Wind driven generator blade icing monitoring method

Abstract

The invention discloses a wind driven generator blade icing monitoring method, and particularly relates to the technical field of wind power generation. According to the method, a multi-dimensional primary ice condition feature tensor is constructed, thermal anomaly spot density, stress amplitude change rate and disturbance frequency anomaly indexes are extracted, a blade surface multi-scale feature topology network is constructed based on node feature relevancy, and a potential freezing point coupling aggregation area is identified; further, through an energy change entropy calculation and regression prediction model, outputting a dynamic probability value generated by a microscopic ice nucleus, constructing a recessive ice condition propagation path map, generating an icing warning curve, and realizing real-time grading judgment on the blade state; if it is judged that the freezing state is the hidden freezing state, local temperature rise control is triggered, and directional temperature control operation is executed on the corresponding heating units; the method has the high-time-efficiency hidden ice condition recognition and response capacity, and the operation safety and the intelligent level of the wind generating set in the cold region can be remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, and more specifically to a method for monitoring icing on wind turbine blades. Background Technology

[0002] As a crucial component of renewable energy, the stability and efficiency of wind power generation are particularly critical under extreme climatic conditions. In cold regions or high-altitude environments, wind turbine blades are highly susceptible to icing due to meteorological conditions such as low temperatures, strong winds, and high humidity during operation. This leads to a sharp decline in the aerodynamic performance of the blades, resulting in a series of safety hazards, including reduced power generation efficiency, increased structural vibration, and even blade breakage.

[0003] Existing methods for monitoring blade icing mainly rely on temperature and humidity sensors, visual recognition systems, or abnormal blade vibration frequency monitoring, but these methods have the following drawbacks: Conventional sensor deployment is limited by the complexity of blade structure and high-speed rotation during operation, making it difficult to achieve comprehensive coverage and high-density monitoring of the entire blade surface, and easily leading to the risk of unidentified local ice layers. Existing methods struggle to establish a dynamic causal mapping between minute thermal anomalies, airflow disturbances, and potential icing behavior in monitoring data, lacking the ability to predict microscopic icing inducing factors and resulting in poor early warning timeliness. Furthermore, most ice condition models rely too heavily on training data and are difficult to adapt to complex scenarios with coupled changes in multiple factors such as terrain, wind direction, temperature, and humidity, resulting in poor generalization ability of icing prediction results. Summary of the Invention

[0004] The purpose of this invention is to provide a method for monitoring icing on wind turbine blades, in order to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring icing on wind turbine blades, comprising: Acquire multi-source weak signal distribution data on the blade surface during wind turbine operation, including infrared thermal imaging matrix, blade stress wave signal, and local air disturbance velocity field data; Based on multi-source weak signal distribution data, a multidimensional primary ice condition feature tensor is constructed, which includes thermal anomaly spot density distribution, stress waveform amplitude change rate and disturbance frequency anomaly index. Based on the node correlation degree of the multidimensional primary ice condition feature tensor, a multi-scale feature topology network of the blade surface is established to identify potential ice point coupling clustering regions. By taking hot spot outliers in potential ice point coupling and aggregation regions as high-risk observation nodes, calculating the energy change entropy of the nodes in different time windows, and outputting a dynamic prediction set of the probability of micro ice nucleus formation; A multidimensional recursive analysis is performed on the dynamic prediction value set to construct a hidden ice condition propagation path map. Based on the synchronization rate of thermal disturbance and stress feedback at the nodes in the propagation path, an icing warning curve for potential hidden ice areas is generated. Based on the icing warning curve, a real-time icing condition classification index is output, and the leaf condition is divided into four levels: normal, warning, visible icing and hidden icing. When a latent icing level is present, a heating control command is triggered to locally increase the temperature of the heating element of the blade corresponding to the icing area.

[0006] Preferably, the construction of the multidimensional primary ice condition feature tensor includes: By performing spatial grayscale gradient enhancement and regional connectivity analysis on the infrared thermal imaging matrix, thermal anomaly spots on the leaf surface were extracted and their unit area distribution density was calculated as the thermal anomaly spot density distribution. The time-frequency features of the stress wave signal were extracted using the sliding window Fourier transform algorithm to obtain the rate of change of stress waveform amplitude in different time periods. Spatial vector field stability analysis was performed on local air disturbance velocity field data, and the root mean square deviation of velocity changes between adjacent time frames was calculated as an indicator of disturbance frequency anomalies. After aligning the density distribution of the thermal anomaly spots, the rate of change of stress waveform amplitude, and the anomaly index of disturbance frequency according to the time sequence of monitoring nodes, a multidimensional primary ice condition feature tensor is constructed based on a unified blade spatial coordinate mapping rule.

[0007] Preferably, the step of establishing a multi-scale feature topology network on the blade surface to identify potential ice point coupling clustering regions includes: For each monitoring node in the multidimensional primary ice condition feature tensor, calculate the feature correlation between any two nodes for thermal anomaly spot density, stress waveform amplitude change rate and disturbance frequency anomaly index, and generate a three-dimensional feature correlation matrix between nodes. Hierarchical clustering analysis was performed on the three-dimensional feature correlation matrix to divide highly correlated nodes into multiple local subsets, and a multi-scale feature topology network of the blade surface was constructed. Based on the distribution density of the multi-scale feature topology network on the blade surface in the spatial coordinate system, the density peak clustering algorithm is used to identify potential ice point coupling clustering regions.

[0008] Preferably, the entropy of energy change of computing nodes within different time windows includes: Points in the potential freezing point coupling cluster region whose temperature values ​​deviate from the regional temperature mean by more than two standard deviations are marked as hot spot outliers and defined as high-risk observation nodes. For each high-risk observation node, the temperature change curve corresponding to that node in the infrared thermal imaging matrix is ​​extracted within a continuous time window, and the rate of change of thermal radiation energy of the node within each time window is calculated. Based on the rate of change of thermal radiation energy, the entropy of energy change of each node within a fixed time window is calculated. The energy change entropy of each node is input into the support vector regression model, which outputs a dynamic predicted set of values ​​for the probability of microscopic ice nucleus formation.

[0009] Preferably, based on the synchronization rate of node thermal disturbance and stress feedback in the propagation path, an icing warning curve for potential hidden icing areas is generated, including: The dynamic prediction set of the probability of micro ice nucleus formation at each high-risk observation node is smoothed by a time-series weighted moving average algorithm, and its trend feature vector is extracted. Based on the spatial distance between nodes and the similarity of predicted value changes, an impact matrix for the propagation of icing probability is constructed, and possible hidden ice condition propagation paths are identified. In the identified path, extract the sequence of thermal anomaly spot change rate and stress waveform amplitude change rate corresponding to the path node, and calculate the Pearson synchronization rate between the two as the icing coordination strength. Based on continuous path segments where the icing coordination intensity exceeds a set threshold, an icing warning curve for potential hidden ice zones is generated, and its risk level is dynamically labeled and output.

[0010] Preferably, based on the icing warning curve, a real-time ice condition classification index is output, including: Extract the characteristic parameters of the icing warning curve, including maximum value, average growth rate, duration, and fluctuation amplitude; An ice condition risk discrimination model is constructed based on the extracted feature parameters, and a classification criterion is set. The maximum value is greater than 0.7 and the duration is more than 15 seconds, which is judged as latent icing. The maximum value is greater than 0.6 and the fluctuation range is more than 0.3, which is judged as overt icing. When the maximum value is between 0.4 and 0.6 and the average growth rate is consistently positive, it is considered to be in a warning state; if all characteristic parameters are below the set threshold, it is considered to be in a normal state. Based on the assessment results, the current state of the blades is classified into four levels in real time: normal, warning, visible icing, or hidden icing.

[0011] Preferably, localized temperature control is applied to the heating elements of the blades in the icing area, including: Based on the location coordinates of the icing warning curve under the latent icing level, locate the corresponding high-risk observation nodes and map them to the blade spatial coordinate system; Call the control instruction table of the partitioned heating element embedded in the blade structure and match the heating unit number within the positioning coordinate range; Send a local heating control command to the matched heating unit, set the target heating temperature to the current node average temperature plus 8 degrees Celsius, and the heating duration is not less than 30 seconds; During the heating process, the density change of thermal anomaly spots in the area is continuously monitored. If it falls below the preset safety threshold, the heating control is lifted and normal operation is restored.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention integrates multi-source weak signals such as infrared thermal imaging, stress wave analysis, and air disturbance detection to construct a multi-dimensional primary icing feature tensor. Based on node correlation, it builds a multi-scale feature topology network on the blade surface, enabling highly sensitive identification of early-stage latent icing behavior in wind turbine blades. By introducing an energy change entropy index and a support vector regression model, it can dynamically predict the probability of microscopic ice nucleus formation, overcoming the problems of lagging identification and inaccurate response in existing technologies, and achieving trend warning and path prediction of latent icing conditions.

[0013] 2. Upon determining a latent icing level, this invention automatically triggers a localized heating control strategy, precisely locating the heating units corresponding to high-risk areas and implementing targeted thermal intervention. Simultaneously, it monitors real-time changes in the density of thermal anomalies, achieving closed-loop regulation of the thermal control response. This control mechanism not only significantly improves de-icing efficiency and energy utilization, avoiding energy waste caused by indiscriminate heating of the entire blade, but also ensures the safe and stable operation of the wind power generation system under extreme weather conditions, demonstrating excellent engineering adaptability. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] For examples, please refer to Figure 1As shown in this embodiment, a method for monitoring icing on wind turbine blades includes: Acquire multi-source weak signal distribution data on the blade surface during wind turbine operation, including infrared thermal imaging matrix, blade stress wave signal, and local air disturbance velocity field data.

[0018] First, multiple monitoring nodes are pre-set along the radial and chordal directions on the surface of each blade of the wind turbine, and the following three types of non-contact sensing devices are deployed accordingly: A high-sensitivity mid-wave infrared camera module (wavelength range 3~5μm) was used to continuously acquire thermal radiation images of the blade surface at a frame rate of 30 Hz. By using a multi-point thermal imaging matrix arranged from the blade tip to the root, the surface temperature field of the blade under different rotation phases could be reconstructed to obtain infrared thermal imaging matrix data.

[0019] A stress wave sensing array, constructed by embedding piezoelectric ceramic thin films (PZT) within the blade structure, is used to acquire high-frequency micro-stress signals induced by ice loading or structural disturbances during blade operation. The stress waveforms are recorded at a sampling rate of 100 kHz using a high-speed data acquisition card, and their dominant frequency distribution, energy density, and variability characteristics are calculated to form a blade stress wave signal sequence. These signals can reflect the impact of the initial formation of micro-ice layers on the mechanical state of the blade structure.

[0020] A microparticle image velocimetry (μPIV) sensor is deployed near the blade tip, and combined with a laser light sheet projection system, to record the near-surface airflow disturbance behavior during blade rotation. By using high-speed photography and image flow field reconstruction, local air disturbance velocity field data around the leading and trailing edges of the blade are obtained, and fluid anomalies such as boundary layer separation and enhanced local eddies caused by ice adhesion are identified.

[0021] The three types of sensors mentioned above collect subtle abnormal changes caused by early icing on the blade surface from three dimensions: thermal, mechanical, and fluid. The collected multi-source weak signal distribution data are synchronously time-series calibrated and uniformly mapped to the blade coordinate system to form a multimodal spatial-temporal joint sensing dataset, providing an input basis for subsequent ice condition feature extraction and risk prediction.

[0022] It should be noted that the sensor layout density, sampling frequency and data synchronization strategy in this embodiment can be flexibly adjusted according to the actual wind farm blade size, operating speed and environmental conditions to ensure that the monitoring system is adaptable and deployable in engineering.

[0023] Based on multi-source weak signal distribution data, a multidimensional primary ice condition feature tensor is constructed, which includes thermal anomaly spot density distribution, stress waveform amplitude change rate, and disturbance frequency anomaly indicators.

[0024] First, a two-dimensional spatial gray-level gradient enhancement method is employed to improve the contrast of edge regions in the infrared thermal image, targeting the infrared thermal imaging matrix. Specifically, the original infrared image is high-pass filtered using the Laplacian operator to extract regions with abrupt gray-level changes. Subsequently, region connectivity analysis is applied to the enhanced thermal image, identifying locally enclosed high-temperature regions based on the eight-neighbor connectivity rule; these regions are defined as "thermal anomaly spots."

[0025] Next, the number of thermal anomaly spots per unit area (per square meter) is counted as the thermal anomaly spot density value. To exclude false hot spots caused by non-icing factors such as direct sunlight and structural shadows, a hot spot identification threshold is set as the mean background temperature plus three times the standard deviation; areas below this threshold are not included in the statistics. Finally, the thermal anomaly spot density distribution data is mapped onto the blade spatial coordinate system to form the first type of feature input.

[0026] For the stress wave signal inside the blade, it is first segmented into time windows of 1 second each. For the stress wave signal sequence within each time window, a sliding window Fourier transform algorithm is used to extract time-frequency domain features. The sliding window length is set to 256 points, the sliding step size is 64 points, and the frequency resolution is 10 Hz.

[0027] In the frequency domain, the amplitude value sequence corresponding to the dominant frequency is extracted, and the ratio of the absolute value of the difference between the dominant frequency amplitudes of two adjacent time windows to the amplitude of the previous window is calculated. This ratio is defined as the "stress waveform amplitude change rate". For example, when the dominant frequency amplitude of the nth time window is... The amplitude of the dominant frequency in the (n+1)th time window is The rate of change is calculated as follows: Rate of change = If the amplitude change rate exceeds 20% for three consecutive time windows, the monitoring node is marked as a "high-risk node for structural disturbance". This change rate sequence is then interpolated and mapped to the blade spatial coordinates to form the second type of feature input.

[0028] For air disturbance velocity field data, anomaly identification of disturbance frequencies is achieved by comparing velocity vector fields between frames. First, the Euclidean distance is calculated for the velocity vector differences between each pair of adjacent time frames to obtain the root mean square deviation of the velocity change. Specifically, for each pixel, the sum of the squares of the velocity vector differences between frame t and frame t+1 is averaged, and then the square root is taken to obtain the root mean square deviation value.

[0029] If the root mean square deviation of a certain region is consistently more than twice the stable value of the wind field background in consecutive time frames, then the region is judged to have a "high-frequency disturbance anomaly". This disturbance frequency anomaly index is projected to the corresponding blade monitoring node and synchronously registered with the two types of features mentioned above.

[0030] The aforementioned thermal anomaly spot density distribution, stress waveform amplitude change rate, and disturbance frequency anomaly indicators were aligned according to timestamps and monitoring node numbers, and projected onto a unified blade spatial coordinate system. The coordinate system reference point was set at the blade root, with the chordal and radial directions serving as two-dimensional spatial mapping axes, and the time series as the third dimension.

[0031] Finally, the three types of feature data are fused using a tensor structure to form a three-dimensional data volume of the form (M×N×T), where M represents the number of monitoring nodes, N represents the number of feature types (thermal, mechanical, and fluid), and T represents the continuous sampling time step. This primary ice condition feature tensor serves as input for subsequent ice condition topology network identification and evolution trend modeling.

[0032] Based on the node correlation degree of the multidimensional primary ice condition feature tensor, a multi-scale feature topology network of the blade surface is established to identify potential ice point coupling clustering regions.

[0033] For each monitoring node in the primary ice condition characteristic tensor, three types of characteristic values ​​are defined: thermal anomaly spot density, stress waveform amplitude change rate, and disturbance frequency anomaly index. For any two nodes, the correlation between them on the three types of characteristics is calculated, and the Pearson correlation coefficient is used as the measurement method.

[0034] Specifically, for each feature class, the linear correlation between the feature value sequences of two nodes at T consecutive time points is calculated. The Pearson correlation coefficient is calculated by subtracting the mean from the feature value sequences of the two nodes, multiplying the results by the product of their respective standard deviations. The result ranges from -1 to 1, with the absolute value closer to 1 indicating a stronger correlation.

[0035] The above process calculates the three types of features independently, and finally forms a three-dimensional feature correlation matrix containing the correlation of the three types of features between node pairs, denoted as A(m,n,f), where m and n represent any two nodes, and f represents one of the three feature dimensions.

[0036] After constructing the three-dimensional feature correlation matrix, hierarchical clustering analysis is used to structurally partition the matrix to identify node sets with high correlation across multiple feature dimensions. The specific steps are as follows: First, the three-dimensional feature correlation matrix is ​​expanded into a set of feature correlation vectors by node pairs; Next, Euclidean distance is used as the metric for similarity between nodes, and the minimum distance method is used as the merging criterion to perform bottom-up clustering operations; A similarity threshold of 0.85 is set. When the average correlation coefficient of two nodes across all three types of features is greater than this threshold, they are classified into the same primary cluster. During the clustering process, the optimal clustering level is determined based on the hierarchical structure of the dendrogram, and each set of highly relevant nodes is used as a local feature subset to construct a preliminary connection graph of the multi-scale feature topology network.

[0037] Each local feature subset represents a set of spatially adjacent nodes that exhibit a coordinated changing trend in the three categories of thermal, mechanical, and fluid features, providing a basis for subsequent identification of potential icing aggregation trends.

[0038] After the multi-scale feature topology network is constructed, each local feature subset is mapped to the spatial coordinate system of the blade. Taking the root of the blade as the origin, two-dimensional planar coordinates are established along the chord and radial directions respectively, and time series are superimposed to form a three-dimensional space-time distribution map.

[0039] To identify potential freezing point coupling clustering regions, a density peak clustering algorithm is introduced. This algorithm identifies nodes that simultaneously possess high local point density and a large minimum distance as cluster centers, achieving spatial clustering identification without relying on a preset number of clusters.

[0040] Density peak clustering includes the following specific steps: Calculate the density value for each node, where density is defined as the number of neighboring nodes within a given distance threshold (set to 0.2 meters); For each node, calculate its minimum distance to all nodes with higher density to form a density-distance graph; Nodes that are high in both density and distance dimensions are selected as cluster centers; The remaining nodes are assigned to the cluster of the nearest higher-density node, thus completing the clustering region division.

[0041] Ultimately, all clustering results are mapped back into the blade surface spatial model, and high-density areas are identified as potential freezing point coupling clusters, which are used to guide subsequent risk warning and response strategies.

[0042] By using hot spot outliers in potential ice point coupling aggregation regions as high-risk observation nodes, the energy change entropy of the nodes in different time windows is calculated, and a dynamic prediction set of the probability of microscopic ice nucleus formation is output.

[0043] First, within each potential freezing point coupling cluster region, the surface temperature values ​​of all monitoring nodes in that region are statistically analyzed at the same time frame, and the mean temperature and standard deviation of the temperature for that region are calculated. The temperature value of a single monitoring node is compared with the mean temperature of the region. When the temperature of a node deviates from the mean temperature of the region by more than twice the standard deviation, the thermal anomaly spot corresponding to that node is marked as a thermal anomaly point.

[0044] The aforementioned hotspot anomalies were defined as high-risk observation nodes, meaning that these nodes exhibit significant thermal radiation anomalies that differ from the surrounding environment, posing a potential risk of forming a localized icing initial core. The spatial coordinates and temporal markers of all high-risk observation nodes were recorded for subsequent time-series analysis.

[0045] For each high-risk observation node, the temperature change curve corresponding to that node in the infrared thermal imaging matrix is ​​extracted within a continuous time window. The time window length is set to 5 seconds, and the time windows slide in 1-second increments to ensure the continuity of the temporal features.

[0046] Based on the correlation between infrared radiation and temperature, the node temperature value is converted into an equivalent thermal radiation energy value, and the rate of change of the node's thermal radiation energy within adjacent time windows is calculated. The rate of change is calculated as follows: subtract the thermal radiation energy value of the previous time window from the current time window value, take the absolute value, and then divide it by the thermal radiation energy value of the previous time window. This rate of change is used to characterize the intensity of fluctuations in the node's thermal state over short timescales.

[0047] After obtaining the thermal radiation energy change rate sequence of high-risk observation nodes, a statistical distribution analysis is performed on the sequence within a fixed time window. Specifically, the energy change rate sequence is divided into several intervals, and the probability value of the energy change rate occurring in each interval is calculated.

[0048] Based on the above probability distribution, the energy change entropy value is calculated using the definition of Shannon entropy, which involves multiplying each probability value by its logarithm, summing the results, and taking the negative value. This energy change entropy is used to quantify the uncertainty of the change in nodal thermal radiation energy. The higher the entropy value, the more unstable the thermal state of the node, and the stronger the randomness of the micro-icing behavior.

[0049] The energy change entropy value corresponding to each high-risk observation node is used as an input feature and fed into the support vector regression model to output the prediction result of the probability of micro ice nucleus generation at that node in the future time period.

[0050] The support vector regression model uses a radial basis function as its kernel function. During model training, historically labeled icing occurrence samples are used as training data. The optimal model parameters are determined by minimizing a trade-off function between prediction error and model complexity. The model output is a continuous numerical value ranging from 0 to 1, representing the probability of a corresponding node forming a micro-ice nucleus within the prediction time window.

[0051] The prediction results of all high-risk observation nodes are arranged in chronological order to form a dynamic prediction set of the probability of micro-ice nucleus formation, providing a basis for subsequent analysis of the evolution of latent ice conditions and active response control.

[0052] A multidimensional recursive analysis is performed on the dynamic predicted value set to construct a hidden ice condition propagation path map. Based on the synchronization rate of thermal disturbance and stress feedback at the nodes in the propagation path, an icing warning curve for potential hidden ice areas is generated.

[0053] First, the dynamic predicted value set of the probability of microscopic ice nucleus formation for each high-risk observation node is subjected to time-series smoothing. A time-series weighted moving average algorithm is used for noise suppression. The algorithm sets the sliding window length to 5 time steps, and assigns higher weights to the weight factors according to the most recent time point. The specific weight distribution is 0.4, 0.3, 0.15, 0.1, and 0.05.

[0054] The smoothed sequence is used to eliminate the interference of abrupt changes in short periods. Then, its first-order difference result is extracted as a feature vector of change trend, which represents the direction and rate of change of the predicted value over time, and provides dynamic feature input for subsequent path construction.

[0055] Based on the spatial coordinates and trend feature vectors of all high-risk observation nodes, a propagation impact matrix for icing probability is constructed. Each element of this matrix represents the propagation probability between two nodes. Specifically, the Euclidean spatial distance between nodes is calculated first, followed by the cosine similarity of the corresponding trend feature vectors. These two factors are combined to form the propagation impact weight. The closer the spatial distance and the more similar the trends, the higher the propagation weight.

[0056] Based on this, a path recursion algorithm based on depth-first search is adopted, with the minimum propagation weight threshold set to 0.75. Starting from any node with a sudden increase in ice core probability, continuous high-weight connection paths are recursively searched to form a set of implicit ice condition propagation paths.

[0057] For each propagation path, the sequences of thermal anomaly spot change rate and stress waveform amplitude change rate for each node in the path within the same time period are extracted. The synchronization rate of these two types of features in the time dimension is calculated using the Pearson correlation coefficient, serving as a quantitative indicator of icing coordination intensity.

[0058] The Pearson synchronization rate is calculated as follows: The two feature sequences are centered separately, multiplied by a dot, and the result is then divided by the product of their respective standard deviations. The resulting value falls within the range of... The value is between 1 and 2, with a larger value indicating stronger synchronization.

[0059] If the synchronization rate of consecutive nodes within a certain path segment is higher than a set threshold (set to 0.85), then the path segment is marked as a high-risk area for collaboration.

[0060] All high-risk collaborative path segments are merged in chronological order to generate an icing warning curve for potential hidden ice areas. The warning curve is plotted with time on the horizontal axis and the superposition value of ice nucleus formation probability or collaborative intensity on the vertical axis, reflecting the spatial and temporal propagation trend of hidden ice conditions.

[0061] Based on the icing warning curve, a real-time icing condition classification index is output, dividing the leaf condition into four levels: normal, warning, visible icing, and hidden icing.

[0062] First, time series analysis was performed on the icing warning curve corresponding to each high-risk transmission path to extract the following four types of characteristic parameters: Maximum value: refers to the peak value of the icing warning curve within the set analysis time period, indicating the highest intensity of icing risk in a certain area of ​​the blade; Average growth rate: The average of the first difference of the curve over a continuous time step reflects the risk growth trend; Duration: Calculates the duration of the icing warning curve above the risk initiation threshold (set to 0.4), in seconds; Fluctuation amplitude: defined as the difference between the maximum and minimum values ​​of the curve within the analysis period, used to characterize the dynamic instability of ice condition risk.

[0063] The above parameters serve as the basic inputs for determining ice condition status.

[0064] Based on the extracted feature parameters, a rule-driven ice condition risk assessment model is constructed. This model employs a hierarchical threshold logic for judgment, is independent of training samples, and possesses real-time performance and interpretability. The model sets the following state division rules: If the maximum value of the icing warning curve is greater than 0.7 and the duration exceeds 15 seconds, it is judged as a latent icing state, indicating that there is a continuous and highly coordinated icing evolution trend inside the blade, even though no obvious ice layer appears on the surface. If the maximum value is greater than 0.6 and the fluctuation range exceeds 0.3, it is judged as an explicit icing state, indicating that the ice nucleation behavior has formed a local physical characterization on the surface, which poses a structural safety risk. If the maximum value is in the range of 0.4 to 0.6 and the average growth rate is consistently positive, it is judged as a warning state, indicating that the initial trend of ice conditions has been formed but has not yet reached a dangerous level; If all parameters are below the above-mentioned threshold, it is judged to be in a normal state, and the blade surface and structure do not show signs of icing.

[0065] Based on the results output by the ice condition risk assessment model, the operating status of the blades in each monitoring cycle is divided into four levels in real time: "normal", "early warning", "visible icing" or "latent icing".

[0066] The state classification results are stored in association with the blade spatial coordinates and continuously updated with timestamps for use by the control logic. Simultaneously, the classification results are pushed to the upper-level monitoring interface to alert maintenance personnel, trigger heating control commands, or implement shutdown protection strategies.

[0067] When a latent icing level is present, a heating control command is triggered to locally increase the temperature of the heating element of the blade corresponding to the icing area.

[0068] First, based on the continuous high-risk path segments identified as having latent icing levels from the generated icing warning curves, the corresponding node spatial coordinate information is extracted. These coordinates are based on a three-dimensional spatial reference frame of the blade structure, using standardized coordinate values ​​in the chord, radial, and thickness directions to ensure positioning accuracy within 0.1 meters.

[0069] The coordinates are input into the blade spatial mapping function to locate the corresponding high-risk observation node group. Then, a pre-configured blade zone heating element control instruction table is invoked, matching the heating unit number containing the corresponding area based on the node's spatial location. Each heating unit covers a specific blade area, with its number corresponding one-to-one with its spatial coordinates, supporting precise local thermal control.

[0070] For the matched heating unit, a local heating control command is sent, which includes two control parameters: the target heating temperature and the heating duration.

[0071] Target temperature rise: Set at 8 degrees Celsius above the average temperature of the current high-risk observation node over the last 5 seconds. This temperature rise is a safe value that, after experimental verification, can effectively eliminate the early ice core state.

[0072] Heating duration: The default is no less than 30 seconds, ensuring that the local material temperature exceeds the ice layer's stability point, causing potential icing structures to become unstable and desorbed.

[0073] The heating control command is applied to the selected heating unit through current driving, activating the flexible thermistor material or carbon fiber heating element in the area to achieve rapid heating.

[0074] After the temperature control is activated, the dynamic changes in the density of thermal anomalies in the target area are continuously monitored using an infrared thermal imaging matrix on the blade surface. This density is defined as the number of anomalies per unit area whose temperature is more than three standard deviations above the average value.

[0075] The safety threshold is set at no more than 5 thermal anomaly spots per square meter. If the monitoring value is lower than this threshold for 5 consecutive seconds, it is determined that the local ice nucleus has been effectively removed.

[0076] When the above conditions are met, a termination heating command is immediately sent, the power supply of the corresponding heating unit is turned off, and the temperature change data, heating time, hot spot density response, etc. during this heating response process are recorded in the blade ice condition control database as the basic data source for subsequent model dynamic calibration and strategy optimization.

[0077] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method of monitoring icing on a wind turbine blade, characterized by: The method comprises the following steps: acquiring multi-source weak signal distribution data on the surface of the blade during operation of the wind turbine, including an infrared thermal imaging matrix, blade stress wave signals, and local air disturbance velocity field data; based on the multi-source weak signal distribution data, constructing a multi-dimensional preliminary ice condition feature tensor containing thermal anomaly spot density distribution, stress wave amplitude change rate, and disturbance frequency anomaly index; based on the node correlation degree of the multi-dimensional preliminary ice condition feature tensor, establishing a blade surface multi-scale feature topology network, and identifying a potential ice point coupling aggregation area; taking the thermal spot outlier points in the potential ice point coupling aggregation area as high-risk observation nodes, calculating the energy change entropy of the nodes in different time windows, and outputting a dynamic prediction value set of the possibility of micro-ice nucleus generation; performing multi-dimensional recursive analysis on the dynamic prediction value set to construct an implicit ice condition propagation path map, and generating an icing warning curve for the potential implicit ice area based on the synchronization rate of thermal disturbance and stress feedback in the propagation path; according to the icing warning curve, outputting real-time ice condition grading indexes, and dividing the blade state into four levels: normal, warning, explicit icing, and implicit icing; when there is an implicit icing level state, triggering a temperature rise control command to locally control the heating elements of the blade corresponding to the icing area.

2. A method of ice detection on a wind turbine blade according to claim 1, characterised in that: Wherein, constructing a multi-dimensional preliminary ice condition feature tensor comprises: performing spatial gray scale gradient enhancement and regional connectivity analysis on the infrared thermal imaging matrix to extract thermal anomaly spots on the surface of the blade and calculate the unit area distribution density thereof as the thermal anomaly spot density distribution; using a sliding window Fourier transform algorithm to perform time-frequency feature extraction on the stress wave signals to obtain the stress wave amplitude change rate in different time periods; performing spatial vector field stability analysis on the local air disturbance velocity field data to calculate the root mean square deviation of the velocity change between adjacent time frames as the disturbance frequency anomaly index; aligning the thermal anomaly spot density distribution, stress wave amplitude change rate, and disturbance frequency anomaly index in time sequence based on a unified blade spatial coordinate mapping rule, and fusing to construct a multi-dimensional preliminary ice condition feature tensor.

3. A method of ice detection on a wind turbine blade according to claim 1, characterised in that: The method for establishing a blade surface multi-scale feature topology network and identifying a potential ice point coupling aggregation area comprises: calculating the feature correlation degree of the thermal anomaly spot density, stress wave amplitude change rate, and disturbance frequency anomaly index between any two nodes in the multi-dimensional preliminary ice condition feature tensor to generate a three-dimensional feature correlation matrix between the nodes; performing hierarchical clustering analysis on the three-dimensional feature correlation matrix to divide high-correlation-degree nodes into multiple local subsets, and constructing a blade surface multi-scale feature topology network; based on the distribution density of the blade surface multi-scale feature topology network in the spatial coordinate system, identifying a potential ice point coupling aggregation area using a density peak clustering algorithm.

4. A method of ice detection on wind turbine blades as claimed in claim 1, wherein: Wherein, based on the energy change entropy of the calculation nodes in different time windows comprises: marking points with a temperature value deviating from the mean value of the regional temperature by more than two standard deviations as thermal spot outlier points in the potential ice point coupling aggregation area, and defining them as high-risk observation nodes; For each high-risk observation node, extract the temperature change curve corresponding to the node in the infrared thermal imaging matrix within a continuous time window, and calculate the thermal radiation energy change rate of the node in each time window; Based on the thermal radiation energy change rate, calculate the energy change entropy of each node in a fixed time window; Input the energy change entropy of each node into the support vector regression model, and output the dynamic prediction value set of the micro-ice nucleus generation possibility.

5. A method of ice detection on a wind turbine blade according to claim 4, characterised in that: Based on the synchronization rate of thermal disturbance and stress feedback of nodes in the propagation path, generate the icing warning curve of the potential hidden ice area, including: For the dynamic prediction value set of the micro-ice nucleus generation possibility of each high-risk observation node, use the time series weighted moving average algorithm to smooth process, and extract its change trend feature vector; Based on the spatial distance between nodes and the prediction value change similarity, construct the icing probability propagation influence matrix, and identify the possible hidden ice propagation path; In the identified path, extract the thermal anomaly spot change rate and stress waveform amplitude change rate sequence corresponding to the path nodes, and calculate the Pearson synchronization rate between them as the icing coordination strength; According to the continuous path segment with icing coordination strength exceeding the set threshold, generate the icing warning curve of the potential hidden ice area, and dynamically label and output its risk level.

6. A method of ice detection on a wind turbine blade according to claim 5, characterised in that: According to the icing warning curve, output the real-time ice condition grading index, including: Extract the feature parameters of the icing warning curve, including the maximum value, average growth rate, duration and fluctuation amplitude; Based on the extracted feature parameters, construct an ice condition risk discrimination model and set the grading criteria, where a maximum value greater than 0.7 and a duration greater than 15 seconds are determined as hidden icing, and a maximum value greater than 0.6 and a fluctuation amplitude greater than 0.3 are determined as explicit icing; When the maximum value is between 0.4 and 0.6 and the average growth rate is continuously positive, it is determined to be in a warning state; if all feature parameters are below the set threshold, it is determined to be in a normal state; According to the determination result, the current state of the blade is real-time divided into four levels: normal, warning, explicit icing or hidden icing.

7. A method of ice detection on wind turbine blades as claimed in claim 1, wherein: Local temperature control of the heating elements of the blade corresponding to the icing area, including: According to the position coordinates of the icing warning curve under the hidden icing level state, locate the corresponding high-risk observation node and map it to the blade spatial coordinate system; Call the partition heating element control instruction table embedded in the blade structure, match the heating unit number in the located coordinate range; Send a local temperature control command to the matched heating unit, set the target temperature to the current node average temperature plus 8 degrees Celsius, and the temperature rise duration is not less than 30 seconds; During the temperature rise process, continuously monitor the change of the thermal anomaly spot density in the region, if it is lower than the preset safety threshold, then remove the temperature rise control and restore normal operation.

Images