A wind turbine blade crack monitoring method based on resistance feature learning

Abstract

The application discloses a kind of wind power blade crack monitoring methods based on resistance characteristic learning, comprising the following steps: constructing conductive coating network, resistance time series data of each line section is collected and preprocessed;Resistance feature set is output based on the construction of multi-class feature of conductive coating network and is labeled;Improved graph wavelet network is trained with labeled resistance feature set, and crack monitoring model is output by optimizing parameters;Resistance time series is collected in real time and inputs crack monitoring model inference, generates crack diagnosis information;Crack location estimation result is mapped to the physical coordinate system of wind power blade, and crack positioning information is generated;Risk comparison processing is carried out based on crack diagnosis information and crack positioning information, and crack warning information is generated and output together with crack positioning information.The application realizes real-time intelligent monitoring of wind power blade crack by using improved graph wavelet network, with the advantages of accurate positioning, timely warning and high operation reliability.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, and in particular to a method for monitoring cracks in wind turbine blades based on resistance feature learning. Background Technology

[0002] Current wind turbine blade structures commonly employ manual inspection, ultrasonic testing, or optical image detection to identify cracks. However, manual methods are limited by environmental factors, cost, and timeliness, making continuous monitoring difficult. While ultrasonic or optical methods can achieve a degree of automation, they are constrained by the large blade size, complex surface, varying lighting conditions, and installation requirements, making it difficult to maintain stable performance under high-speed rotation and long-term operation. Furthermore, traditional strain gauge-based monitoring methods are complex to deploy, have weak damage resistance, cannot cover a large area of ​​the blade surface, and most methods rely on only a single point or a small number of sensors, making it difficult to reflect the crack propagation trend and spatial distribution characteristics.

[0003] Existing structural health monitoring methods based on resistance changes mostly rely on single-path measurements or simple statistical features for crack detection. They cannot obtain complete spatial resistance distribution information through multi-path conductive networks, nor can they effectively model the complex relationship between resistance and temporal and spatial coupling. Furthermore, existing deep learning models generally lack the ability to model features specific to the topology of conductive coating networks, making it difficult to achieve high-precision prediction of crack states and locations across multiple scales and in a spatiotemporally integrated manner.

[0004] Therefore, how to provide a method for monitoring cracks in wind turbine blades based on resistance feature learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for monitoring cracks in wind turbine blades based on resistance feature learning. This invention utilizes an improved graph wavelet network to achieve real-time intelligent monitoring of cracks in wind turbine blades, which has the advantages of accurate positioning, timely early warning, and high operational reliability.

[0006] A method for monitoring cracks in wind turbine blades based on resistance feature learning according to an embodiment of the present invention includes the following steps: A conductive coating network was constructed on the surface of the wind turbine blade, and resistance time series data of each line segment were collected, preprocessed, and a standardized resistance time series was generated. Based on the topology of the conductive coating network, the standardized resistance time series is constructed into multiple types of features and output as a labeled resistance feature set corresponding to the crack state label and crack location label; The labeled resistance feature set is input into the improved graph wavelet network for training. The parameters of the improved graph wavelet network are optimized through supervised learning, and the crack monitoring model is output. During the online operation of the wind turbine, the resistance time series is collected in real time, and forward reasoning is performed through the crack monitoring model to generate crack diagnosis information corresponding to each monitored line segment. Based on crack diagnosis information and the topological coordinates of the conductive coating network, the crack location estimation results are mapped to the physical coordinate system of the wind turbine blade to generate crack location information. Risk comparison processing is performed based on crack diagnosis information and crack location information to generate crack early warning information, which is then output together with crack location information through the communication interface of the wind turbine SCADA system.

[0007] Optionally, the resistance time series data includes the initial resistance of each line segment, multi-channel resistance time series during operation, correlation data of resistance changes with temperature, and multi-path resistance differential data. The preprocessing includes outlier removal, missing point interpolation, temperature compensation, and sequence segmentation processing based on the wind turbine operating conditions.

[0008] Optionally, the output of the labeled resistor feature set specifically includes: Based on the topology of the conductive coating network, the standardized resistance time series is arranged in the order of line segment numbers to generate a time change sequence composed of each line segment at continuous sampling time and form a time feature matrix. Based on the spatial adjacency relationship of the conductive coating network, the difference between the standardized resistance value of each line segment and the standardized resistance value of the adjacent line segments is calculated at each sampling time to generate an adjacency difference feature matrix. Based on the multi-path structure of conductive coating network, the standardized resistance values ​​of the line segments in each path are connected in series according to the node order of radial path, circumferential path and corresponding combined path to generate multi-path joint feature matrix. The time feature matrix, the adjacency difference feature matrix, and the multi-path joint feature matrix are concatenated to obtain resistance feature samples. The resistance feature samples are then mapped to crack state labels and crack location labels to output a labeled resistance feature set.

[0009] Optionally, the output of the crack monitoring model specifically includes: Receive the labeled resistor feature set, divide it into batches according to the sample number, and construct a resistor feature mapping relationship for the resistor feature samples in each training batch; Based on the resistance feature mapping relationship, the spatiotemporal operator structure of the improved graph wavelet network is used to process the resistance feature vectors of each time sampling point in time index order and generate a spatiotemporal operator hidden state sequence. The improvements of the improved graph wavelet network compared to the original graph wavelet network include: replacing the dilated causal convolution structure in the original graph wavelet network with a spatiotemporal operator structure, replacing the adaptive graph convolution structure with a graph neural operator structure based on learnable operator kernels, and adjusting the graph wavelet multi-scale feature extraction structure to a multi-scale projection structure that matches the graph neural operator structure. The hidden state sequence of the spatiotemporal operator is input into the graph neural operator structure. The resistance hidden states of each line segment are weighted and aggregated according to the relationship between the line segments and generate the hidden state sequence of the graph neural operator. The hidden state sequence of graph neural operators is input into the multi-scale projection structure of the improved graph wavelet network to perform projection processing on the resistive hidden states at each scale and generate multi-scale fusion features. The multi-scale fusion features are input into the crack monitoring output layer to perform classification and localization operations, generating crack state discrimination results and crack location estimation results. The trainable parameters of the improved graph wavelet network are iteratively updated based on the supervised learning loss function. After multiple rounds of parameter updates and reaching the loss convergence condition, the improved graph wavelet network in the converged parameter state is output as the crack detection model. The supervised learning loss function is a single loss scalar obtained by weighting and summing the difference between the crack state discrimination result and the crack state label and the difference between the crack location estimation result and the crack location label according to a preset weight coefficient. The crack state difference measure is the difference value calculated for each category of the probability distribution of crack categories in each line segment and the corresponding crack state label, and then summed over all categories. The crack location difference measure is the difference value calculated for each element of the numerical difference between the crack location area, the position coordinate along the blade span, the position coordinate along the blade chord, and the crack length estimation value and the corresponding crack location label, and then summed over all position elements. Finally, the crack state difference measure and the crack location difference measure are obtained by weighting and combining them according to a preset weight. The loss convergence condition is that the amount of loss change in the supervised learning loss function in a series of consecutive iterations is less than a preset loss change threshold. The amount of loss change is the difference between a single loss scalar obtained by weighting the crack state difference measure and the crack position difference measure according to preset weights between adjacent iteration steps.

[0010] Optionally, the generation of the crack diagnostic information specifically includes: During the online operation of the wind turbine, the resistance time series data of each monitoring line section are collected in real time and preprocessed to generate a real-time standardized resistance time series. Based on the topology of the conductive coating network, the real-time standardized resistance time series is organized according to the time sampling point index and the line segment index. For each time sampling point, the standardized resistance value of each line segment, the resistance difference value of adjacent line segments, and the multi-path joint resistance change are combined to construct the real-time resistance feature mapping relationship. According to the monitoring line segment number, the real-time resistance feature mapping relationship is compressed into the corresponding real-time resistance feature vector, and the real-time resistance feature vector of each monitoring line segment is input into the crack monitoring model. Forward reasoning is performed on the real-time resistance feature vector to generate crack diagnosis information corresponding to each monitored line segment. The crack diagnosis information includes crack occurrence probability, crack type discrimination result, and crack location estimation result.

[0011] Optionally, the generation of the crack location information specifically includes: Record the spanwise and chordwise coordinates of the intersection nodes of each radial trunk line and ring line in the conductive coating network in the physical coordinate system of the wind turbine blade, and generate a topological coordinate index table including the physical coordinates of the starting and ending nodes of each monitoring line segment based on the node coordinates. Extract the crack location estimation results of each monitoring line segment from the crack diagnosis information, obtain the physical coordinates of the starting node and ending node of the corresponding monitoring line segment according to the topological coordinate index table, and perform linear interpolation processing on the physical coordinates of the starting node and ending node according to the relative position ratio in the crack location estimation results to generate the spanwise coordinates and chordwise coordinates of the crack center point along the blade. Based on the actual length of the monitored line segment in the physical coordinate system of the wind turbine blade and the crack length ratio in the crack location estimation results, a length conversion process is performed to generate a crack length estimate. Crack location information is generated by combining the position coordinates of the crack center point along the blade spanwise, the position coordinates along the blade chordwise, the estimated crack length, and the blade region to which the monitored line segment belongs.

[0012] Optionally, the generation of the crack early warning information specifically includes: Rules for setting crack risk classification thresholds based on the operational safety requirements of wind turbine blades; The crack occurrence probability and crack severity level are extracted from the crack diagnosis information, and the crack length estimate is extracted from the crack location information. The above three types of data are combined and processed to generate a crack risk input vector. The crack risk input vector is compared with the upper and lower limits of each risk score interval in the crack risk classification threshold rule to determine the risk score interval where the crack risk score value is located, and obtain the corresponding crack risk level identifier. The crack risk level identifier is combined with the crack location blade area, the position coordinate along the blade span, the position coordinate along the blade chord and the crack length estimate to generate crack early warning information for the corresponding monitoring line segment. After encapsulating the crack location information and crack early warning information according to the preset message format, they are output through the communication interface with the wind turbine SCADA system.

[0013] The beneficial effects of this invention are: This invention achieves continuous and distributed sensing of the blade surface structure by constructing a conductive coating network combining radial main lines and ring lines on the blade surface and collecting multi-channel resistance time series data. Compared with traditional technologies that rely on manual inspection, local sensors, or optical methods, this invention significantly improves the monitoring range and continuity. The conductive path covers most of the critical stress areas of the blade, enabling the reflection of crack occurrence and propagation trends through changes in the resistance field before cracks extend to the visible scale. This overcomes the shortcomings of existing technologies, such as large monitoring blind spots and difficulty in achieving long-term real-time monitoring. Through preprocessing of the resistance data, including anomaly removal, temperature compensation, and multi-path differential analysis, this invention obtains a stable and reliable standardized resistance time series, providing high-quality input for subsequent intelligent crack identification.

[0014] This invention further comprehensively expresses the spatial topological relationship and spatiotemporal variation law of the conductive coating network by constructing resistance time features, adjacency difference features, and multi-path joint features, and proposes an improved graph wavelet network as a crack monitoring model. Compared with existing methods that rely on dilated convolution or ordinary graph convolution, this invention introduces a spatiotemporal operator structure, a graph neural operator structure, and a multi-scale projection structure, enabling the model to perform high-precision modeling of the temporal evolution, spatial dependence, and multi-scale features in the resistance data. This structure improves the joint identification ability of crack occurrence, development stage, crack type, and crack location, significantly enhancing the accuracy and robustness of crack diagnosis. Under the supervised learning framework, by jointly optimizing the crack state label and crack location label, this invention obtains a crack monitoring model with high generalization ability, enabling the model to maintain stable identification performance under complex operating conditions and diverse crack patterns.

[0015] This invention also proposes a crack location method based on the physical coordinate system of wind turbine blades. By mapping the crack location estimation results to the spanwise and chordwise physical coordinates of the blade, it achieves accurate reconstruction of the crack's spatial location and geometric dimensions, providing directly usable engineering location information for subsequent crack assessment and maintenance. Based on this, the invention constructs crack risk classification threshold rules, combining crack occurrence probability, severity level, and crack length estimation to generate a crack risk score and output crack early warning information. This can trigger warnings at the early crack initiation stage, assisting in operation and maintenance decisions. Overall, this invention achieves full automation, real-time operation, and intelligence in the entire process of wind turbine blade crack detection, location, and risk warning, significantly improving the operational safety of wind turbine blades and reducing maintenance costs, demonstrating significant engineering application value and promising prospects for widespread adoption. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a wind turbine blade crack monitoring method based on resistance feature learning proposed in this invention; Figure 2 This is a schematic diagram of an improved wavelet network structure for a wind turbine blade crack monitoring method based on resistance feature learning proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1-2 A method for monitoring cracks in wind turbine blades based on resistance feature learning includes the following steps: A conductive coating network was constructed on the surface of the wind turbine blade, and resistance time series data of each line segment were collected, preprocessed, and a standardized resistance time series was generated. Based on the topology of the conductive coating network, the standardized resistance time series is constructed into multiple types of features and output as a labeled resistance feature set corresponding to the crack state label and crack location label; The labeled resistance feature set is input into the improved graph wavelet network for training. The parameters of the improved graph wavelet network are optimized through supervised learning, and the crack monitoring model is output. During the online operation of the wind turbine, the resistance time series is collected in real time, and forward reasoning is performed through the crack monitoring model to generate crack diagnosis information corresponding to each monitored line segment. Based on crack diagnosis information and the topological coordinates of the conductive coating network, the crack location estimation results are mapped to the physical coordinate system of the wind turbine blade to generate crack location information. Risk comparison processing is performed based on crack diagnosis information and crack location information to generate crack early warning information, which is then output together with crack location information through the communication interface of the wind turbine SCADA system.

[0019] In this embodiment, the construction of the conductive coating network includes coating the wind turbine blade surface with a conductive coating according to a preset topology combining radial trunk lines and ring lines. The trunk lines, ring lines, and intersection nodes are numbered to form measurable conductive paths. The data acquisition includes using a multi-channel resistance measurement module to synchronously acquire the resistance value and initial resistance reference value of each line segment at a preset sampling period, and combining it with a temperature sensor to acquire blade surface temperature data. The resistance time series data includes the initial resistance of each line segment, the multi-channel resistance time series during operation, the correlation data of resistance with temperature change, and the multi-path resistance differential data. The preprocessing includes outlier removal, missing point interpolation, temperature compensation, and sequence segmentation processing based on wind turbine operating conditions.

[0020] In this embodiment, the output of the labeled resistor feature set specifically includes: Based on the topology of the conductive coating network, the standardized resistance time series is arranged in the order of line segment numbers to generate a time change sequence composed of each line segment at continuous sampling time and form a time feature matrix. The time feature matrix is ​​a matrix with the time sampling point as the first dimension and the line segment number as the second dimension, and is formed by arranging the standardized resistance values ​​of each line segment in time order. Based on the spatial adjacency relationship of the conductive coating network, the difference between the standardized resistance value of each line segment and the standardized resistance value of the adjacent line segment is calculated at each sampling time, and an adjacency difference feature matrix is ​​generated. The adjacency difference feature matrix is ​​composed of the standardized resistance difference of the adjacent line segments at the same sampling point, with time index and line segment pair index as the two dimensions of the matrix. Based on the multi-path structure of conductive coating network, the standardized resistance values ​​of the line segments in each path are connected in series according to the node order of radial path, circumferential path and corresponding combined path to generate multi-path joint feature matrix. The time feature matrix, adjacency difference feature matrix, and multi-path joint feature matrix are concatenated to obtain resistance feature samples. These resistance feature samples are then mapped to crack status labels and crack location labels to output a labeled resistance feature set. The crack status labels and crack location labels are added based on crack status, crack location, and crack length information recorded in historical maintenance records, as well as crack status and crack location information generated by the crack simulation system. The crack status includes crack existence status, crack development stage status, crack type status, and crack severity level status. The crack type status includes spanwise cracks, chordal cracks, and composite cracks. The crack location information corresponding to the crack location label includes the blade region where the crack is located, the position coordinates along the blade spanwise direction, the position coordinates along the blade chordal direction, and the estimated crack length.

[0021] In this embodiment, the output of the crack monitoring model specifically includes: The system receives a set of labeled resistance features, divides them into batches according to the sample number, and constructs a resistance feature mapping relationship for the resistance feature samples in each training batch. The resistance feature mapping relationship uses the time sampling point index and the line segment index as independent variables, the resistance feature vector of the corresponding line segment as the function value, and is associated with the crack status label and the crack location label. Based on the resistance feature mapping relationship, the spatiotemporal operator structure of the improved graph wavelet network is used to process the resistance feature vectors of each time sampling point in time index order and generate a spatiotemporal operator hidden state sequence. The improvements of the improved graph wavelet network compared to the original graph wavelet network include: replacing the dilated causal convolution structure in the original graph wavelet network with a spatiotemporal operator structure, replacing the adaptive graph convolution structure with a graph neural operator structure based on learnable operator kernels, and adjusting the graph wavelet multi-scale feature extraction structure to a multi-scale projection structure that matches the graph neural operator structure. The generation of the spatiotemporal operator hidden state sequence includes: in the spatiotemporal operator structure, receiving resistance feature vectors composed of standardized resistance values ​​of each line segment in time index order, inputting the resistance feature vectors into the time domain operator unit, performing linear transformation and time mapping processing based on a learnable time kernel between adjacent time sampling points, and outputting the time mapping result; inputting the time mapping result into the spatial operator unit, performing spatial propagation of the resistance features of each line segment based on the node connection relationship of the conductive coating network and performing spatial mapping processing based on a learnable adjacency operator kernel, and outputting the spatial mapping result; through the spatiotemporal combination unit, performing sequential combination and dimension alignment processing on the time mapping result and the spatial mapping result to generate the spatiotemporal operator hidden state vectors corresponding to the time sampling points, and arranging them in time index order to generate the spatiotemporal operator hidden state sequence; The hidden state sequence of the spatiotemporal operator is input into the graph neural operator structure. The resistance hidden states of each line segment are weighted and aggregated according to the relationship between the line segments and generate the hidden state sequence of the graph neural operator. The generation of the graph neural operator hidden state sequence includes: in the graph neural operator structure, receiving the spatiotemporal operator hidden state sequence of corresponding time sampling points according to the line segment number order; inputting the spatiotemporal operator hidden state sequence into the node feature projection unit; performing feature mapping processing on the hidden state of each line segment based on the learnable node projection kernel and outputting the node projection result; inputting the node projection result into the graph adjacency operator unit; performing weighted aggregation processing on the hidden state of each line segment according to the line segment pair relationship based on the node connection relationship of the conductive coating network and the learnable adjacency operator kernel, and outputting the aggregation result; after obtaining the aggregation result, inputting the aggregation result into the node update unit; performing state update processing on the hidden state of each line segment based on the learnable update kernel; generating the graph neural operator hidden state vector of the corresponding line segment; and arranging it according to the line segment number order to form the graph neural operator hidden state sequence. The hidden state sequence of graph neural operators is input into the multi-scale projection structure of the improved graph wavelet network to perform projection processing on the resistive hidden states at each scale and generate multi-scale fusion features. The generation of the multi-scale fusion features includes: in the multi-scale projection structure, receiving the graph neural operator hidden state sequence according to the line segment number order, inputting the graph neural operator hidden state sequence into the projection branches corresponding to different scales respectively, performing scale transformation processing on the graph neural operator hidden state based on the learnable scale projection kernel in each projection branch and outputting the scale projection result; inputting the projection results of each scale into the scale alignment unit, performing dimension alignment processing on the projection results of different scales according to the line segment number and time index and outputting the scale alignment result; inputting multiple scale alignment results into the feature fusion unit, performing sequence splicing or weighted combination processing on the features of different scales to generate multi-scale fusion features, and outputting the corresponding multi-scale fusion feature sequence according to the line segment number order; The multi-scale fusion features are input into the crack monitoring output layer to perform classification and localization operations, generating crack state discrimination results and crack location estimation results. The generation of the crack status discrimination result includes: in the crack monitoring output layer, receiving multi-scale fusion features in the order of line segment number, inputting the multi-scale fusion features into the crack status classification branch, performing feature projection processing on the multi-scale fusion features based on the learnable classification kernel and outputting the classification projection result; inputting the classification projection result into the classification mapping unit, performing category mapping processing on the classification projection result according to the category index of crack existence status, crack development stage status, crack type status and crack severity level status, and generating the crack status discrimination result for the corresponding line segment; The generation of the crack location estimation result includes: in the crack monitoring output layer, receiving multi-scale fusion features in the order of line segment number, inputting the multi-scale fusion features into the crack location regression branch, performing location projection processing on the multi-scale fusion features based on the learnable regression kernel and outputting the location projection result; inputting the location projection result into the location mapping unit, performing coordinate mapping processing on the location projection result according to the corresponding index of the crack location area on the blade, the location coordinates along the blade spanwise, the location coordinates along the blade chordwise, and the crack length estimation value, and generating the crack location estimation result for the corresponding line segment; The trainable parameters of the improved graph wavelet network are iteratively updated based on the supervised learning loss function. After multiple rounds of parameter updates and reaching the loss convergence condition, the improved graph wavelet network in the converged parameter state is output as the crack detection model. The supervised learning loss function is a single loss scalar obtained by weighting and summing the difference between the crack state discrimination result and the crack state label and the difference between the crack location estimation result and the crack location label according to a preset weight coefficient. The crack state difference measure is the difference value calculated for each category of the probability distribution of crack categories in each line segment and the corresponding crack state label, and then summed over all categories. The crack location difference measure is the difference value calculated for each element of the numerical difference between the crack location area, the position coordinate along the blade span, the position coordinate along the blade chord, and the crack length estimation value and the corresponding crack location label, and then summed over all position elements. Finally, the crack state difference measure and the crack location difference measure are obtained by weighting and combining them according to a preset weight. The loss convergence condition is that the amount of loss change in the supervised learning loss function in a series of consecutive iterations is less than a preset loss change threshold. The amount of loss change is the difference between a single loss scalar obtained by weighting the crack state difference measure and the crack position difference measure according to preset weights between adjacent iteration steps.

[0022] In this embodiment, the generation of crack diagnostic information specifically includes: During the online operation of the wind turbine, the resistance time series data of each monitoring line section are collected in real time and preprocessed to generate a real-time standardized resistance time series. Based on the topology of the conductive coating network, the real-time standardized resistance time series is organized according to the time sampling point index and the line segment index. For each time sampling point, the standardized resistance value of each line segment, the resistance difference value of adjacent line segments, and the multi-path joint resistance change are combined to construct the real-time resistance feature mapping relationship. According to the monitoring line segment number, the real-time resistance feature mapping relationship is compressed into the corresponding real-time resistance feature vector, and the real-time resistance feature vector of each monitoring line segment is input into the crack monitoring model. Forward reasoning is performed on the real-time resistance feature vector to generate crack diagnosis information corresponding to each monitored line segment. The crack diagnosis information includes crack occurrence probability, crack type discrimination result, and crack location estimation result.

[0023] In this embodiment, the generation of the crack location information specifically includes: In the physical coordinate system of the wind turbine blade, the spanwise and chordwise coordinates of the intersection nodes of each radial trunk line and ring line in the conductive coating network are recorded. Based on the node coordinates, a topological coordinate index table including the physical coordinates of the starting and ending nodes of each monitoring line segment is generated. The physical coordinate system of the wind turbine blade is a reference coordinate system preset in the design stage of the wind turbine blade with the spanwise and chordwise directions of the blade as coordinate axes. Extract the crack location estimation results of each monitoring line segment from the crack diagnosis information, obtain the physical coordinates of the starting node and ending node of the corresponding monitoring line segment according to the topological coordinate index table, and perform linear interpolation processing on the physical coordinates of the starting node and ending node according to the relative position ratio in the crack location estimation results to generate the spanwise coordinates and chordwise coordinates of the crack center point along the blade. Based on the actual length of the monitored line segment in the physical coordinate system of the wind turbine blade and the crack length ratio in the crack location estimation results, a length conversion process is performed to generate a crack length estimate. Crack location information is generated by combining the position coordinates of the crack center point along the blade spanwise, the position coordinates along the blade chordwise, the estimated crack length, and the blade region to which the monitored line segment belongs.

[0024] In this embodiment, the generation of the crack early warning information specifically includes: Based on the safety requirements for wind turbine blade operation, a crack risk classification threshold rule is set. The crack risk classification threshold rule uses the estimated crack length, the probability of crack occurrence, and the severity level of crack as independent variables to define several risk scoring intervals. Each risk scoring interval corresponds to a crack risk level identifier. The crack occurrence probability and crack severity level are extracted from the crack diagnosis information, and the crack length estimate is extracted from the crack location information. The above three types of data are combined and processed to generate a crack risk input vector. The crack risk input vector is compared with the upper and lower limits of each risk score interval in the crack risk classification threshold rule to determine the risk score interval where the crack risk score value is located, and obtain the corresponding crack risk level identifier. The crack risk level identifier is combined with the crack location blade area, the position coordinate along the blade span, the position coordinate along the blade chord and the crack length estimate to generate crack early warning information for the corresponding monitoring line segment. The generation of the crack risk score includes: in the crack risk scoring function, according to preset weight coefficients, the crack length estimate, crack occurrence probability, and crack severity level in the crack risk input vector are weighted and combined; a linear transformation corresponding to the length weight is performed on the crack length estimate; a linear transformation corresponding to the probability weight is performed on the crack occurrence probability; a linear transformation corresponding to the severity weight is performed on the crack severity level; and the results of the three linear transformations are summed to generate the crack risk score. After encapsulating the crack location information and crack early warning information according to the preset message format, they are output through the communication interface with the wind turbine SCADA system.

[0025] Example 1: To verify the feasibility of this invention in practice, it was applied to the monitoring of wind turbine blade cracks in a wind farm. In this wind farm, wind turbines are constantly exposed to the natural environment, and cracks or microcracks may appear on the blade surface. Traditional manual inspections and vibration monitoring methods are insufficient to detect early cracks in a timely manner, posing safety risks and high maintenance costs. This invention constructs a conductive coating network on the surface of the wind turbine blades, forming a measurable conductive path that combines radial main lines and ring lines, and collects the resistance time series data of each line segment. Through outlier removal, missing point interpolation, temperature compensation, and sequence segmentation, a standardized resistance time series is generated, providing an accurate data foundation for subsequent crack feature extraction and modeling.

[0026] On wind turbine blades, standardized resistance time series are processed through topology to construct resistance feature samples that include single-line time variation characteristics, differential characteristics between adjacent lines, and joint characteristics of multiple paths. These samples are then combined with historical maintenance records and crack simulation annotations to generate corresponding crack status and location labels, forming an annotated resistance feature set. This feature set is input into an improved graph wavelet network for training. Internally, the network uses spatiotemporal operators to process time and space dependent features, graph neural operators to perform weighted aggregation between lines, and a multi-scale projection structure to project and fuse features at each scale. The outputs crack status judgment results and crack location estimation results, thus forming a trained crack monitoring model. During the online operation of the wind turbine, resistance time series are collected in real time and input into the crack monitoring model for forward inference, generating crack diagnostic information corresponding to each monitored line segment. This allows for the determination of the crack's existence, development stage, type, and severity level, and provides accurate estimations of crack locations along the blade spanwise and chordwise directions, as well as crack length.

[0027] Furthermore, by utilizing the topological coordinates of the conductive coating network, the crack location estimation results are mapped to the physical coordinate system of the wind turbine blade, enabling visualization of crack location information. This information is then combined with crack diagnostic information for risk comparison. Based on crack length, crack occurrence probability, and severity level, a crack risk score is generated and compared with a preset risk score range to determine the crack risk level. Simultaneously, crack location information and crack early warning information are output through the communication interface of the wind turbine SCADA system, enabling remote monitoring and early warning of blade cracks. Throughout the implementation process, continuous multi-round parameter iterative optimization is employed, and a supervised learning loss function is used to weight the crack state and location errors, ensuring the convergence of the improved graph wavelet network and guaranteeing the reliability and stability of the crack monitoring model.

[0028] In the implementation of this invention in the wind farm, it can obtain information on the resistance changes of each line segment of the blade in a timely manner, accurately monitor the early formation, development process and specific location of cracks, significantly improve the operational safety of wind turbine blades, reduce the frequency of manual inspections, and provide accurate data references for subsequent maintenance, thus verifying the feasibility and effectiveness of this invention in practical applications.

[0029] Table 1. Performance Comparison of the Invention and Traditional Wind Turbine Blade Crack Monitoring Methods

[0030] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0031] The accuracy of crack detection has been improved from 65% for traditional manual inspection and 78% for vibration monitoring methods to 95%. This improvement stems from constructing a conductive coating network on the wind turbine blade surface, real-time acquisition of resistance time series for each line segment, and the use of an improved graph wavelet network for deep learning of multi-path, multi-scale resistance features, enabling the capture of minute cracks and their development and changes. In contrast, manual inspection relies on visual judgment and has a long inspection cycle, while vibration monitoring methods, although capable of detecting some anomalies, lack sufficient accuracy in location and type identification.

[0032] Regarding crack location error, traditional manual inspection has an average error of 15 cm, vibration monitoring methods have an error of 8 cm, while the method of this invention reduces it to 2 cm. The improvement lies in utilizing the topological information of the conductive coating network to map the crack location estimation results to the physical coordinate system of the wind turbine blade, enabling precise mapping of the crack location to the spanwise and chordwise coordinates of the blade, achieving high-precision spatial positioning. Simultaneously, the crack type discrimination accuracy has increased from 50%-68% to 93%, thanks to the deep fusion of temporal and spatial correlation features by the improved graph wavelet network's spatiotemporal operator structure and graph neural operator structure, effectively distinguishing spanwise cracks, chordwise cracks, and composite cracks.

[0033] This invention significantly optimizes the detection cycle and real-time monitoring capabilities. Traditional manual inspections take 30 days, and vibration monitoring takes 7 days; this invention shortens the cycle to 1 day. Early warning response time is reduced from 48 hours for manual inspections and 12 hours for vibration monitoring to 1 hour. This improvement primarily relies on real-time data acquisition, online inference, and a closed-loop monitoring mechanism, enabling crack diagnostic information and early warnings to be quickly generated and transmitted to the SCADA system, thus improving maintenance response efficiency.

[0034] Regarding the accuracy of maintenance intervention, traditional methods achieve 60%-72%, while this invention reaches 96%. The performance improvement is attributed to several factors: deep fusion of multi-path resistance characteristics and topological information provides rich spatiotemporal data support; improved graph wavelet networks and graph neural operator structures efficiently model and extract spatiotemporal dynamic features; and real-time data acquisition and online inference form a closed-loop system, enabling accurate, rapid, and continuous crack identification and early warning.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring cracks in wind turbine blades based on resistance feature learning, characterized in that, Includes the following steps: A conductive coating network was constructed on the surface of the wind turbine blade, and resistance time series data of each line segment were collected, preprocessed, and a standardized resistance time series was generated. Based on the topology of the conductive coating network, the standardized resistance time series is constructed into multiple types of features and output as a labeled resistance feature set corresponding to the crack state label and crack location label; The labeled resistance feature set is input into the improved graph wavelet network for training. The parameters of the improved graph wavelet network are optimized through supervised learning, and the crack monitoring model is output. During the online operation of the wind turbine, the resistance time series is collected in real time, and forward reasoning is performed through the crack monitoring model to generate crack diagnosis information corresponding to each monitored line segment. Based on crack diagnosis information and the topological coordinates of the conductive coating network, the crack location estimation results are mapped to the physical coordinate system of the wind turbine blade to generate crack location information. Risk comparison processing is performed based on crack diagnosis information and crack location information to generate crack early warning information, which is then output together with crack location information through the communication interface of the wind turbine SCADA system.

2. The wind turbine blade crack monitoring method based on resistance feature learning according to claim 1, characterized in that, The resistance time series data includes the initial resistance of each line segment, multi-channel resistance time series during operation, correlation data of resistance changes with temperature, and multi-path resistance differential data. The preprocessing includes outlier removal, missing point interpolation, temperature compensation, and sequence segmentation processing based on the wind turbine operating conditions.

3. The wind turbine blade crack monitoring method based on resistance feature learning according to claim 1, characterized in that, The output of the labeled resistor feature set specifically includes: Based on the topology of the conductive coating network, the standardized resistance time series is arranged in the order of line segment numbers to generate a time change sequence composed of each line segment at continuous sampling time and form a time feature matrix. Based on the spatial adjacency relationship of the conductive coating network, the difference between the standardized resistance value of each line segment and the standardized resistance value of the adjacent line segments is calculated at each sampling time to generate an adjacency difference feature matrix. Based on the multi-path structure of conductive coating network, the standardized resistance values ​​of the line segments in each path are connected in series according to the node order of radial path, circumferential path and corresponding combined path to generate multi-path joint feature matrix. The time feature matrix, the adjacency difference feature matrix, and the multi-path joint feature matrix are concatenated to obtain resistance feature samples. The resistance feature samples are then mapped to crack state labels and crack location labels to output a labeled resistance feature set.

4. The method for monitoring cracks in wind turbine blades based on resistance feature learning according to claim 1, characterized in that, The output of the crack monitoring model specifically includes: Receive the labeled resistor feature set, divide it into batches according to the sample number, and construct a resistor feature mapping relationship for the resistor feature samples in each training batch; Based on the resistance feature mapping relationship, the spatiotemporal operator structure of the improved graph wavelet network is used to process the resistance feature vectors of each time sampling point in time index order and generate a spatiotemporal operator hidden state sequence. The improvements of the improved graph wavelet network compared to the original graph wavelet network include: replacing the dilated causal convolution structure in the original graph wavelet network with a spatiotemporal operator structure, replacing the adaptive graph convolution structure with a graph neural operator structure based on learnable operator kernels, and adjusting the graph wavelet multi-scale feature extraction structure to a multi-scale projection structure that matches the graph neural operator structure. The hidden state sequence of the spatiotemporal operator is input into the graph neural operator structure. The resistance hidden states of each line segment are weighted and aggregated according to the relationship between the line segments and generate the hidden state sequence of the graph neural operator. The hidden state sequence of graph neural operators is input into the multi-scale projection structure of the improved graph wavelet network to perform projection processing on the resistive hidden states at each scale and generate multi-scale fusion features. The multi-scale fusion features are input into the crack monitoring output layer to perform classification and localization operations, generating crack state discrimination results and crack location estimation results. The trainable parameters of the improved graph wavelet network are iteratively updated based on the supervised learning loss function. After multiple rounds of parameter updates and reaching the loss convergence condition, the improved graph wavelet network in the converged parameter state is output as the crack detection model. The supervised learning loss function is a single loss scalar obtained by weighting and summing the difference between the crack state discrimination result and the crack state label and the difference between the crack location estimation result and the crack location label according to a preset weight coefficient. The crack state difference measure is the difference value calculated for each category of the probability distribution of crack categories in each line segment and the corresponding crack state label, and then summed over all categories. The crack location difference measure is the difference value calculated for each element of the numerical difference between the crack location area, the position coordinate along the blade span, the position coordinate along the blade chord, and the crack length estimation value and the corresponding crack location label, and then summed over all position elements. Finally, the crack state difference measure and the crack location difference measure are obtained by weighting and combining them according to a preset weight. The loss convergence condition is that the amount of loss change in the supervised learning loss function in a series of consecutive iterations is less than a preset loss change threshold. The amount of loss change is the difference between a single loss scalar obtained by weighting the crack state difference measure and the crack position difference measure according to preset weights between adjacent iteration steps.

5. The method for monitoring wind turbine blade cracks based on resistance feature learning according to claim 1, characterized in that, The generation of the crack diagnostic information specifically includes: During the online operation of the wind turbine, the resistance time series data of each monitoring line section are collected in real time and preprocessed to generate a real-time standardized resistance time series. Based on the topology of the conductive coating network, the real-time standardized resistance time series is organized according to the time sampling point index and the line segment index. For each time sampling point, the standardized resistance value of each line segment, the resistance difference value of adjacent line segments, and the multi-path joint resistance change are combined to construct the real-time resistance feature mapping relationship. According to the monitoring line segment number, the real-time resistance feature mapping relationship is compressed into the corresponding real-time resistance feature vector, and the real-time resistance feature vector of each monitoring line segment is input into the crack monitoring model. Forward reasoning is performed on the real-time resistance feature vector to generate crack diagnosis information corresponding to each monitored line segment. The crack diagnosis information includes crack occurrence probability, crack type identification result, and crack location estimation result.

6. The method for monitoring cracks in wind turbine blades based on resistance feature learning according to claim 1, characterized in that, The generation of the crack location information specifically includes: Record the spanwise and chordwise coordinates of the intersection nodes of each radial trunk line and ring line in the conductive coating network in the physical coordinate system of the wind turbine blade, and generate a topological coordinate index table including the physical coordinates of the starting and ending nodes of each monitoring line segment based on the node coordinates. Extract the crack location estimation results of each monitoring line segment from the crack diagnosis information, obtain the physical coordinates of the starting node and ending node of the corresponding monitoring line segment according to the topological coordinate index table, and perform linear interpolation processing on the physical coordinates of the starting node and ending node according to the relative position ratio in the crack location estimation results to generate the spanwise coordinates and chordwise coordinates of the crack center point along the blade. Based on the actual length of the monitored line segment in the physical coordinate system of the wind turbine blade and the crack length ratio in the crack location estimation results, a length conversion process is performed to generate a crack length estimate. Crack location information is generated by combining the position coordinates of the crack center point along the blade spanwise, the position coordinates along the blade chordwise, the estimated crack length, and the blade region to which the monitored line segment belongs.

7. The method for monitoring wind turbine blade cracks based on resistance feature learning according to claim 1, characterized in that, The generation of the crack early warning information specifically includes: Rules for setting crack risk classification thresholds based on the operational safety requirements of wind turbine blades; The crack occurrence probability and crack severity level are extracted from the crack diagnosis information, and the crack length estimate is extracted from the crack location information. The above three types of data are combined and processed to generate a crack risk input vector. The crack risk input vector is compared with the upper and lower limits of each risk score interval in the crack risk classification threshold rule to determine the risk score interval where the crack risk score value is located, and obtain the corresponding crack risk level identifier. The crack risk level identifier is combined with the crack location blade area, the position coordinate along the blade span, the position coordinate along the blade chord and the crack length estimate to generate crack early warning information for the corresponding monitoring line segment. After encapsulating the crack location information and crack early warning information according to the preset message format, they are output through the communication interface with the wind turbine SCADA system.

Images