A method for identifying damage of fan blades

The multi-dimensional feature extraction of fan signals through stack denoising autoencoder and convolutional denoising autoencoder, combined with full-connection network and transfer learning, solves the problems of single feature indicators and single sensor signals in traditional fan health monitoring systems. , realizes high-precision identification and robust monitoring of fan blade damage.

CN114528869BActive Publication Date: 2024-11-08SOUTHEAST UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202210007502.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-05
Publication Date
2024-11-08
Estimated Expiration
2042-01-05

AI Technical Summary

Technical Problem

Traditional fan health monitoring systems have problems of poor stability and low diagnostic accuracy in complex industrial testing sites, especially due to the singleness of characteristic indicators and the singleness of sensor signals, resulting in poor diagnostic results for fan blade damage.

Method used

The stack denoising autoencoder and convolutional denoising autoencoder are used to extract the vibration and sound pressure signals in a multi-dimensional feature, and the features are fusion and classification are performed through a fully connected network to establish a training model to realize the identification of blade damage. The method includes processing one-dimensional and two-dimensional sample data into multi-dimensional fusion features, optimizing network model parameters through transfer learning, and improving the robustness and accuracy of the model.

Benefits of technology

It significantly improves the accuracy of fan blade damage recognition and the robustness of the model, can effectively identify blade damage in complex environments, reduce the probability of failure, and realize online status monitoring.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114528869B_ABST
    Figure CN114528869B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for identifying damage to fan blades, comprising: collecting vibration signals and sound pressure signals during the operation of the fan, processing the vibration and sound pressure signals into one-dimensional sample data and two-dimensional sample data; establishing a training network model: using a stack denoising autoencoder to extract features from the one-dimensional sample data, and using a convolution denoising autoencoder to extract features from the two-dimensional sample data; fusing the features of the one-dimensional sample and the two-dimensional sample respectively through a first-layer fully connected network to obtain multi-dimensional fusion features, fusing the multi-dimensional fusion features corresponding to the sound pressure signal and the vibration signal through a second-layer fully connected network to obtain a fusion feature matrix, and classifying through a classifier to obtain a classification model; inputting the vibration signal of the fan with a blade crack to be detected into the trained network model, and judging the crack state of the blade according to the output result. The present invention has advantages in terms of blade damage identification accuracy and model robustness.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of fan fault diagnosis, and in particular to a fan blade damage identification method. Background Art

[0002] In the field of machinery, large-scale equipment such as aircraft engines, wind power generation equipment, and industrial production equipment are developing in the direction of high speed, high precision, and high efficiency. Establishing a reliable health monitoring system is a necessary measure to ensure the safe operation of these equipment. Traditional mechanical health monitoring systems use simple time domain and frequency domain indicators to describe the operating status of the equipment. The selection of these specific indicators often comes from professional technicians and expert experience. They only have good diagnostic effects on specific equipment and their faults and are not universally applicable. In complex industrial testing sites, fault information is complex and changeable, often mixed with background noise and multi-sensor measurement noise, and may also include internal and external excitations and the coupling of multiple faults. This makes the diagnostic methods based on these characteristic indicators have poor stability and low diagnostic accuracy.

[0003] At present, the signal feature fusion method based on deep learning can provide a solution to the problem of mismatch between fault features and fault categories. The current deep learning network and feature fusion method can adaptively mine valuable blade damage information from raw data or images, getting rid of the dependence on expert experience and the problem of poor universal applicability of features. However, there are still problems such as single feature dimension and single sensor acquisition signal.

[0004] Therefore, studying how to reduce the redundancy of wind turbine blade operation signal characteristics is of great significance to effectively improve the blade damage diagnosis capability and robustness. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a method for identifying damage to wind turbine blades, the purpose of which is to improve the detection accuracy by increasing the diversity of acquired signals and increasing feature dimensions.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A method for identifying fan blade damage comprises the following steps:

[0008] S1. Collect vibration signals and sound pressure signals during the operation of the fan, and process the vibration and sound pressure signals into one-dimensional sample data and two-dimensional sample data;

[0009] S2. Establish training network model:

[0010] Using a stack denoising autoencoder to extract features from the one-dimensional sample data, and using a convolutional denoising autoencoder to extract features from the two-dimensional sample data;

[0011] The features of the one-dimensional samples and the two-dimensional samples are fused respectively through the first layer of fully connected network to obtain multi-dimensional fusion features, and the multi-dimensional fusion features corresponding to the sound pressure signal and the vibration signal are secondary fused through the second layer of fully connected network to obtain a fusion feature matrix, and the fusion feature matrix is ​​classified through the classifier to obtain a classification model;

[0012] S3. Input the vibration signal of the fan with blade cracks to be detected into the trained network model, and judge the crack status of the blade according to the output result.

[0013] Further technical solutions are:

[0014] In step S2, the convolutional denoising autoencoder takes the minimum difference between the reconstructed signal and the original signal as the objective function, and iterates the optimization of the network weights. The convolutional denoising autoencoder includes two convolutional layers and two deconvolutional layers. The gradient transfer of the deconvolutional layer is as follows:

[0015]

[0016] Where L is the objective function, K is the deconvolution kernel, and y is the output of the deconvolution layer;

[0017] in, x is the input of the deconvolution layer, t is the input signal length of the deconvolution layer, and m is the stacking step size during deconvolution.

[0018] The stacked denoising autoencoder includes at least three denoising autoencoders.

[0019] Step S2 establishes a training network model, and also includes: inputting sample data under different working conditions into the network model for training, screening out the network model parameters obtained by training the sample data of a certain working condition as the optimal network model parameters, and performing transfer learning on the network models trained by the sample data of the remaining working conditions to optimize the parameters of the network model.

[0020] Step S3 includes inputting the vibration signal of the fan whose blade cracks are to be detected into the trained network model, extracting the fusion feature matrix and classification model of the signal, visually displaying and outputting the extraction results, and judging the crack status of the fan blade according to the output results.

[0021] The beneficial effects of the present invention are as follows:

[0022] The present invention makes full use of the characteristics of vibration and sound pressure signals that are easy to collect and contain fault feature information. It does not need to judge the status of the fan blades through manual on-site detection, and can realize online status monitoring during the operation of the fan, fundamentally reducing the probability of blade failure. In addition, the fusion of multi-dimensional information and acoustic vibration signals is considered. The method has significant advantages in blade damage identification accuracy and model robustness, and has the following significant improvements:

[0023] The present invention can perform unsupervised feature extraction from the frequency domain and time-frequency domain of the original signal through the stack denoising autoencoder and the convolution denoising autoencoder, and the extracted feature samples have good separability.

[0024] In the process of establishing and training the network model, through transfer learning, the network weight parameters of the working conditions with high classification accuracy are transferred to the network model of the working conditions with weak classification, which can indirectly expand the samples and achieve the effect of improving the model training accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of the structure of the training network model of an embodiment of the present invention.

[0026] Figure 2 This is sample data of an embodiment of the present invention.

[0027] Figure 3 It is a schematic diagram of the visual structure of the intermediate process of feature extraction in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The specific implementation of the present invention is described below with reference to the accompanying drawings.

[0029] A method for identifying wind turbine blade damage in this embodiment includes the following steps:

[0030] S1. Collect vibration signals and sound pressure signals during the operation of the fan, and process the vibration and sound pressure signals into one-dimensional sample data and two-dimensional sample data. The one-dimensional and two-dimensional sample data are respectively as follows: Figure 2 (a) and (b) show the signal's frequency domain and wavelet time-frequency domain data respectively.

[0031] S2. Establish training network model:

[0032] Using a stack denoising autoencoder (SDAE) to extract features from the one-dimensional sample data, and using a convolutional denoising autoencoder (CDAE) to extract features from the two-dimensional sample data;

[0033] The features of the one-dimensional samples and the two-dimensional samples are fused respectively through the first layer of fully connected network to obtain multi-dimensional fusion features, and the multi-dimensional fusion features corresponding to the sound pressure signal and the vibration signal are secondary fused through the second layer of fully connected network to obtain a fusion feature matrix, and the fusion feature matrix is ​​classified through the classifier to obtain a classification model;

[0034] Common autoencoders are constructed by fully connected networks, which are more sensitive to the feature capture of one-dimensional signals and are far less effective than convolutional neural networks in extracting features of two-dimensional image information. This application adopts a convolutional denoising autoencoder, which performs a deconvolution operation on the convolution pooled signal, uses the pooled features as weights, and performs weighted superposition on the deconvolution kernel to complete the reconstruction of the original signal.

[0035] The solution of the convolutional denoising autoencoder (CDAE) takes the minimum difference between the reconstructed signal and the original signal as the objective function, and optimizes the weights of the network iteratively. The convolutional denoising autoencoder of this embodiment includes two convolutional layers and two deconvolutional layers. The gradient transfer of the deconvolutional layer is as follows:

[0036]

[0037] Where L is the objective function, K is the deconvolution kernel, and y is the output of the deconvolution layer;

[0038] in, x is the input of the deconvolution layer, t is the input signal length, and m is the stacking step size during deconvolution.

[0039] The structure diagram of the training network model is as follows Figure 1 As shown in the figure, the number and size of convolution kernels of the two convolution layers in the convolutional denoising autoencoder are 6×5×5 and 12×5×5 respectively, and the deconvolution layer is similar to the convolution layer. The number of iterations is 1000, and the final output feature sample is 5000×5×5.

[0040] The stacked denoising autoencoder of this embodiment includes at least three denoising autoencoders. Stacking three denoising autoencoders will obtain a stacked denoising autoencoder. When training the model, the output of the previous denoising autoencoder is used as the input of the next denoising autoencoder until all denoising autoencoders are trained, thereby obtaining a stacked deep network. The stacked denoising autoencoder network has strong robustness in extracting high-dimensional, nonlinear input data features.

[0041] The denoising auto encoder (DAE) damages the data in the input layer by adding noise, and encodes and decodes the "damaged" original data, and then restores the original data. The denoising process is to assign zeros to certain nodes in the input layer according to a certain probability, and then use them as the input of the auto encoder for training. The encoding formula from the input layer to the hidden layer is:

[0042]

[0043] Where: f e () is the encoding function from the input layer to the hidden layer; Wxh is the weight value of the encoding function; σ() is the Sigmoid activation function; b xh is its offset value, is the training sample.

[0044] The decoding formula from the hidden layer to the output layer is:

[0045] r = f d (h)=σ(W hr h+b hr )

[0046] In the formula, f d () is the decoding function from hidden layer to output layer; W hr is the weight value of the decoding function; b hr Its offset value.

[0047] The loss function of DAE is:

[0048]

[0049] In the formula, x i is the i-th training sample, r i is the reconstructed representation of the i-th training sample after DAE.

[0050] Specifically, the network structures (number of input and output neurons) of the three denoising autoencoders are 2560-200, 200-100, and 100-50, respectively. The learning rate decreases from large to small as the number of network layers increases, and is 1, 0.5, and 0.3, respectively. The number of iterations is 200, and finally a 5000×50 feature sample set is output.

[0051] In a complex industrial environment, sample data is easily disturbed, unprocessed sound and vibration signals contain a lot of noise, and the signal characteristics of initial faults are weak. Therefore, this application removes most of the noise in the signal through DAE processing and extracts robust features from the hidden layer data.

[0052] The denoising convolutional autoencoder of this embodiment combines the image information capture performance of the convolutional neural network while taking into account the unsupervised capability of the denoising autoencoder.

[0053] In this embodiment, the first fully connected layer, the second fully connected layer and the classifier constitute three fully connected layers, the network structure is 300-50-5, the activation function is "sigm", the learning rate is 0.1, and the number of fine-tuning times is 100.

[0054] The network model training process also includes inputting sample data under different working conditions into the network model for training, screening out the network model parameters obtained by training the sample data of a certain working condition as the optimal network model parameters, and performing transfer learning on the network models trained by the sample data of the remaining working conditions to optimize the parameters of the network model.

[0055] S3. Input the vibration signal of the fan with blade cracks to be detected into the trained network model, extract the fusion feature matrix and classification model of the signal, visualize and output the extraction results, and judge the crack status of the fan blade according to the output results.

[0056] Specifically, the t-SNE algorithm is used for visualization output, such as Figure 3 As shown, this is a schematic diagram of the visualization structure of the intermediate process of feature extraction. Figure 3 In the figure, (a), (b), (c), and (d) are schematic diagrams of one-dimensional sample data, the result of feature extraction of one-dimensional sample data using stack denoising autoencoder, two-dimensional sample data, and the result of feature extraction of two-dimensional sample data using convolutional denoising autoencoder, respectively.

[0057] This application makes full use of multi-sensor and multi-dimensional information, and has higher accuracy and robustness in blade damage diagnosis compared to the diagnosis method based on single-dimensional features and single-modal sensors.

Claims

1. A method for identifying damage to a fan blade, characterized in that: The following steps are involved: S1. Collect vibration signals and sound pressure signals during the operation of the fan, and process the vibration and sound pressure signals into one-dimensional sample data and two-dimensional sample data; S2. Establish training network model: Using a stack denoising autoencoder to extract features from the one-dimensional sample data, and using a convolutional denoising autoencoder to extract features from the two-dimensional sample data; The features of the one-dimensional samples and the two-dimensional samples are fused respectively through the first layer of fully connected network to obtain multi-dimensional fusion features, and the multi-dimensional fusion features corresponding to the sound pressure signal and the vibration signal are secondary fused through the second layer of fully connected network to obtain a fusion feature matrix, and the fusion feature matrix is ​​classified through the classifier to obtain a classification model; S3, inputting the vibration signal of the fan with blade cracks to be detected into the trained network model, and judging the crack state of the blade according to the output result; In step S2, the convolutional denoising autoencoder takes the minimum difference between the reconstructed signal and the original signal as the objective function, and iterates the optimization of the network weights. The convolutional denoising autoencoder includes two convolutional layers and two deconvolutional layers. The gradient transfer of the deconvolutional layer is as follows: Where L is the objective function, K is the deconvolution kernel, and y is the output of the deconvolution layer; in, x is the input of the deconvolution layer, t is the input signal length of the deconvolution layer, and m is the stacking step size during deconvolution.

2. The wind turbine blade damage identification method according to claim 1, characterized in that: The stacked denoising autoencoder includes at least three denoising autoencoders.

3. The wind turbine blade damage identification method according to claim 1, characterized in that: Step S2 establishes a training network model, and also includes: inputting sample data under different working conditions into the network model for training, screening out the network model parameters obtained by training the sample data of a certain working condition as the optimal network model parameters, and performing transfer learning on the network models trained by the sample data of the remaining working conditions to optimize the parameters of the network model.

4. The wind turbine blade damage identification method according to claim 1, characterized in that: Step S3 includes inputting the vibration signal of the fan whose blade cracks are to be detected into the trained network model, extracting the fusion feature matrix and classification model of the signal, visually displaying and outputting the extraction results, and judging the crack status of the fan blade according to the output results.

Citation Information

Patent Citations

  • Multi-domain stacked denoising automatic encoder network-based planetary gearbox fault diagnosis method

    CN110146282A

  • Performance equipment fault diagnosis model construction method based on time convolution noise reduction network

    CN113705424A

  • Bearing defect identification method based on SDAE and improved GWO-svm

    WO2021128510A1