Offshore wind turbine fault identification method and system, storage medium and computer device

Abstract

The application discloses a kind of offshore wind turbine fault identification method and system, storage medium, computer equipment, the method includes: obtaining the historical vibration time series signal of offshore wind turbine gearbox multiple operating states and pre-processing training sample, on the basis of traditional one-dimensional convolutional neural network, integration multi-scale convolution module, SE attention mechanism and residual network structure, construct new network model.Training time, model receives sample by input layer, middle layer extracts feature, weighting, output layer generates probability distribution and outputs state, after reaching standard, it predicts the operating state under the new signal of real-time acquisition.By multi-scale convolution, the features of different time scales are extracted, the adaptive weighting of feature channels is realized by combining the SE attention mechanism module, and the stability of deep network training is enhanced by using the residual structure, thereby realizing the accurate identification of complex vibration signals, effectively improving the diagnostic accuracy and robustness, and providing a scientific basis for the operation and maintenance of offshore wind turbines.

Description

Technical Field

[0001] This application relates to the field of fault detection technology, and in particular to a method and system for identifying faults in offshore wind turbines, a storage medium, and a computer device. Background Technology

[0002] With the continuous growth of global energy demand, increasing attention is being focused on clean and renewable new marine energy sources. Among them, offshore wind turbines are an emerging system for developing marine energy, characterized by their flexible design and construction of the main wind power generation devices according to specific mission requirements.

[0003] Among existing wind turbine fault monitoring methods, vibration monitoring has the advantages of accurate fault location, simple testing, high testing efficiency, and good real-time performance. However, due to the non-stationary and nonlinear characteristics of vibration signals, their frequency components are complex and easily affected by noise and changes in operating conditions, resulting in low accuracy in fault diagnosis. Summary of the Invention

[0004] In view of this, this application provides a method and system for identifying offshore wind turbine faults, a storage medium, and a computer device. To address the problems of insufficient feature extraction capability and limited classification accuracy of traditional fault diagnosis models, a one-dimensional convolutional neural network (1D-CNN) integrating multi-scale convolution, SE channel attention mechanism (squeeze-and-excitation module), and residual structure (Residual Block) is designed, namely the novel network model (MS-SE-Res1DCNN), which is used to achieve efficient feature extraction and intelligent identification from the vibration signals of offshore wind turbines.

[0005] According to one aspect of this application, a method for identifying faults in offshore wind turbines is provided, the method comprising: Historical vibration time-series signals of offshore wind turbine gearboxes under various operating conditions are acquired. The historical vibration time-series signals are preprocessed to obtain training samples. The operating conditions include healthy conditions and various fault modes. Based on the traditional one-dimensional convolutional neural network, a new network model is constructed by integrating multi-scale convolutional modules, SE attention mechanism and residual network structure. The new network model includes an input layer, an intermediate layer and an output layer. The intermediate layer corresponds to multiple deep convolutional blocks connected by residual networks. Each deep convolutional block deploys a multi-scale convolutional module and an SE attention mechanism. The multi-scale convolutional module contains convolutional kernels of various sizes. When training a novel network model using training samples, the novel network model receives training samples through the input layer and converts the training samples into tensors that are fed into the first deep convolutional block of the intermediate layer. Inside the first deep convolutional block, a multi-scale convolutional module uses convolutional kernels of various sizes to simultaneously extract features from different frequency domains in the tensor corresponding to the same training sample. The extracted frequency domain features are then concatenated along the channel dimension to generate frequency domain fusion features. The SE attention mechanism adaptively learns the correlation between channels to generate weight vectors. These weight vectors are then used to weight the frequency domain fusion features to obtain initial key features. These initial key features are then passed between deep convolutional blocks via a residual network until the last deep convolutional block outputs the final key features to the output layer. By using the fully connected layer to map key features in the output layer, a probability distribution for various operating states is generated, and based on the probability distribution, the predicted operating state of the offshore wind turbine gearbox is output. When the predicted results of the operating status and the actual operating status in the training samples are used to determine that the new network model has reached the training standard, new vibration time-series signals of the offshore wind turbine gearbox are collected in real time so that the new network model that has reached the training standard can predict the operating status of the offshore wind turbine gearbox based on the new vibration time-series signals collected in real time.

[0006] According to another aspect of this application, an offshore wind turbine fault identification system is provided, the system comprising: The training sample construction module is used to acquire historical vibration time-series signals of offshore wind turbine gearboxes under various operating conditions, preprocess the historical vibration time-series signals to obtain training samples, wherein the operating conditions include healthy states and various fault modes. A novel model building and training module is used to integrate multi-scale convolutional modules, SE attention mechanisms, and residual network structures on the basis of traditional one-dimensional convolutional neural networks to construct a novel network model. This novel network model includes an input layer, intermediate layers, and an output layer. The intermediate layers correspond to multiple deep convolutional blocks connected by residual networks. Each deep convolutional block deploys multi-scale convolutional modules and SE attention mechanisms, with convolutional kernels of various sizes within each module. When training the novel network model using training samples, the model receives training samples through the input layer and converts them into tensors, which are then fed to the first deep convolutional block of the intermediate layer. Within this first deep convolutional block, multi-scale convolutions... The module utilizes convolutional kernels of various sizes to simultaneously extract features from different frequency domains in the tensors corresponding to the same training sample. It then concatenates these extracted frequency domain features along the channel dimension to generate frequency domain fusion features. Furthermore, it adaptively learns the correlation between channels through the SE attention mechanism to generate weight vectors. These weight vectors are used to weight the frequency domain fusion features, yielding initial key features. These initial key features are then passed through a residual network between deep convolutional blocks until the final deep convolutional block outputs the final key features to the output layer. The output layer uses fully connected layers to map these key features, generating probability distributions for various operating states. Based on these probability distributions, it outputs predictions of the operating state of the offshore wind turbine gearbox. The offshore wind turbine fault identification module is used to collect new vibration time-series signals of the offshore wind turbine gearbox in real time when the new network model reaches the training standard by using the prediction results of the operating status and the actual operating status in the training samples. This enables the new network model that has reached the training standard to predict the operating status of the offshore wind turbine gearbox based on the new vibration time-series signals collected in real time.

[0007] According to another aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described offshore wind turbine fault identification method.

[0008] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described offshore wind turbine fault identification method.

[0009] By employing the above technical solution, this application provides a method and system for identifying offshore wind turbine faults, a storage medium, and a computer device. First, the collected raw vibration signals are processed to obtain a fixed-length training sample input. Then, fault features at different time scales are captured through an initial (first) deep convolutional layer and multi-scale convolutional modules. The SE attention mechanism is used to adaptively weight key channel features, strengthening useful features and suppressing redundant information. Simultaneously, a residual connection structure is combined to ensure effective training of the deep network and avoid gradient vanishing. Finally, a fully connected layer is used to achieve accurate classification of multiple fault states. This effectively improves the stability and accuracy of offshore wind turbine fault diagnosis and possesses strong generalization ability.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a method for identifying faults in offshore wind turbines according to an embodiment of this application is shown. Figure 2 This illustration shows a schematic diagram of an SE attention mechanism structure provided in an embodiment of this application; Figure 3 A flowchart illustrating another offshore wind turbine fault identification method provided in an embodiment of this application is shown; Figure 4 This illustration shows a schematic diagram of an accuracy curve provided in an embodiment of this application; Figure 5 This illustration shows a schematic diagram of a loss curve provided in an embodiment of this application; Figure 6 This application provides a fault diagnosis result confusion matrix according to an embodiment of the present application. Figure 7 A schematic diagram of the structure of an offshore wind turbine fault identification system provided in an embodiment of this application is shown. Detailed Implementation

[0012] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0013] This embodiment provides a method for identifying faults in offshore wind turbines, such as... Figure 1 As shown, the method includes: Step 101: Obtain historical vibration time-series signals of offshore wind turbine gearbox under various operating conditions, preprocess the historical vibration time-series signals to obtain training samples, wherein the operating conditions include healthy state and various fault modes.

[0014] In the embodiments described above, a vibration acceleration sensor can be used to collect historical vibration time-series signals under different operating conditions and store them as digital signals for analysis and processing. Furthermore, the effectiveness of the method can be verified using the gearbox fault dataset publicly available from Southeast University. This dataset contains two sub-datasets: bearing data and gear data, both of which were acquired using a transmission system dynamics simulator (DDS).

[0015] Optionally, the fault modes include bearing roller faults, bearing inner ring faults, bearing outer ring faults, bearing compound faults, gear defects, gear tooth breakage, tooth root cracks, and tooth surface wear. In step 101, the historical vibration time series signal is preprocessed to obtain training samples, including: Step 1011: The historical vibration time series signal is segmented using the fixed-length sliding window method to obtain multiple signal samples, wherein each signal sample contains a complete fault characteristic cycle. Step 1012: Standardize the signal samples to obtain training samples, wherein each standardized signal sample corresponds to one training sample.

[0016] In the embodiments described above, the collected signals undergo data cleaning and normalization to ensure comparability of signals under different operating conditions at the same scale. For example, a fixed-length segmentation method can be used, with 1024 sampling points as a sample window, dividing the long time-series signal into multiple training samples to ensure that each sample contains complete periodic feature information. Furthermore, training, validation, and test sets can be divided according to temporal continuity to ensure reasonable data distribution.

[0017] Step 102: Based on the traditional one-dimensional convolutional neural network, a new network model is constructed by integrating a multi-scale convolutional module, an SE attention mechanism, and a residual network structure. The new network model includes an input layer, an intermediate layer, and an output layer. The intermediate layer corresponds to multiple deep convolutional blocks connected by a residual network. Each deep convolutional block is equipped with a multi-scale convolutional module and an SE attention mechanism. The multi-scale convolutional module contains convolutional kernels of various sizes.

[0018] Next, based on the traditional one-dimensional convolutional neural network (1D-CNN), the following improved structure is introduced to form a new network model, namely the MS-SE-Res1DCNN model.

[0019] Multi-scale convolution module: Different convolution kernels such as 3, 5, and 7 are used in parallel to extract features in the same layer. Multi-band features are fused by channel splicing to enhance the model’s sensitivity to features of different fault modes.

[0020] SE attention mechanism: A one-dimensional Squeeze-and-Excitation module (SE attention mechanism) is introduced after the convolution output to adaptively adjust the weights of each channel using global information, thereby strengthening key features and suppressing redundant information.

[0021] Residual network connection structure (residual network structure): Residual shortcuts are used to connect deep convolutional blocks to ensure that deep features can be effectively transmitted, alleviate the gradient vanishing problem, and improve training convergence speed and model stability.

[0022] Step 103: When training a new network model using training samples, the new network model receives training samples through the input layer and converts the training samples into tensors to feed into the first deep convolutional block of the intermediate layer. Step 104: Inside the first deep convolutional block, the multi-scale convolutional module uses convolutional kernels of various sizes to simultaneously extract features in different frequency domains of the tensor corresponding to the same training sample. The extracted features in each frequency domain are then concatenated along the channel dimension to generate frequency domain fusion features. The correlation between each channel is adaptively learned through the SE attention mechanism to generate a weight vector. The frequency domain fusion features are then weighted using the weight vector to obtain initial key features. These initial key features are then passed between deep convolutional blocks via the residual network until the last deep convolutional block outputs the final key features to the output layer. Step 105: The key features are mapped by the fully connected layer through the output layer to generate probability distributions for various operating states, and the predicted operating state of the offshore wind turbine gearbox is output based on the probability distributions.

[0023] Next, taking the specific failure mode of "tooth root crack" in offshore wind turbine gearboxes as an example, the input layer converts the vibration acceleration sequence (historical vibration time-series signal, such as a sampling point of length 1024) collected by sensors into tensors. In the first convolutional block, three types of convolutional kernels, 3×1, 5×1, and 7×1, work simultaneously: the short kernel (3×1) keenly captures the high-frequency subtle impacts on the tooth surface, while the long kernel (7×1) identifies low-frequency periodic vibration fluctuations. These multi-dimensional features are spliced ​​together in the channel dimension to form a "frequency domain fusion feature" containing information in the entire frequency domain.

[0024] Because tooth root cracks have the strongest response at specific frequencies, the SE attention mechanism automatically identifies them, i.e., it adaptively learns channel correlations, assigning high weights (e.g., 0.9) to channels representing crack features and low weights (e.g., 0.1) to background noise channels. The weighted "initial key features" are then reinforced layer by layer by the residual network to filter out interference. Finally, the output layer maps deep features to probabilities, for example: healthy state 0.02, broken gear tooth 0.05, tooth root crack 0.93. Based on this, the model determines that the gearbox is in a cracked operating state.

[0025] Optionally, the size of the convolution kernel includes 3×1, 5×1, and 7×1. In step 104, the multi-scale convolution module utilizes convolution kernels of various sizes to simultaneously extract features from different frequency domains in the tensor corresponding to the same training sample, and concatenates the extracted frequency domain features along the channel dimension to generate frequency domain fusion features, including: Step 1041: The tensors corresponding to the same training sample are synchronously distributed to parallel convolution branches with convolution kernels of sizes of 3×1, 5×1, and 7×1 respectively through the multi-scale convolution module. Step 1042: For the high-frequency region of the tensor corresponding to the training sample, extract the short-term transient impact features of the offshore wind turbine gearbox caused by tooth surface wear and bearing roller failure through the convolution branch of the convolution kernel with a size of 3×1. Step 1043: For the mid-frequency region of the tensor corresponding to the training samples, extract the medium-periodic vibration characteristics of the offshore wind turbine gearbox caused by gear defects, broken teeth and tooth root cracks through the convolution branch of the convolution kernel with a size of 5×1. Step 1044: For the low-frequency region of the tensor corresponding to the training samples, extract the long-term periodic fluctuation characteristics of the offshore wind turbine gearbox caused by bearing inner ring failure, bearing outer ring failure and bearing compound failure through the convolution branch of the convolution kernel with a size of 7×1. Step 1045: The features output by the three convolutional branches are spliced ​​together in the channel dimension using a multi-scale convolution module to generate frequency domain fusion features.

[0026] In the above embodiments of this application, the main architecture of the convolutional neural network consists of five parts: input layer, convolutional layer, pooling layer, fully connected layer and output layer. The convolutional layer, pooling layer and fully connected layer can also be classified as intermediate layers.

[0027] Convolutional layers and pooling layers work together to form the key components responsible for feature extraction in the network. Multi-scale convolution, as an important optimization of convolutional layers, further improves the comprehensiveness and accuracy of feature extraction. Specifically: During model operation, various defect images (or vibration signals obtained through preprocessing) are first input to the front-end input layer of the model, and then enter the deep convolutional block. Here, multi-scale convolution operation is introduced. By using convolution kernels of different sizes (such as 3×1, 5×1, 7×1, etc.) to extract features in parallel, small-sized convolution kernels capture local detail features, while medium and large-sized convolution kernels capture global trends and overall features. Then, features of different scales are fused to make up for the shortcomings of single-scale convolution in capturing defect features of different sizes.

[0028] After alternating processing of multiple layers and scales of convolution and pooling operations, key features are effectively preserved and data dimensionality is reduced, ultimately yielding high-level abstract features that can accurately represent defect information, namely frequency domain fusion features, which provide reliable support for classification and recognition in subsequent fully connected layers.

[0029] Specifically, the input layer is used to receive raw data or pre-processed data, which can be different types of data such as images, audio, and waveforms.

[0030] Furthermore, the convolutional layer, as the core structure of a CNN, is responsible for performing operations on the input data through convolutional kernels, mapping the raw data to a higher-dimensional feature space, thereby achieving effective feature extraction. The size and number of convolutional kernels are important hyperparameters of this layer, directly affecting the accuracy of feature extraction and the model complexity. The design of the convolutional layer directly affects the feature extraction capability and computational efficiency. The formula for convolution operation is as follows: , In the formula: For activation functions; For input; Indicates the output; For convolution kernel; For bias terms; This is a convolution operation.

[0031] Pooling layers are typically placed after convolutional layers and primarily compress feature maps through downsampling operations, such as max pooling or average pooling. This layer mainly aims to reduce the number of parameters and computational complexity. Max pooling preserves salient features and suppresses noise by calculating the maximum value within a region; average pooling smooths features by calculating the mean value within a region. Max pooling is more commonly used in practice because it preserves the most salient features, and its formula is as follows: , In the formula: For output; For activation functions; The multiplication weights of the feature maps are typically set to 1. A function that takes the maximum value; For bias.

[0032] Optionally, in step 104, the SE attention mechanism corresponds to a first fully connected layer and a second fully connected layer. The SE attention mechanism adaptively learns the correlation between channels to generate a weight vector. This weight vector is then used to weight the frequency domain fusion features to obtain initial key features, including: Step 1046: The two-dimensional features of each channel in the frequency domain fusion features are mapped and compressed into a single global statistical value through the SE attention mechanism, forming a one-dimensional feature vector that covers the global feature response of each channel. Step 1047: The one-dimensional feature vector is reduced in dimensionality by the first fully connected layer in the SE attention mechanism, and then nonlinearly transformed by the ReLU activation function to obtain a compressed feature vector containing nonlinear interaction information of each channel. Step 1048: The compressed feature vector is up-dimensioned through the second fully connected layer in the SE attention mechanism, and the feature values ​​are mapped to the 0 to 1 interval by the Sigmoid activation function to generate an adaptive weight vector. Step 1049: The generated weight vector and the frequency domain fusion feature are multiplied channel by channel using the SE attention mechanism to generate the initial key features.

[0033] In the embodiments described above, the attention mechanism (AM) in the human visual system enables observers to focus their perceptual resources on regions of interest. In recent years, similar attention mechanisms have been introduced into computer vision tasks. Numerous studies have confirmed that integrating attention mechanism modules into convolutional neural networks can effectively enhance the extraction of target features from specific regions of an image, thereby significantly improving the overall performance of the network model.

[0034] The Squeeze-and-Excitation block (SEblock) is a channel-domain-based attention mechanism, with a typical structure as follows: Figure 2 As shown, SEblock (SE attention mechanism) mainly consists of two key operations: compression and activation. Through compression, the network generates global statistics reflecting the importance of each feature channel. Subsequently, in the activation phase, the attention weights of each channel are optimized using a loss function, thereby amplifying effective feature channels and suppressing inefficient or redundant feature channels. This mechanism not only improves the model's ability to express key features but also significantly enhances the accuracy in image classification tasks.

[0035] Specifically, for example, for a height, width, and number of channels respectively... , andC Input feature map First, it is compressed using the Global Average Pooling (GAP) algorithm. Transform into The output feature map is given by the formula: , in, This represents the output feature map. This indicates the coordinate position on the feature map.

[0036] The activation operation utilizes a dual fully connected layer network structure to achieve the fusion and interaction of information from each channel in the feature map, thereby reducing the computational cost of the model. The dimension is reduced to 1 / r of the original. The original dimensions are restored, and the channel weights are subjected to a non-linear transformation using the sigmoid function to finally obtain the attention weight s for each channel, as shown in the formula: .

[0037] in, Represents the ReLU activation function. This represents the Sigmoid activation function.

[0038] After obtaining the adjustment parameters between channels, these adjustment parameters are then multiplied element-wise with the original feature map to complete the recalibration of the feature map and the adaptive adjustment of the feature weights. The specific calculation formula is as follows: , in, This represents the adjusted output feature map.

[0039] Specifically, to further enhance the training stability and feature extraction capabilities of deep networks, the model introduces a one-dimensional residual module to construct a residual network structure. This module uses multi-scale convolution as the core feature extraction unit, combined with the SE channel attention mechanism and identity shortcut connections to form an integrated structure. The residual module (corresponding to the residual network structure) directly passes the input feature map (the tensor corresponding to the input training samples) across layers to the output, and performs element-wise weighted fusion with the features extracted by multi-scale convolution to construct a residual learning path. This effectively solves the gradient vanishing and gradient exploding problems that easily occur during the training of deep convolutional neural networks, ensuring the complete transmission of the original effective defect features. In the feature forward propagation, the input features first extract fault features of different granularities in parallel through multi-scale convolution and complete channel fusion. Then, the SE attention mechanism adaptively adjusts the weights of the feature channels, strengthening the fault-sensitive feature channels and suppressing invalid and redundant features. Finally, the attention-optimized features are residually superimposed with the original input features of the shortcut connections, and a nonlinear mapping is completed through an activation function. The model achieves stable training of deep networks and refined extraction of high-order fault features by stacking multiple layers of residual modules, significantly improving the network's ability to represent and identify complex bearing fault modes. , In the formula: For residual mapping, To facilitate quick connection mapping, either an identity mapping or a 1×1 convolution mapping is selected based on the matching degree of the input and output channels. This is the output.

[0040] Optionally, in step 105, key features are mapped using the fully connected layer through the output layer to generate probability distributions for various operating states. Based on the probability distributions, the predicted operating state of the offshore wind turbine gearbox is output, including: Step 1051: The output layer uses a fully connected layer to perform nonlinear mapping on the final key features to obtain a feature vector that matches the number of all operating state categories of the offshore wind turbine gearbox. Step 1052: The output layer uses an activation function to map the values ​​of the feature vectors to the interval between 0 and 1, generating a probability distribution for the running state category. In the probability distribution, each probability value corresponds to the confidence level of a running state category. Step 1053: Compare the probability values ​​corresponding to each operating state in the probability distribution, and select the operating state category with the highest probability value as the prediction result of the operating state of the offshore wind turbine gearbox.

[0041] In the embodiments described above, the fully connected layer is generally located after the convolutional and pooling layers. Its main function is to globally integrate and non-linearly map the high-level features extracted earlier. This layer expands the multi-dimensional features into a one-dimensional vector, and through a combination of weighted summation and activation functions, it achieves further fusion and expression of the features, thereby extracting key information with class discrimination capabilities. Its expression is as follows: , In the formula: For the first l Layer output; For the first l Layer weights; For bias.

[0042] As the end of the network structure, it is used to map the features output by the fully connected layer to the final prediction result. In multi-class classification tasks, the Softmax function is often used to normalize the output, outputting probability values ​​with the sum of the probabilities of each class being 1, thus obtaining the probability that the input belongs to each class.

[0043] Step 106: When the predicted results of the operating status and the actual operating status in the training samples are used to determine that the new network model has reached the training standard, new vibration time-series signals of the offshore wind turbine gearbox are collected in real time so that the new network model that has reached the training standard can predict the operating status of the offshore wind turbine gearbox based on the new vibration time-series signals collected in real time.

[0044] Next, the optimal trained model can be deployed in the online monitoring and diagnostic system of offshore wind turbines, combining it with signals collected in real time by vibration sensors for fault identification. The system can automatically determine the health status and various fault types of offshore wind turbines, and promptly identify and issue early warnings when potential faults are detected, thereby improving the intelligence and reliability of equipment operation and maintenance.

[0045] Optionally, in step 106, the predicted running state and the actual running state in the training samples are used to determine whether the new network model has met the training criteria, including: Step 1061: Using precision, recall, F1 score and overall accuracy as evaluation metrics, calculate each evaluation metric based on the prediction results of the running state output by the new network model for the training samples and the actual running state of the original records in the training samples. Step 1062: When each evaluation indicator reaches its respective preset threshold, the new network model is judged to have met the training standard.

[0046] In the embodiments described above, when training samples are input into the MS-SE-Res1DCNN model, cross-entropy can be used as the loss function, the Adam optimizer can be used for parameter updates, and an adaptive learning rate scheduling strategy (ReduceLROnPlateau) can be introduced to dynamically adjust the learning rate to avoid getting trapped in local optima. During training, the generalization ability of the model is monitored through the performance of the validation set, and the optimal model parameters are saved when the validation accuracy is highest. The test set is input into the trained model, and the classification prediction results are output. The model performance is evaluated from multiple dimensions using the confusion matrix, classification report (precision, recall, F1 score), and overall accuracy, and the predicted probability distribution is saved for further analysis of model interpretability. At the same time, the loss curve and accuracy curve of the training process are plotted to intuitively show the convergence effect and stability of the model.

[0047] Furthermore, the table below shows the performance of each class of training samples in identifying different fault types using the MS-SE-Res1DCNN model, covering nine states: healthy working state, ballfault, inner ring fault, outer ring fault, combination fault on both inner and outer rings, chipped tooth, missing tooth, root fault, and surface fault. The Precision, Recall, and F1-score results for each class are shown in Table 1, and the classification results of the MS-SE-Res1DCNN model on the test set are shown in Table 2.

[0048] Table 1

[0049] Table 2

[0050] As shown in Tables 1 and 2, the MS-SE-Res1DCNN model achieves an overall accuracy of 97.82% on the test set, with Macro-Precision, Macro-Recall, and Macro-F1 scores all approaching 98%, indicating that the model is balanced and reliable across various fault identification types. Specifically, the F1 scores for most categories (such as Ball, Health, Chipped, Comb, Inner, and Outer) exceed 97%, with the Comb class achieving a perfect score of 1.0, demonstrating the model's extremely accurate recognition of these features. However, the precision for the Surface class (89.69%) is relatively low, indicating some false positives. The recall rates for the Miss and Root classes (92.16% and 94.12%, respectively) are slightly lower, suggesting some overlap between these fault types and other categories. Overall, this model effectively improves the accuracy and robustness of bearing fault diagnosis through the combination of multi-scale convolution, SE attention, and residual structures.

[0051] Specifically, the following parameters are included: label, fault type, precision, recall, F-score (also known as F1 score), number of test samples, and specific fault type. Ball: ball bearing; Health: healthy; Chipped: chipped; Comb: composite; Inner: inner ring; Miss: missing; Outer: outer ring; Root: tooth root; Surface: surface; Table 2: Accuracy (%): Accuracy (%) Macro-Precision (%): Macro average precision (%) Macro-Recall (%): Macro average recall rate (%); Macro-F1(%): Macro average F1 value (%).

[0052] Furthermore, the accuracy curve is as follows: Figure 4 As shown, Figure 4 In the middle, Accuracy History: the history of accuracy, the horizontal and vertical axis labels are: Epoch: round (in machine learning, it refers to the process of performing a complete forward and backward propagation calculation on the entire training dataset), Accuracy: accuracy.

[0053] Training Accuracy: Training accuracy; Validation Accuracy: Validation accuracy Loss curve as shown Figure 5As shown, Figure 5 In this context, Loss History refers to the history (changes / records) of the loss value, Loss refers to the loss value, Training Loss refers to the training loss value, and Validation Loss refers to the validation loss value.

[0054] Fault diagnosis result confusion matrix as follows Figure 6 As shown, Figure 6 In this context, Confusion Matrix: True (true category) and Predicted (predicted category).

[0055] In one specific embodiment, such as Figure 3 As shown, vibration time-series signals of the gearbox under various conditions, including healthy, bearing failure (roller / inner ring / outer ring), and gear failure (missing / broken / cracked / worn), were collected. These signals were preprocessed using denoising techniques such as wavelet transform and normalization to transform them into standardized training samples. The preprocessed signals were then divided into a training set (70%), a validation set (20%), and a test set (10%) to construct a dataset containing multiple failure modes for model training and evaluation.

[0056] Based on a novel network model: Multi-scale convolution module: 3×1 kernels capture high-frequency impacts, such as tooth surface wear; 5×1 kernels extract mid-frequency features; and 7×1 kernels extract low-frequency periodic vibrations, such as broken gear teeth. The three convolution kernels simultaneously extract features in different frequency domains, and the channel dimensions are spliced ​​to generate frequency domain fusion features.

[0057] SE attention mechanism: Adaptively learns channel weights for fused features and obtains initial key features through weighted summation.

[0058] Residual network structure: Key features are passed through residual connections, avoiding gradient vanishing in deep networks and enhancing feature representation.

[0059] The model is trained using the training set, and the parameters are optimized through backpropagation using precision, recall, F1 score, and accuracy as indicators until the indicators are met (e.g., F1 ≥ 0.95).

[0060] Use the test set to verify the model's generalization ability and ensure that it can still accurately identify unknown fault samples (e.g., bearing composite fault identification rate ≥90%).

[0061] Real-time acquisition of gearbox vibration signals, input into a compliance model, outputting a probability distribution of operating status, and realizing real-time fault early warning.

[0062] The entire process, from data preparation to online application, uses multi-scale feature capture and attention weighting to accurately identify complex fault modes and ensure the safe operation of offshore wind turbines.

[0063] By applying the technical solution of this embodiment, based on the residual structure framework, multi-scale convolutional units are first constructed in parallel using three types of convolutional kernels: 3×1, 5×1, and 7×1, to simultaneously extract local details, intermediate waveforms, and global trend features. Then, an SE channel attention mechanism is introduced to adaptively weight the multi-scale fused features, strengthening the sensitive feature channels related to faults. Finally, the input information is directly transmitted to the output through residual shortcut connections, and element-wise superimposed with the attention-optimized features to construct the residual learning process. This structure organically integrates the three advantages of multi-scale feature extraction, attention weighting, and residual stable training. It not only solves the problem of limited receptive field of a single convolutional kernel but also enhances the expressive power of effective fault features, while avoiding gradient vanishing in deep networks. This significantly improves the model's accuracy and robustness in recognizing complex fault modes, and has the following beneficial effects: 1. Introduce a multi-scale convolution module to improve feature extraction capabilities: Traditional 1D-CNNs typically use fixed convolutional kernels to extract features at a single scale, easily overlooking key features at different frequencies and scales. This invention employs a multi-scale convolutional module, utilizing multiple convolutional kernels such as 3, 5, and 7 in parallel to extract features, and then fuses them through channel concatenation to achieve comprehensive capture of multi-frequency features of wind turbine vibration fault signals, improving the model's robustness and recognition accuracy under complex operating conditions.

[0064] 2. Enhance feature weight allocation capability by incorporating SE attention mechanism: Traditional 1D-CNNs use the same feature weights for each channel, making it impossible to distinguish the importance of different features. This invention introduces a one-dimensional SE attention module, which adaptively learns the importance of features in each channel using global information, weights and strengthens key features, suppresses redundant features, and thus improves the model's discriminative ability.

[0065] 3. Optimize deep feature transfer capability using residual network structure: Traditional 1D-CNNs are prone to gradient vanishing or degradation problems as the number of layers increases, leading to unstable training. This invention introduces residual connections on top of multi-scale convolutions and SE modules to ensure that information is not lost during the training of deep networks, effectively alleviating the gradient vanishing problem and improving the convergence speed and stability of the model.

[0066] Furthermore, as Figure 1 In terms of specific implementation, this application provides an offshore wind turbine fault identification system, such as... Figure 4 As shown, the system includes: The training sample construction module 201 is used to acquire historical vibration time-series signals of offshore wind turbine gearboxes under various operating conditions, preprocess the historical vibration time-series signals to obtain training samples, wherein the operating conditions include healthy conditions and various fault modes. The novel model construction and training module 202 is used to integrate multi-scale convolutional modules, SE attention mechanisms, and residual network structures on the basis of traditional one-dimensional convolutional neural networks to construct a novel network model. The novel network model includes an input layer, intermediate layers, and an output layer. The intermediate layers correspond to multiple deep convolutional blocks connected by residual networks. Each deep convolutional block deploys a multi-scale convolutional module and an SE attention mechanism. The multi-scale convolutional module contains convolutional kernels of various sizes. When training the novel network model using training samples, the novel network model receives training samples through the input layer and converts the training samples into tensors, which are then fed to the first deep convolutional block of the intermediate layer. Within the first deep convolutional block, multi-scale convolutional modules are used to... The integration module utilizes convolutional kernels of various sizes to simultaneously extract features from different frequency domains in the tensor corresponding to the same training sample. It then concatenates these extracted frequency domain features along the channel dimension to generate frequency domain fusion features. Furthermore, it adaptively learns the correlation between channels through the SE attention mechanism to generate weight vectors. These weight vectors are used to weight the frequency domain fusion features, yielding initial key features. These initial key features are then passed through the residual network between deep convolutional blocks until the final deep convolutional block outputs the final key features to the output layer. The output layer uses fully connected layers to map these key features, generating probability distributions for various operating states. Based on these probability distributions, it outputs predictions of the operating state of the offshore wind turbine gearbox. The offshore wind turbine fault identification module 203 is used to collect new vibration time-series signals of the offshore wind turbine gearbox in real time when the new network model reaches the training standard by using the prediction results of the operating status and the actual operating status in the training samples. This enables the new network model that has reached the training standard to predict the operating status of the offshore wind turbine gearbox based on the new vibration time-series signals collected in real time.

[0067] It should be noted that other corresponding descriptions of the functional units involved in the offshore wind turbine fault identification system provided in this application embodiment can be found in the following references. Figure 1 The corresponding descriptions in the method will not be repeated here.

[0068] Based on the above, Figure 1 Accordingly, this application also provides a storage medium storing a computer program, which, when executed by a processor, implements the above-described method. Figure 1 The method for identifying faults in offshore wind turbines is shown.

[0069] Based on this understanding, the technical solution of this application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, or portable hard drive), and includes several instructions to cause a computer device (such as a personal computer, server, or network device) to execute the methods described in the various implementation scenarios of this application.

[0070] Based on the above, Figure 1 The method shown, and Figure 7 To achieve the above objectives, the virtual system embodiment shown in this application also provides a computer device, specifically a personal computer, server, network device, etc. This computer device includes a storage medium and a processor; the storage medium stores a computer program; the processor executes the computer program to achieve the above-described objectives. Figure 1 The method for identifying faults in offshore wind turbines is shown.

[0071] Optionally, the computer device may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB ports, card reader ports, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Bluetooth interfaces, Wi-Fi interfaces), etc.

[0072] Those skilled in the art will understand that the computer device structure provided in this embodiment does not constitute a limitation on the computer device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0073] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages and stores the hardware and software resources of a computer device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software within the physical device.

[0074] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware platforms, or it can be implemented in hardware by combining multi-scale convolution, SE attention mechanism and residual network structure in 1DCNN, proposing a fault diagnosis intelligent identification method suitable for offshore wind turbine systems. This method can fully extract key information at different scales and adaptively enhance important feature channels, thereby improving the model's robustness and identification accuracy of vibration signals under complex operating conditions and noise interference. At the same time, by utilizing the improved deep convolutional neural network structure and optimized training strategy, the model's feature extraction capability and classification performance can be dynamically adjusted according to actual data, providing reliable technical support for intelligent monitoring and condition assessment of offshore wind turbines.

[0075] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the system of the embodiment scenario can be distributed throughout the system of the embodiment scenario as described, or they can be modified to reside in one or more systems different from this embodiment scenario. The modules of the above-described embodiment scenario can be combined into one module, or further divided into multiple sub-modules.

[0076] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any modifications that can be made by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for identifying faults in offshore wind turbines, characterized in that, The method includes: Historical vibration time-series signals of offshore wind turbine gearboxes under various operating conditions are acquired. The historical vibration time-series signals are preprocessed to obtain training samples. The operating conditions include healthy conditions and various fault modes. Based on the traditional one-dimensional convolutional neural network, a new network model is constructed by integrating multi-scale convolutional modules, SE attention mechanism and residual network structure. The new network model includes an input layer, an intermediate layer and an output layer. The intermediate layer corresponds to multiple deep convolutional blocks connected by residual networks. Each deep convolutional block deploys a multi-scale convolutional module and an SE attention mechanism. The multi-scale convolutional module contains convolutional kernels of various sizes. When training a novel network model using training samples, the novel network model receives training samples through the input layer and converts the training samples into tensors that are fed into the first deep convolutional block of the intermediate layer. Inside the first deep convolutional block, a multi-scale convolutional module uses convolutional kernels of various sizes to simultaneously extract features from different frequency domains in the tensor corresponding to the same training sample. The extracted frequency domain features are then concatenated along the channel dimension to generate frequency domain fusion features. The SE attention mechanism adaptively learns the correlation between channels to generate weight vectors. These weight vectors are then used to weight the frequency domain fusion features to obtain initial key features. These initial key features are then passed between deep convolutional blocks via a residual network until the last deep convolutional block outputs the final key features to the output layer. By using the fully connected layer to map key features in the output layer, a probability distribution for various operating states is generated, and based on the probability distribution, the predicted operating state of the offshore wind turbine gearbox is output. When the predicted results of the operating status and the actual operating status in the training samples are used to determine that the new network model has reached the training standard, new vibration time-series signals of the offshore wind turbine gearbox are collected in real time so that the new network model that has reached the training standard can predict the operating status of the offshore wind turbine gearbox based on the new vibration time-series signals collected in real time.

2. The method according to claim 1, characterized in that, The convolutional kernel sizes include 3×1, 5×1, and 7×1. A multi-scale convolution module utilizes kernels of various sizes to simultaneously extract features from different frequency domains of the tensor corresponding to the same training sample. The extracted frequency domain features are then concatenated along the channel dimension to generate frequency domain fusion features, including: The tensors corresponding to the same training sample are synchronously distributed to parallel convolution branches with convolution kernels of sizes 3×1, 5×1, and 7×1 respectively through the multi-scale convolution module. For the high-frequency region of the tensor corresponding to the training samples, the short-term transient impact characteristics of offshore wind turbine gearbox caused by tooth surface wear and bearing roller failure are extracted by using a convolution branch with a 3×1 kernel. For the mid-frequency region of the tensor corresponding to the training samples, the mid-periodic vibration characteristics of offshore wind turbine gearbox caused by gear defects, broken teeth and tooth root cracks are extracted by using a convolution branch with a 5×1 kernel. For the low-frequency region of the tensor corresponding to the training samples, the long-term periodic fluctuation characteristics of offshore wind turbine gearbox caused by bearing inner ring failure, bearing outer ring failure and bearing compound failure are extracted by the convolution branch of the convolution kernel with a size of 7×1. The multi-scale convolution module concatenates the features output from the three convolution branches along the channel dimension to generate frequency domain fusion features.

3. The method according to claim 1, characterized in that, The SE attention mechanism corresponds to a first fully connected layer and a second fully connected layer. It adaptively learns the correlation between channels to generate a weight vector. This weight vector is then used to weight the frequency domain fusion features to obtain initial key features, including: The SE attention mechanism maps and compresses the two-dimensional features of each channel in the frequency domain fusion features into a single global statistical value, forming a one-dimensional feature vector that covers the global feature response of each channel. The one-dimensional feature vector is reduced in dimensionality by the first fully connected layer in the SE attention mechanism, and then nonlinearly transformed by the ReLU activation function to obtain a compressed feature vector containing nonlinear interaction information of each channel. The compressed feature vector is up-dimensioned through the second fully connected layer in the SE attention mechanism, and the feature values ​​are mapped to the 0 to 1 range by the Sigmoid activation function to generate an adaptive weight vector. The generated weight vector is multiplied with the frequency domain fusion features through a channel-by-channel weighted multiplication operation using the SE attention mechanism to generate initial key features.

4. The method according to claim 1, characterized in that, By mapping key features through the fully connected layer in the output layer, probability distributions for various operating states are generated, including: By using a fully connected layer to perform nonlinear mapping on the final key features in the output layer, a feature vector matching the number of all operating state categories of the offshore wind turbine gearbox is obtained. The output layer uses an activation function to map the values ​​of the feature vectors to the interval between 0 and 1, generating a probability distribution for each running state category. In this probability distribution, each probability value corresponds to the confidence level of a running state category.

5. The method according to claim 4, characterized in that, Based on the probability distribution, the predicted operating status of the offshore wind turbine gearbox is output, including: By comparing the probability values ​​corresponding to each operating state in the probability distribution, the operating state category with the highest probability value is selected as the prediction result of the operating state of the offshore wind turbine gearbox.

6. The method according to claim 1, characterized in that, Using the predicted running states and the actual running states in the training samples, we determine whether the novel network model has met the training criteria, including: Precision, recall, F1 score and overall accuracy are used as evaluation metrics. The evaluation metrics are calculated based on the prediction results of the running state output by the new network model for the training samples and the actual running state of the original records in the training samples. When each evaluation indicator reaches its respective preset threshold, the new network model is judged to have met the training standard.

7. The method according to any one of claims 1 to 6, characterized in that, Fault modes include bearing roller failure, bearing inner ring failure, bearing outer ring failure, combined bearing failure, gear damage, gear tooth breakage, tooth root cracks, and tooth surface wear. Historical vibration time-series signals are preprocessed to obtain training samples, including: The historical vibration time series signal was segmented by using the fixed-length sliding window method to obtain multiple signal samples, each of which contains a complete fault characteristic period. The signal samples are standardized to obtain training samples, where each standardized signal sample corresponds to one training sample.

8. A fault identification system for offshore wind turbines, characterized in that, The system includes: The training sample construction module is used to acquire historical vibration time-series signals of offshore wind turbine gearboxes under various operating conditions, preprocess the historical vibration time-series signals to obtain training samples, wherein the operating conditions include healthy states and various fault modes. A novel model building and training module is used to integrate multi-scale convolutional modules, SE attention mechanisms, and residual network structures on the basis of traditional one-dimensional convolutional neural networks to construct a novel network model. This novel network model includes an input layer, intermediate layers, and an output layer. The intermediate layers correspond to multiple deep convolutional blocks connected by residual networks. Each deep convolutional block deploys multi-scale convolutional modules and SE attention mechanisms, with convolutional kernels of various sizes within each module. When training the novel network model using training samples, the model receives training samples through the input layer and converts them into tensors, which are then fed to the first deep convolutional block of the intermediate layer. Within this first deep convolutional block, multi-scale convolutions... The module utilizes convolutional kernels of various sizes to simultaneously extract features from different frequency domains in the tensors corresponding to the same training sample. It then concatenates these extracted frequency domain features along the channel dimension to generate frequency domain fusion features. Furthermore, it adaptively learns the correlation between channels through the SE attention mechanism to generate weight vectors. These weight vectors are used to weight the frequency domain fusion features, yielding initial key features. These initial key features are then passed through a residual network between deep convolutional blocks until the final deep convolutional block outputs the final key features to the output layer. The output layer uses fully connected layers to map these key features, generating probability distributions for various operating states. Based on these probability distributions, it outputs predictions of the operating state of the offshore wind turbine gearbox. The offshore wind turbine fault identification module is used to collect new vibration time-series signals of the offshore wind turbine gearbox in real time when the new network model reaches the training standard by using the prediction results of the operating status and the actual operating status in the training samples. This enables the new network model that has reached the training standard to predict the operating status of the offshore wind turbine gearbox based on the new vibration time-series signals collected in real time.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for identifying offshore wind turbine faults as described in any one of claims 1 to 7.

10. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for identifying offshore wind turbine faults as described in any one of claims 1 to 7.

Images