier corresponding to each data item.

[0038] The real-time operational data sets of the blade components are arranged in chronological order of acquisition, starting with the earliest acquired data and proceeding sequentially. During this arrangement, each data item is labeled with a corresponding operational scenario identifier; for example, the normal power generation operational scenario is labeled "FC01," and the strong wind feathering operational scenario is labeled "FC02." This allows for rapid identification of the operational scenario for each data item during subsequent data analysis and processing, facilitating the establishment of a correspondence between data and operational scenarios.

[0039] Step S120: Establish a time-series correlation network between the real-time running data set and the running scenarios and component running states. The time-series correlation network is used to characterize the component running states corresponding to the changes in the real-time running data set over time under different running scenarios.

[0040] After collecting and organizing the real-time operating data set of the blade components, a time-series correlation network is established between the real-time operating data set and the operating scenarios and component operating states. This time-series correlation network can show how the real-time operating data set of the blade components changes over time under different operating scenarios, and the corresponding blade operating states.

[0041] Step S121: Obtain the historical operation data set, historical operation scenario records and corresponding historical component operation status records accumulated by each component of the wind power generation equipment during historical operation.

[0042] By accessing the historical database of wind power equipment, a historical operational data set for the blade components was extracted. This set contains physical state data and environmental impact data of the blades under various operating scenarios at different times in the past. Simultaneously, historical operational scenario records were acquired, detailing the various operating scenarios the blades experienced in the past, such as operating scenarios under different wind speed ranges and wind directions, as well as the time and duration of each scenario. Corresponding historical component operational status records were also obtained, including various status information of the blades during historical operation, such as normal operation, minor fault status, and severe fault status, along with the start time, end time, and status description for each status.

[0043] Step S122: Classify the historical operation data set, historical operation scenario records, and historical component operation status records according to component type to form a historical data scenario status dataset corresponding to each component.

[0044] The acquired historical data was categorized by component type. Since the current research focuses on blade components, all historical operational data sets, historical operational scenario records, and historical component operational status records related to blades were extracted and classified into the blade component category. This data was then integrated to form a historical data scenario status dataset specifically for blade components. This dataset includes historical operational data, the operational scenarios in which they occurred, and their corresponding operational status.

[0045] Step S123: Standardize the physical state data and environmental impact data in the historical operation data set.

[0046] The units for vibration changes in physical state data might be meters per second squared, force intensity data in Pascals, temperature changes in degrees Celsius, and motion velocity data in meters per second. Similarly, the units for airflow velocity changes in environmental impact data might be meters per second, airflow direction changes in degrees Celsius, ambient temperature stability data in degrees Celsius, ambient humidity changes in percentages, and ambient impurity concentration data in milligrams per cubic meter. These dimensional differences can affect subsequent data analysis and model building, necessitating standardization. During standardization, for each data type, the data value is subtracted from its historical average, and then divided by its historical standard deviation. This process transforms all data to a range with a mean of 0 and a standard deviation of 1, eliminating dimensional differences and ensuring that different types of data have the same scale, facilitating comparison and correlation analysis.

[0047] Step S124: For the historical data scene status dataset corresponding to each component, extract the operation scene features from the historical operation scene records. The operation scene features include the core attributes of the historical physical action form and the core attributes of the historical environmental interaction mode.

[0048] The core attributes of historical physical interaction patterns include the transmission contact methods of the blades in historical operating scenarios (such as the tightness level of bolted connections), the force distribution pattern (such as the range of maximum force values ​​at the blade root), and the characteristics of the motion trajectory (such as the range of rotational angular velocity). The core attributes of historical environmental interaction patterns include the contact angle range between the blades and the airflow, the heat exchange efficiency level with surrounding components, and the intensity level of interaction with environmental impurities. By analyzing and extracting historical operating scenario records, these core attributes are quantified or ranked to form the operating scenario characteristics of the blade components. These characteristics can accurately reflect the key information of historical operating scenarios.

[0049] Step S125: Extract data features from the standardized historical operational data set. The data features include the changing trends of historical physical state data, the fluctuation patterns of historical environmental impact data, and the correlation between different data items.

[0050] For the historical physical state data, we analyze whether vibration changes gradually increase or decrease over a period of time, the rate of change of stress intensity data, whether temperature changes show an upward or downward trend, and the stability of motion velocity data. Regarding the fluctuation patterns of historical environmental impact data, we study the magnitude and frequency of fluctuations in airflow velocity changes, the periodicity of airflow direction changes, the range of fluctuations in stable environmental temperature data, the degree of fluctuation in environmental humidity data, and the frequency of peak values ​​in environmental impurity concentration data. The correlations between different data items are determined by analyzing the correlation between physical state data and environmental impact data. For example, we analyze whether there is a positive correlation between airflow velocity changes and blade vibration changes, and the degree of correlation between stable environmental temperature data and blade temperature changes. These trends, fluctuation patterns, and correlations are then extracted as data features.

[0051] Step S126: Extract the status features from the historical component operation status records. The status features include the duration of the historical operation status, the status transition conditions, and the component performance corresponding to the status.

[0052] The duration of historical operating states refers to the length of time the blade is in each operating state, such as how many hours the normal operating state lasted or how many minutes the minor fault state lasted. State transition conditions refer to the conditions that must be met to transition from one operating state to another. For example, when blade vibration data exceeds a certain threshold and persists for a certain period, the blade transitions from the normal operating state to the minor fault state. The corresponding component performance characteristics include changes in performance indicators such as blade power generation efficiency and power output under different operating states. For example, power generation efficiency is higher under normal operating conditions and decreases under minor fault conditions. This information is extracted as state features.

[0053] Step S127: Analyze the temporal correlation between the characteristics of the operating scenario and the data characteristics, determine the pattern of data characteristics changing over time under different operating scenario characteristics, and establish the temporal correspondence between scenario data.

[0054] The extracted operational scenario features and data features are aligned chronologically to observe how the data features change over time under different operational scenario characteristics. For example, under normal power generation operational scenario characteristics (stable historical physical action patterns and favorable historical environmental interactions), the vibration change data of the blades shows a stable trend with small fluctuations, and the correlation between data items is stable. However, under strong wind and feathering operational scenario characteristics, the vibration change data may show larger fluctuations, with stronger fluctuations, and the correlation between data items may also change. Through the analysis of a large amount of historical data, the patterns of data feature changes over time under different operational scenario characteristics are determined. For example, under certain operational scenario characteristics, the stress intensity data shows a pattern of first increasing and then stabilizing over time. Based on this, a temporal correspondence relationship of scenario data is established, which describes the temporal evolution pattern of data features under specific operational scenario characteristics.

[0055] Step S128: Analyze the temporal correlation between data features and state features, determine the state features corresponding to the changes of different combinations of data features over time, and establish the temporal correspondence between data states.

[0056] Similarly, by aligning data features and state features chronologically, we can study how different combinations of data features change over time and correspond to different state features. For example, when the vibration change data features of a blade show a continuous increasing trend, and the stress intensity data features also exceed a certain threshold, and this combination of data features persists for a period of time, the corresponding state feature may change from a normal operating state to a minor fault state. The duration of this state will also change accordingly, and performance will decline. By analyzing the correspondence between data feature combinations and state features in historical data, we can determine the state features corresponding to different combinations of data features over time, establishing a temporal correspondence between data and state features. This temporal correspondence clarifies the rules governing the correspondence between the temporal change patterns of data features and state features.

[0057] Step S129: Based on the temporal correspondence between scene data and the temporal correspondence between data state, construct the initial structure of the temporal association network. The initial structure includes running scene nodes, data nodes, state nodes, and temporal connection edges between nodes.

[0058] In this initial structure, operating scenario nodes are set, each representing a specific operating scenario characteristic of the blade; data nodes are set, each representing a specific data characteristic; and state nodes are set, each representing a specific state characteristic. Then, based on the temporal correspondence between scenario data, temporal connection edges are added between operating scenario nodes and data nodes. These temporal connection edges represent the relationship between data characteristics and time under a specific operating scenario. Similarly, based on the temporal correspondence between data states, temporal connection edges are added between data nodes and state nodes. These connection edges represent the correspondence between combinations of data characteristics changing over time and state characteristics. For example, the "normal power generation operating scenario node" points to data nodes such as the "stable vibration data node" and the "uniform stress data node" via temporal connection edges, and these data nodes, in turn, point to the "normal operating state node" via temporal connection edges.

[0059] Step S1210: Add time weights to the connection edges in the initial structure of the temporal association network, wherein the time weights are determined based on the duration of the association between the corresponding nodes.

[0060] The time weights are determined based on the duration of the association between corresponding nodes. The longer the association lasts, the greater the time weight, indicating stronger temporal stability. For example, in a normal power generation scenario, the association between a node and a stable vibration data node has a long duration in historical data, thus assigning a larger time weight to their connection edge; conversely, in a specific operating scenario, the association between a node and a certain abnormal data node has a short duration, thus assigning a smaller time weight. By adding time weights, the temporal correlation network can better reflect the importance of different associations in the time dimension.

[0061] Step S1211: Select some historical scenario state data that did not participate in the initial structure construction, input them into the initial structure of the time series correlation network for verification, and observe whether the component operation status output by the network matches the historical component operation status records.

[0062] The aforementioned validation data is input into the initial structure of the temporal correlation network. Based on the input operational scenario and data characteristics, the network infers the corresponding component operational status through the connection edges between nodes and time weights. Then, the component operational status output by the network is compared with the actual historical component operational status records in the validation data to observe whether they match, thereby evaluating the accuracy and reliability of the initial structure of the temporal correlation network.

[0063] Step S1212: Adjust the node division, connection edge settings and time weights in the initial structure of the temporal correlation network according to the verification results, so that the initial temporal correlation network can represent the component operation status corresponding to the change of real-time operation data set under different operation scenarios over time, and form the final temporal correlation network.

[0064] If the component operating status output by the network matches the historical component operating status records with a high degree of similarity, the initial structure is considered reasonable, and only minor adjustments to the time weights of some connection edges are needed. If the matching degree is low, it is necessary to check whether the node partitioning is reasonable, whether there are any missing operating scenario nodes, data nodes, or status nodes, and to repartition and supplement the nodes accordingly; check whether the connection edge settings are correct, whether there are any redundant or incorrect connection edges, and to delete or add connection edges accordingly; at the same time, the time weights of the connection edges should be re-evaluated and adjusted according to the duration of the correlation relationships in the validation data. After multiple validations and adjustments, the time-series correlation network can accurately represent the component operating status corresponding to the changes in the real-time operating data set of the blade component under different operating scenarios over time, ultimately forming the time-series correlation network of the blade component.

[0065] Step S130: Call the hierarchical machine learning inference model to perform collaborative inference processing on the real-time running data set and time series association network to generate real-time status evaluation results of each component of the wind power generation equipment.

[0066] After constructing the temporal correlation network of the blade components, a hierarchical machine learning inference model is invoked to perform collaborative inference processing on the real-time operating data set and temporal correlation network of the blade. Through the collaboration of each level of the model, the operating status of the blade is accurately evaluated, and real-time status evaluation results are generated.

[0067] Step S131: Obtain the structural information of the hierarchical machine learning inference model, which includes the functional information of the perception layer, the association layer, the inference layer, and the data transmission relationship between the layers.

[0068] Step S1311: Access the machine learning model storage system and extract the design specification file of the hierarchical machine learning inference model.

[0069] Access to the machine learning model storage system deployed on the local server is via a dedicated data interface. This system employs a distributed file storage architecture, encrypting and version-controlling the model files. Based on the unique identifier of the blade component condition monitoring task, the corresponding hierarchical machine learning inference model design specification file is retrieved and extracted. The file format is XML, containing core content such as the model architecture, parameter configurations for each level, and data interaction protocols. Identity authentication and permission verification are required during the extraction process to ensure the security of model file access.

[0070] Step S1312: Read the functional information of the perception layer from the design specification file. The functional information includes the input data type of the perception layer, the feature extraction algorithm type, the output feature vector dimension, and the step division of feature extraction.

[0071] The XML-formatted design specification file is parsed to locate the perception layer configuration node. The input data type is a real-time blade operation data set, which includes multi-dimensional time-series data such as vibration change data and stress intensity data. The feature extraction algorithm is an improved convolutional neural network, containing 3 convolutional layers and 2 pooling layers, with convolutional kernel sizes of 3×3, 5×5, and 7×7, and the activation function is ReLU. The output feature vector has a dimension of 256, of which the physical state data feature components account for 128 dimensions, and the environmental influence data feature components account for 128 dimensions. The feature extraction steps are divided into four sub-steps: data preprocessing (denoising and normalization), shallow feature extraction (convolutional layers 1-2), deep feature extraction (convolutional layer 3), and feature dimensionality reduction (global average pooling). The execution order and dependencies of each sub-step are described in the file through flowchart nodes.

[0072] Step S1313: Read the functional information of the association layer from the design specification file. The functional information includes the input parameter type of the association layer, the type of association analysis algorithm, the calculation logic of scene matching degree and historical similarity, and the output result format.

[0073] Continuing to analyze the design specification document, we obtain the configuration information for the association layer's functions. Input parameter types include a 256-dimensional real-time data feature vector output from the perception layer, a 128-dimensional runtime scene feature vector output from the temporal association network, and a 128-dimensional association data feature vector retrieved from the historical database. The association analysis algorithm is a feature fusion network based on a multi-head attention mechanism, containing 8 attention heads, each with a feature dimension of 32. The scene matching degree calculation logic uses cosine similarity weighted summation, with weights dynamically generated through a self-attention mechanism, involving a dimension-by-dimensional comparison between the real-time data feature vector and the runtime scene feature vector. The historical similarity calculation logic uses a dynamic time warping algorithm to measure the distance between the real-time data feature vector and the association data feature vector after time axis alignment. The output results are in the form of two scalar values ​​(scene matching degree and historical similarity) and one optimization instruction set, which includes feature extraction weight adjustment coefficients and a feature dimension selection mask.

[0074] Step S1314: Read the functional information of the inference layer from the design specification file. The functional information includes the input feature type of the inference layer, the state inference algorithm type, the calling method of the timing correspondence rule, and the type classification of the output evaluation result.

[0075] The system locates the configuration node of the inference layer and extracts functional parameters. The input feature type is a 256-dimensional data feature vector optimized by the association layer; the state inference algorithm type is a bidirectional long short-term memory network (Bi-LSTM), containing two hidden layers with 128 neurons per layer, a dropout ratio of 0.3, and using the Adam optimizer; the temporal correspondence rule invocation method is dynamic matching based on a rule engine, with the rule base stored in the relational database of the temporal association network, and fast indexing is performed through the hash value of the scene feature vector during invocation; the output evaluation result type is divided into 5 levels, namely normal state (S0), slight abnormality (S1), moderate abnormality (S2), severe abnormality (S3), and emergency fault (S4), each level corresponding to a different confidence threshold and fault feature template.

[0076] Step S1315: Extract the data transmission relationship between the perception layer and the association layer from the design specification document. The data transmission relationship includes the path for transmitting the output feature vector of the perception layer to the association layer, the format requirements for the transmitted data, and the triggering conditions for data transmission.

[0077] In the data flow diagram node of the design specification document, the interaction configuration between the perception layer and the association layer is analyzed. The data transmission path is that the feature output port (PortID: F_OUT_001) of the perception layer is connected to the feature input port (PortID: F_IN_001) of the association layer via the internal bus (BusID: IBUS_02). The bus bandwidth is configured to be 100Mbps, and the transmission latency is required to be less than 10ms. The data transmission format adopts Google Protocol Buffers (Protobuf) encoding, and the dimension, data type (float32), timestamp, and checksum fields of the feature vector are defined. The data transmission trigger condition is that after the perception layer completes the feature extraction of the current window, the receiving process of the association layer is triggered by the interrupt signal (IRQ_0x05), and the interrupt priority is set to medium priority (Level 3).

[0078] Step S1316: Extract the data transmission relationship between the association layer and the inference layer from the design specification document. The data transmission relationship includes the path for transmitting the scene matching degree and historical similarity output of the association layer to the inference layer, the verification method of the transmitted data, and the frequency of data transmission.

[0079] Locate the communication configuration node between the association layer and the inference layer. The data transmission path is from the decision output port (PortID: D_OUT_002) of the association layer to the decision input port (PortID: D_IN_002) of the inference layer via an external data link (LinkID: ELINK_03). The link uses TCP / IP protocol encapsulation, and the port number is 5005. The data transmission verification method is Cyclic Redundancy Check (CRC32), and the verification field covers the data payload and header information. A retransmission mechanism is triggered when the receiving end fails the verification. The data transmission frequency is consistent with the feature extraction frequency of the perception layer. The default configuration is 1Hz, which can be adjusted to 0.5Hz or 2Hz through dynamic configuration commands, with an adjustment step of 0.1Hz.

[0080] Step S1317: Extract the feedback transmission relationship between the association layer and the perception layer from the design specification document. The feedback transmission relationship includes the path for the association layer to output adjustment instructions to the perception layer, the parsing method of the feedback data, and the execution flow of the feedback instructions.

[0081] The feedback channel configuration from the association layer to the perception layer is analyzed. The transmission path connects the feedback output port (PortID: FB_OUT_003) of the association layer to the feedback input port (PortID: FB_IN_003) of the perception layer via a control bus (BusID: CBUS_01), using the CANopen protocol. The feedback data is parsed in XML format, containing a feature weight adjustment matrix (256×1 floating-point array) and a feature dimension mask (256-bit binary array). The parsing process verifies the integrity of the XML tags and the validity of the data type. The feedback instruction execution flow is as follows: after receiving the instruction, the perception layer first performs instruction validity verification (signature verification), then updates the convolutional kernel weight parameters of the feature extraction network, and finally masks invalid dimensions using a feature selection mask. The entire process is controlled by a state machine (containing five states: idle→receive→parse→update→ready).

[0082] Step S1318: Extract the feedback transmission relationship between the inference layer and the association layer from the design specification document. The feedback transmission relationship includes the path for the preliminary evaluation result output by the inference layer to be transmitted to the association layer, the analysis method of the feedback data, and the judgment criteria of the feedback result.

[0083] Configure the feedback mechanism from the inference layer to the association layer. The transmission path is from the evaluation output port (PortID: E_OUT_004) of the inference layer to the evaluation input port (PortID: E_IN_004) of the association layer via the internal data bus (BusID: IBUS_02). It shares the same physical bus with the perception layer data and avoids conflicts through a time-division multiplexing mechanism. The feedback data analysis method is evaluation result verification based on a rule tree. The rule tree contains 1024 decision rules, which involve the logical relationship judgment between the preliminary evaluation result and the scene matching degree and historical similarity. The feedback result judgment standard is a preset reasonableness threshold range. For example, when the scene matching degree is >0.7 and the historical similarity is >0.6, the reasonableness score of the preliminary evaluation result is ≥0.8. Otherwise, a re-inference process is triggered. The threshold range can be dynamically adjusted through the model configuration file.

[0084] Step S1319: Integrate the functional information of the perception layer, association layer, and reasoning layer, as well as the data transmission relationships between the layers, to form the structural information of the hierarchical machine learning reasoning model.

[0085] The extracted functional information and data transmission relationships at each level are structurally integrated to construct a model structure information graph. A directed graph data structure is used to represent the dependencies between levels. Nodes contain attributes such as level identifier, functional description, and parameter configuration, while edges contain attributes such as transmission path, data format, and triggering conditions. A model architecture diagram is generated using a graph visualization tool, annotating the input / output interfaces, data flow, and key parameters of each level, forming a complete structure information document.

[0086] Step S132: Input the real-time running data set into the perception layer of the hierarchical machine learning inference model. The perception layer extracts features from the real-time running data set to generate a real-time data feature vector. The real-time data feature vector includes physical state data feature components and environmental influence data feature components.

[0087] Step S1321: Divide the real-time running data set into a physical state data subset and an environmental impact data subset according to the data type. The physical state data subset includes vibration change data, force intensity data, temperature change data, and motion speed data. The environmental impact data subset includes airflow speed change data, airflow direction change data, environmental temperature stability data, environmental humidity change data, and environmental impurity concentration data.

[0088] The physical state data subset includes vibration changes, stress intensity, temperature changes, and velocity data generated by the blade during operation. These data directly reflect the physical state of the blade itself. The environmental impact data subset includes airflow velocity changes, airflow direction changes, ambient temperature stability, ambient humidity changes, and ambient impurity concentration data of the environment in which the blade is located. These data reflect the impact of the environment on the blade.

[0089] Step S1322: Standardize the physical state data subset and the environmental impact data subset respectively to eliminate the dimensional differences of different physical quantities and make the data have the same scale.

[0090] The physical state data subset and the environmental impact data subset were standardized separately. The standardization method was similar to that for historical data. For each data type in the physical state data subset, such as vibration change data, each data value was subtracted from the average value of that data type in the real-time running dataset, and then divided by the standard deviation of that data type. Similarly, a similar process was performed on each data type in the environmental impact data subset. This method eliminated the dimensional differences between different physical quantities, ensuring that the physical state data subset and the environmental impact data subset had the same scale, which facilitated subsequent feature extraction.

[0091] Step S1323: Extract features from the standardized physical state data subset, including vibration period features and vibration amplitude features from vibration change data, peak force features and force distribution features from force intensity data, temperature rise and fall rate features and temperature stability range features from temperature change data, and velocity change gradient features and velocity stability duration features from motion velocity data.

[0092] For vibration variation data, the vibration period (i.e., the time required for one vibration) and vibration amplitude (i.e., the distance between the peak and trough of the vibration waveform) are extracted by analyzing the waveform. For force intensity data, the maximum value in the data is identified as the peak force characteristic, and the distribution of force intensity at different locations is analyzed to extract the force distribution characteristic. For temperature variation data, the rate of temperature rise and fall is calculated per unit time, and the range of temperature fluctuation within a certain range is identified as the temperature stability range characteristic. For motion velocity data, the rate of change of velocity over time is calculated to obtain the velocity gradient characteristic, and the length of time the velocity remains within a certain range is statistically analyzed as the velocity stability duration characteristic.

[0093] Step S1324: Extract features from the standardized environmental impact data subset, including the velocity fluctuation amplitude and velocity change frequency features of the airflow velocity change data, the direction deflection angle and direction stabilization duration features of the airflow direction change data, the temperature fluctuation range and temperature mean features of the stable environmental temperature data, the humidity change rate and humidity stabilization interval features of the environmental humidity change data, and the concentration change amplitude and concentration peak features of the environmental impurity concentration data.

[0094] In airflow velocity variation data, the difference between the maximum and minimum velocity values ​​is used to obtain the velocity fluctuation amplitude characteristic, and the number of velocity changes per unit time is used to obtain the velocity change frequency characteristic. In airflow direction variation data, the angle of deflection of the direction from one value to another is used as the direction deflection angle characteristic, and the length of time the direction remains unchanged is used as the direction stability duration characteristic. In environmental temperature stability data, the range of maximum and minimum temperature values ​​is determined as the temperature fluctuation range characteristic, and the average value of all temperature data is calculated as the temperature mean characteristic. In environmental humidity variation data, the change in humidity per unit time is calculated as the humidity change rate characteristic, and the range of humidity fluctuation within a certain range is determined as the humidity stability range characteristic. In environmental impurity concentration data, the difference between the maximum and minimum concentration values ​​is calculated as the concentration change amplitude characteristic, and the maximum value in the concentration data is identified as the concentration peak value characteristic.

[0095] Step S1325: Arrange the features extracted from the physical state data subset in a preset order to form physical state data feature components, with each feature corresponding to one dimension of the physical state data feature component.

[0096] The vibration period features, vibration amplitude features, peak force features, force distribution features, temperature rise and fall rate features, temperature stability range features, velocity change gradient features, and velocity stability duration features extracted from a subset of blade physical state data are arranged in a preset order (e.g., vibration-related features, force-related features, temperature-related features, and velocity-related features). After arrangement, each feature corresponds to one dimension of the physical state data feature components, thus forming the physical state data feature components of the blade.

[0097] Step S1326: Arrange the features extracted from the environmental impact data subset in a preset order to form environmental impact data feature components, with each feature corresponding to one dimension of the environmental impact data feature component.

[0098] Similarly, features extracted from the blade environmental impact data subset, such as velocity fluctuation amplitude, velocity change frequency, directional deflection angle, directional stability duration, temperature fluctuation range, average temperature, humidity change rate, humidity stability interval, concentration change amplitude, and peak concentration, are arranged in a preset order (e.g., airflow-related features, environmental temperature and humidity-related features, and environmental impurity-related features). Each feature corresponds to one dimension of the environmental impact data feature component, forming the blade environmental impact data feature component.

[0099] Step S1327: Adjust the dimensions of the physical state data feature components and the environmental impact data feature components, and combine the adjusted physical state data feature components and the environmental impact data feature components to form a real-time data feature vector.

[0100] The dimensions of the physical state data feature components and the environmental impact data feature components of the blade are adjusted to ensure that their dimensions can be combined. If the physical state data feature components are 8-dimensional and the environmental impact data feature components are 10-dimensional, their dimensions are adjusted to be the same or compatible (e.g., through feature mapping). Then, the adjusted physical state data feature components and environmental impact data feature components are concatenated sequentially to form the real-time data feature vector of the blade, which integrates the feature information of both the physical state and environmental impact of the blade.

[0101] Step S133: Extract the running scenario feature vector and the historical time period related data feature vector from the temporal correlation network corresponding to the current running scenario. Input the running scenario feature vector and the related data feature vector into the correlation layer of the hierarchical machine learning inference model. The correlation layer performs correlation analysis on the real-time data feature vector, the running scenario feature vector, and the related data feature vector, calculates the scenario matching degree between the real-time data feature vector and the running scenario feature vector, and calculates the historical similarity between the real-time data feature vector and the related data feature vector.

[0102] In the normal power generation operation scenario of the blade component, after the association layer receives the 256-dimensional real-time data feature vector output by the perception layer, it needs to perform deep association analysis by combining the operation scenario feature vector and historical data feature vector provided by the time-series association network.

[0103] Step S1331: Obtain the running scenario feature vector corresponding to the current running scenario from the temporal correlation network. The running scenario feature vector includes the physical action form feature component and the environmental interaction mode feature component under the current running scenario.

[0104] The association layer calls the feature extraction module of the temporal association network through the API interface, passing in the current running scene identifier (“FC01”) and timestamp (“2023-10-20 T15:30:00Z”). The temporal association network retrieves the scene feature library based on the identifier and extracts the running scene feature vector containing physical action form feature components (64-dimensional) and environmental interaction mode feature components (64-dimensional). The physical action form feature components include bolt connection tightness (10-dimensional), blade root force distribution (20-dimensional), rotation trajectory deviation (15-dimensional), and vibration modal parameters (19-dimensional); the environmental interaction mode feature components include airflow angle of attack (12-dimensional), heat exchange coefficient (18-dimensional), salt spray corrosion rate (15-dimensional), and sandstorm impact intensity (19-dimensional). Each component is encoded by unique thermal encoding and normalized to a floating-point value in the range [0, 1].

[0105] Step S1332: Obtain the associated data feature vectors from the temporal correlation network that are the same as the historical time period and the current operating scenario. The associated data feature vectors include historical physical state data feature components and historical environmental influence data feature components.

[0106] The association layer sends a historical data retrieval request to the temporal association network. The request parameters include the current running scene identifier (“FC01”), the time window (last 30 days), and the similarity threshold (0.85). The temporal association network locates the “FC01” scene dataset in the historical database through the scene index. It uses the K-nearest neighbor algorithm (K=5) to retrieve the 5 historical records most similar to the current real-time data feature vector. For each record, the physical state data feature components (128 dimensions) and environmental impact data feature components (128 dimensions) are weighted and averaged (weights are calculated based on the time decay factor) to generate a 128-dimensional association data feature vector. The historical physical state data feature components include historical vibration spectrum (32 dimensions), historical strain distribution (32 dimensions), historical temperature field (32 dimensions), and historical rotational speed curve (32 dimensions); the historical environmental impact data feature components include historical wind speed spectrum (32 dimensions), historical wind direction change (32 dimensions), historical temperature and humidity sequence (32 dimensions), and historical impurity concentration (32 dimensions).

[0107] Step S1333: Align the real-time data feature vector with the running scene feature vector in terms of dimensions, and calculate the similarity between the physical state data feature components in the real-time data feature vector and the physical action form feature components in the running scene feature vector. Use the correspondence analysis method between features to determine the matching degree of each dimension feature, and calculate the physical scene matching score by weighting according to the dimension weight.

[0108] The association layer first aligns the real-time data feature vector (256-dimensional) and the runtime scene feature vector (128-dimensional). The runtime scene feature vector is then upscaled from 128-dimensional to 256-dimensional using a feature mapping matrix generated by a pre-trained autoencoder. The first 128 dimensions of physical state data features are extracted from the aligned real-time data feature vector, and the first 64 dimensions of physical action form features are extracted from the upscaled runtime scene feature vector (achieved through a feature selection mask). Dimensional similarity is calculated for the two 64-dimensional feature components using the cosine similarity formula: cos(a, b) = a·b / (||a||·||b||), resulting in 64 similarity values. Dimensional weights are dynamically generated using a multilayer perceptron (input is 64-dimensional similarity values, output is a 64-dimensional weight vector). The weight vector is normalized using softmax and then multiplied element-wise with the similarity values, and the sum is used to obtain the physical scene matching score (range [0, 1]).

[0109] Step S1334: Calculate the similarity between the environmental impact data feature component in the real-time data feature vector and the environmental interaction mode feature component in the running scene feature vector. Use the same feature correspondence analysis method to determine the matching degree of each dimension feature and calculate the environmental scene matching score by weighting according to the dimension weight.

[0110] The 128-dimensional environmental impact data feature components are extracted from the aligned real-time data feature vector, and the 64-dimensional environmental interaction mode feature components are extracted from the upgraded running scene feature vector. Using the same cosine similarity calculation method as in step S1333, 64 environmental feature similarity values ​​are obtained. Dimensional weights are generated through a separate multilayer perceptron, with the environmental feature similarity values ​​as input and a 64-dimensional weight vector as output. The weight vector is multiplied element-wise by the similarity values ​​and then summed to obtain the environmental scene matching score (value range [0, 1]).

[0111] Step S1335: Integrate the physical scene matching score and the environmental scene matching score according to a preset ratio to obtain the scene matching degree between the real-time data feature vector and the running scene feature vector.

[0112] The default integration ratio of physical scene matching score to environmental scene matching score is 6:4 (this can be adjusted via the model configuration file), using a weighted summation formula: Scene Matching Degree = 0.6 × Physical Scene Matching Score + 0.4 × Environmental Scene Matching Score. During the calculation, the two matching scores undergo min-max normalization to ensure consistent dimensions. For example, if the physical scene matching score is 0.85 and the environmental scene matching score is 0.75, then the scene matching degree = 0.6 × 0.85 + 0.4 × 0.75 = 0.81.

[0113] Step S1336: Align the real-time data feature vector with the associated data feature vector in terms of dimensions, calculate the similarity between the physical state data feature components in the real-time data feature vector and the historical physical state data feature components in the associated data feature vector, analyze the consistency of the changing trends of each dimension feature, and calculate the physical history similarity score by weighting according to the dimension weight.

[0114] The real-time data feature vector (256-dimensional) and the associated data feature vector (128-dimensional) are dimensionally aligned. A feature reduction matrix is ​​then used to reduce the dimensionality of the real-time data feature vector from 256 to 128 dimensions. This reduction matrix is ​​generated through principal component analysis (PCA) training. The first 64 dimensions of physical state data feature components are extracted from the reduced real-time data feature vector, and the first 64 dimensions of historical physical state data feature components are extracted from the associated data feature vector. A dynamic time warping algorithm is used to calculate the trend similarity between the two feature components. The feature sequences of each dimension are aligned along the time axis, and the cumulative distance is calculated and normalized to obtain 64 trend similarity values. The dimensional weights are optimized through backpropagation of historical matching accuracy and weighted summed with the trend similarity values ​​to obtain the physical historical similarity score (range [0, 1]).

[0115] Step S1337: Calculate the similarity between the environmental impact data feature components in the real-time data feature vector and the historical environmental impact data feature components in the associated data feature vector, analyze the consistency of the fluctuation patterns of each dimension feature, and calculate the historical environmental similarity score by weighting according to the dimension weight.

[0116] The environmental impact data feature components of the last 64 dimensions were extracted from the feature vector of the real-time data after dimensionality reduction, and the historical environmental impact data feature components of the last 64 dimensions were extracted from the feature vector of the associated data. The consistency of the fluctuation pattern was analyzed by power spectral density comparison. The power spectrum was obtained by performing Fourier transform on the feature sequence of each dimension, and the spectral similarity was calculated (using KL divergence measure) to obtain 64 fluctuation similarity values. The dimension weights were optimized by genetic algorithm, and the weighted sum with the fluctuation similarity values ​​was used to obtain the environmental historical similarity score (value range [0, 1]).

[0117] Step S1338: Integrate the physical history similarity score and the environmental history similarity score according to a preset ratio to obtain the historical similarity between the real-time data feature vector and the associated data feature vector, thereby outputting the scene matching degree and historical similarity.

[0118] The preset integration ratio of physical history similarity score and environmental history similarity score is 5:5, using a weighted summation formula: Historical Similarity = 0.5 × Physical History Similarity Score + 0.5 × Environmental History Similarity Score. Min-max normalization is also performed; for example, if the physical history similarity score is 0.78 and the environmental history similarity score is 0.82, then the historical similarity score is 0.5 × 0.78 + 0.5 × 0.82 = 0.80. The final association layer outputs two scalar results: scene matching degree (e.g., 0.81) and historical similarity score (e.g., 0.80).

[0119] Step S134: The association layer feeds back the scene matching degree and historical similarity to the perception layer. The perception layer adjusts the extraction weights of the physical state data feature components and the environmental influence data feature components according to the scene matching degree, and adjusts the dimension of feature extraction according to the historical similarity to generate an optimized data feature vector.

[0120] The association layer feeds back the calculated blade scene matching degree and historical similarity to the perception layer. Upon receiving this feedback, the perception layer adjusts the extraction weights of the physical state data feature components and environmental impact data feature components based on the scene matching degree. If the scene matching degree is high, it indicates that the currently extracted features match the operating scene well, and the extraction weights can be maintained or fine-tuned appropriately. If the scene matching degree is low, it may be that the extraction weight of a certain feature component is inappropriate; for example, if the environmental impact data feature component is more important in the current operating scene, then the extraction weight of the environmental impact data feature component should be increased, and the extraction weight of the physical state data feature component should be decreased. Simultaneously, the feature extraction dimensions are adjusted based on historical similarity. If historical similarity is high, it indicates that the current feature dimensions can reflect historical data features well, and the dimensions can remain unchanged. If historical similarity is low, it may be necessary to increase or decrease certain feature dimensions, such as increasing feature dimensions that differ significantly from historical data, to improve feature discriminative power. After adjustment, an optimized data feature vector is generated.

[0121] Step S135: Input the optimized data feature vector into the inference layer of the hierarchical machine learning inference model. The inference layer calls the temporal correspondence rules between data features and states in the current running scenario from the temporal association network, and performs state inference on the optimized data feature vector based on the temporal correspondence rules. After generating the preliminary running state evaluation results of the component, the preliminary running state evaluation results are fed back to the association layer. The association layer checks the rationality of the preliminary running state evaluation results by combining the scene matching degree and historical similarity.

[0122] For example, in step S1351: the optimized data feature vector is input into the inference layer of the hierarchical machine learning inference model. The inference layer extracts the historical data feature sequence corresponding to the current running scenario and the component running status sequence corresponding to the historical data feature sequence from the temporal correlation network.

[0123] The optimized data feature vector (256 dimensions) is transmitted to the input buffer of the inference layer via the data bus. The Bi-LSTM network of the inference layer first retrieves the historical data feature sequence and component running state sequence corresponding to the current running scenario from the rule base of the temporal correlation network through the scenario identifier ("FC01"). The historical data feature sequence is a three-dimensional tensor (number of samples × time step × feature dimension), where the number of samples = 1000, the time step = 128, and the feature dimension = 256, containing feature data under the "FC01" scenario over the past year; the component running state sequence is a two-dimensional tensor (number of samples × time step × state dimension), with a state dimension of 5 (corresponding to five state levels S0-S4), and the state at each time step is represented by one-hot encoding.

[0124] Step S1352: Compare the feature dimensions of the historical data feature sequence and the optimized data feature vector to determine the common feature dimensions and the range of variation of each common feature dimension in the historical data feature sequence.

[0125] The feature comparison module of the inference layer performs a dimension-by-dimensional matching between the optimized data feature vector (256 dimensions) and the feature dimensions of the historical data feature sequence. It determines the common feature dimensions (200 dimensions) and the non-common feature dimensions (56 dimensions) by mapping feature names and physical meanings. For the common feature dimensions, it calculates the maximum, minimum, mean, and standard deviation of each dimension in the historical data feature sequence to determine the range of variation ([min-3σ, max+3σ]). For example, if the vibration amplitude feature dimension in the historical data has min=0.1g, max=2.5g, and σ=0.3g, then the range of variation = [0.1-0.9, 2.5+0.9] = [-0.8g, 3.4g]. Values ​​outside this range are marked as abnormal fluctuations.

[0126] Step S1353: Call the time-series correspondence rules of data features and states in the current running scenario from the time-series correlation network. The time-series correspondence rules include the correspondence between the range of change of each feature dimension and the running state of the component.

[0127] The time-series rule base uses a production rule representation and contains 1024 rules. Each rule includes a precondition (range of variation in feature dimensions) and a conclusion (component operating status). For example, rule R105: IF vibration amplitude ∈ [1.8g, 2.5g] AND blade root strain ∈ [300με, 400με] AND temperature gradient ∈ [5℃ / min, 10℃ / min] THEN state = S2 (moderate anomaly). The rule base is optimized using a decision tree algorithm to form a hierarchical rule index, which can quickly locate matching rules based on the current feature dimension.

[0128] Step S1354: Compare the values ​​of each common feature dimension in the optimized data feature vector with the range of change of the corresponding feature dimension in the historical data feature sequence to determine the range of change of each common feature dimension value in the optimized data feature vector.

[0129] The inference layer performs interval matching on each of the 200 common feature dimensions of the optimized data feature vector, comparing the current value of each dimension with its historical range of variation to determine its interval position (normal interval, slightly abnormal interval, moderately abnormal interval, severely abnormal interval). For example, if the current vibration amplitude is 2.0g, and the historical range of variation is [-0.8g, 3.4g], where the normal interval is [0.1g, 1.5g], the slightly abnormal interval is [1.5g, 2.0g], the moderately abnormal interval is [2.0g, 2.5g], and the severely abnormal interval is [2.5g, 3.4g], then the current vibration amplitude is in the moderately abnormal interval. After marking all common feature dimensions with intervals, the number of feature dimensions in each interval is counted.

[0130] Step S1355: Based on the correspondence between the change range and the component operating state in the time sequence correspondence rule, generate a set of candidate operating states corresponding to the optimized data feature vector.

[0131] The inference layer inputs the interval labeling results of the common feature dimensions into the rule engine. The engine uses a forward inference strategy to match the corresponding rules in the time sequence and generate a set of candidate running states. For example, among 200 common feature dimensions, 30 are in the mildly abnormal interval, 50 are in the moderately abnormal interval, and 5 are in the severe abnormal interval. The rule engine matches rules R210, R305, and R412, and the corresponding candidate states are S1 (30% confidence), S2 (50% confidence), and S3 (15% confidence), forming a set of candidate running states = {S1, S2, S3}.

[0132] Step S1356: Extract historical runtime data corresponding to each candidate running state in the candidate running state set under the current running scenario from the temporal correlation network. The historical runtime data includes the duration of each candidate running state in the historical scenario.

[0133] The inference layer extracts the historical runtime data of each state under the "FC01" scenario from the historical state database of the temporal correlation network based on the candidate runtime state set {S1, S2, S3}. The historical runtime data of state S1 consists of 100 records, with a mean of 120 minutes, a median of 90 minutes, and a standard deviation of 45 minutes; the historical runtime data of state S2 consists of 80 records, with a mean of 60 minutes, a median of 50 minutes, and a standard deviation of 20 minutes; the historical runtime data of state S3 consists of 30 records, with a mean of 20 minutes, a median of 15 minutes, and a standard deviation of 10 minutes.

[0134] Step S1357: Perform statistical analysis on the historical runtime data corresponding to each candidate running state to determine the probability of each candidate running state occurring in the current running scenario.

[0135] The inference layer uses kernel density estimation (KDE) to estimate the probability density function of historical runtime data, calculating the probability of each candidate state occurring in the current runtime scenario. The KDE curve for state S1 peaks at 120 minutes, with a probability density of 0.02 / minute; state S2 peaks at 60 minutes, with a probability density of 0.03 / minute; and state S3 peaks at 20 minutes, with a probability density of 0.05 / minute. The cumulative probabilities of each state are calculated through integration, yielding the following probabilities: P(S1) = 0.3, P(S2) = 0.5, and P(S3) = 0.2.

[0136] Step S1358: Sort the candidate running state set according to the occurrence probability of each candidate running state, and select the candidate running state with the highest occurrence probability as the first candidate state.

[0137] The candidate running states are sorted from highest to lowest probability of occurrence: S2 (0.5) > S1 (0.3) > S3 (0.2), and S2 is selected as the first candidate state.

[0138] Step S1359: Extract the values ​​of non-common feature dimensions from the optimized data feature vector and analyze the similarity between the values ​​of non-common feature dimensions and the historical values ​​of non-common feature dimensions corresponding to the first candidate state.

[0139] Fifty-six non-shared feature dimensions (such as the surface charge distribution and acoustic impedance of blades collected by a novel sensor) were extracted from the optimized data feature vector. Historical non-shared feature dimension data (56 dimensions) corresponding to state S2 were retrieved from the temporal correlation network. The similarity between the current non-shared feature vector and the historical non-shared feature vector was calculated using Manhattan distance: distance = Σ|current value - historical mean| / historical standard deviation, resulting in a similarity score (range [0, 1], with a higher score for smaller distances). For example, if the Manhattan distance between the current non-shared feature vector and the historical mean is 12.5, and the sum of the historical standard deviations is 25, then the similarity score is 1 - 12.5 / 25 = 0.5.

[0140] Step S13510: If the similarity meets the preset standard, the first candidate state is determined as the preliminary operating status evaluation result of the component.

[0141] The preset similarity threshold is 0.6. The current similarity score is 0.5 < 0.6, which does not meet the standard. Proceed to step S13511. If the similarity score is ≥ 0.6, then S2 is directly determined as the preliminary operational status assessment result.

[0142] Step S13511: If the similarity does not meet the preset standard, select the candidate running state with the second highest probability of occurrence from the candidate running state set as the second candidate state, and repeat the analysis of the similarity between the non-common feature dimension value and the historical non-common feature dimension value corresponding to the second candidate state.

[0143] S1, which has the second highest probability of occurrence, is selected as the second candidate state. Historical non-shared feature dimension data corresponding to the S1 state are retrieved. The similarity score between the current non-shared feature vector and the historical data of S1 is calculated. If the similarity score is 0.7 ≥ 0.6, which meets the preset standard, S1 is determined as the preliminary operational status assessment result.

[0144] Step S136: If the preliminary operating status assessment result meets the reasonableness requirements, the association layer transmits the preliminary operating status assessment result to the inference layer, and the inference layer outputs the preliminary operating status assessment result as the real-time status assessment result of the component.

[0145] If the association layer finds that the preliminary operational status assessment result of the blade component meets the reasonableness requirements (e.g., the preliminary operational status assessment result is consistent with the changing trends of scene matching degree and historical similarity, and is within the preset reasonableness threshold range), the association layer transmits the preliminary operational status assessment result to the inference layer. Upon receiving it, the inference layer outputs the preliminary operational status assessment result as the real-time status assessment result of the blade component.

[0146] Step S137: If the preliminary operational status assessment result does not meet the reasonableness requirements, the association layer adjusts the calculation method of scene matching degree and historical similarity, feeds back to the perception layer to re-extract features, and the inference layer infers again based on the new feature vector until a real-time status assessment result that meets the reasonableness requirements is generated.

[0147] If the association layer finds that the preliminary operational status assessment of the blade component does not meet the reasonableness requirements, it will adjust the calculation method of scene matching degree and historical similarity, such as changing the allocation of dimension weights or adjusting the algorithm parameters for similarity calculation. The adjusted calculation method is then fed back to the perception layer, which re-extracts features from the real-time operational data set based on the new calculation method, generating a new optimized data feature vector. The inference layer then performs state inference again based on the new feature vector, generating a new preliminary operational status assessment result, which is fed back to the association layer for reasonableness checking. This process may be repeated multiple times until a real-time status assessment result that meets the reasonableness requirements is generated.

[0148] Step S138: Summarize the real-time status assessment results of each component according to component type to form the real-time status assessment results of each component of the wind power generation equipment.

[0149] For other components of the wind power generation equipment (such as gearboxes and generators), the same processing procedure as for the blade components is adopted to generate their own real-time status assessment results. Then, the assessment results are summarized according to component type (blades, gearboxes, generators, etc.), and the real-time status assessment results of all components are integrated to form the real-time status assessment results of all components of the entire wind power generation equipment.

[0150] Step S140: Generate a component operation control scheme based on the real-time status assessment results. The component operation control scheme includes component operation specifications for different operating states.

[0151] Step S141: Receive the real-time status assessment results of each component of the wind power generation equipment, classify the real-time status assessment results according to component type and operating status category, and determine the current operating status of each component.

[0152] For example, the evaluation results can be categorized by component type, such as blades, gearboxes, and generators. Within each component type, they can be further classified according to their operating status (normal operation, minor fault, severe fault, etc.). This classification clarifies the specific operating status of each component.

[0153] Step S142: For the current operating state of each component, analyze the factors affecting the component's performance under this operating state. The factors include the component's operating parameter settings, the degree of external environmental interference, and the degree of component wear.

[0154] Taking blade components as an example, if they are currently in a state of minor failure, analyze the factors affecting their performance. Regarding operating parameter settings, are the blade rotation speed and angle of attack settings reasonable? Excessive rotation speed may exacerbate wear on the faulty part. Regarding the degree of external environmental interference, is the current wind speed and direction stable? Strong winds or frequent changes in wind direction may exert greater impact on the faulty blade. Regarding component wear, are there any wear or corrosion on the blade surface? The degree of wear directly affects the aerodynamic performance and structural strength of the blade. For gearbox components, if they are in normal operating condition, influencing factors may include the gearbox's lubricating oil temperature, gear speed, and other operating parameter settings; the impact of external environmental temperature and humidity on gearbox heat dissipation; and the degree of gear wear.

[0155] Step S143: Obtain historical control records of the component corresponding to the same operating state during historical operation. The historical control records include the content of the historical control operation parameter adjustment, control operation steps, and changes in the component's operating state after control.

[0156] For example, for minor faults in blade components, the historical database is used to search for control records when the blades were in a minor fault state in the past. These records include the adjustments made to the blade operating parameters (such as rotation speed and windward angle) during the historical control, the specific control operation steps (such as how to gradually adjust the windward angle), and the changes in the blade operating state after control (such as whether the fault was alleviated and whether the power generation efficiency was restored).

[0157] Step S144: Analyze the correspondence between the adjustment of operating parameters in the historical control records and the changes in the operating status of components, and determine the effectiveness of different operating parameter adjustments in improving the operating status of components.

[0158] Taking the blade component as an example, historical control records show that when the rotational speed was adjusted from one value to another, and the angle of attack was adjusted from one angle to another, minor blade malfunctions were alleviated, and the operating condition returned to normal. Analysis of multiple sets of similar historical data determined the effectiveness of different operating parameter adjustments in improving component performance. For example, reducing the rotational speed by 20% was more effective in alleviating minor malfunctions than reducing it by 10%, and adjusting the angle of attack by 5 degrees was more effective than adjusting it by 3 degrees, thus clarifying which operating parameter adjustments were more effective.

[0159] Step S145: Determine the adjustment direction of the component operating parameters under the current operating state based on the degree of effectiveness. The adjustment direction includes the direction of increasing, decreasing or maintaining the parameter value.

[0160] Based on the effectiveness of adjusting operating parameters in improving the operational status of components, determine the direction of parameter adjustment for the current operating state. For minor faults in blade components, if reducing rotational speed and adjusting the angle of attack are effective control measures, then the current direction of parameter adjustment is to reduce rotational speed and adjust the angle of attack in a specific direction (determining the most effective direction based on historical data). If certain operating parameters have not been significantly effective in improving the operational status or have had a negative impact in historical adjustments, then their adjustment direction is determined to be maintained, and no adjustment is made.

[0161] Step S146: Based on the adjustment direction and the range of values ​​of operating parameters in the historical control records, determine the target adjustment value of the component operating parameters under the current operating state, so that the target adjustment value is within the parameter range for safe operation of the component.

[0162] For example, if the blade rotation speed is adjusted to decrease, and the historical control records show that the rotation speed values ​​fall within a certain range, and the values ​​within this range ensure safe blade operation, then, based on the current operating status and effectiveness analysis, a suitable target adjustment value is selected within this range. This could involve adjusting the rotation speed from its current value to a historically effective value that falls within a safe range, ensuring that the target adjustment value does not exceed the parameter range for safe component operation.

[0163] Step S147: Set the component operation specifications for the current operating state. The component operation specifications include the order of adjusting operating parameters, key monitoring points during the adjustment process, and the required stabilization time after adjustment.

[0164] Regarding the order of adjusting operating parameters, for example, for blade components, adjust the angle of attack first, then the rotation speed, to avoid sudden changes in blade stress that might occur if the rotation speed is adjusted first. Key monitoring points during the adjustment process include real-time monitoring of blade vibration changes, stress intensity data, and temperature changes to ensure the blade condition does not deteriorate further during adjustment. The required stabilization time after adjustment means that after adjustment, the operating parameter needs to be maintained for a period of time (e.g., 30 minutes) to observe whether the blade's operating condition stabilizes at the improved state before deciding whether to proceed with the next step.

[0165] Step S148: For different operating states that the same component may experience, set the corresponding operating parameter adjustment content and component operation specifications respectively.

[0166] For different operating states that the same component may experience, such as normal operation, minor fault, and severe fault of the blade component, corresponding operating parameter adjustments and component operation procedures should be set. Under normal operation, the operating parameter adjustments may be minor to optimize power generation efficiency, and the operation procedures are relatively simple. Under minor fault, parameter adjustments and procedures should be formulated according to the above steps. Under severe fault, more stringent parameter adjustments (such as emergency shutdown) and detailed operation procedures (such as safety inspection procedures after shutdown) may be required.

[0167] Step S149: Integrate the operating parameter adjustment contents and component operation specifications corresponding to different operating states of each component according to component type to form a component operation control plan.

[0168] The operating parameters and operating specifications for each component (blades, gearbox, generator, etc.) under different operating conditions are integrated together. Each component's control measures are documented separately, outlining the operating steps and parameter requirements for different operating conditions, forming a complete component operation control plan. This plan provides clear and specific operational guidance for different operating conditions of various wind power equipment components, facilitating execution by equipment operators.

[0169] Step S150: Transmit the real-time status assessment results and component operation control scheme to the equipment control terminal, and update the inference parameters of the time-series correlation network and hierarchical machine learning inference model in combination with the component control results fed back by the equipment control terminal.

[0170] The real-time status assessment results of each component of the wind power generation equipment and the formulated component operation control plan are transmitted to the equipment control terminal. After receiving this information, the equipment control terminal adjusts the operating parameters of each component according to the control plan and feeds back the control results. Based on the feedback control results, the inference parameters of the time-series correlation network and hierarchical machine learning inference model are updated to improve the accuracy of subsequent status monitoring and the effectiveness of the control plan.

[0171] Step S151: Transmit the real-time status assessment results and component operation control plan to the equipment control terminal, so that after receiving the real-time status assessment results and component operation control plan, the equipment control terminal will perform component operation parameter adjustment operations according to the component operation specifications in the component operation control plan.

[0172] Real-time status assessment results and component operation control plans are transmitted to the equipment control terminal via wired or wireless communication. Upon receiving this data, the receiving module of the equipment control terminal parses and verifies the data to ensure its integrity and accuracy. Then, based on the operating specifications for each component in the component operation control plan, the control module of the equipment control terminal generates specific control commands and sends them to the corresponding actuators (such as the blade pitch mechanism and the gearbox speed control mechanism) to execute component operating parameter adjustments, such as adjusting the blade's angle of attack and the gearbox's output speed.

[0173] Step S152: During the adjustment operation, the equipment control terminal collects real-time operating data and changes in operating status of the components to form component control results. The component control results include the adjusted operating parameter values, component status change data during the adjustment process, and the adjusted stable operating status of the components.

[0174] For example, during the adjustment of the blade's windward angle, real-time operational data such as blade vibration changes, stress intensity, and rotational speed are collected, while simultaneously monitoring the change in the blade's operating status from a minor fault state to a normal state. The adjusted operating parameter values ​​(such as the final adjusted windward angle), the component status change data during the adjustment process (such as the vibration amplitude change curve with angle adjustment), and the stable operating state achieved by the component after adjustment (such as stable power generation and temperature) are integrated to form the component control result.

[0175] Step S153: Receive the component control results, extract the adjusted operating parameter values, adjusted operating data, adjusted component operating status and corresponding operating scenarios from the component control results, add the adjusted operating data, corresponding operating scenarios and adjusted component operating status to the time-series correlation network, supplement the network nodes and connection edges, and update the time weights between nodes.

[0176] Step S1531: Receive the component control results fed back by the device control terminal. The component control results include component identification, control operation time, adjusted operating parameter values, component status change data during the adjustment process, and adjusted stable operating status of the component.

[0177] The equipment control terminal feeds back the control results data of the blade components through the 5G communication module, which is encapsulated in JSON format and includes the following fields: Component Identifier = "YB001", Control Operation Time = "2023-10-20 T16:00:00Z", Adjusted Operating Parameter Values ​​= {"Angle of Attack": 12.5°, "Rotation Speed": 15.2rpm, "Pitch Rate": 2.0° / s}, Component Status Change Data during Adjustment = 3D Array (Time Step = 100, Feature Dimension = 256), Adjusted Stable Operating Status of the Component = {"Status Level": S0, "Stability Duration": 30 minutes, "Performance Indicators": {"Power Generation": 1.8MW, "Vibration Amplitude": 0.5g, "Temperature Deviation": 2℃}}.

[0178] Step S1532: Classify the component control results according to component identification, and group the control results corresponding to the same component identification into the same category to facilitate subsequent component-based analysis.

[0179] The data processing module reads the "component identifier" field from the control results and stores the control results corresponding to "YB001" in the blade component data buffer using a hash table. This buffer is stored separately from the control results of other components such as the gearbox ("CL001") and generator ("FD001"). The buffer adopts a circular queue structure with a capacity of 1000 records, arranged in ascending order by control operation timestamp, and supports fast retrieval by component identifier.

[0180] Step S1533: Extract the adjusted operating parameter values, component status change data during the adjustment process, adjusted stable operating status of components and corresponding operating scenarios from the classified control results. Merge the component status change data during the adjustment process and the operating data corresponding to the adjusted stable operating status of components to form adjusted operating data.

[0181] The "adjusted operating parameter values" field is extracted from the blade component data buffer and stored as a key-value pair structure. A three-dimensional array of "component state change data during the adjustment process" is extracted and converted into a time-series data frame (timestamp + feature vector). The "adjusted stable operating state of the component" field is extracted, and the state level, stability duration, and performance indicators are parsed. The corresponding operating scenario description (retrieved from the scenario database) is associated with the "control operation time." The state change data during the adjustment process (100 time steps) and the adjusted stable operating data (30 time steps, 1 minute interval between steps) are merged to form 130 time steps of adjusted operating data, each time step containing a 256-dimensional feature vector.

[0182] Step S1534: Parse the description of the running scenario corresponding to the adjusted running data, and determine the feature vector of the running scenario corresponding to the adjusted running data.

[0183] The operational scenario description includes a description of the physical action pattern ("bolted connection is tight, blade root is under uniform force, rotation trajectory deviation <0.5°") and a description of the environmental interaction method ("wind speed 12m / s, wind direction is stable, salt spray concentration 0.01mg / m³"). The description text is converted into structured data using a Natural Language Processing (NLP) module and mapped to a 128-dimensional operational scenario feature vector. The physical action pattern feature component is generated through fuzzy logic reasoning, while the environmental interaction method feature component is calibrated using sensor data.

[0184] Step S1535: Add the adjusted running data, the corresponding running scenario feature vector, and the adjusted stable running status of the components to the time-series correlation network, and supplement the network nodes and connection edges.

[0185] The update module of the temporal correlation network adds the adjusted running data (130×256 feature matrix) as a new temporal data node to the network's input layer; adds the running scene feature vector (128 dimensions) to the scene feature layer; and adds the adjusted stable operating state (S0) of the component to the state node layer. Based on the data flow, it supplements the connection edges between the input layer and the scene feature layer, between the scene feature layer and the state node layer, and between the input layer and the state node layer, with each connection edge associated with the corresponding timestamp and feature dimension.

[0186] Step S1536: Based on the adjusted running data and the duration of the stable running state of the adjusted components, update the time weights of the corresponding connection edges in the time-series correlation network.

[0187] The temporal correlation network uses a time-decay weighted algorithm to update the edge weights. The weight calculation formula is: W_new = W_old × exp(-λ × Δt) + α × T, where λ = 0.01 (decay coefficient), Δt = adjustment interval (hours), α = 0.1 (learning rate), and T = stable operating time after adjustment (hours). The weights are updated to the entire network through backpropagation to ensure that the temporal weights reflect the latest adjustment results.

[0188] Step S1537: Output the updated temporal correlation network structure and weight parameters for the next state evaluation.

[0189] The updated temporal correlation network is serialized in protobuf format and stored in a distributed database. The version number is equal to the original version number plus 1. The update log records information such as adjustment time, component identifier, and weight change. Simultaneously, a network structure visualization file (SVG format) is generated, annotating newly added nodes and connections, facilitating network evolution monitoring for operations personnel.

[0190] Step S154: Analyze the feature differences between the adjusted running data and the real-time running data set before adjustment, and adjust the feature extraction weights and dimensions of the perception layer of the hierarchical machine learning inference model according to the feature differences.

[0191] For example, the vibration amplitude characteristics of the blade vibration change data after adjustment were lower than before adjustment, and the peak force characteristics of the stress intensity data were reduced. Based on these feature differences, the feature extraction weights of the perception layer in the hierarchical machine learning inference model were adjusted. For feature dimensions with significant differences, their extraction weights were increased, so that the perception layer would pay more attention to these features that can reflect the control effect in subsequent feature extraction. At the same time, the dimensions of feature extraction were adjusted according to the feature differences. If some new feature dimensions showed importance in the adjusted data, these dimensions were increased; if some original dimensions had no significant differences and had little impact on state assessment, these dimensions were decreased.

[0192] Step S155: Analyze the correspondence between the adjusted component operating status and the real-time status evaluation results, adjust the scene matching degree and historical similarity calculation logic of the hierarchical machine learning inference model association layer according to the correspondence, and adjust the timing correspondence rule calling method of the inference layer.

[0193] Analyze the correspondence between the adjusted component operating status and the real-time status assessment results. If the adjusted component operating status is consistent with the real-time status assessment results (e.g., assessed as a minor fault state, restored to normal after adjustment), it indicates that the scene matching degree and historical similarity calculation logic of the association layer and the time-series correspondence rule calling method of the inference layer are reasonable. If they are inconsistent, the calculation logic of the association layer needs to be adjusted, such as changing the dimension weights or similarity calculation method in the scene matching degree calculation to make the association analysis more accurate. At the same time, the time-series correspondence rule calling method of the inference layer should be adjusted. For example, in the current operating scenario, the time-series correspondence rules related to the latest control results should be called first to improve the accuracy of state inference.

[0194] Step S156: Apply the updated temporal correlation network and the adjusted model inference parameters to the next wind power equipment condition monitoring process.

[0195] After updating the temporal correlation network and model inference parameters, the system automatically triggers the model deployment process. The updated temporal correlation network file is pushed to each node in the state monitoring service cluster via the internal API, replacing the old version file. The adjusted model inference parameters (such as the feature weights of the perception layer, the attention coefficients of the correlation layer, and the LSTM weights of the inference layer) are broadcast to each inference node through the parameter server. After loading the new parameters, the nodes undergo a warm-up process to ensure that the inference latency is <50ms. At the start of the next state monitoring cycle (default 5 minutes), the real-time operating data of the blades collected by the data acquisition module is directly input into the updated system for processing, realizing online iterative optimization of the model.

[0196] In one exemplary embodiment, a wind power generation equipment condition monitoring system combining the Internet of Things (IoT) and machine learning is provided. This system can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2As shown, this wind power equipment condition monitoring system combining IoT and machine learning includes a processor, memory, input / output interfaces, a communication interface, a display unit, and input devices. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input devices are also connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interfaces are used for information exchange between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by a processor, this computer program implements a method for monitoring the condition of wind power generation equipment that combines the Internet of Things (IoT) and machine learning. The display unit of this IoT-integrated machine learning wind power generation equipment condition monitoring system is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device for this IoT-integrated machine learning wind power generation equipment condition monitoring system can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the system's casing, or an external keyboard, touchpad, or mouse.

[0197] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for condition monitoring of wind power generation equipment combining the Internet of Things and machine learning, characterized in that, The method includes: Based on the dynamic configuration of IoT sensing nodes for the operation scenarios of various components of wind power generation equipment, a set of real-time operation data is collected. The operation scenarios include the physical action mode and environmental interaction mode of the components during operation, and the set of real-time operation data includes the physical state data and environmental impact data generated by the components under different operation scenarios. A time-series correlation network is established between the real-time operating data set and the operating scenarios and component operating states. The time-series correlation network is used to characterize the component operating states corresponding to the changes of the real-time operating data set over time under different operating scenarios. A hierarchical machine learning inference model is invoked to perform collaborative inference processing on the real-time running data set and time series association network to generate real-time status evaluation results for each component of the wind power generation equipment. A component operation control plan is generated based on the real-time status assessment results. The component operation control plan includes component operation specifications for different operating states. The real-time status assessment results and component operation control schemes are transmitted to the equipment control terminal. Based on the component control results fed back by the equipment control terminal, the inference parameters of the time-series correlation network and hierarchical machine learning inference model are updated.

2. The method for monitoring the condition of wind power generation equipment combining the Internet of Things and machine learning according to claim 1, characterized in that, The method involves dynamically configuring IoT sensing nodes based on the operational scenarios of various components of wind power equipment to collect real-time operational data sets, including: The physical interaction patterns of each component of a wind power generation device during operation are analyzed, including the transmission contact mode between components, the force distribution pattern, and the characteristics of the motion trajectory. The environmental interaction mode of each component during operation is analyzed, including the contact angle between the component and the airflow, the heat exchange path with surrounding components, and the interaction mode with environmental impurities. By integrating physical action patterns and environmental interaction methods, the operating scenarios corresponding to each component are divided, and the operating characteristics of the component under each operating scenario are determined. Based on the physical characteristics of the components in each operating scenario, select the matching IoT sensing node type. The IoT sensing node type must match the component's transmission contact method, force distribution pattern, and motion trajectory characteristics. Based on the characteristics of the environmental interaction methods of components in each operating scenario, determine the installation location of IoT sensing nodes; Deploy selected types of IoT sensing nodes at the determined installation locations, and set the data collection interval of the sensing nodes according to the frequency of changes in the operating scenario; The IoT sensing node is activated to collect physical state data generated by the component in different operating scenarios during the corresponding time periods. The physical state data includes vibration change data, force intensity data, temperature change data, and movement speed data of the component during transmission. The environmental impact data of the environment in which the component is located during different operating scenarios and time periods includes airflow speed change data, airflow direction change data, ambient temperature stability data, ambient humidity change data, and ambient impurity concentration data. The collected physical state data and environmental impact data are categorized and organized according to the operating scenarios. The physical state data and environmental impact data of the same time period under the same operating scenario are associated to form a real-time operating data set. The real-time running data set is arranged in order of collection time, and the running scenario identifier corresponding to each data item is labeled.

3. The method for monitoring the condition of wind power generation equipment combining the Internet of Things and machine learning according to claim 1, characterized in that, The establishment of a time-series correlation network between the real-time operating data set and the operating scenarios and component operating states includes: Acquire historical operating data sets, historical operating scenario records, and corresponding historical component operating status records accumulated by each component of the wind power generation equipment during historical operation; The historical operation data set, historical operation scenario record, and historical component operation status record are classified according to component type to form the historical data scenario status dataset corresponding to each component; The physical state data and environmental impact data in the historical operation data set are standardized. For the historical data scene state dataset corresponding to each component, the operation scene features in the historical operation scene records are extracted. The operation scene features include the core attributes of the historical physical action form and the core attributes of the historical environmental interaction mode. Extract data features from the standardized historical operational data set. These data features include the changing trends of historical physical state data, the fluctuation patterns of historical environmental impact data, and the correlations between different data items. Extract the status features from the historical component operation status records. The status features include the duration of the historical operation status, the status transition conditions, and the component performance corresponding to the status. Analyze the temporal correlation between operational scenario characteristics and data characteristics, determine the pattern of data characteristics changing over time under different operational scenario characteristics, and establish a temporal correspondence between scenario data; Analyze the temporal correlation between data features and state features, determine the state features corresponding to the changes of different combinations of data features over time, and establish a temporal correspondence between data states. Based on the temporal correspondence between scene data and the temporal correspondence between data state, an initial structure of a temporal association network is constructed. The initial structure includes running scene nodes, data nodes, state nodes, and temporal connection edges between nodes. Add time weights to the connection edges in the initial structure of the temporal association network, the time weights being determined based on the duration of the association between the corresponding nodes; Select some historical scenario state data that did not participate in the initial structure construction, input them into the initial structure of the time series correlation network for verification, and observe whether the component operation status output by the network matches the historical component operation status records; Based on the verification results, the node division, connection edge settings, and time weights in the initial structure of the temporal correlation network are adjusted so that the initial temporal correlation network can represent the component operating status corresponding to the change of real-time operating data set over time under different operating scenarios, thus forming the final temporal correlation network.

4. The method for monitoring the condition of wind power generation equipment combining the Internet of Things and machine learning according to claim 1, characterized in that, The hierarchical machine learning inference model is invoked to perform collaborative inference processing on the real-time operating data set and time-series association network to generate real-time status evaluation results for each component of the wind power generation equipment, including: Obtain the structural information of the hierarchical machine learning inference model, which includes the functional division of the perception layer, the association layer, and the inference layer, as well as the data transmission relationship between the layers. The real-time running data set is input into the perception layer of the hierarchical machine learning inference model. The perception layer extracts features from the real-time running data set to generate a real-time data feature vector. The real-time data feature vector includes physical state data feature components and environmental influence data feature components. The current running scenario feature vector and the historical time period related data feature vector are extracted from the temporal correlation network. The running scenario feature vector and the related data feature vector are input into the correlation layer of the hierarchical machine learning inference model. The correlation layer performs correlation analysis on the real-time data feature vector, the running scenario feature vector and the related data feature vector, calculates the scenario matching degree between the real-time data feature vector and the running scenario feature vector, and calculates the historical similarity between the real-time data feature vector and the related data feature vector. The association layer feeds back the scene matching degree and historical similarity to the perception layer. The perception layer adjusts the extraction weights of the physical state data feature components and the environmental influence data feature components according to the scene matching degree, and adjusts the dimension of feature extraction according to historical similarity to generate an optimized data feature vector. The optimized data feature vector is input into the inference layer of the hierarchical machine learning inference model. The inference layer calls the temporal correspondence rules between data features and states in the current running scenario from the temporal association network, and performs state inference on the optimized data feature vector based on the temporal correspondence rules. After generating the preliminary running status assessment results of the component, the preliminary running status assessment results are fed back to the association layer. The association layer checks the rationality of the preliminary running status assessment results by combining the scene matching degree and historical similarity. If the preliminary operational status assessment results meet the reasonableness requirements, the association layer will transmit the preliminary operational status assessment results to the inference layer, and the inference layer will output the preliminary operational status assessment results as the real-time status assessment results of the component. If the initial operational status assessment results do not meet the requirements of reasonableness, the association layer adjusts the calculation method of scene matching degree and historical similarity, feeds back to the perception layer to re-extract features, and the reasoning layer reasons again based on the new feature vector until a real-time status assessment result that meets the requirements of reasonableness is generated. The real-time status assessment results of each component are summarized according to component type to form the real-time status assessment results of each component of the wind power generation equipment.

5. The method for monitoring the condition of wind power generation equipment combining the Internet of Things and machine learning according to claim 4, characterized in that, The acquisition of structural information of the hierarchical machine learning inference model includes: Access the machine learning model storage system and extract the design specification file of the hierarchical machine learning inference model; Read the functional information of the perception layer from the design specification file. The functional information includes the input data type of the perception layer, the feature extraction algorithm type, the output feature vector dimension, and the step division of feature extraction. Read the functional information of the association layer from the design specification file. The functional information includes the input parameter type of the association layer, the type of association analysis algorithm, the calculation logic of scene matching degree and historical similarity, and the output result format. Read the functional information of the inference layer from the design specification document. The functional information includes the input feature type of the inference layer, the state inference algorithm type, the calling method of the timing correspondence rule, and the type classification of the output evaluation result. Extract the data transmission relationship between the perception layer and the association layer from the design specification document. The data transmission relationship includes the path for transmitting the output feature vector of the perception layer to the association layer, the format requirements for the transmitted data, and the triggering conditions for data transmission. Extract the data transmission relationship between the association layer and the inference layer from the design specification document. The data transmission relationship includes the path for transmitting the scene matching degree and historical similarity output of the association layer to the inference layer, the verification method of the transmitted data, and the frequency of data transmission. Extract the feedback transmission relationship between the association layer and the perception layer from the design specification document. The feedback transmission relationship includes the path for the association layer to transmit the adjustment command to the perception layer, the parsing method of the feedback data, and the execution flow of the feedback command. Extract the feedback transmission relationship between the inference layer and the association layer from the design specification document. The feedback transmission relationship includes the path for the preliminary evaluation results output by the inference layer to be transmitted to the association layer, the analysis method of the feedback data, and the judgment criteria of the feedback results. By integrating the functional information of the perception layer, association layer, and reasoning layer, as well as the data transmission relationships between the layers, the structural information of the hierarchical machine learning reasoning model is formed.

6. The method for monitoring the condition of wind power generation equipment combining the Internet of Things and machine learning according to claim 4, characterized in that, The process of inputting a real-time running dataset into the perception layer of a hierarchical machine learning inference model, and the perception layer extracting features from the real-time running dataset to generate a real-time data feature vector, includes: The real-time operational data set is divided into a physical state data subset and an environmental impact data subset according to data type. The physical state data subset includes vibration change data, force intensity data, temperature change data, and motion speed data. The environmental impact data subset includes airflow speed change data, airflow direction change data, environmental temperature stability data, environmental humidity change data, and environmental impurity concentration data. The physical state data subset and the environmental impact data subset are standardized respectively. Feature extraction is performed on the standardized physical state data subset. Vibration period features and vibration amplitude features are extracted from vibration change data. Peak force features and force distribution features are extracted from force intensity data. Temperature rise and fall rate features and temperature stable range features are extracted from temperature change data. Velocity change gradient features and velocity stability duration features are extracted from motion velocity data. Feature extraction is performed on the standardized environmental impact data subset. The features of velocity fluctuation amplitude and velocity change frequency are extracted from the airflow velocity change data; the features of directional deflection angle and directional stability duration are extracted from the airflow direction change data; the features of temperature fluctuation range and temperature mean are extracted from the stable environmental temperature data; the features of humidity change rate and humidity stability interval are extracted from the environmental humidity change data; and the features of concentration change amplitude and concentration peak are extracted from the environmental impurity concentration data. The features extracted from the physical state data subset are arranged in a preset order to form physical state data feature components, with each feature corresponding to a dimension of the physical state data feature component. The features extracted from the subset of environmental impact data are arranged in a preset order to form environmental impact data feature components, with each feature corresponding to a dimension of the environmental impact data feature component. Adjust the dimensions of the physical state data feature components and the environmental impact data feature components, and combine the adjusted physical state data feature components and environmental impact data feature components to form a real-time data feature vector.

7. The method for monitoring the condition of wind power generation equipment combining the Internet of Things and machine learning according to claim 4, characterized in that, The association layer performs association analysis on the real-time data feature vector, the runtime scenario feature vector, and the associated data feature vector, calculates the scenario matching degree between the real-time data feature vector and the runtime scenario feature vector, and calculates the historical similarity between the real-time data feature vector and the associated data feature vector, including: The runtime scenario feature vector corresponding to the current runtime scenario is obtained from the temporal correlation network. The runtime scenario feature vector includes physical action form feature components and environmental interaction mode feature components under the current runtime scenario. The feature vectors of related data that are the same as the historical time period and the current operating scenario are obtained from the temporal correlation network. The feature vectors of related data include historical physical state data feature components and historical environmental impact data feature components. The real-time data feature vector and the running scene feature vector are aligned in dimension, and the similarity between the physical state data feature components in the real-time data feature vector and the physical action form feature components in the running scene feature vector is calculated. The correspondence analysis method between features is used to determine the matching degree of each dimension feature, and the physical scene matching score is calculated by weighting according to the dimension weight. Calculate the similarity between the environmental impact data feature components in the real-time data feature vector and the environmental interaction mode feature components in the operational scenario feature vector. Use the same feature correspondence analysis method to determine the matching degree of each dimension feature. Calculate the environmental scenario matching score by weighting according to the dimension weight. The physical scene matching score and the environmental scene matching score are integrated according to a preset ratio to obtain the scene matching degree between the real-time data feature vector and the running scene feature vector; The real-time data feature vector and the associated data feature vector are dimensionally aligned, and the similarity between the physical state data feature components in the real-time data feature vector and the historical physical state data feature components in the associated data feature vector is calculated. The consistency of the changing trend of each dimension feature is analyzed, and the physical history similarity score is obtained by weighting according to the dimension weight. Calculate the similarity between the environmental impact data feature components in the real-time data feature vector and the historical environmental impact data feature components in the associated data feature vector, analyze the consistency of the fluctuation patterns of each dimension feature, and calculate the historical environmental similarity score by weighting according to the dimension weight; The physical history similarity score and the environmental history similarity score are integrated according to a preset ratio to obtain the historical similarity between the real-time data feature vector and the associated data feature vector, thereby outputting the scene matching degree and historical similarity.

8. The method for monitoring the condition of wind power generation equipment combining the Internet of Things and machine learning according to claim 1, characterized in that, The process of generating a component operation control plan based on real-time status assessment results includes: Receive real-time status assessment results of each component of the wind power generation equipment, classify the real-time status assessment results according to component type and operating status category, and determine the current operating status of each component; For each component's current operating status, analyze the factors affecting the component's performance under this operating status. These factors include the component's operating parameter settings, the degree of external environmental interference, and the degree of component wear. Obtain historical control records of the component corresponding to the same operating state during historical operation. The historical control records include the content of the historical control of the operating parameter adjustment, the control operation steps, and the changes in the component's operating state after the control. Analyze the correspondence between the adjustment of operating parameters in historical control records and the changes in the operating status of components to determine the effectiveness of different operating parameter adjustments in improving the operating status of components; The adjustment direction of the component's operating parameters under the current operating state is determined based on the degree of effectiveness. The adjustment direction includes the direction of increasing, decreasing, or maintaining the parameter value. Based on the adjustment direction and the range of values ​​of operating parameters in historical control records, determine the target adjustment value of the component operating parameters under the current operating state, so that the target adjustment value is within the parameter range for safe operation of the component; Establish component operation specifications for the current operating state. The component operation specifications include the order of adjusting operating parameters, key monitoring points during the adjustment process, and the required stabilization time after adjustment. For different operating states that the same component may experience, corresponding operating parameter adjustments and component operation specifications should be set. The operating parameter adjustment content and operating specifications corresponding to different operating states of each component are integrated according to component type to form a component operation control plan.

9. The method for monitoring the condition of wind power generation equipment combining the Internet of Things and machine learning according to claim 1, characterized in that, The process of transmitting real-time status assessment results and component operation control schemes to the equipment control terminal, and updating the inference parameters of the time-series correlation network and hierarchical machine learning inference model based on the component control results fed back by the equipment control terminal, includes: The real-time status assessment results and component operation control plan are transmitted to the equipment control terminal so that after receiving the real-time status assessment results and component operation control plan, the equipment control terminal can perform component operation parameter adjustment operations according to the component operation specifications in the component operation control plan. During the adjustment operation, the equipment control terminal collects real-time operating data and changes in operating status of the components to form component adjustment results. The component adjustment results include the adjusted operating parameter values, component status change data during the adjustment process, and the adjusted stable operating status of the components. Receive the component control results, extract the adjusted operating parameter values, adjusted operating data, adjusted component operating status and corresponding operating scenarios from the component control results, add the adjusted operating data, corresponding operating scenarios and adjusted component operating status to the time series correlation network, supplement the network nodes and connection edges, and update the time weights between nodes. Analyze the feature differences between the adjusted running data and the real-time running data set before the adjustment, and adjust the feature extraction weights and dimensions of the perception layer of the hierarchical machine learning inference model according to the feature differences; Analyze the correspondence between the adjusted component operating status and the real-time status evaluation results, adjust the scene matching degree and historical similarity calculation logic of the hierarchical machine learning inference model association layer according to the correspondence, and adjust the time sequence correspondence rule calling method of the inference layer. The updated temporal correlation network and the adjusted model inference parameters will be applied to the next wind power equipment condition monitoring process.

10. A condition monitoring system for wind power generation equipment combining the Internet of Things and machine learning, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the wind power generation equipment condition monitoring method combining Internet of Things and machine learning as described in any one of claims 1 to 9 by executing the machine-executable instructions.

Images