Fan blade fault prediction method and system based on EEMD-HHT and CBCNN

Abstract

The invention relates to the technical field of fan blade fault detection, and particularly discloses a fan blade fault prediction method and system based on EEMD-HHT and CBCNN, and the method comprises the steps: obtaining an original vibration signal, obtaining an effective IMF based on EEMD technology processing, and carrying out the HHT transformation of the effective IMF to obtain a Hilbert spectrum; the method comprises the following steps: pre-constructing a CNN network and adding a CBAM attention module to form a CBCNN fault prediction model; dividing all Hilbert spectrums to obtain a verification set, a test set and a training set for training a CBCNN fault prediction model, and performing model training optimization processing until a final CBCNN fault prediction model is obtained; and obtaining a vibration signal used for prediction, inputting the vibration signal to the final CBCNN fault prediction model, and outputting a prediction result. According to the method, integrated empirical mode decomposition (EEMD) is adopted to improve HHT, a CBMA attention module is introduced into a CNN to enhance a CNN model, training is performed based on a training set containing a detection result, and fault type prediction is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of fan blade fault detection, and in particular to a fan blade fault prediction method and system based on EEMD-HHT and CBCNN. Background Art

[0002] Modern wind turbine blade fault detection systems usually rely on a large number of multi-type sensors installed on the blades, which provide critical real-time data on the operating status of the wind turbine. The widespread use of sensor technology has laid the foundation for fault detection, which can capture subtle changes in the blades during operation. However, wind turbine blade failures often originate from extremely small local problems such as cracks, wear, fatigue, etc. These early fault signals are usually very weak and easily masked by environmental noise. In particular, structural damage inside the blades, such as interlayer separation and fatigue cracks, are difficult to perceive with traditional surface sensors. Existing detection technologies are difficult to detect these faults before they expand to an obvious stage, resulting in a lag in diagnosis. In order to further improve the accuracy of fault detection, machine learning and artificial intelligence technologies have also been gradually introduced into the fault diagnosis of wind turbine blades. The advantage of machine learning algorithms is that they can handle large-scale complex nonlinear data sets, especially multi-source heterogeneous data generated during wind turbine operation. Through training on historical data, these algorithms can automatically identify the failure mode of the blades and adaptively optimize the diagnostic model, thereby improving the accuracy of fault detection and classification.

[0003] However, despite the excellent performance of machine learning methods in wind turbine blade fault detection, their performance is still limited by the quality and quantity of training data. In particular, in practical applications, the failure modes of wind turbine blades are complex and diverse, and are affected by many factors, including differences in wind turbine models, operating environments such as changes in wind speed, temperature, humidity, and operating conditions such as changes in operating loads. These factors increase the diversity and uncertainty of failure modes, making the fault detection model insufficient in generalization ability.

[0004] With the development of the times, machine learning is used in many fields, but the same or similar technologies are applied to different fields with very different effects, such as wind turbine blades and bearings; in addition, the operating environment of wind turbine blades is complex and often faces harsh natural conditions, such as high temperature, low temperature, strong wind and marine salt spray corrosion, etc. These conditions will bring a lot of noise interference to sensor data. How to extract useful feature information from massive data and effectively eliminate noise interference is an important challenge of current technology. Noise data not only affects the accuracy of fault detection, but also may cause deviations in machine learning models and reduce the reliability of predictions. Therefore, those skilled in the art are in urgent need of developing more sophisticated data preprocessing methods to improve the cleanliness and effectiveness of data. Summary of the invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the defects existing in the above-mentioned prior art, thereby providing a fan blade fault prediction method and system based on EEMD-HHT and CBCNN.

[0006] A fan blade fault prediction method based on EEMD-HHT and CBCNN includes the following steps: Obtain the original vibration signal, obtain the effective IMF based on EEMD technology, and perform HHT transformation on the effective IMF to obtain the Hilbert spectrum; Pre-build the CNN network and add the CBAM attention module to form the CBCNN fault prediction model; By dividing all Hilbert spectra, a validation set, a test set, and a training set for training the CBCNN fault prediction model are obtained, and the model training optimization process is performed until the final CBCNN fault prediction model is obtained; The vibration signal used for prediction is obtained and input into the final CBCNN fault prediction model to output the prediction result.

[0007] Preferably, in the process of obtaining valid IMF based on EEMD technology, two repetition steps are provided, namely: repeating steps S101-S106 to obtain an IMF component that meets the IMF condition; repeating steps S101-S108 to obtain n IMFs and a residual term; specifically including: S101-105. Obtain IMF components; S106. Check the first candidate IMF component Whether the IMF conditions are met: If not, As the original vibration signal after the first update, the steps S101-106 of obtaining the IMF component are repeated until an IMF component satisfying the IMF condition is obtained. ; is the first IMF component, denoted as: ; If satisfied, then the first candidate IMF component For the first IMF component, directly execute step S107; S107. Obtain the residual signal: subtract the first IMF component from the original vibration signal or the original vibration signal after the first update to obtain the residual signal obtained for the first time ; S108. The residual signals obtained from the 1st to nth times correspond to the original vibration signals updated from the 2nd to the n-1th time, and the steps S101-108 of obtaining the IMF components are repeated to obtain the second IMF: , and so on, until the remaining signal can no longer be decomposed into IMFs, and finally n IMFs and a residual term are obtained , , represents the jth IMF obtained after the i-th noise addition, ; , t represents time.

[0008] Preferably, the determination condition of the valid IMF in obtaining the valid IMF based on the EEMD technology is: Use the average energy as the set threshold to select valid IMFs: calculate the energy of each IMF and the total energy of the signal, and the IMFs whose ratio exceeds the set threshold are selected as valid IMFs.

[0009] Preferably, the attention weight s of the CBAM attention module is expressed as: ; in, , is the weight of the fully connected layer, δ is the ReLU activation function, is the Sigmoid activation function, Z max and Z avg They are global maximum pooling and average pooling respectively.

[0010] Preferably, the overall network structure of the CNN network used and the CBAM attention module added is: input layer-first convolution layer-first pooling layer-second convolution layer-second pooling layer-third convolution layer-third pooling layer-attention layer-fully connected layer-output layer; Among them, the first pooling layer, the second pooling layer and the third pooling layer all use the same 2×2 pooling operation; The first convolutional layer extracts features from the image transmitted by the input layer, generates 8 feature channels, and each feature channel extracts features through a 3×3 convolution kernel sliding.

[0011] Preferably, the method further comprises: drawing and normalizing the Hilbert spectrum, specifically: The Hilbert spectrum is drawn using the scatter plot method; Normalize the colors and scales in the Hilbert spectrum.

[0012] A fan blade fault prediction system based on EEMD-HHT and CBCNN, comprising: an acquisition module, which acquires original vibration signals based on multiple types of sensors arranged on the fan blades, and is implemented based on the fan blade fault prediction method. The fan blade fault prediction system also comprises: The signal processing module obtains the original vibration signal from the acquisition module, obtains the effective IMF based on the EEMD technology, and performs HHT transformation on the effective IMF to obtain the Hilbert spectrum; Fault prediction model building module, pre-builds CNN network and adds CBAM attention module to form CBCNN fault prediction model; The model training optimization module divides all Hilbert spectra into a validation set, a test set, and a training set for training the CBCNN fault prediction model, and performs model training optimization processing until the final CBCNN fault prediction model is obtained; The prediction module obtains the vibration signal used for prediction and inputs it into the final CBCNN fault prediction model, and outputs the prediction result.

[0013] The technical solution of the present invention has the following advantages: 1. The present invention proposes a technical solution for predicting crack and icing faults in wind turbine blades. In the solution, a convolutional neural network combining EEMD-HHT and a convolutional attention module is used, and integrated empirical mode decomposition EEMD is used to improve the HHT transformation. The CNN convolutional neural network model is enhanced by introducing a CBMA attention module into the convolutional neural network and training is performed based on a training set containing detection results to achieve prediction of fault types.

[0014] 2. In order to be further adapted to the fault detection of fan blades, the present invention adds HHT transformation on the basis of using integrated empirical mode decomposition EEMD; the most obvious difference between the fault detection of fan blades and the detection of rolling elements such as gears or bearings is that the impact component of the signal is not obvious, so it is necessary to perform feature analysis in the frequency domain or time-frequency domain on the signal reconstructed by EEMD. Therefore, adding HHT transformation on the basis of the existing EEMD is more suitable for the fault detection of fan blades.

[0015] 3. In order to improve the accuracy, the present invention also adds a CBMA attention module to enhance the CNN convolutional neural network model. At the same time, in order to improve the unsatisfactory influence of the Hilbert spectrum itself on the training of the CBCNN fault prediction model, the Hilbert spectrum is drawn by a scatter plot method and the color scale is normalized to achieve improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 It is a schematic diagram of the overall process of the method of the present invention; Figure 2 It is a schematic diagram of integrated empirical mode decomposition of the present invention; Figure 3 It is a schematic diagram of the network structure of the present invention; Figure 4 It is a schematic diagram of the structure of the convolutional attention of the present invention; Figure 5 This is a schematic diagram of the effect of the EEMD-HHT+CNN method; Figure 6 This is a schematic diagram of the effect of the EEMD-HHT+CBCNN method; Figure 7 Schematic diagram of the effect of the STFT-CBCNN method. DETAILED DESCRIPTION

[0018] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0020] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0021] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0022] Embodiment 1: A fan blade fault prediction method based on EEMD-HHT and CBCNN includes the following steps: Obtain the original vibration signal, obtain the effective IMF based on EEMD technology, and perform HHT transformation on the effective IMF to obtain the Hilbert spectrum; Pre-build the CNN network and add the CBAM attention module to form the CBCNN fault prediction model; like Figure 1 By dividing all Hilbert spectra, a validation set, a test set, and a training set for training the CBCNN fault prediction model are obtained, and the model training optimization process is performed until the final CBCNN fault prediction model is obtained; the ratios of the training set, validation set, and test set used in this embodiment are set to 60%, 20%, and 20%. The CBCNN fault prediction model built by training these data sets; The vibration signal used for prediction is obtained and input into the final CBCNN fault prediction model to output the prediction result.

[0023] Specifically: like Figure 2 The original vibration signal Perform integrated empirical mode decomposition (EEDM), decompose the signal into a series of IMFs according to its characteristics, and screen effective IMFs, including the following steps: S101. Generate white noise: Generate a series of white noise sequences of different levels. Sub-white noise, for each white noise, a white noise sequence is generated ; S102. Noise processing: Add the generated white noise sequence to the original vibration signal to obtain the noisy signal ; represents the i-th white noise; S103. Determine local maximum and minimum values: Determine the signal These extreme values ​​are connected using cubic spline interpolation to form the upper envelope. and lower envelope ; S104. Calculate the mean of the upper and lower envelopes: ; S105. Obtaining approximate IMF: From the original signal Subtract the mean from the value to get the approximate value of the first intrinsic mode function IMF ; S106. Check IMF conditions: OK Whether the IMF conditions are met: IMF conditions: a) Extreme values: The number of maximum and minimum values ​​must be equal, or differ by at most one; b) At any point, the average value of the envelope defined by the local maximum and local minimum points is zero.

[0024] Check the first candidate IMF component Whether the IMF conditions are met; if not, As the original vibration signal after the first update, the steps S101-106 of obtaining the IMF component are repeated until an IMF component satisfying the IMF condition is obtained. ; is the first IMF component, denoted as: ; If satisfied, then the first candidate IMF component For the first IMF component, directly execute step S107; S107. Obtain the residual signal: subtract the first IMF component from the original vibration signal or the original vibration signal after the first update to obtain the residual signal obtained for the first time ; ; S108. The residual signals obtained from the 1st to nth times correspond to the original vibration signals updated from the 2nd to the n-1th time, and the steps S101-S108 of obtaining the IMF components are repeated to obtain the second IMF: , and so on, until the remaining signal can no longer be decomposed into IMFs, and finally n IMFs and a residual term are obtained , , represents the jth IMF obtained after the i-th noise addition, ; , t represents time.

[0025] S109. Take the average value of IMF: average the IMFs decomposed by EMD N times to obtain the final IMF. IMF, the average result is ; The determination conditions for obtaining effective IMF based on EEMD technology are as follows: S110. Use the average energy as a set threshold to select valid IMFs: calculate the energy of each IMF and the total energy of the signal, and the IMFs whose ratio exceeds the set threshold are selected as valid IMFs.

[0026] The HHT transform is performed on the effective IMF to obtain the Hilbert spectrum. Specifically, the HHT transform is performed on the effective IMF to obtain the instantaneous frequency and instantaneous amplitude, thereby constructing the Hilbert spectrum. Specifically, the following steps are included: S201. Hilbert transform: Perform Hilbert transform on each IMF signal , where PV represents the Cauchy principal value; Indicates signal The Hilbert transform of represents the integration variable.

[0027] S202. Obtaining parsing signal: ,in is the instantaneous amplitude, is the instantaneous phase; S203. Calculate the instantaneous frequency: ; S204. Hilbert spectrum: The Hilbert spectrum represents the energy distribution of a signal at different times and frequencies. The instantaneous frequency and instantaneous amplitude of all IMFs are combined to obtain the Hilbert spectrum. .

[0028] It should be noted that in actual application, when drawing and normalizing the Hilbert spectrum, it is found that it is difficult to see obvious features of the Hilbert spectrum when drawing an ordinary Hilbert spectrum, and the color scale in the picture is not unified, resulting in unsatisfactory results when training the model based on the CNN convolutional neural network. Therefore, this embodiment also includes: drawing and normalizing the Hilbert spectrum, so that the time spectrum is clean and clear, the features are more obvious, and the color scale is fixed, which is conducive to the learning of features by the model based on the CNN convolutional neural network, specifically: The Hilbert spectrum is drawn using the scatter plot method; Normalize the colors and scales in the Hilbert spectrum.

[0029] In this embodiment, the attention weight s of the CBAM attention module is expressed as: ; in, , is the weight of the fully connected layer, δ is the ReLU activation function, is the Sigmoid activation function.

[0030] like Figure 3 In this embodiment, the overall network structure of the CNN network used in the present invention and adding the CBAM attention module is: input layer-first convolution layer-first pooling layer-second convolution layer-second pooling layer-third convolution layer-third pooling layer-attention layer-fully connected layer-output layer; Among them, the first pooling layer, the second pooling layer and the third pooling layer all use the same 2×2 pooling operation; The first convolutional layer extracts features from the image transmitted by the input layer, generates 8 feature channels, and each feature channel extracts features through a 3×3 convolution kernel sliding.

[0031] Embodiment 2: First, modern wind turbine blade fault detection systems usually rely on a large number of multi-type sensors installed on the blades, which provide critical real-time data on the operating status of the wind turbine. These sensors include accelerometers, strain gauges, acoustic sensors, and temperature sensors, which are used to monitor physical parameters such as blade vibration, structural strain, acoustic wave changes, and temperature changes. Through the synergy of these sensors, the health status of wind turbine blades can be recorded and analyzed in detail. Especially during the long-term operation of the wind turbine, sensors can continuously track the dynamic behavior of the blades to help detect abnormalities and potential faults in a timely manner. However, it is impossible to accurately detect damage in the early stage through sensors and existing technologies alone; damage that is not detected in the early stage may accumulate during the long-term operation of the wind turbine, eventually causing serious failures, increasing the complexity and cost of maintenance.

[0032] Secondly, with the continuous accumulation of wind turbine operation data, big data analysis technology plays an increasingly important role in fault prediction. By analyzing the massive historical data collected by sensors, engineers can build a health model for wind turbine blades and identify potential failure trends. Using these models, more intelligent fault prediction and preventive maintenance can be achieved to prevent faults from developing to an unrepairable stage. Data-driven predictive maintenance strategies can help operation and maintenance personnel intervene before problems expand, greatly reducing the unplanned downtime and expensive maintenance costs of wind turbines, and improving the operating efficiency and economy of wind turbines.

[0033] Algorithms such as deep learning are particularly outstanding in fault classification and prediction. By continuously learning and adjusting model parameters, they can discover hidden patterns in data and predict and classify new fault signals. This not only helps detect existing faults, but also predicts potential faults that may occur.

[0034] In summary, in response to the above problems, this embodiment studies the vibration signal characteristics of the blade in detail, uses integrated empirical mode decomposition EEMD to reduce the impact of noise, suppress the modal aliasing problem, and then improves the Hilbert Huang transform HHT. Then, the convolutional attention module CBAM is used to enhance the attention to the important features of the Hilbert spectrum feature map, and finally the convolutional neural network CNN is used for fault diagnosis and classification to achieve early identification and warning of blade faults. The following contents are specifically disclosed: A fan blade fault prediction system based on EEMD-HHT and CBCNN includes: a collection module, which collects original vibration signals based on multiple types of sensors set on the fan blades, and is implemented based on the fan blade fault prediction method of embodiment 1. The fan blade fault prediction system also includes: The signal processing module obtains the original vibration signal from the acquisition module, obtains the effective IMF based on the EEMD technology, and performs HHT transformation on the effective IMF to obtain the Hilbert spectrum; Fault prediction model building module, pre-builds CNN network and adds CBAM attention module to form CBCNN fault prediction model; The model training optimization module divides all Hilbert spectra into a validation set, a test set, and a training set for training the CBCNN fault prediction model, and performs model training optimization processing until the final CBCNN fault prediction model is obtained; The prediction module obtains the vibration signal used for prediction and inputs it into the final CBCNN fault prediction model, and outputs the prediction result.

[0035] Specifically: Model principle: 1. Convolutional layer: The convolution layer in this embodiment is used to extract local features from the input image. The convolution layer performs a convolution operation on the input feature map through a convolution kernel to generate a multi-channel feature map. The convolution operation is expressed as: ; in, is the output feature map at position No. The value of the channel, is the input image, is the convolution kernel weight, is the bias term, represents the height of the convolution kernel, represents the width of the convolution kernel, Indicates the number of input channels, Represents the pixel value of the input image at a certain position, Represents the weight value of the corresponding position of the convolution kernel; m represents the height of the convolution kernel The index of the input channel. The index of .

[0036] 2. Pooling layer: The pooling layer is used to downsample the output feature map of the convolutional layer, reduce the size of the feature map, and retain the main feature information. The pooling operation can use maximum pooling or average pooling, which is expressed as: ; in, Represents the value of the kth channel of the feature map value at position (a, b) after pooling, Represents the value of the kth channel of the feature map output by the convolution at position (a+m,b+l), Represents the height of the pooling window, Represents the width of the pooling window; M represents the offset of the pooling window in the height direction during the pooling operation; L represents the offset of the pooling window in the width direction during the pooling operation.

[0037] This pooling operation can extract the most significant features in each region, reduce the amount of computation and enhance the generalization ability of the network.

[0038] 3. CBAM module: like Figure 4 The CBAM module is an important component of this embodiment, which includes two parts: channel attention and spatial attention, which respectively enhance the attention to important features in a weighted manner.

[0039] Channel attention extracts the global features of each channel through a global pooling operation and generates a channel weight matrix through a two-layer fully connected network. The formula is as follows: Global maximum pooling and average pooling: , ; Y represents the output feature map of the convolutional layer; Generate attention weights through the fully connected layer: ; in, , is the weight of the fully connected layer, δ is the ReLU activation function, is the Sigmoid activation function.

[0040] Channel weighting: ; Where s is the generated channel weight, It is the feature map after channel weighting.

[0041] Spatial attention generates global features in space through maximum pooling and average pooling, and generates a spatial weighted matrix through convolution operation. The formula is as follows: Max pooling and average pooling: , ; Concatenate the pooling results and generate a spatial weighting matrix through convolution: ; Spatial Weighting:

[0042] The CBAM module can effectively highlight the important feature areas in the image and suppress irrelevant or minor information by dual weighting of channels and space.

[0043] 4. Fully connected layer: The fully connected layer maps the convolutional features to the output categories, and the calculation formula is as follows: ; Among them, X is the input feature, W is the weight matrix of the fully connected layer, O is the bias term, and Q is the output category.

[0044] 5. Model building: In this embodiment, there are two convolutional layers, namely: the first convolutional layer and the second convolutional layer; there are two pooling layers, namely: the first pooling layer and the second pooling layer. The overall network structure of the CNN network used in this embodiment and adding the CBAM attention module is: input layer-first convolutional layer-first pooling layer-second convolutional layer-second pooling layer-third convolutional layer-third pooling layer-attention layer-fully connected layer-output layer; Figure 3 ,The original vibration signal is a one-dimensional vibration signal, which is processed by EEMD-HHT to obtain a time-frequency graph and input into the first convolutional layer; 1) The input layer accepts an RGB image of size 32 × 32 × 3. This image is a color image containing 32 × 32 pixels with 3 channels corresponding to the colors red, green, and blue.

[0045] 2) The first convolutional layer will perform feature extraction on the input RGB image and generate 8 feature channels. Each channel extracts features by sliding a 3×3 convolution kernel with a step size of 1×1 to ensure that the convolution operation covers every pixel of the image.

[0046] 3) The first pooling layer performs a 2×2 pooling operation, usually a maximum pooling operation, which is used to reduce the size of the feature map while retaining important features. Through the pooling operation, the size of the feature map will be halved.

[0047] 4) The second convolutional layer further extracts the pooled feature map. The convolution kernel is 3×3, generating 16 feature channels to continue capturing deeper features in the image.

[0048] 5) The second pooling layer is also a 2×2 pooling operation, which again reduces the size of the feature map by half, further reducing the computational complexity and retaining important features.

[0049] 6) The third convolutional layer continues to extract higher-level features through a 3×3 convolution kernel, generating 32 channels. This step provides more details for further feature expression.

[0050] 7) The third pooling layer performs another 2×2 pooling operation, further reducing the size of the feature map by half and extracting streamlined global features.

[0051] 8) The network introduces the CBAM module, and enhances the representation ability of features through channel attention and spatial attention modules: channel attention uses global pooling and two fully connected layers to weight each channel, and spatial attention focuses on important positions in the image through 2×2 convolution. Finally, after the fully connected layer, the network combines and classifies the features.

[0052] Comparison of experimental results: The wind turbine blade fault identification technology used in this embodiment is EEMD-HHT+CBCNN, and the performance of the technology is verified by comparing it with STFT+CBCNN technology and EEMD-HHT+CNN technology. The software used in the experiment is PyCharm software and PyTorch framework.

[0053] The data set used in this embodiment is the Zenodo public data set, which contains three types of vibration signals of wind turbine blades, namely normal vibration signal, crack vibration signal and icing vibration signal. The vibration signal is collected by a network composed of 8 accelerometers with a working range of ±10g.

[0054] The performance of the EEMD-HHT+CBCNN model is verified by comparing it with the EEMD-HHT+CNN method. Figure 5 , Figure 6 and Figure 7 It can be seen that CBCNN has a better fitting effect than S-CNN, and the accuracy on the test set is improved and the Loss value is lower.

[0055]

[0056] It can also be seen from the table that the accuracy of the EEMD-HHT+CBCNN model is better than that of the other two models, indicating that the model performance provided by this embodiment is better than that of the traditional algorithm.

[0057] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the invention.

Claims

1. A fan blade fault prediction method based on EEMD-HHT and CBCNN, characterized in that: The following steps are involved: Obtain the original vibration signal, obtain the effective IMF based on EEMD technology, and perform HHT transformation on the effective IMF to obtain the Hilbert spectrum; Pre-build the CNN network and add the CBAM attention module to form the CBCNN fault prediction model; By dividing all Hilbert spectra, a validation set, a test set, and a training set for training the CBCNN fault prediction model are obtained, and the model training optimization process is performed until the final CBCNN fault prediction model is obtained; The vibration signal used for prediction is obtained and input into the final CBCNN fault prediction model to output the prediction result.

2. The wind turbine blade fault prediction method according to claim 1, characterized in that: In the process of obtaining effective IMF based on EEMD technology, there are two repeated steps, namely: repeating steps S101-S106 to obtain an IMF component that meets the IMF condition; repeating steps S101-S108 to obtain n IMFs and a residual term; specifically including: S101-105. Obtain IMF components; S106. Check the first candidate IMF component Whether the IMF conditions are met: If not, As the original vibration signal after the first update, the steps S101-106 of obtaining the IMF component are repeated until an IMF component satisfying the IMF condition is obtained. ; is the first IMF component, denoted as: ; If satisfied, then the first candidate IMF component For the first IMF component, directly execute step S107; S107. Obtain the residual signal: subtract the first IMF component from the original vibration signal or the original vibration signal after the first update to obtain the residual signal obtained for the first time ; S108. The residual signals obtained from the 1st to nth times correspond to the original vibration signals updated from the 2nd to the n-1th time, and the steps S101-108 of obtaining the IMF components are repeated to obtain the second IMF: , and so on, until the remaining signal can no longer be decomposed into IMFs, and finally n IMFs and a residual term are obtained , , represents the jth IMF obtained after the i-th noise addition, ; , t represents time.

3. The wind turbine blade fault prediction method according to claim 1, characterized in that: The determination conditions for obtaining effective IMF based on EEMD technology are as follows: Use the average energy as the set threshold to select valid IMFs: calculate the energy of each IMF and the total energy of the signal, and the IMFs whose ratio exceeds the set threshold are selected as valid IMFs.

4. The wind turbine blade fault prediction method according to claim 1, characterized in that: The attention weight s of the CBAM attention module is expressed as: ； in, , is the weight of the fully connected layer, δ is the ReLU activation function, is the Sigmoid activation function, Z max and Z avg They are global maximum pooling and average pooling respectively.

5. The wind turbine blade fault prediction method according to claim 1, characterized in that: The overall network structure of the CNN network used in the present invention and adding the CBAM attention module is: input layer-first convolution layer-first pooling layer-second convolution layer-second pooling layer-third convolution layer-third pooling layer-attention layer-fully connected layer-output layer; Among them, the first pooling layer, the second pooling layer and the third pooling layer all use the same 2×2 pooling operation; The first convolutional layer extracts features from the image transmitted by the input layer, generates 8 feature channels, and each feature channel extracts features through a 3×3 convolution kernel sliding.

6. The method for predicting wind turbine blade failure according to claim 1, characterized in that: Also includes: The Hilbert spectrum is drawn and normalized, specifically: The Hilbert spectrum is drawn using the scatter plot method; Normalize the colors and scales in the Hilbert spectrum.

7. A fan blade fault prediction system based on EEMD-HHT and CBCNN, comprising: The acquisition module acquires original vibration signals based on multiple types of sensors arranged on the fan blades, wherein the fan blade fault prediction method according to any one of claims 1 to 6 is implemented, and the fan blade fault prediction system further comprises: The signal processing module obtains the original vibration signal from the acquisition module, obtains the effective IMF based on the EEMD technology, and performs HHT transformation on the effective IMF to obtain the Hilbert spectrum; Fault prediction model building module, pre-builds CNN network and adds CBAM attention module to form CBCNN fault prediction model; The model training optimization module divides all Hilbert spectra into a validation set, a test set, and a training set for training the CBCNN fault prediction model, and performs model training optimization processing until the final CBCNN fault prediction model is obtained; The prediction module obtains the vibration signal used for prediction and inputs it into the final CBCNN fault prediction model, and outputs the prediction result.

Images