CN115761398A - Bearing fault diagnosis method based on lightweight neural network and dimension expansion - Google Patents
Bearing fault diagnosis method based on lightweight neural network and dimension expansion Download PDFInfo
- Publication number
- CN115761398A CN115761398A CN202211342324.4A CN202211342324A CN115761398A CN 115761398 A CN115761398 A CN 115761398A CN 202211342324 A CN202211342324 A CN 202211342324A CN 115761398 A CN115761398 A CN 115761398A
- Authority
- CN
- China
- Prior art keywords
- neural network
- vibration signal
- bearing vibration
- dimensional
- bearing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Complex Calculations (AREA)
Abstract
The invention relates to the field of machine vision and fault diagnosis, in particular to a bearing fault diagnosis method based on a lightweight neural network and dimension expansion, which comprises the steps of obtaining a historical bearing vibration signal, normalizing the obtained bearing vibration signal and then carrying out polar coordinate coding; converting the polar coordinate coded bearing vibration signal into a two-dimensional bearing vibration signal based on a gram angle sum field, a gram angle difference field and a Markov transition field; constructing a lightweight neural network, and training the neural network by using a two-dimensional bearing vibration signal; converting a bearing vibration signal to be detected into a two-dimensional bearing vibration signal, and inputting the two-dimensional bearing vibration signal into a light weight neural network which completes training to obtain a diagnosis result; the invention can effectively realize the visualization of vibration signals and provide RGB three channels for neural network learning, thereby realizing the accurate identification and diagnosis of the rolling bearing fault by a machine vision method with relatively small samples.
Description
Technical Field
The invention relates to the field of machine vision and fault diagnosis, in particular to a bearing fault diagnosis method based on a lightweight neural network and dimension expansion.
Background
At present, fault diagnosis methods based on rolling bearing vibration signals are mainly divided into three categories: the first type is that the extraction of the characteristic frequency of a fault after the noise reduction of an original signal is realized by using envelope spectrum analysis or signal decomposition or the combination of the envelope spectrum analysis and the signal decomposition, and then the characteristic frequency of the fault is compared with the typical characteristic frequency of each fault to judge the fault. However, the method can only realize qualitative fault diagnosis and cannot realize quantitative fault diagnosis. And the second type is that the mode of combining fault feature extraction and shallow machine learning is adopted, noise reduction and fault feature extraction are realized through a time domain, frequency domain and time-frequency domain signal processing method, and then fault classification is realized through the extracted features through a machine learning method. However, under the conditions of complex working conditions and more fault types, the method has poor diagnosis effect and complicated steps. The third type is a deep learning-based method, which has a deep structure and strong nonlinear feature extraction capability, can directly realize fault feature extraction and pattern recognition in bearing vibration signals, and can particularly meet end-to-end fault diagnosis under complex working conditions. However, a large number of data samples are needed for training the deep learning model, and bearing fault samples are often difficult to obtain (belong to small samples), so how to implement fault diagnosis under the small samples is a problem to be solved.
Deep learning is realized by constructing a multilayer network, computers automatically learn on the network and obtain the relation of data hidden in the network, and higher-dimensional and more abstract data are extracted, so that the learned characteristics have more expressive power. The deep network structure of the system can directly learn the most essential characteristics from the vibration signals so as to realize fault diagnosis, and the process that various signal processing methods need to manually extract the characteristics is avoided. The Convolutional Neural Network (CNN) is a special deep feedforward network, and the CNN model mainly includes an input layer, a convolutional layer, a pooling layer, a full-link layer, and an output layer. However, in the network structure, in order to make the output more accurate and the feature extraction more abundant, a network model combining multiple convolutional layers and multiple pooling layers is usually used in the network model, and more classical CNN models include LeNet-5, alexNet, ZF-Net, VGGNet, google LeNet, resNet, and densneet. However, most of the existing deep neural networks occupy huge memory and computing resources, and cannot be deployed on lighter-weight equipment, so that the applicability is reduced. Meanwhile, enough fault samples are difficult to obtain in an actual industrial environment, so that the established optimization model is difficult to apply practically, and therefore, the realization of breakthrough of deep learning under small samples in the aspect of fault diagnosis is a major current problem.
The processing of the vibration signal mostly uses signal processing methods, such as short-time fourier transform, wavelet transform, modal decomposition, etc. However, the signals processed by these methods are difficult to train through deep neural networks, especially convolutional neural networks for machine vision. How to reasonably convert one-dimensional signal data to facilitate efficient training of convolutional neural networks is a problem to be solved.
Disclosure of Invention
The invention provides a bearing fault diagnosis method based on a lightweight neural network and dimension expansion, which specifically comprises the following steps:
acquiring a historical bearing vibration signal, normalizing the acquired bearing vibration signal and then carrying out polar coordinate coding;
converting the polar coordinate coded bearing vibration signal into a two-dimensional bearing vibration signal based on a gram angle sum field, a gram angle difference field and a Markov transition field;
constructing a lightweight neural network, and training the neural network by using a two-dimensional bearing vibration signal;
and converting the bearing vibration signal to be detected into a two-dimensional bearing vibration signal, and inputting the two-dimensional bearing vibration signal into the lightweight neural network which completes training to obtain a diagnosis result.
Further, fourier transform and wavelet transform are respectively carried out on the obtained bearing vibration signals, polar coordinate coding is carried out on the bearing vibration signals and the bearing vibration signals subjected to Fourier transform and wavelet transform, and data obtained after the polar coordinate coding is carried out on the bearing vibration signals are converted into two-dimensional data from one-dimensional data on the basis of a gram angle difference field; converting the bearing vibration signal subjected to wavelet transformation from one-dimensional data to two-dimensional data based on a gram angle and a field; converting the bearing vibration signal from one-dimensional data to two-dimensional data based on a Markov transition field; and splicing the obtained three two-dimensional data on the channel and then using the spliced three-dimensional data as the input of the lightweight neural network.
Further, the process of polar encoding is represented as:
wherein phi is i Is a' i Based on the polar coordinates encoded angle cosine; a' i For normalized acceleration values, A 'is the set of all normalized acceleration values, a' i Is the ith element in the set A'; r is i The radius is coded based on the polar coordinates; t is t i Is a timestamp and N' is a constant factor.
Further, the bearing vibration signal which is converted into two dimensions after being subjected to polar coordinate coding based on the gram angle sum field and the gram angle difference field comprises:
converting the angle cosine coded by the polar coordinates into two-dimensional data according to the gram angle and the field, wherein the elements of the ith row and the jth column of the two-dimensional data are represented as follows:
GASF(i,j)＝sin(Φ i -Φ j )
converting the angle cosine coded by the polar coordinates into two-dimensional data according to the gram angle difference field, wherein the elements of the ith row and the jth column of the two-dimensional data are represented as follows:
GADF(i,j)＝cos(Φ i +Φ j )
wherein phi is i Representing the angle cosine of the ith normalized signal after being coded based on the polar coordinates; phi j Express the j normalized postambleThe number is based on the polar-coded angle cosine.
Further, the method for converting the polar coordinate coded bearing vibration signal into a two-dimensional bearing vibration signal based on the Markov transition field comprises the following steps: dividing the data coded by polar coordinates into Q quantile boxes to construct a QxQ Markov transition matrix, wherein the quantile boxes of the data on the time stamp i and the time stamp j are Q i And q is j The element in the ith row and the jth column in the Markov transition matrix represents q i →q j The transition probability of (2).
Further, before the two-dimensional signal is input into the lightweight neural network, the channels of the image are normalized one by adopting Gaussian distribution with the mean value of 0 and the standard deviation of 1.
Furthermore, the lightweight neural network comprises sixteen layers of stacked structures, the first layer is a 3 × 3 convolutional layer, the second to twelfth layers are feature extraction layers, the thirteenth layer and the fifteenth layer are 1 × 1 convolutional layers, the fourteenth layer is a pooling layer, and the sixteenth layer is a full connection layer.
Furthermore, each feature extraction layer is formed by an Antisym module or the Antisym module and a separation convolution; the data input into the feature extraction layer is superposed with the input of the feature extraction layer as the output of the feature extraction layer after sequentially passing through an Antisym module and the output after separation and convolution processing; when the step length of the feature extraction layer is 1, the feature extraction layer is formed by cascading two anti modules, and when the step length of the feature extraction layer is 2, the feature extraction layer is formed by cascading the anti modules, the separation convolution layer and the anti modules.
Furthermore, the Antisym module comprises a forward branch and a reverse branch, the input data respectively passes through the forward branch and the direction branch to extract features, the two branches are spliced together to be used as the output of the Antisym module, and the method specifically comprises the following steps:
if the number of two-dimensional data channels of data input into the light weight neural network is C, the height is H and the width is W, the number of output characteristic channels of the light weight neural network is C ', the height is H ' and the width is W ', 1 × 1 point convolution and 3 × 3 separation convolution are adopted in a forward branch and a backward branch to process the data, and the processing process of the forward branch is represented as:
Y 1 ＝Φ(Y′ 1 )；
the process of the backward branch is represented as:
Y′ 2 ＝Φ(X)；
where Φ () represents a separate convolution operation; x represents data input to the lightweight neural network; PConv1 (i) represents a point convolution operation with the ith convolution kernel in the forward branch; b 1 Representing a bias term; PConv2 (i) represents a point convolution operation with the ith convolution kernel in the backward branch.
The rolling bearing fault diagnosis method based on the dimension expansion vibration signal and the machine vision can effectively realize the visualization of the vibration signal, and provide RGB three channels for neural network learning, thereby realizing the accurate identification and diagnosis of the rolling bearing fault through the machine vision method with relatively small samples.
Drawings
FIG. 1 is a schematic diagram of a lightweight neural network according to the present invention;
FIG. 2 is a schematic structural diagram of the Antisym bottleeck of the present invention;
FIG. 3 is a schematic diagram of the comparison of the accuracy of TOP-1 on the MiniImageNet dataset for the present invention and the prior art;
FIG. 4 is a schematic diagram illustrating a sliding manner of a window according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of the process of expanding the dimension of the original data according to the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention provides a bearing fault diagnosis method based on a lightweight neural network and dimension expansion, which specifically comprises the following steps:
acquiring a historical bearing vibration signal, normalizing the acquired bearing vibration signal and then carrying out polar coordinate coding;
converting the polar coordinate coded bearing vibration signal into a two-dimensional bearing vibration signal based on a gram angle sum field, a gram angle difference field and a Markov transition field;
constructing a lightweight neural network, and training the neural network by using a two-dimensional bearing vibration signal;
and converting the bearing vibration signal to be detected into a two-dimensional bearing vibration signal, and inputting the two-dimensional bearing vibration signal into the lightweight neural network which completes training to obtain a diagnosis result.
Since the rolling bearing data belongs to a small sample, in the present invention, a large visual data set ImageNet2012 or larger is not used. Instead, we benchmark on a smaller scale dataset MiniImageNet. It is desirable to explore the behavior of neural networks designed for locks on a smaller set of data (in the case of small samples), with the results shown in table 2; it can be seen that AntisymNet can reduce the number of parameters more than the latest lightweight neural network while ensuring the accuracy.
Although deep learning can directly process one-dimensional data, current deep learning methods mainly process two-dimensional structural data, especially in the fields of Computer Vision (CV) and Natural Language Processing (NLP). In addition, the effect of directly classifying the one-dimensional signals by using deep learning is not good. Therefore, it is necessary to convert the one-dimensional sequence data into two-dimensional (image) data for processing by the convolutional neural network.
The vibration signal of the rolling bearing belongs to a one-dimensional signal, although the CNN has strong image recognition capability, the one-dimensional signal cannot be directly recognized, and the invention considers that the time-frequency sequence of the vibration signal is encoded into an image so as to allow the machine to recognize, classify and learn the structure and the mode in a visual mode.
The invention adopts a window sliding mode to carry out pre-interception on an original vibration signal sample, as shown in figure 4. In order to facilitate calculation and identification of the lightweight neural network, the window size is set to 2048.
For the pre-intercepted original time domain signal, three preprocessing modes are adopted: 1. fast Fourier transform, 2, wavelet transform, 3, do not process.
The preprocessed signals are normalized using gaussian distributions (mean 0, standard deviation 1) to speed up model convergence. The normalization process is represented as:
wherein output [ channel ] represents a normalized output signal; input [ channel ] represents a preprocessed signal, namely a signal which needs to be normalized; mean [ channel ] represents the mean value of the signal; std [ channel ] represents the standard deviation of the signal.
Converting the preprocessed time-frequency signals into two-dimensional arrays by adopting the modes of GASF, GADF and MTF respectively to obtain three single-channel images; and finally, performing channel fusion on the three single-channel images to finally obtain a two-dimensional image for mapping the one-dimensional signal information, as shown in fig. 5.
And making the finally processed image data into an image data set, training by using the designed AntisymNet lightweight neural network, and using the trained network for diagnosis and identification.
In the actual working environment, the obtained fault characteristic signals of the rolling bearing are mostly non-stationary sequence signals, so that the noise with a large degree cannot be avoided, and the original fault characteristic signals are covered. Therefore, preprocessing of the signal is very important. Since the Fourier Transform (FT) is for a continuous signal, it is not applicable to a discrete signal. The actually measured vibration signal is discrete, and a Discrete Fourier Transform (DFT) is generally used to perform a fourier transform on the discrete signal. The forward and inverse fourier transforms include:
wherein DFT [ x (N) ] denotes performing a discrete fourier transform on an input sequence x (N), N denotes the number of transform points of the fourier transform; IDFT [ X (k) ] represents performing an inverse discrete fourier transform on sequence X (k).
The discrete wavelet transform resembles the discrete fourier transform for a discrete signal of N points (N = 2) J ) Then, there are:
wherein:
wherein J =0,1, \8230;, J-1, and k =0,1, \8230;, 2 j -1; f (x) represents data obtained after wavelet transform,
a scale function representing a wavelet transform,
a scale function representing a Haar wavelet; psi j,k (x) A wavelet function representing a wavelet transform;
representing a scale function for a wavelet transform
The conjugation operation is carried out, and the operation,
indicating that the wavelet function of the wavelet transform is subjected to a conjugation operation, the conjugation operation may be omitted if the scale function and the wavelet function are real functions.
The DFT is mathematically feasible, but the computer is more complex, especially in embedded devices where the source is limited. Therefore, the invention adopts the fast Fourier transform which is more friendly to hardware equipment and the discrete wavelet transform.
The preprocessed time-frequency signals are coded into two-dimensional images in a dimension expansion mode, so that the machine vision is allowed to identify, classify and learn structures and modes.
The vibration signal of the rolling bearing is generally a one-dimensional time series, A = { a = { (a) } 1 ,a 2 ,…,a n }，a n The acceleration value of the nth sampling point is obtained; normalizing the values of the one-dimensional vibration signal sequence to make all the values in the A sample set be [ -1,1]As shown in the formula (1),
or between [0,1], as shown in equation (2),
in a Cartesian coordinate, a horizontal axis of a one-dimensional vibration signal is time, and a vertical axis of the one-dimensional vibration signal is an acceleration value; converting the horizontal coordinate code of the normalized vibration signal sequence into a radius in a polar coordinate, converting the vertical coordinate code into an angle cosine in the polar coordinate, and recoding the time sequence in the polar coordinate by using a formula (3);
wherein, a' i For normalized acceleration values, A' is for all normalized acceleration values; r is i Representing the radius in the polar coordinates after conversion; phi is a i For the coded angle cosine, t i Is the timestamp and N' is a constant factor to regularize the span of the polar coordinate system. The code mapping of equation (3) has two important properties, first, it is bijective, because when it comes to
∈[0，π]Time of flight
Is monotonic, secondly, given a time series, the proposed mapping yields one and only one result in a polar coordinate system with a unique inverse mapping.
By taking into account trigonometric functions and/or differences between the points after conversion of the scaled time series to a polar coordinate system, the angular view can easily be used to identify the time dependencies within the different time intervals. On the basis of three modes of a Gram Angle Sum Field (GASF), a Gram Angle Difference Field (GADF) and a Markov Transition Field (MTF), as many data sets as possible are obtained through simple improvement and data processing for training and identification of the light weight neural network, which are as follows:
(1) Glatiramer angle sum/difference field
GAF contains a temporal correlation because G (i, j | | i-j | = k) represents a relative correlation of the superposition/difference with respect to the direction of time interval k. The main diagonal Gi, i is a special case when k =0, which contains the original value/angle information. From the main diagonal, the time series can be reconstructed by using the high-level features learned by the deep neural network. GAF is divided into GASF and GADF;
GASF (i, j) is expressed as the glam angle and the i row, j column,
GASF(i,j)＝sin(Φ i -Φ j ) (4)
GADF (i, j) is represented as the i-th row, j-th column,
GADF(i,j)＝cos(Φ i +Φ j ) (5)
(2) Markov transition field
By dividing the data (magnitude) into Q quantile bins, a QxQ Markov transition matrix (W) is constructed. The quantile bin containing the data at time stamps i and j (time axis) is q i And q is j (q∈[1,Q]). M in MTF ij Denotes q i →q j The transition probability of (2). That is, the matrix W containing the transition probability on the amplitude axis is expanded into the MTF matrix by considering the time position.
In the embodiment, data obtained by polar coordinate encoding of a bearing vibration signal is converted from one-dimensional data to two-dimensional data based on a gram angular difference field; converting the bearing vibration signal subjected to wavelet transformation from one-dimensional data to two-dimensional data based on a gram angle and a field; converting the bearing vibration signal from one-dimensional data to two-dimensional data based on a Markov transition field; and splicing the obtained three two-dimensional data on the channel to be used as the input of the lightweight neural network.
(3) Window smoothing
The original vibration signal samples are valued in a sliding window mode, the window size is set to be 2048 for calculation and identification of a lightweight neural network, and the size of the converted picture is set to be 224 multiplied by 224. The step size of the window sliding is 400 (or an integer multiple of the sampling period) in order to get as many training samples as possible.
(4) Image normalization
The method can be used for converting the one-dimensional vibration signals by adopting the Gramami angle and/or difference field or the Markov transition field, and the specific use mode is determined according to the actual situation, and experiments prove that better precision can be achieved by adopting the MTF. During neural network training, data standardization needs to be performed on the converted images again, and the images are standardized one by adopting Gaussian distribution (the mean value is 0, and the standard deviation is 1) so as to accelerate model convergence.
The specific formula (6) is as follows:
(5) Training and use of neural networks
Dividing the finally processed image data into a data set and a test set, wherein the proportion is 8; training was performed using the AntisymNet lightweight neural network. And using the trained network for diagnosis and identification.
And aiming at the vibration signal data with the fault state label, attaching the corresponding fault label to the two-dimensional image which realizes the dimension expansion through the algorithm. And selecting the designed Antisym Net lightweight neural network for training. And setting training parameters, and training Antisym Net used for visualizing the vibration signal so as to realize fault diagnosis of the rolling bearing.
Deep convolutional neural networks consist of a large number of convolutions, which can result in a large computational cost. With the need to deploy neural networks on embedded devices, a series of methods for studying compact deep neural networks, such as network pruning, low-level quantization, knowledge distillation, etc., have been proposed in recent years. In addition, the redesign of the convolution module has great potential in establishing a lightweight neural network with fewer parameters and calculations. Generally speaking, for vibration signals, more advanced features are needed to improve the accuracy of classification and identification, and the invention designs an efficient convolution module, which adopts a shallower and wider architecture. As shown in fig. 1, the present invention is called a convolution module (or called Antisym module) of a lightweight neural network. As in fig. 1, the antilymmodule is divided into two branches, a forward branch and a backward branch.
Giving input data in practice
Where C is the number of input channels, and H and W are the height and width of the input profile. Presume this time output characteristic diagram
Where C ' is the number of channels output, and H ' and W ' are the height and width of the output signature. To achieve this effectively, we use a 1 × 1 point convolution for the normal convolution of the two branches, and the size of the separation convolution is set to 3 × 3. Expressed as:
Y 1 ＝Φ(Y′ 1 )
wherein Y' 1 Representing the inherent characteristics of the generation, i represents the i-th convolution kernel, PConv1 represents the point convolution used by the forward branch, phi represents the separate convolution operation, Y 1 Representing the output profile of the mapping branch.
The reverse branch is represented as:
Y′ 2 ＝Φ(X)
wherein Y' 2 Representing intermediate features generated by separate convolutions, PConv2 representing the point convolution used by the inverse branch, Y 2 Representing the output profile of the backward branch.
Thus, the output signature graph may be represented as:
wherein
Represents the stitching operation (feature map stitching) of the feature map. Note that the number of point convolutions used by both branches is C'/2.
Utilizing the advantages of the Antisym module, an Antisym bottomblock (abbreviated as a-bneck) specially designed for small-sized CNNs is designed, as shown in fig. 2, and is similar to basic residual block of ResNet. By using the stack design of the Ghost bottleeck for reference, the first Antisym module serves as an expansion layer to increase the number of channels; the second Antisym module is used to reduce the number of channels. Similarly, shortcut is used between the input and output of the two anti modules.
The left side in fig. 2 shows the structure of a feature extraction layer (Antisym bottomenck, or a-bneck for short) when the step size is 1, that is, when the step size is 1, the feature extraction layer is formed by cascading two Antisym modules, and the output of the last Antisym module is superposed with the input of the first Antisym module and then used as the output of the Antisym module with the step size of 1; the right side of fig. 2 shows the structure of the Antisym module when the step size is 2, that is, when the step size is 2, the feature extraction layer is formed by cascading the Antisym module, the separation convolutional layer and the Antisym module, and the output of the last Antisym module is superposed with the input of the first Antisym module to be used as the output of the Antisym module with the step size of 2.
According to the designed Antisym bottleeck, the Antisym Net shown in the table 1 is built in the embodiment. The first column Input in table 1 represents the change in shape of the output profile of each layer of the AntisymNet; the second column Operator represents the block structure and the size of the structure to be experienced by each feature layer, e.g., "Conv2d,3 × 3" denotes Conv2d passing 3 × 3; the third column and the fourth column respectively represent the number of channels after the rising of the inverse residual error structure in the bneck and the number of channels of the characteristic layer when the inverse residual error structure in the bneck is input; the fifth column SE represents whether or not attention is drawn in this layer, the value representing the sensitivity of the attention; the seventh column s represents the step size used for each block structure.
TABLE 1 network architecture
In this example, the first layer of the AntisymNet is a standard convolutional layer, using a common 3 × 3 convolution, with 16 filters. And then 11 layers of feature extraction layers are arranged, the number of channels is increased step by step, each layer is composed of anti system bottleeck, and anti system modules are used as building blocks. Thanks to the advantages of the anti module designed herein, there is no need to adopt an overly deep network architecture, which not only reduces the amount of computation, but also reduces the unnecessary computation accompanying the feed forward and back propagation processes.
TABLE 2 comparison of parameters and accuracy
To prove the advantages of the network designed in this embodiment, the designed network is compared with the lightweight neural network which is popular at present, and MiniImageNet is used as a data set, and the results are shown in table 2.
For visual presentation, this example also uses a scatter plot as shown in FIG. 3. All test results are typical results.
As can be seen from the results, generally larger flo results in higher accuracy, which suggests its effectiveness; the AntisymNet provided by the present embodiment is consistently superior to other competitors at various levels of computational complexity, thanks to the superiority of the Antisym module.
In order to better illustrate the effect achieved by the present invention, the proposed bearing fault diagnosis method based on the dimension expansion vibration signal and machine vision is further described below with reference to examples.
For bearing fault diagnosis, the bearing data of the Kaiser Sichu university are used as raw data, and the raw data are one-dimensional vibration data under different rotating speeds or loads. 10 groups of high-quality sample data are selected as test data to ensure the effectiveness of the experiment. The data length of the data storage device comprises 9 fault data and 1 normal data, and each data length is about 480000 sample values. The original vibration signal sample is pre-intercepted in a window sliding mode. In order to facilitate calculation and identification of the lightweight neural network, the window size is set to 2048.
The processing of the pre-intercepted sample signal is divided into 3 steps: firstly, preprocessing signals, and respectively performing fast Fourier transform, wavelet transform and no processing on the original signals; then carrying out dimensionality expansion on the three preprocessed signals to respectively obtain three single-channel images; and finally, carrying out channel fusion on the three single-channel images to finally obtain a two-dimensional image for mapping the one-dimensional signal information.
The size of the GAF or MTF converted picture is 224 × 224. To get as many training samples as possible, the step size of the window sliding is set to 400, approximately one cycle. 10000 pictures are finally generated as a bearing fault diagnosis data set, 1000 pictures are generated for each category, and a training set and a test set are calculated according to the following steps of 9:1 division.
For this case, the test results are shown in table 3, and it can be seen that the accuracy of our method is as high as 99.79% in the ten types of fault diagnosis.
TABLE 3 Fault diagnosis Performance
In the description of the present invention, it is to be understood that the terms "coaxial", "bottom", "one end", "top", "middle", "other end", "upper", "one side", "top", "inner", "outer", "front", "center", "both ends", and the like, indicate orientations or positional relationships based on those shown in the drawings, and are only for convenience of description and simplicity of description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and thus, are not to be construed as limiting the present invention.
In the present invention, unless otherwise expressly stated or limited, the terms "mounted," "disposed," "connected," "fixed," "rotated," and the like are to be construed broadly, e.g., as meaning fixedly connected, detachably connected, or integrally formed; can be mechanically or electrically connected; the terms may be directly connected or indirectly connected through an intermediate, and may be communication between two elements or interaction relationship between two elements, unless otherwise specifically limited, and the specific meaning of the terms in the present invention will be understood by those skilled in the art according to specific situations.
Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.
Claims (9)
1. The bearing fault diagnosis method based on the lightweight neural network and the dimension expansion is characterized by comprising the following steps:
acquiring a historical bearing vibration signal, normalizing the acquired bearing vibration signal and then carrying out polar coordinate coding;
converting the polar coordinate coded bearing vibration signal into a two-dimensional bearing vibration signal based on a gram angle sum field, a gram angle difference field and a Markov transition field;
constructing a lightweight neural network, and training the neural network by using a two-dimensional bearing vibration signal;
and converting the bearing vibration signal to be detected into a two-dimensional bearing vibration signal, and inputting the two-dimensional bearing vibration signal into the lightweight neural network which completes training to obtain a diagnosis result.
2. The bearing fault diagnosis method based on the lightweight neural network and the dimension expansion as claimed in claim 1, wherein the fourier transform and the wavelet transform are performed on the obtained bearing vibration signals, the bearing vibration signals and the bearing vibration signals subjected to the fourier transform and the wavelet transform are subjected to polar coordinate coding, and data obtained by subjecting the bearing vibration signals to the polar coordinate coding is converted from one-dimensional data to two-dimensional data based on a gram angle difference field; converting the bearing vibration signal subjected to wavelet transformation from one-dimensional data to two-dimensional data based on a gram angle and a field; converting the bearing vibration signal from one-dimensional data to two-dimensional data based on a Markov transition field; and splicing the obtained three two-dimensional data on the channel to be used as the input of the lightweight neural network.
3. The bearing fault diagnosis method based on the light weight neural network and the dimension expansion as claimed in claim 1 or 2, wherein the process of polar coordinate encoding is represented as:
wherein phi is i Is a' i Encoding the angle cosine based on the polar coordinates; a' i For normalized acceleration values, A 'is the set of all normalized acceleration values, a' i Is the ith element in the set A'; r is i The radius is coded based on the polar coordinates; t is t i Is a time stamp and N' is a constant factor.
4. The bearing fault diagnosis method based on the lightweight neural network and the dimension expansion as claimed in claim 1 or 2, wherein the step of converting the polar coordinate-coded bearing vibration signal into the two-dimensional bearing vibration signal based on the gram angle sum field and the gram angle difference field comprises:
converting the angle cosine coded by the polar coordinates into two-dimensional data according to the gram angle and the field, wherein the elements of the ith row and the jth column of the two-dimensional data are represented as follows:
GASF(i,j)＝sin(Φ i -Φ j )
converting the angle cosine coded by the polar coordinates into two-dimensional data according to the gram angle difference field, wherein the elements of the ith row and the jth column of the two-dimensional data are represented as follows:
GADF(i,j)＝cos(Φ i +Φ j )
wherein phi i Representing the angle cosine of the ith normalized signal after being coded based on the polar coordinates; phi j Indicating that the j-th normalized signal is based on the polar-coded angle cosine.
5. The bearing fault diagnosis method based on the lightweight neural network and the dimension expansion as claimed in claim 1 or 2, wherein the step of converting the polar coordinate-coded bearing vibration signal into a two-dimensional bearing vibration signal based on the markov transition field comprises:
dividing the data coded by polar coordinates into Q quantile boxes to construct a QxQ Markov transition matrix, wherein the quantile boxes of the data on the time stamp i and the time stamp j are Q i And q is j The element in the ith row and the jth column in the Markov transition matrix represents q i →q j The transition probability of (2).
6. The bearing fault diagnosis method based on the lightweight neural network and the dimension expansion as claimed in claim 1, wherein before the two-dimensional signal is input into the lightweight neural network, the channels of the image are normalized one by using a gaussian distribution with a mean value of 0 and a standard deviation of 1.
7. The bearing fault diagnosis method based on the lightweight neural network and the dimension expansion as claimed in claim 1, wherein the lightweight neural network comprises sixteen layers stacked, the first layer is a 3 x 3 convolution layer, the second to twelfth layers are feature extraction layers, the thirteenth layer and the fifteenth layer are both 1 x 1 convolution layers, the fourteenth layer is a pooling layer, and the sixteenth layer is a full-connection layer.
8. The bearing fault diagnosis method based on the light-weight neural network and the dimension expansion as claimed in claim 1, wherein each feature extraction layer is composed of an Antisym module or an Antisym module and a separation convolution; the data input into the feature extraction layer is superposed with the input of the feature extraction layer as the output of the feature extraction layer after sequentially passing through an Antisym module and the output after separation and convolution processing; and when the step length of the feature extraction layer is 1, the feature extraction layer is cascaded by two anti modules, and when the step length of the feature extraction layer is 2, the feature extraction layer is cascaded by the anti modules, the separation convolutional layer and the anti modules.
9. The bearing fault diagnosis method based on the light-weight neural network and the dimension expansion as claimed in claim 8, wherein the anti module comprises a forward branch and a reverse branch, the input data respectively passes through the forward branch and the reverse branch to extract features, and the two branches are spliced together to be used as the output of the anti module, and the method specifically comprises the following processes:
if the number of two-dimensional data channels of data input into the light weight neural network is C, the height is H and the width is W, the number of output characteristic channels of the light weight neural network is C ', the height is H ' and the width is W ', 1 × 1 point convolution and 3 × 3 separation convolution are adopted in a forward branch and a backward branch to process the data, and the processing process of the forward branch is represented as:
Y 1 ＝Φ(Y′ 1 )；
the process of the backward branch is represented as:
Y′ 2 ＝Φ(X)；
where Φ () represents a separate convolution operation; x represents data input to the lightweight neural network; PConv1 (i) represents a point convolution operation with the ith convolution kernel in the forward branch; b is a mixture of 1 Representing a bias term; PConv2 (i) represents a point convolution operation with the ith convolution kernel in the backward branch.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211342324.4A CN115761398A (en) | 2022-10-31 | 2022-10-31 | Bearing fault diagnosis method based on lightweight neural network and dimension expansion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211342324.4A CN115761398A (en) | 2022-10-31 | 2022-10-31 | Bearing fault diagnosis method based on lightweight neural network and dimension expansion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115761398A true CN115761398A (en) | 2023-03-07 |
Family
ID=85354280
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211342324.4A Pending CN115761398A (en) | 2022-10-31 | 2022-10-31 | Bearing fault diagnosis method based on lightweight neural network and dimension expansion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115761398A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115979548A (en) * | 2023-03-16 | 2023-04-18 | 北京理工大学 | Method and system for diagnosing leakage of hydrogen system for vehicle, electronic device, and storage medium |
| CN116109897A (en) * | 2023-04-14 | 2023-05-12 | 中国科学院自动化研究所 | Robot fish sensor fault detection method and device based on airspace image fusion |
| CN117874503A (en) * | 2024-03-11 | 2024-04-12 | 沃德传动（天津）股份有限公司 | CNN-based bearing fault diagnosis method, electronic equipment and storage medium |
| CN117951604A (en) * | 2024-02-26 | 2024-04-30 | 兰州理工大学 | Rolling bearing fault high-efficiency diagnosis method and system |
-
2022
- 2022-10-31 CN CN202211342324.4A patent/CN115761398A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115979548A (en) * | 2023-03-16 | 2023-04-18 | 北京理工大学 | Method and system for diagnosing leakage of hydrogen system for vehicle, electronic device, and storage medium |
| CN116109897A (en) * | 2023-04-14 | 2023-05-12 | 中国科学院自动化研究所 | Robot fish sensor fault detection method and device based on airspace image fusion |
| CN116109897B (en) * | 2023-04-14 | 2023-08-15 | 中国科学院自动化研究所 | Robot fish sensor fault detection method and device based on airspace image fusion |
| CN117951604A (en) * | 2024-02-26 | 2024-04-30 | 兰州理工大学 | Rolling bearing fault high-efficiency diagnosis method and system |
| CN117874503A (en) * | 2024-03-11 | 2024-04-12 | 沃德传动（天津）股份有限公司 | CNN-based bearing fault diagnosis method, electronic equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Shao et al. | A deep learning approach for fault diagnosis of induction motors in manufacturing | |
| CN115761398A (en) | Bearing fault diagnosis method based on lightweight neural network and dimension expansion | |
| CN110490946B (en) | Text image generation method based on cross-modal similarity and antagonism network generation | |
| CN107316013B (en) | Hyperspectral image classification method based on NSCT (non-subsampled Contourlet transform) and DCNN (data-to-neural network) | |
| CN112651973A (en) | Semantic segmentation method based on cascade of feature pyramid attention and mixed attention | |
| CN110657984B (en) | Planetary gearbox fault diagnosis method based on reinforced capsule network | |
| Liao et al. | Wavelet transform based convolutional neural network for gearbox fault classification | |
| CN103400143A (en) | Data subspace clustering method based on multiple view angles | |
| CN104077742B (en) | Human face sketch synthetic method and system based on Gabor characteristic | |
| CN115083435B (en) | Audio data processing method and device, computer equipment and storage medium | |
| Kim et al. | Label-preserving data augmentation for mobile sensor data | |
| CN110929080A (en) | Optical remote sensing image retrieval method based on attention and generation countermeasure network | |
| CN115294563A (en) | 3D point cloud analysis method and device based on Transformer and capable of enhancing local semantic learning ability | |
| CN111401156A (en) | Image identification method based on Gabor convolution neural network | |
| CN115859142A (en) | Small sample rolling bearing fault diagnosis method based on convolution transformer generation countermeasure network | |
| CN114612902A (en) | Image semantic segmentation method, device, equipment, storage medium and program product | |
| CN116628605A (en) | Method and device for electricity stealing classification based on ResNet and DSCAttention mechanism | |
| CN115393690A (en) | Light neural network air-to-ground observation multi-target identification method | |
| CN117475216A (en) | Hyperspectral and laser radar data fusion classification method based on AGLT network | |
| CN117634459A (en) | Target content generation and model training method, device, system, equipment and medium | |
| Trevino-Sanchez et al. | Hybrid pooling with wavelets for convolutional neural networks | |
| Chung et al. | Filter pruning by image channel reduction in pre-trained convolutional neural networks | |
| CN115035408A (en) | Unmanned aerial vehicle image tree species classification method based on transfer learning and attention mechanism | |
| CN114677545A (en) | Lightweight image classification method based on similarity pruning and efficient module | |
| CN114627370A (en) | Hyperspectral image classification method based on TRANSFORMER feature fusion |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |