Fan blade broken line degree detection method based on electrical parameters

Abstract

The invention discloses a fan blade broken line degree detection method based on electrical parameters, and relates to the technical field of wind power, and the method comprises the following steps: S100, collecting electrical parameters; s200, extracting a feature vector; s300, constructing a graph data structure; s400, electrical parameter characteristics are extracted; and S500, detecting the wire breaking degree. According to the method, the problem that a traditional single feature extraction method is insufficient in robustness under complex working conditions is solved, the circuit state is comprehensively reflected, and the accuracy and stability of parameter decoding are remarkably improved; according to the method, the wire breaking degree is quantified, the wire breaking fault and the severity thereof are accurately identified, and the misjudgment risk is effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power, and particularly to a method for detecting the degree of broken wires of a wind turbine blade based on electrical parameters. Background Art

[0002] As one of the key components of a wind turbine, the performance and condition of the wind turbine blade directly affect the power generation efficiency and overall operation safety of the wind turbine. Due to the long-term influence of natural factors such as strong wind, temperature difference, and ultraviolet radiation during the operation of the blade, problems such as fatigue, aging, and fracture may occur in the blade material. In particular, the cables, sensors, and connecting wires of the control system inside the wind turbine blade are at risk of breaking. These broken wire problems often affect the normal operation of the equipment and may even cause the equipment to shut down, bringing higher maintenance costs and equipment losses to the wind farm. Therefore, timely and effectively detecting broken wire faults is of great significance for ensuring the normal operation of the system, reducing maintenance costs, and improving reliability.

[0003] Currently, the detection of broken wires in wind turbine blades mainly relies on traditional methods such as manual inspection, vibration monitoring, and visual inspection. Although these methods can detect faults and damages of the blade to a certain extent, there are still some obvious limitations. Specifically: Manual inspection method: It relies on manual patrol. During the inspection process, tools such as telescopes are used to observe the surface of the blade to find cracks or damages. However, manual inspection has low efficiency, is prone to missing minor cable broken wire problems, and there are safety risks during high-altitude operations and bad weather.

[0004] Vibration monitoring method: By monitoring the vibration characteristics of the blade through vibration sensors to infer whether the blade has damages or faults, but this method mainly targets structural faults and is not sensitive to broken wire problems in electrical systems such as sensor cables and control signal lines. In addition, vibration monitoring is easily affected by external interference, resulting in false alarms or missed alarms.

[0005] Visual inspection method: Using high-definition cameras or drones to take images of the blade for appearance inspection can find surface cracks or damages, but for faults in the internal electrical system, especially cable breaks and signal problems, visual inspection cannot provide effective identification.

[0006] Most of these traditional methods rely on manual operation, have limited detection ranges, and low efficiency. In particular, it is difficult to detect minor broken wire problems in the internal electrical system of the blade, resulting in the failure to detect faults in a timely manner, affecting the operation efficiency and safety of the wind turbine.

[0007] Therefore, those skilled in the art are committed to developing a method for detecting the degree of broken wires of a wind turbine blade based on electrical parameters. Summary of the Invention

[0008] In view of the above defects of the prior art, the technical problem to be solved by the present invention is how to effectively detect the disconnection fault of the connecting wire of the wind turbine blade, ensure the normal operation of the wind turbine, reduce the maintenance cost and improve the reliability.

[0009] The applicant's research shows that by extracting the time-domain characteristics of electrical parameter information from the high-frequency pulse signals collected in real time, capturing the subtle changes in the signals, constructing a hybrid architecture of a graph convolutional network (GCN) and a graph recurrent unit (GRU), through cascaded feature extraction, feature fusion, variational feature encoding, decoding, and attention feature enhancement, for the high-precision extraction of electrical parameters, introducing an architecture based on a dynamic threshold generation network DTGN (Dynamic Threshold Generation Network), combining the extracted electrical parameter features and environmental factor feature data, and with a physical constraint optimization strategy, to quantitatively calculate the degree of disconnection.

[0010] In one embodiment of the present invention, a method for detecting the degree of disconnection of a wind turbine blade based on electrical parameters is provided, including: S100, Electrical parameter acquisition: Real-time acquisition of the electrical parameters of the cable of the target wind turbine blade to obtain a high-frequency pulse signal with proper attenuation, simply referred to as a high-frequency pulse signal; S200, Feature vector extraction: Divide the above high-frequency pulse signal into multiple fixed time windows, and extract feature vectors with the length of the fixed time window; S300, Construct a graph data structure: Based on the feature vectors, construct temporal edges, add similarity edges, and fuse the temporal edges and similarity edges to obtain a complete graph data structure; S400, Extract electrical parameter features: Build a hybrid architecture based on a graph convolutional network (GCN) and a graph recurrent unit (GRU), perform local spatio-temporal feature fusion and temporal dynamic fusion with the complete graph data structure as the input, and extract electrical parameter features; S500, Detect the degree of disconnection: Concatenate the electrical parameter features and environmental factor feature data, input them into the dynamic threshold generation network (DTGN), fuse the electrical parameter and environmental factor feature data, calculate the predicted value of the dynamic threshold of the electrical parameters, compare it with the measured resistance change amount in the wind turbine blade fault, and determine the degree of disconnection according to the disconnection degree classification.

[0011] Optionally, in the method for detecting the degree of disconnection of a wind turbine blade based on electrical parameters in the above embodiment, the high-frequency pulse signal with proper attenuation shows a characteristic of stable attenuation of the amplitude over time under normal working conditions, and the attenuation law reflects the health status of the circuit.

[0012] Optionally, in the method for detecting the degree of broken wire of a wind turbine blade based on electrical parameters in any of the above embodiments, the length T of the fixed time window is set between 10 ms and 100 ms.

[0013] Preferably, in the method for detecting the degree of broken wire of a wind turbine blade based on electrical parameters in the above embodiment, the length of the fixed time window is T = 50 ms, taking into account the instantaneous response characteristics of high-frequency pulse signals and a sufficient time-domain coverage range, and improving the integrity and robustness of feature extraction.

[0014] Optionally, in the method for detecting the degree of broken wire of a wind turbine blade based on electrical parameters in any of the above embodiments, the feature vector includes: Time-domain features, including voltage, current, and their first and second derivatives; Frequency-domain features, and energy features are extracted using short-time Fourier transform (STFT) in a preset power frequency and harmonic frequency band, a preset resonance and switching noise frequency band, and a preset high-frequency pulse signal frequency band.

[0015] Further, in the method for detecting the degree of broken wire of a wind turbine blade based on electrical parameters in the above embodiment, the preset power frequency and harmonic frequency band is 50 Hz - 1 kHz.

[0016] Preferably, in the method for detecting the degree of broken wire of a wind turbine blade based on electrical parameters in the above embodiment, the preset power frequency and harmonic frequency band is 500 Hz.

[0017] Further, in the method for detecting the degree of broken wire of a wind turbine blade based on electrical parameters in the above embodiment, the preset resonance and switching noise frequency band is 1 kHz - 100 kHz.

[0018] Preferably, in the method for detecting the degree of broken wire of a wind turbine blade based on electrical parameters in the above embodiment, the preset resonance and switching noise frequency band is 50 kHz.

[0019] Further, in the method for detecting the degree of broken wire of a wind turbine blade based on electrical parameters in the above embodiment, the preset high-frequency pulse signal frequency band is 100 kHz - 3 MHz.

[0020] Preferably, in the method for detecting the degree of broken wire of a wind turbine blade based on electrical parameters in the above embodiment, the preset high-frequency pulse signal frequency band is 1.5 MHz.

[0021] Optionally, in the method for detecting the degree of broken wire of a wind turbine blade based on electrical parameters in any of the above embodiments, step S300 includes: S310. Construct a time-series edge set, add undirected edges to the nodes corresponding to the fixed time window, and form a time-series edge set , representing the continuity characteristics of the signal in the time dimension; S320. Construct a similarity edge set. Represent the high-frequency pulse signals corresponding to each fixed time window as node feature vectors to obtain a set of node feature vectors. Calculate the cosine similarity between any two node feature vectors. When the cosine similarity is greater than the similarity threshold, add a similarity edge between the corresponding nodes to construct a similarity edge set. , forming node pairs that are discontinuous in time series but similar in feature dimension; S330. Construct a graph data structure for the time series edge set and the similarity edge set to perform a union operation to obtain a complete edge set. Combine the complete edge set with the set of node feature vectors to form a graph data structure.

[0022] Further, in the method for detecting the degree of broken wire of a fan blade based on electrical parameters in the above embodiment, the similarity threshold is 0.85.

[0023] Optionally, in the method for detecting the degree of broken wire of a fan blade based on electrical parameters in any of the above embodiments, step S400 includes: S410. Build a hybrid architecture, building a hybrid architecture based on a graph convolutional network (GCN) and a graph recurrent unit (GRU); S420. Temporal dynamic fusion. Use the graph convolutional network (GCN) to update the features of each node. By aggregating the information of its own node and neighbor nodes, realize the fusion of local spatio-temporal features, extract local structure information, perform temporal modeling on the nodes with updated features, construct a time series input tensor according to the length of the fixed time window, and use the graph recurrent unit (GRU) to capture the dynamic changes and dependencies of high-frequency pulse signals in the time series for temporal dynamic fusion; S430. Obtain a hidden state vector. The graph recurrent unit (GRU) processes the time series input tensors of each fixed time window in sequence to calculate the hidden state vector; S440. Introduce attention to form a temporal attention representation and a frequency domain attention representation; S450. Form a global feature vector. Perform global pooling on the local spatio-temporal features to form a global feature vector, reflecting the original global characteristics; S460. Construct a comprehensive spatio-temporal feature representation. Use the graph cyclic variational inference encoder in the graph recurrent unit (GRU) to encode the global feature vector to generate a latent variable. Introduce joint temporal attention, including temporal attention and frequency domain attention, splice the latent variables into vectors, and output a comprehensive spatio-temporal feature representation; S470. Calculate the joint attention weight vector, concatenate the temporal attention representation and the frequency-domain attention representation to obtain a fusion vector, input the fusion vector into a fully connected network for transformation to obtain the joint attention weight vector; S480. Form the node embedding representation, use the joint attention weight vector to weight the comprehensive spatio-temporal feature representation, and calculate the structured embedding feature vector of each node; S490. Extract the electrical parameter features, input the structured embedding feature vector into the variational decoder of the graph recurrent unit (GRU), and map and generate the electrical parameters.

[0024] Furthermore, in the method for detecting the degree of broken wires of a wind turbine blade based on electrical parameters in the above embodiment, the electrical parameters include resistance R, capacitance C, and inductance L.

[0025] Furthermore, in the method for detecting the degree of broken wires of a wind turbine blade based on electrical parameters in the above embodiment, step S420 includes: S421. Extract the local structure information, use the graph convolutional network (GCN) to update the features of each node, and realize the fusion of local spatio-temporal features by aggregating the information of its own node and neighbor nodes, and extract the local structure information; S422. Temporal dynamic fusion, perform temporal modeling on the nodes with updated features, construct a time series input tensor according to the length of a fixed time window, and use the graph recurrent unit (GRU) to capture the dynamic changes and dependencies of high-frequency pulse signals in the time series for temporal dynamic fusion.

[0026] Furthermore, in the method for detecting the degree of broken wires of a wind turbine blade based on electrical parameters in the above embodiment, step S440 includes: S441. Introduce temporal attention, perform a linear transformation on the hidden state vector, calculate the correlation score between contexts, obtain the attention weight after Softmax normalization, and weighted aggregate the hidden state vector to form the temporal attention representation of the current node; S442. Introduce frequency-domain attention, perform a short-time Fourier transform (STFT) on the high-frequency pulse signal, use the extracted energy features as the input, and adopt a simple two-layer fully connected attention module to output the frequency-domain attention vector to form a frequency-domain attention representation with differentiated frequency band response capabilities.

[0027] Optionally, in the method for detecting the degree of broken wires of a wind turbine blade based on electrical parameters in any of the above embodiments, the dynamic threshold generation network (DTGN) includes a first-layer head fully connected neural network, a second-layer head fully connected neural network, a variational inference encoder and decoder, and a tail fully connected neural network.

[0028] Further, in the method for detecting the degree of broken wire of a wind turbine blade based on electrical parameters in the above embodiment, step S500 includes: S510. Extract and strengthen key information, integrate the characteristics of electrical parameters and the characteristic data of environmental factors, perform high-dimensional feature encoding, splice by dimension, and then input it into a dynamic threshold generation network (DTGN) to obtain a feature encoding result , including the key features of electrical parameters and environmental information; S520. Obtain the dynamic threshold prediction value, input the feature encoding result into the terminal fully connected neural network through linear transformation, map it to the dynamic threshold space, and obtain the dynamic threshold prediction value of the electrical parameters; S530. Calculate the degree of broken wire, compare the dynamic threshold prediction value with the resistance change amount in the measured wind turbine blade fault to obtain the resistance offset degree , and determine the degree of broken wire according to the classification of the degree of broken wire.

[0029] Further, in the method for detecting the degree of broken wire of a wind turbine blade based on electrical parameters in the above embodiment, the classification of the degree of broken wire includes: The resistance offset degree ≤0.1, indicating that the blade connection is normal and there is no broken wire; 0.1 < the resistance offset degree ≤0.3, indicating a slight broken wire; The resistance offset degree >0.3, indicating a serious broken wire.

[0030] Further, in the method for detecting the degree of broken wire of a wind turbine blade based on electrical parameters in the above embodiment, step S510 includes: S511. Integrate the characteristics of electrical parameters and the characteristic data of environmental factors, perform high-dimensional feature encoding on the integrated electrical parameter and environmental factor characteristic data, splice by dimension, and form a fusion vector; S512. Calculate the high-dimensional feature representation in the first stage, input the fusion vector into the first-layer head fully connected neural network of the dynamic threshold generation network (DTGN), and perform nonlinear processing through the first-stage ReLU non-linear activation function to obtain the high-dimensional feature representation in the first stage; S513. Calculate the feature encoding result, input the high-dimensional feature representation in the first stage into the second-layer head fully connected neural network of the dynamic threshold generation network (DTGN), and perform nonlinear processing through the second-stage ReLU non-linear activation function to obtain the feature encoding result , including the key features of electrical parameters and environmental information.

[0031] Further, in the method for detecting the degree of broken wires of a wind turbine blade based on electrical parameters in the above embodiments, the ReLU non-linear activation function in the first stage is the same as the ReLU non-linear activation function in the second stage.

[0032] Optionally, in the method for detecting the degree of broken wires of a wind turbine blade based on electrical parameters in any of the above embodiments, the key features of the electrical parameters and the environmental information include the time-domain features and frequency-domain features of the electrical parameters, the environmental factor feature data, and the statistical features of the blade health state.

[0033] Optionally, in the method for monitoring the degree of broken wires of a wind turbine blade based on electrical parameters in any of the above embodiments, the environmental factor feature data includes temperature (°C), humidity (%RH), and electromagnetic interference intensity (μV / m).

[0034] The present invention adopts a deep learning architecture combining a graph convolutional network and a graph recurrent unit to accurately extract electrical parameters from a high-frequency pulse signal with proper attenuation. Through spatio-temporal feature fusion, joint temporal attention and frequency-domain attention are introduced to capture local and global information, making up for the problem of insufficient robustness of traditional single feature extraction methods under complex working conditions, comprehensively reflecting the circuit state, and significantly improving the accuracy and stability of parameter decoding. On the other hand, the present invention introduces a dynamic threshold generation network (DTGN), fuses electrical parameters and environmental factor feature data, and uses physical constraint optimization to realize the quantification of the degree of broken wires, accurately identify the broken wire fault and its severity, effectively reduce the risk of misjudgment, and ensure the safe operation of the power system and equipment. The present invention provides a new solution for complex time-frequency signal analysis, can be widely applied to fields such as intelligent manufacturing, fault diagnosis, and predictive maintenance, and has extremely high practical value and application prospects.

[0035] The following will further illustrate the concept, specific structure, and technical effects of the present invention with reference to the accompanying drawings to fully understand the purpose, features, and effects of the present invention. Description of the Drawings

[0036] Figure 1 is a flowchart showing a method for detecting the degree of broken wires of a wind turbine blade based on electrical parameters according to an exemplary embodiment; Figure 2 is a flowchart showing the construction of a graph data structure according to an exemplary embodiment; Figure 3 is a flowchart showing the extraction of electrical parameter features according to an exemplary embodiment; Figure 4 is a flowchart showing the detection of the degree of broken wires according to an exemplary embodiment. Detailed Embodiments

[0037] The following describes several preferred embodiments of the present invention with reference to the accompanying drawings of the specification, making its technical content clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0038] In the drawings, components with the same structure are denoted by the same reference numerals, and components with similar structures or functions everywhere are denoted by similar reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. To make the illustration clearer, the thickness of some parts in the drawings is schematically exaggerated appropriately.

[0039] The applicant has designed a method for detecting the degree of broken wires of fan blades based on electrical parameters, as Figure 1 shown, including the following steps: S100. Electrical parameter acquisition, real-time acquisition of the electrical parameters of the cable of the blades of the target wind turbine. The electrical parameters include resistance R, capacitance C, and inductance L, and a high-frequency pulse signal with proper attenuation is obtained, simply referred to as the high-frequency pulse signal.

[0040] S200. Extracting feature vectors, dividing the above high-frequency pulse signal into multiple fixed-time windows, and extracting feature vectors with the length of the fixed-time window. The high-frequency pulse signal with proper attenuation shows the characteristic that the amplitude decays stably with time under the normal working state, and the decay law reflects the health state of the circuit. The length of the fixed-time window is T = 50 ms, taking into account the instantaneous response characteristics of the high-frequency pulse signal and sufficient time-domain coverage range, and improving the integrity and robustness of feature extraction. The feature vectors include: Time-domain features, including voltage, current, and their first and second derivatives; Frequency-domain features, using short-time Fourier transform (STFT) to extract energy features in the preset power frequency and harmonic frequency bands, the preset resonance and switching noise frequency bands, and the preset high-frequency pulse signal frequency bands. The preset power frequency and harmonic frequency bands are 500 Hz, the preset resonance and switching noise frequency bands are 50 kHz, and the preset high-frequency pulse signal frequency bands are 1.5 MHz.

[0041] S300. Constructing a graph data structure, constructing a time-series edge based on the feature vector, adding a similarity edge, and fusing the time-series edge and the similarity edge to obtain a complete graph data structure; as Figure 2 shown, specifically including: S310. Constructing a time-series edge set, adding an undirected edge to the node corresponding to the fixed-time window to form a time-series edge set , representing the continuity feature of the signal in the time dimension; S320. Construct a similarity edge set. Represent the high-frequency pulse signals corresponding to each fixed time window as node feature vectors to obtain a set of node feature vectors. Calculate the cosine similarity between any two node feature vectors. When the cosine similarity is greater than the similarity threshold, add a similarity edge between the corresponding nodes to construct a similarity edge set. , forming node pairs that are discontinuous in time series but similar in feature dimension, with a similarity threshold of 0.85. S330. Construct a graph data structure. Perform a union operation on the time series edge set and the similarity edge set to obtain a complete edge set. Combine the complete edge set with the set of node feature vectors to form a graph data structure.

[0042] S400. Extract electrical parameter features. Build a hybrid architecture based on a graph convolutional network (GCN) and a graph recurrent unit (GRU). Use the complete graph data structure as the input for local spatio-temporal feature fusion and temporal dynamic fusion to extract electrical parameter features; as Figure 3 shown, specifically including: S410. Build a hybrid architecture. Build a hybrid architecture based on a graph convolutional network (GCN) and a graph recurrent unit (GRU). S420. Temporal dynamic fusion. Use the graph convolutional network (GCN) to update the features of each node. By aggregating the information of its own node and neighbor nodes, achieve the fusion of local spatio-temporal features, extract local structure information, perform temporal modeling on the nodes with updated features, construct a time series input tensor according to the length of the fixed time window, and use the graph recurrent unit (GRU) to capture the dynamic changes and dependencies of high-frequency pulse signals in the time series for temporal dynamic fusion; specifically including: S421. Extract local structure information. Use the graph convolutional network (GCN) to update the features of each node. By aggregating the information of its own node and neighbor nodes, achieve the fusion of local spatio-temporal features, extract local structure information. S422. Temporal dynamic fusion. Perform temporal modeling on the nodes with updated features, construct a time series input tensor according to the length of the fixed time window, and use the graph recurrent unit (GRU) to capture the dynamic changes and dependencies of high-frequency pulse signals in the time series for temporal dynamic fusion. S430. Obtain the hidden state vector. The graph recurrent unit (GRU) processes the time series input tensors of each fixed time window in sequence to calculate the hidden state vector. S440. Introduce attention to form a temporal attention representation and a frequency domain attention representation; specifically including: S441. Introduce temporal attention, perform a linear transformation on the hidden state vector, calculate the correlation score between contexts, obtain the attention weights through Softmax normalization, and perform weighted aggregation on the hidden state vector to form the temporal attention representation of the current node; S442. Introduce frequency-domain attention, perform a short-time Fourier transform (STFT) on the high-frequency pulse signal, use the extracted energy features as input, adopt a simple two-layer fully-connected attention module, output the frequency-domain attention vector, and form the frequency-domain attention representation with the ability to respond to different frequency bands; S450. Form the global feature vector, perform global pooling on the local spatio-temporal features to form the global feature vector, which reflects the original global characteristics; S460. Construct the comprehensive spatio-temporal feature representation. Use the graph recurrent variational inference encoder in the graph recurrent unit (GRU) to encode the global feature vector, generate the latent variable, introduce the joint temporal attention, including temporal attention and frequency-domain attention, splice the latent variables, and output the comprehensive spatio-temporal feature representation; S470. Calculate the joint attention weight vector. Splice the temporal attention representation and the frequency-domain attention representation to obtain the fusion vector, input the fusion vector into a fully-connected network for transformation, and obtain the joint attention weight vector; S480. Form the node embedding representation. Use the joint attention weight vector to weight the comprehensive spatio-temporal feature representation and calculate the structured embedding feature vector of each node; S490. Extract the electrical parameter features. Input the structured embedding feature vector into the decoder of the graph recurrent unit (GRU) to map and generate the electrical parameters.

[0043] S500. Detect the degree of wire breakage. Splice the electrical parameter features and the environmental factor feature data, and input them into the dynamic threshold generation network (DTGN). The dynamic threshold generation network (DTGN) includes the first-layer head fully-connected neural network, the second-layer head fully-connected neural network, the variational inference encoder and decoder, and the end fully-connected neural network. Integrate the electrical parameter and environmental factor feature data, calculate the dynamic threshold prediction value of the electrical parameters, compare it with the resistance change amount in the measured wind turbine blade failure, and determine the degree of wire breakage according to the degree classification of wire breakage; as Figure 4 shown, specifically including: S510. Extract and strengthen the key information, integrate the features of the electrical parameters and the environmental factor feature data, perform high-dimensional feature encoding, splice by dimension, and then input them into the dynamic threshold generation network (DTGN) to obtain the feature encoding result , including the key features of electrical parameters and environmental information; specifically including: S511. Integrate the characteristic data of electrical parameters and environmental factors, perform high-dimensional feature encoding on the integrated electrical parameters and environmental factor characteristic data, and splice them by dimension to form a fusion vector; S512. Calculate the high-dimensional feature representation in the first stage. Input the fusion vector into the first-layer head full-connected neural network of the dynamic threshold generation network (DTGN), and perform non-linear processing through the ReLU non-linear activation function in the first stage to obtain the high-dimensional feature representation in the first stage; S513. Calculate the feature encoding result. Input the high-dimensional feature representation in the first stage into the second-layer head full-connected neural network of the dynamic threshold generation network (DTGN), and perform non-linear processing through the ReLU non-linear activation function in the second stage to obtain the feature encoding result , including the key features of electrical parameters and environmental information. The key features of electrical parameters and environmental information include the time-domain features and frequency-domain features of electrical parameters, environmental factor characteristic data, and statistical features of the blade health state. The environmental factor characteristic data includes temperature (°C), humidity (%RH), and electromagnetic interference intensity (μV / m); The ReLU non-linear activation function in the first stage is the same as the ReLU non-linear activation function in the second stage; S520. Obtain the dynamic threshold prediction value. Input the feature encoding result through a linear transformation into the end full-connected neural network, map it to the dynamic threshold space, and obtain the dynamic threshold prediction value of the electrical parameter; S530. Calculate the degree of wire breakage. Compare the dynamic threshold prediction value with the resistance change amount in the measured fan blade fault to obtain the resistance offset degree , classify according to the degree of wire breakage to determine the degree of wire breakage; The classification of the degree of wire breakage includes: Resistance offset degree ≤0.1, indicating that the blade connection is normal and there is no wire breakage; 0.1 < Resistance offset degree ≤0.3, indicating a slight wire breakage; Resistance offset degree >0.3, indicating a serious wire breakage.

[0044] The above has described in detail the preferred specific embodiments of the present invention. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of the present invention based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should be within the protection scope determined by the claims.

Claims

1. A method for detecting the degree of broken wire of a fan blade based on electrical parameters, characterized in that, Including: S100, Electrical parameter acquisition: Real-time acquisition of the electrical parameters of the cables of the blades of the target wind turbine to obtain a high-frequency pulse signal with proper attenuation, simply referred to as the high-frequency pulse signal; S200, Feature vector extraction: Divide the high-frequency pulse signal into multiple fixed time windows, and extract feature vectors based on the length of the fixed time window; S300, Constructing a graph data structure: Based on the feature vectors, construct temporal edges and add similarity edges. The temporal edges and the similarity edges are fused to obtain a complete graph data structure; S400, Extracting electrical parameter features: Build a hybrid architecture based on the graph convolutional network (GCN) and the graph recurrent unit (GRU). Use the complete graph data structure as the input for local spatio-temporal feature fusion and temporal dynamic fusion to extract electrical parameter features; S500, Detecting the degree of wire breakage: Concatenate the electrical parameter features with the environmental factor feature data, input them into the dynamic threshold generation network (DTGN), fuse the electrical parameters and the environmental factor feature data, calculate the dynamic threshold prediction value of the electrical parameters, compare it with the resistance change amount in the actual measured wind turbine blade failure, and determine the degree of wire breakage according to the wire breakage degree classification; 2. The method for detecting the degree of broken wire of a fan blade based on electrical parameters according to claim 1, wherein The length T of the fixed time window is set between 10 ms and 100 ms.

3. The method for detecting the degree of broken wire of a fan blade based on electrical parameters according to claim 2, wherein The feature vector includes: Time domain features, including voltage, current, and their first and second derivatives; Frequency domain features: Use short-time Fourier transform (STFT) to extract energy features in the preset power frequency and harmonic frequency bands, the preset resonance and switching noise frequency bands, and the preset high-frequency pulse signal frequency bands.

4. The method for detecting the degree of broken wire of a fan blade based on electrical parameters according to claim 3, wherein, The step S300 includes: S310. Construct a set of temporal edges. Add undirected edges to the nodes corresponding to the fixed time window to form a set of temporal edges, which represents the continuous feature of the signal in the time dimension. , representing the continuous feature of the signal in the time dimension. S320. Construct a similarity edge set. Represent the high-frequency pulse signals corresponding to each fixed time window as node feature vectors to obtain a set of node feature vectors. Calculate the cosine similarity between any two node feature vectors. When the cosine similarity is greater than the similarity threshold, add a similarity edge between the corresponding nodes to construct a similarity edge set , forming node pairs that are discontinuous in time series but similar in feature dimension; S330. Construct a graph data structure for the set of temporal edges and the set of similarity edges to perform a union operation to obtain a complete set of edges, and combine the complete set of edges with the set of node feature vectors to form a graph data structure.

5. The method for detecting the degree of broken wire of a fan blade based on electrical parameters according to claim 3, wherein The step S400 includes: S410, Building a hybrid architecture: Build a hybrid architecture based on the graph convolutional network (GCN) and the graph recurrent unit (GRU); S420, Temporal dynamic fusion: Use the graph convolutional network (GCN) to update the features of each node. By aggregating the information of its own node and neighbor nodes, realize the fusion of local spatio-temporal features, extract local structure information, perform temporal modeling on the nodes with feature updates, construct a time series input tensor according to the length of the fixed time window, and use the graph recurrent unit (GRU) to capture the dynamic changes and dependencies of the high-frequency pulse signal in the time series for temporal dynamic fusion; S430, Obtaining the hidden state vector: The graph recurrent unit (GRU) processes the time series input tensors of each fixed time window in sequence to calculate the hidden state vector; S440, Introducing attention: Form temporal attention representation and frequency domain attention representation; S450, Forming a global feature vector: Perform global pooling on the local spatio-temporal features to form a global feature vector, reflecting the original global characteristics; S460, Constituting a comprehensive spatio-temporal feature representation: Use the graph recurrent variational inference encoder in the graph recurrent unit (GRU) to encode the global feature vector, generate latent variables, introduce joint temporal attention, including temporal attention and frequency domain attention, splice the latent variables into vectors, and output a comprehensive spatio-temporal feature representation; S470. Calculate the joint attention weight vector, concatenate the temporal attention representation and the frequency-domain attention representation to obtain a fusion vector, input the fusion vector into a fully connected network for transformation to obtain the joint attention weight vector; S480. Form the node embedding representation, use the joint attention weight vector to weight the comprehensive spatio-temporal feature representation, and calculate the structured embedding feature vector of each node; S490. Extract the electrical parameter features, input the structured embedding feature vector into the variational decoder of the graph recurrent unit (GRU) to map and generate the electrical parameters.

6. The method for detecting the degree of broken wire of a fan blade based on electrical parameters according to claim 5, wherein, The step S420 includes: S421. Extract the local structure information, use the graph convolutional network (GCN) to update the features of each node, and realize the fusion of local spatio-temporal features by aggregating the information of its own node and neighbor nodes, and extract the local structure information; S422. Temporal dynamic fusion, perform temporal modeling on the nodes with updated features, construct a time series input tensor according to the length of the fixed time window, and use the graph recurrent unit (GRU) to capture the dynamic changes and dependencies of the high-frequency pulse signal in the time series for temporal dynamic fusion.

7. The method for detecting the degree of broken wire of a fan blade based on electrical parameters according to claim 6, wherein, The step S440 includes: S441. Introduce temporal attention, perform a linear transformation on the hidden state vector, calculate the correlation score between contexts, obtain the attention weight after Softmax normalization, and weighted aggregate the hidden state vector to form the temporal attention representation of the current node; S442. Introduce frequency-domain attention, perform a short-time Fourier transform (STFT) on the high-frequency pulse signal, use the extracted energy features as input, and adopt a simple two-layer fully connected attention module to output the frequency-domain attention vector to form a frequency-domain attention representation with differentiated frequency band response capabilities.

8. The method for detecting the degree of broken wire of a fan blade based on electrical parameters according to claim 3, wherein The dynamic threshold generation network (DTGN) includes a first-layer head fully connected neural network, a second-layer head fully connected neural network, a variational inference encoder and decoder, and a tail fully connected neural network.

9. The method for detecting the degree of broken wire of a fan blade based on electrical parameters according to claim 8, characterized in that The step S500 includes: S510. Extract and strengthen key information, integrate the characteristics of the electrical parameters and the characteristic data of environmental factors, perform high-dimensional feature encoding, splice by dimension, and then input into the dynamic threshold generation network (DTGN) to obtain the feature encoding result , including the key features of electrical parameters and environmental information; S520. Obtain the dynamic threshold prediction value, input the feature encoding result into the tail fully connected neural network through a linear transformation, map it to the dynamic threshold space to obtain the dynamic threshold prediction value of the electrical parameter; S540. Calculate the degree of wire breakage. Compare the predicted value of the dynamic threshold with the measured change in resistance in the wind turbine blade fault to obtain the degree of resistance deviation. Classify according to the degree of wire breakage to determine the degree of wire breakage.

10. The method for detecting the degree of broken wire of a fan blade based on electrical parameters according to claim 8, wherein, The step S510 includes: S511. Form a fusion vector, integrate the features of the electrical parameter and the environmental factor feature data, perform high-dimensional feature encoding on the integrated electrical parameter and environmental factor feature data, and splice them by dimension to form a fusion vector; S512. Calculate the high-dimensional feature representation in the first stage, input the fusion vector into the first-layer head fully connected neural network of the dynamic threshold generation network (DTGN), and perform nonlinear processing through the ReLU nonlinear activation function in the first stage to obtain the high-dimensional feature representation in the first stage; S513. Calculate the feature encoding result, input the high-dimensional feature representation in the first stage into the first fully-connected neural network at the head end of the second layer of the dynamic threshold generation network (DTGN), and perform non-linear processing through the ReLU non-linear activation function in the second stage to obtain the feature encoding result , including key features of electrical parameters and environmental information.

Images