Blade health degree detection method and system based on transfer learning

Abstract

The invention discloses a blade health degree detection method and system based on transfer learning. The blade health degree detection method comprises the following steps: acquiring real-time vibration data of a fan unit blade; and converting the real-time vibration data into a time sequence oscillogram. And converting the real-time vibration data into a two-dimensional pixel image by adopting a Grubrum angle field algorithm. And identifying the time sequence oscillogram according to a pre-constructed first identification model. And identifying the two-dimensional pixel image according to a pre-constructed second identification model. And fusing the classification result of the first identification model and the classification result of the second identification model based on a weighted voting strategy to obtain a final classification result of the health degree of the fan unit blade. According to the method, high-precision, low-cost and real-time detection of the health degree of the fan blade is achieved, and the defects that in the prior art, only basic structure monitoring is concerned, calculation complexity is high, and dependence on data is high are overcome.

Description

Technical Field

[0001] This invention relates to the field of wind farm equipment maintenance technology, and in particular to a method and system for detecting blade health based on transfer learning. Background Technology

[0002] Wind energy, as a clean and renewable energy source, plays a vital role in global energy transition and electricity production. Wind turbines are typically deployed in areas rich in wind resources but with harsh environments, such as coastal areas, high mountains, and deserts. During long-term operation, the blades are susceptible to erosion from dust particles and extreme weather conditions, leading to wear, cracks, and other damage. Failure to promptly detect and assess the blade's health can not only result in decreased power generation efficiency but also pose a risk of major safety accidents.

[0003] In the prior art, monitoring methods based on multi-source signals and deep learning have been proposed. For example, patent CN115456012A discloses a condition monitoring system and method for large components of wind turbines in wind farms. This technical solution collects acoustic emission signals and vibration signals of the wind turbine foundation structure through sensors. The acoustic emission signals are transformed into images using Gram angle field transform and then input into a convolutional neural network to extract features. Simultaneously, frequency domain statistical features are extracted from the vibration signals and processed in conjunction with a time-series encoder. Then, an image encoder is used to extract vibration waveform features. Finally, a classification feature matrix is ​​obtained through joint encoding and feature fusion, and the wind turbine structural condition result is output through a classifier.

[0004] However, the aforementioned existing technologies are mainly aimed at monitoring the foundation structure of offshore wind turbines. They rely on the multimodal fusion of acoustic emission and vibration signals, resulting in high system complexity and strong dependence on large-scale training data and computing resources. Therefore, the applicability of directly using these methods in blade health diagnosis is limited. Summary of the Invention

[0005] To address the shortcomings of existing technologies, which focus solely on basic structure monitoring, suffer from high computational complexity, and are highly dependent on data, the present invention aims to provide a blade health detection method and system based on transfer learning. This method achieves high-precision, low-cost, and real-time detection of wind turbine blade health, overcoming the deficiencies of existing technologies that focus solely on basic structure monitoring, suffer from high computational complexity, and are highly dependent on data.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] In a first aspect, the present invention provides a leaf health detection method based on transfer learning, the leaf health detection method comprising the following steps:

[0008] Acquire real-time vibration data of wind turbine blades;

[0009] The real-time vibration data is converted into a time-series waveform.

[0010] The Gram angle field algorithm is used to convert the real-time vibration data into a two-dimensional pixel image;

[0011] The time-series waveform is identified according to a pre-built first identification model, and a first classification result is output. The first identification model is a residual neural network architecture.

[0012] The two-dimensional pixel image is identified according to a pre-built second recognition model, and a second classification result is output. The second recognition model is a deep convolutional neural network architecture.

[0013] The first and second classification results are fused using a weighted voting strategy to obtain the final classification result of the wind turbine blade health.

[0014] In conjunction with the first aspect, the present invention also provides a first specific implementation of the first aspect, wherein the first identification model is a ResNet18 model based on transfer learning; and / or, the second identification model is a VGG16 model based on transfer learning.

[0015] In conjunction with the first aspect, the present invention also provides a second specific implementation of the first aspect, wherein the final classification result is one of healthy, sub-healthy, mild malfunction, and malfunction;

[0016] The first classification result is a four-dimensional probability distribution among healthy, sub-healthy, mildly faulty, and faulty individuals. ;

[0017] The second classification result is a four-dimensional probability distribution among healthy, sub-healthy, mildly disabled, and disabled individuals. .

[0018] In conjunction with the first aspect, the present invention also provides a third specific implementation of the first aspect, specifically, fusing the first classification result and the second classification result based on a weighted voting strategy, specifically as follows:

[0019] The four categories are, in order: healthy, sub-healthy, minor malfunction, and malfunction. The final probability is obtained by performing probability fusion, and the calculation formula is as follows: ;

[0020] In the formula, It is a category The final probability, The second recognition model predicts the category. The probability, The first recognition model predicts the category as The probability, To assign weight values ​​to the second recognition model, To assign weight values ​​to the first recognition model, and satisfy... .

[0021] In conjunction with the first aspect, the present invention also provides a fourth specific implementation of the first aspect, specifically, converting the real-time vibration data into a time-series waveform diagram, specifically, using the sampling time of the real-time vibration data as the horizontal axis and the corresponding vibration amplitude as the vertical axis, and plotting the time-series waveforms sequentially according to the time order to obtain a time-series waveform diagram characterizing the change of vibration amplitude over time.

[0022] In conjunction with the first aspect, the present invention also provides a fifth specific implementation of the first aspect, specifically, employing the Gram angle field algorithm to convert the real-time vibration data into a two-dimensional pixel image, specifically including:

[0023] The vibration amplitude of the real-time vibration data is normalized to convert the numerical range of the vibration amplitude into the [0,1] interval;

[0024] The normalized real-time vibration data is mapped to the angle component θ and the radius component r in the polar coordinate system through a polar coordinate encoding mechanism; wherein, a linear mapping strategy is used to encode the vibration amplitude to the angle dimension θ and to associate the sampling time with the radius component r.

[0025] Based on polar coordinate encoding, construct the Gram angle and field matrix and the Gram angle difference field matrix respectively;

[0026] The matrix element values ​​of the Gram angle sum field matrix and the Gram angle difference field matrix are linearly mapped to the 0-255 grayscale range or the RGB color space to generate a two-dimensional pixel image with temporal characteristics.

[0027] In conjunction with the first aspect, the present invention also provides a sixth specific embodiment of the first aspect, specifically, the mathematical functions of the Gram angle and the field matrix are as follows: ;

[0028] The mathematical function of the Gram angular difference field matrix is: ;

[0029] In the above function, For the th in the Gram matrix Line number The matrix elements in the column represent time points. Between The angular relationship.

[0030] In conjunction with the first aspect, the present invention also provides a seventh specific embodiment of the first aspect, wherein the leaf health detection method includes the following steps:

[0031] The real-time vibration data is preprocessed to remove abnormal and blank data.

[0032] Secondly, the present invention also provides a leaf health detection system based on transfer learning, the leaf health detection system comprising:

[0033] The acquisition module is used to acquire real-time vibration data of the wind turbine blades;

[0034] The first conversion module is used to convert the real-time vibration data into a time-series waveform.

[0035] The second conversion module is used to convert the real-time vibration data into a two-dimensional pixel image using the Gram angle field algorithm;

[0036] A first recognition model is used to recognize the time-series waveform and output a first classification result. The first recognition model is a residual neural network architecture.

[0037] The second recognition model is used to recognize the two-dimensional pixel image and output a second classification result. The second recognition model is a deep convolutional neural network architecture.

[0038] The classification module is used to fuse the first classification result and the second classification result based on a weighted voting strategy to obtain the final classification result of the wind turbine blade health.

[0039] In conjunction with the second aspect, the present invention also provides a first specific implementation of the second aspect, wherein the first identification model is a ResNet18 model based on transfer learning; and / or, the second identification model is a VGG16 model based on transfer learning.

[0040] Compared with the prior art, the present invention has at least the following beneficial effects:

[0041] This invention provides a blade health detection method based on transfer learning, comprising the following steps: acquiring real-time vibration data of wind turbine blades; converting the real-time vibration data into a time-series waveform; using the Gram angle field algorithm to convert the real-time vibration data into a two-dimensional pixel image; identifying the time-series waveform according to a pre-constructed first recognition model, where the first recognition model is a residual neural network architecture; identifying the two-dimensional pixel image according to a pre-constructed second recognition model, where the second recognition model is a deep convolutional neural network architecture; and fusing the first and second classification results based on a weighted voting strategy to obtain the final classification result of the wind turbine blade health.

[0042] This invention achieves efficient and accurate detection of wind turbine blade health by dual representation of real-time vibration data using time-series waveform diagrams and Gram angle field images, combined with a weighted fusion of a residual neural network and a deep convolutional neural network recognition model. This method enables high-precision, low-cost, and real-time detection of wind turbine blade health, overcoming the shortcomings of existing technologies that focus only on foundation structure monitoring, have high computational complexity, and are highly dependent on data. It reduces reliance on large-scale training data and computational resources, and compared to existing technologies, it offers advantages such as strong applicability, high diagnostic accuracy, high computational efficiency, and good real-time performance for wind turbine blade health inspection. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating a leaf health detection method based on transfer learning according to the present invention. Figure 1 ;

[0044] Figure 2 This is a flowchart illustrating a leaf health detection method based on transfer learning according to the present invention. Figure 2 ;

[0045] Figure 3 This is a technical roadmap for a leaf health detection method based on transfer learning according to the present invention.

[0046] Figure 4 This is a schematic diagram of the network structure of the ResNet18 model based on transfer learning in the leaf health detection method of the present invention;

[0047] Figure 5 This is a schematic diagram of the network structure of the VGG16 model based on transfer learning in the leaf health detection method of the present invention.

[0048] Figure 6 This is a model training roadmap for a leaf health detection method based on transfer learning according to the present invention. Detailed Implementation

[0049] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0050] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0051] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0052] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0053] Figure 1 A flowchart of a leaf health detection method based on transfer learning according to an embodiment of this application is shown. The leaf health detection method can be implemented by a leaf health detection system based on transfer learning. Specifically, the leaf health detection system can be implemented using both hardware and software.

[0054] Reference Figure 1 The diagram shown is a flowchart of a leaf health detection method based on transfer learning. This leaf health detection method includes at least steps S110 to S140, which are described in detail below:

[0055] Step S110: Obtain real-time vibration data of the wind turbine blades.

[0056] In this embodiment, vibration sensors, such as accelerometers or fiber optic sensors, are installed on the wind turbine blades to acquire real-time vibration signals during operation. These vibration signals contain dynamic response information of the blades under different operating conditions, and their sampling frequency should meet the requirements of the blade vibration characteristic frequency band to ensure the integrity and validity of the vibration data.

[0057] Understandably, real-time vibration data includes sampling time and corresponding vibration amplitude. Vibration amplitude, i.e., the vibration signal, can reflect the structural health status of the blade, including defects such as cracks, wear, or loosening.

[0058] Step S120: Convert the real-time vibration data into a time-series waveform diagram.

[0059] In some possible implementations, the acquired vibration data is plotted as a time-series waveform, with the horizontal axis representing time and the vertical axis representing vibration amplitude. For example, the sampling time of the vibration data is used as the horizontal axis, and the corresponding vibration amplitude is used as the vertical axis; by connecting each sampling point sequentially, a time-series waveform representing the change of vibration amplitude over time is obtained.

[0060] Time-series waveforms can intuitively reflect the dynamic changes in blade vibration over time, preserving the amplitude and time dependence information of the signal, providing basic input data for subsequent time-series-based feature extraction and residual neural network identification. Simultaneously, this graph can be used to analyze the trend, abrupt changes, and abnormal peaks of blade vibration signals, thereby assisting in early fault detection and health assessment.

[0061] Step S130: The Gram angle field algorithm is used to convert the real-time vibration data into a two-dimensional pixel image.

[0062] In some possible implementations, the normalized vibration data is mapped to a polar coordinate system using the Gramian Angular Field (GAF) algorithm to obtain the angular component θ and the radius component r, which are then converted into a two-dimensional matrix. This matrix is ​​then used to generate grayscale or color images through pixel mapping. The GAF algorithm can encode the temporal dependencies of a one-dimensional time series into a two-dimensional image, preserving the overall structural features of the time series. This allows deep convolutional neural networks to better extract features from the spatial dimension, improving the accuracy of classification and recognition.

[0063] Step S140: Identify the time-series waveform diagram according to the pre-built first identification model and output the first classification result. The first identification model is a residual neural network architecture.

[0064] In some possible implementations, the time-series waveform generated in step S120 is input into a pre-trained first recognition model for classification. This model is a Residual Neural Network (ResNet) architecture. After transfer learning, the model can quickly adapt to the characteristics of the blade vibration data, achieve a preliminary judgment on the health status of the blade, and output the first classification result.

[0065] In this embodiment, the residual neural network effectively solves the gradient vanishing problem in deep network training by introducing residual connections, and can extract deep feature information from time-series waveforms, thereby improving the sensitivity and recognition accuracy of health anomalies.

[0066] Step S150: Recognize the two-dimensional pixel image according to the pre-constructed second recognition model and output the second classification result. The second recognition model is a deep convolutional neural network architecture.

[0067] In some possible implementations, the two-dimensional pixel image generated in step S130 is input into a pre-trained second recognition model for classification. This model is a deep convolutional neural network (CNN) architecture, which can automatically extract local and global features from the two-dimensional image and output a second classification result.

[0068] In this embodiment, deep CNN excels at extracting texture, edge, and spatial pattern features from images. For two-dimensional images generated by GAF, it can effectively capture potential abnormal patterns in blade vibration signals, improving the accuracy and robustness of detection.

[0069] Step S160: The first classification result and the second classification result are fused based on a weighted voting strategy to obtain the final classification result of the wind turbine blade health.

[0070] In some possible implementations, the first classification result from step S140 and the second classification result from step S150 are weighted and fused. Specifically, weights are assigned based on the accuracy of each model on the training set, and a weighted voting strategy is used to obtain the final classification result of leaf health.

[0071] In this application, the fusion strategy can fully utilize the complementarity of temporal domain features and spatial image features to improve the reliability and accuracy of the overall classification and reduce the risk of misjudgment that may be caused by a single model.

[0072] In some possible implementations, after acquiring real-time vibration data of the wind turbine blades, the vibration data is preprocessed to ensure the accuracy and reliability of subsequent analysis. Specifically, this includes the following operations:

[0073] (1) Abnormal data processing: detect abnormal points in the vibration signal that deviate significantly from the normal fluctuation range, such as sudden peaks, noise interference or sensor acquisition errors, and delete them or correct them by interpolation.

[0074] (2) Blank data processing: Identify missing sampling points or blank data in the vibration signal and supplement them by linear interpolation, spline interpolation or forward filling methods.

[0075] In this embodiment, abnormal and blank data can interfere with temporal feature extraction and subsequent neural network classification, potentially leading to misclassification results. Preprocessing ensures the continuity, integrity, and consistency of the input data, improves the feature representation quality of temporal waveforms and two-dimensional pixel images, thereby enhancing the recognition accuracy and robustness of the classification model.

[0076] like Figure 2 As shown, this is a detailed flowchart of step S130, which uses the Gram angle field algorithm to convert the real-time vibration data into a two-dimensional pixel image, specifically including:

[0077] Step S210: Normalize the vibration amplitude of the real-time vibration data and convert the numerical range of the vibration amplitude to the [0,1] interval.

[0078] In this embodiment, normalization processing can eliminate the dimensional differences in vibration amplitude under different blades, different wind turbines, or different sampling conditions, thereby unifying the standard of subsequent image encoding processes and ensuring the stability of neural network training and recognition. The vibration amplitude of the real-time vibration data obtained in step S110 is normalized, and its numerical range is linearly mapped to the [0,1] interval.

[0079] In one possible implementation, the normalization process is calculated as follows:

[0080] ;

[0081] in, This is a data sequence of real-time vibration data. For the first One data value, and These are the maximum and minimum values ​​in the data sequence, respectively. This is the normalized value.

[0082] Step S220: Through a polar coordinate encoding mechanism, the normalized real-time vibration data is mapped to the angle component θ and the radius component r in the polar coordinate system; wherein, a linear mapping strategy is used to encode the vibration amplitude to the angle dimension θ, and the sampling time is associated with the radius component r.

[0083] In this embodiment, polar coordinate encoding can transform the sequential characteristics of time series signals into spatial distribution features, enabling the subsequent two-dimensional matrix to simultaneously retain vibration amplitude information and time series relationships, thereby facilitating the extraction of temporal features from the spatial dimension by deep convolutional neural networks.

[0084] In some possible implementations, the angular component θ is calculated as follows:

[0085] GASF (Gram Point and Field) is ;

[0086] GADF (Gram Angular Difference Field) is ;

[0087] The radius component r is calculated as follows:

[0088] in, For the first The angles mapped to each data point in the polar coordinate system For the first The radius of each data point, where N is the length of the time series in the real-time vibration data.

[0089] Step S230: Based on the polar coordinate encoding, construct the Gram angle and field matrix and the Gram angle difference field matrix respectively.

[0090] In this embodiment, GASF and GADF reflect the interdependence and changing trend of vibration amplitudes in the time series, respectively, and can transform the dynamic characteristics of one-dimensional time series signals into two-dimensional spatial structures, providing richer feature information for deep convolutional neural networks.

[0091] In one specific implementation, the mathematical functions of the Gram angle and the field matrix are: The mathematical function of the Gram angular difference field matrix is: .

[0092] In the above function, For the th in the Gram matrix Line number The matrix elements in the column represent time points. Between The angular relationship is shown. The temporal dependencies of the time series are encoded by the generated N×N symmetric matrix.

[0093] Step S240: Linearly map the matrix element values ​​of the Gram angle sum field matrix and the Gram angle difference field matrix to the 0-255 grayscale range or RGB color space to generate a two-dimensional pixel image with temporal characteristics.

[0094] In this application, the element values ​​of the GASF and GADF matrices are linearly mapped to the 0–255 grayscale range, or mapped to the RGB color space, to generate a two-dimensional pixel image with temporal characteristics. By mapping matrix elements to pixel values, a two-dimensional image that can be input into a convolutional neural network is obtained, preserving the global features and local patterns of the time-series signal, which can improve the recognition accuracy and robustness of the leaf health classification model.

[0095] Understandably, the elements of the Gram matrix are mapped to colors. Generally, the larger the value, the darker the color. The mapped pixel values ​​are then arranged according to the rows and columns of the matrix to generate a two-dimensional image.

[0096] In some specific implementations, step S160 selects the category with the highest probability after fusion as the result of determining the leaf health, and the final classification result output is one of the following: healthy, sub-healthy, slightly faulty, or faulty.

[0097] For example, Healthy: The blade structure is intact, with no obvious damage, cracks, or wear, and the vibration signal is within the normal range. In terms of characteristics: the vibration amplitude is stable, and no abnormal peaks or patterns appear in the waveform and 2D pixel plot. This result indicates that the blade is functioning normally and requires no maintenance intervention.

[0098] Sub-healthy: The blades exhibit minor abnormalities or early wear, but this has not yet affected normal operation. Characteristically, vibration signals may show slight fluctuations, abnormal peaks may occasionally appear in the waveform, and abnormal textures may be present in localized areas of the 2D image. Monitoring and potentially preventative maintenance are required to prevent the condition from developing into more serious failures.

[0099] Mild Fault: Obvious defects appear on the blades, such as small cracks, localized wear, or loosening, which may have some impact on the operation of the wind turbine. Characteristically, the vibration waveform shows periodic or random abnormal peaks, and the two-dimensional pixel graph displays obvious abnormal patterns. Maintenance or repair should be carried out as soon as possible to prevent further damage.

[0100] Fault: Severe damage to the blade structure, such as large cracks, fractures, or significant wear, has affected or may affect the safe operation of the wind turbine. Characteristically, the vibration signal is abnormally severe, and obvious abnormal features appear in the waveform and 2D images, with the classification probability concentrated in the fault category. Operation must be stopped immediately for repair or replacement to ensure the safety of the wind turbine and personnel.

[0101] In some specific implementations, the first classification result is a four-dimensional probability distribution among healthy, sub-healthy, mildly disabled, and severely disabled individuals. . Corresponding to the probability of being healthy, Corresponding to the probability of sub-health, The probability of a minor fault. The probability of a corresponding failure. Specifically, each element in the four-dimensional probability distribution represents the confidence level that the blade belongs to the corresponding category, and the sum of the probabilities is 1.

[0102] The second classification result is a four-dimensional probability distribution among healthy, sub-healthy, mildly disabled, and disabled individuals. , Corresponding to the probability of being healthy, Corresponding to the probability of sub-health, The probability of a minor fault. The probability of a corresponding failure. Specifically, each element in the four-dimensional probability distribution represents the confidence level that the blade belongs to the corresponding category, and the sum of the probabilities is 1.

[0103] In one specific implementation, step S160 involves fusing the first classification result and the second classification result based on a weighted voting strategy, specifically as follows:

[0104] The four categories are, in order: healthy, sub-healthy, minor malfunction, and malfunction. The final probability is obtained by performing probability fusion, and the calculation formula is as follows: ;

[0105] In the formula, It is a category The final probability, The second recognition model predicts the category. The probability, The first recognition model predicts the category as The probability, To assign weight values ​​to the second recognition model, To assign weight values ​​to the first recognition model, and satisfy... .

[0106] By using probabilistic weighting in this invention, the confidence level of each category can be quantified, and the contribution ratio of different models can be adjusted by weighting, thereby improving the flexibility and robustness of model fusion.

[0107] like Figure 4 As shown, the first recognition model is a ResNet18 model based on transfer learning. Specifically, ResNet18 is a shallower model in the ResNet series, containing 18 layers. Due to the presence of residual connections, ResNet can effectively train very deep networks while maintaining good performance.

[0108] In one implementation, the input image size of the ResNet18 model is 224×224. Layer 1 begins with a 7×7 convolutional layer with a stride of 2, followed by a max-pooling layer with a 3×3 pooling window and a stride of 2. This layer is typically used to reduce the size of the feature maps while extracting preliminary features. Residual Blocks: Each residual block consists of two 3×3 convolutional layers, each followed by batch normalization and a ReLU activation function. At the beginning of each residual block, there is a 1×1 convolutional layer used to adjust the number of channels to match the dimensions of the input and output, implementing residual connections. Residual connections are the core idea of ​​ResNet; they allow the network to learn residual mappings instead of directly learning unmapped features, which helps alleviate the vanishing gradient problem, allowing the network to be trained deeper.

[0109] In terms of network structure, Layer 1 contains two residual blocks. Layer 2 contains two residual blocks, with the number of input channels in each block adjusted using a 1×1 convolution, increasing the number of output channels. Layer 3 contains two residual blocks, again adjusting the number of input channels using a 1×1 convolution, further increasing the number of output channels. Layer 4 contains two residual blocks, with the number of output channels increasing again.

[0110] Specifically, the Global Average Pooling layer: After the last residual block, a global average pooling layer is used to compress the spatial dimension of the feature map to 1×1, which helps reduce the number of parameters and prevent overfitting. The fully connected layer (FC): Finally, a fully connected layer is used to map the pooled feature map to the final output class.

[0111] like Figure 5 As shown, the second recognition model is a VGG16 model based on transfer learning. Specifically, in transfer learning, the convolutional layers of the VGG16 model can be fine-tuned based on pre-training to adapt to new datasets or tasks. This method can significantly reduce training time and improve model performance, as the pre-trained convolutional layers have already learned general feature representations.

[0112] The input image size for the input layer of the VGG16 model is 224×224. The convolutional layers consist of Stages 1 through 6. Stage 2 contains two 3x3 convolutional layers, typically followed by a max-pooling layer (not shown in the diagram). Stage 3 contains three 3x3 convolutional layers, followed by a max-pooling layer. Stage 4 contains three 3x3 convolutional layers, followed by a max-pooling layer. Stage 5 contains three 3x3 convolutional layers, followed by a max-pooling layer. The fully connected layers (Stage 6) consist of three fully connected (FC) layers, used to map the features extracted by the convolutional layers to the final output class.

[0113] refer to Figure 6 The present invention also provides a training method for the first recognition model and the second recognition model. Specifically, it includes the following steps:

[0114] I. Data Acquisition and Merging

[0115] In this embodiment, the dataset used is derived from Zenodo's publicly available blade crack experimental dataset. This dataset was obtained by inputting white noise and sinusoidal signals under different room temperatures and crack lengths, and monitoring them using accelerometers, strain gauges, and temperature and humidity sensors arranged on the blade surface.

[0116] For example, sinusoidal signal experimental data under room temperature (35℃) conditions were selected and categorized into four types based on the blade crack condition:

[0117] (1) Health status: No cracks in the leaves;

[0118] (2) Sub-healthy state: The leaf has a 5 cm crack;

[0119] (3) Mild fault condition: The blade has three 5 cm cracks;

[0120] (4) Fault status: There are three 15 cm cracks on the blade.

[0121] Specifically, the original data was stored in LVM format, and then converted to CSV format for easier computation. Subsequently, the four data categories—healthy, sub-healthy, minor malfunctions, and malfunctions—were merged into a unified dataset, and one-hot encoding was applied to each category. The merged dataset was then divided into training and testing sets in an 8:2 ratio, with the training set used for model learning and the testing set used for performance validation.

[0122] II. Feature Filtering

[0123] Initial model training: Random Forest is used as the baseline classification model to train on the merged dataset to obtain an initial classifier.

[0124] Feature importance calculation: The SHAP (SHapley Additive exPlanations) method is used for each sample. The features are interpreted, and the SHAP value of each feature is calculated. This method obtains the SHAP value by recursively calculating the feature path weights in the tree nodes, with a time complexity of O(n log n). Where T is the number of trees, L is the number of leaf nodes, and D is the tree depth.

[0125] Global Feature Importance Summary: The SHAP values ​​of all samples are summarized and sorted in descending order of feature importance score.

[0126] Feature selection: Select the top 4 features as the main input features, including: displacement signals from strain gauges closest to the three cracks; amplitude signals from the blade tip accelerometer.

[0127] III. Data Transformation

[0128] Time slicing: The original one-dimensional vibration sequence is divided into non-overlapping sliding time windows. The sampling interval is 0.6ms, and the data is divided into time segments of 1s in length to obtain a standardized signal segment sequence.

[0129] Two-dimensional feature space construction: Arrange the multidimensional sampled values ​​in each time segment into matrix rows / columns according to the variables to construct a two-dimensional feature space.

[0130] For image processing, the specific operation process is described in steps 120 and 130 above.

[0131] IV. Model Training

[0132] In this embodiment, ResNet18 and VGG16 are retrained using transfer learning with a pre-training and fine-tuning approach. The pre-trained weight files are obtained by training these two models on the ImageNet dataset (containing over 14 million images across 1000 categories). Fine-tuning the model structure ensures high accuracy with a small dataset without compromising the pre-trained features.

[0133] (1) Transfer learning framework: ResNet18 and VGG16 deep convolutional neural networks are used as recognition models. Pre-trained weights based on the ImageNet dataset are loaded and transferred learning and fine-tuning are performed.

[0134] (2) Data set partitioning: The visualized data is divided into training set, validation set and test set in an 8:1:1 ratio.

[0135] (3) Model adjustment: The input was uniformly scaled to 224×224; all convolutional layers were frozen, and only the output dimension of the fully connected layers was modified to 4 to correspond to the four categories of leaf health. When the validation set loss did not converge, the learning rate was reduced and training continued; when the validation set loss converged, more layers were gradually unfrozen (ResNet18 unfrozen Layer4, VGG16 unfrozen stage5), and fine-tuning was performed using learning rate decay.

[0136] V. Model Saving

[0137] After the model evaluation metrics are met, the trained model is saved as an .h5 file for later use and deployment. Model evaluation metrics include:

[0138] Accuracy: The degree to which the predicted result matches the actual result;

[0139] Precision: The proportion of predictions that are actually positive out of those predictions.

[0140] Recall: The proportion of all positive classes that are correctly predicted;

[0141] F1-score: The harmonic mean of precision and recall.

[0142] This invention also provides a leaf health detection system based on transfer learning. The leaf health detection system implements the above-mentioned leaf health detection method based on transfer learning in hardware and / or software. The leaf health detection system includes:

[0143] The acquisition module is used to acquire real-time vibration data of the wind turbine blades;

[0144] The first conversion module is used to convert the real-time vibration data into a time-series waveform.

[0145] The second conversion module is used to convert the real-time vibration data into a two-dimensional pixel image using the Gram angle field algorithm;

[0146] A first recognition model is used to recognize the time-series waveform and output a first classification result. The first recognition model is a residual neural network architecture.

[0147] The second recognition model is used to recognize the two-dimensional pixel image and output a second classification result. The second recognition model is a deep convolutional neural network architecture.

[0148] The classification module is used to fuse the first classification result and the second classification result based on a weighted voting strategy to obtain the final classification result of the wind turbine blade health.

[0149] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, it will be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A leaf health detection method based on transfer learning, characterized in that, The leaf health detection method includes the following steps: Acquire real-time vibration data of wind turbine blades; The real-time vibration data is converted into a time-series waveform. The Gram angle field algorithm is used to convert the real-time vibration data into a two-dimensional pixel image; The time-series waveform is identified according to a pre-built first identification model, and a first classification result is output. The first identification model is a residual neural network architecture. The two-dimensional pixel image is identified according to a pre-built second recognition model, and a second classification result is output. The second recognition model is a deep convolutional neural network architecture. The first and second classification results are fused using a weighted voting strategy to obtain the final classification result of the wind turbine blade health.

2. The leaf health detection method based on transfer learning as described in claim 1, characterized in that: The first recognition model is a ResNet18 model based on transfer learning; and / or, The second recognition model is the VGG16 model based on transfer learning.

3. The leaf health detection method based on transfer learning as described in claim 1, characterized in that: The final classification result is one of the following: healthy, sub-healthy, mildly disabled, and disabled; The first classification result is a four-dimensional probability distribution among healthy, sub-healthy, mildly faulty, and faulty individuals. ; The second classification result is a four-dimensional probability distribution among healthy, sub-healthy, mildly disabled, and disabled individuals. .

4. The leaf health detection method based on transfer learning as described in claim 3, characterized in that, The first classification result and the second classification result are fused based on a weighted voting strategy, specifically as follows: The four categories are, in order: healthy, sub-healthy, minor malfunction, and malfunction. The final probability is obtained by performing probability fusion, and the calculation formula is as follows: ; In the formula, It is a category The final probability, The second recognition model predicts the category. The probability, The first recognition model predicts the category as The probability, To assign weight values ​​to the second recognition model, To assign weight values ​​to the first recognition model, and satisfy... .

5. A leaf health detection method based on transfer learning as described in any one of claims 1 to 4, characterized in that: The real-time vibration data is converted into a time-series waveform. Specifically, the sampling time of the real-time vibration data is used as the horizontal axis and the corresponding vibration amplitude is used as the vertical axis. The time-series waveforms are plotted sequentially according to the time order to obtain a time-series waveform that represents the change of vibration amplitude over time.

6. A leaf health detection method based on transfer learning as described in any one of claims 1 to 4, characterized in that, The Gram angle field algorithm is used to convert the real-time vibration data into a two-dimensional pixel image, specifically including: The vibration amplitude of the real-time vibration data is normalized to convert the numerical range of the vibration amplitude into the [0,1] interval; The normalized real-time vibration data is mapped to the angle component θ and the radius component r in the polar coordinate system through a polar coordinate encoding mechanism; wherein, a linear mapping strategy is used to encode the vibration amplitude to the angle dimension θ and to associate the sampling time with the radius component r. Based on polar coordinate encoding, construct the Gram angle and field matrix and the Gram angle difference field matrix respectively; The matrix element values ​​of the Gram angle sum field matrix and the Gram angle difference field matrix are linearly mapped to the 0-255 grayscale range or the RGB color space to generate a two-dimensional pixel image with temporal characteristics.

7. The leaf health detection method based on transfer learning as described in claim 6, characterized in that: The mathematical functions of the Gram angle and the field matrix are: ; The mathematical function of the Gram angular difference field matrix is: ; In the above function, For the th in the Gram matrix Line number The matrix elements in the column represent time points. Between The angular relationship.

8. The leaf health detection method based on transfer learning as described in claim 1, characterized in that, The leaf health detection method includes the following steps: The real-time vibration data is preprocessed to remove abnormal and blank data.

9. A leaf health detection system based on transfer learning, characterized in that, The blade health detection system includes: The acquisition module is used to acquire real-time vibration data of the wind turbine blades; The first conversion module is used to convert the real-time vibration data into a time-series waveform. The second conversion module is used to convert the real-time vibration data into a two-dimensional pixel image using the Gram angle field algorithm; A first recognition model is used to recognize the time-series waveform and output a first classification result. The first recognition model is a residual neural network architecture. The second recognition model is used to recognize the two-dimensional pixel image and output a second classification result. The second recognition model is a deep convolutional neural network architecture. The classification module is used to fuse the first classification result and the second classification result based on a weighted voting strategy to obtain the final classification result of the wind turbine blade health.

10. The leaf health detection system based on transfer learning as described in claim 9, characterized in that: The first recognition model is a ResNet18 model based on transfer learning; and / or, The second recognition model is the VGG16 model based on transfer learning.

Images