Bearing fault diagnosis method based on lightweight neural network and dimension expansion

Abstract

The application relates to the fields of machine vision and fault diagnosis, and particularly relates to a bearing fault diagnosis method based on a light-weight neural network and dimension expansion, which comprises the following steps: acquiring a historical bearing vibration signal, performing polar coordinate coding on the acquired bearing vibration signal after normalization; converting the bearing vibration signal subjected to the polar coordinate coding into a two-dimensional bearing vibration signal based on a Grahm angle field, a Grahm angle difference field and a Markov transition field; constructing a light-weight neural network, and training the neural network by using the two-dimensional bearing vibration signal; converting a bearing vibration signal to be detected into a two-dimensional bearing vibration signal, inputting the two-dimensional bearing vibration signal into the light-weight neural network which has been trained, and obtaining a diagnosis result; the application can effectively realize the visualization of the vibration signal, provide an RGB three-channel for neural network learning, and thus realize the accurate identification and diagnosis of rolling bearing faults through a relatively small sample machine vision method.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of machine vision and fault diagnosis, and particularly relates to a bearing fault diagnosis method based on a light neural network and dimension expansion. BACKGROUND

[0002] At present, the fault diagnosis methods based on rolling bearing vibration signals mainly include three types: the first type is to use envelope spectrum analysis or signal decomposition or a combination of both to extract the characteristic frequency of the fault after denoising the original signal, and then compare it with the typical characteristic frequency of each fault to determine the fault. This method has high applicability in engineering. However, this method can only achieve qualitative fault diagnosis and cannot achieve quantitative fault diagnosis. The second type is to use a combination of fault feature extraction and shallow machine learning to achieve denoising and fault feature extraction through time domain, frequency domain and time-frequency domain signal processing methods, and then use machine learning methods to achieve fault classification. However, in the case of complex working conditions and multiple fault types, the diagnosis effect of this method is not good, and the steps are complex. The third type is a deep learning-based method. Deep learning methods have deep structures and strong non-linear feature extraction capabilities, and can directly achieve fault feature extraction and pattern recognition in bearing vibration signals, especially for complex working conditions and end-to-end fault diagnosis. However, deep learning models require a large number of data samples for training, and bearing fault samples are often difficult to obtain (small sample), so how to achieve fault diagnosis under small samples is a problem to be solved.

[0003] Deep learning can automatically learn the internal relationship of data and extract higher-dimensional and more abstract data by building a multi-layer network, so that the learned features are more expressive. Its deep network structure can learn the most essential features from vibration signals to achieve fault diagnosis, avoiding the need for manual feature extraction in various signal processing methods. Among them, the convolutional neural network (CNN) is a special deep feedforward network. The CNN model mainly includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. However, in the network structure, in order to make the output more accurate and the feature extraction more rich, multiple convolutional layers and multiple pooling layers are usually used in the network model, and the more classic CNN models include LeNet-5, AlexNet, ZF-Net, VGGNet, GoogLeNet, ResNet, and DenseNet. However, most deep neural networks currently require large amounts of memory and computing resources, which cannot be deployed on more lightweight devices, reducing their applicability. At the same time, it is difficult to obtain enough fault samples in actual industrial environments, which makes it difficult to apply the optimized model, so achieving a breakthrough in deep learning for fault diagnosis under small samples is the main problem currently facing.

[0004] The processing of the vibration signal mostly adopts a signal processing method, for example, short-time Fourier, wavelet transform, modal decomposition and the like. However, the signal processed through these methods is difficult to train through a deep neural network, especially a convolutional neural network for machine vision. How to reasonably convert one-dimensional signal data so as to facilitate efficient training of the convolutional neural network is a problem to be solved. SUMMARY

[0005] The application provides a bearing fault diagnosis method based on a light-weight neural network and dimension expansion, and specifically comprises the following steps:

[0006] The historical bearing vibration signal is acquired, and the acquired bearing vibration signal is normalized and then polar coordinate coded;

[0007] The bearing vibration signal coded in the polar coordinates is converted into a two-dimensional bearing vibration signal based on a Gram angle and field, a Gram angle difference field and a Markov transition field;

[0008] A light-weight neural network is constructed, and the two-dimensional bearing vibration signal is used to train the neural network;

[0009] The bearing vibration signal to be detected is converted into a two-dimensional bearing vibration signal, and input into the trained light-weight neural network to obtain a diagnosis result.

[0010] Further, the acquired bearing vibration signal is respectively subjected to Fourier transform and wavelet transform, and the bearing signal and the bearing vibration signal subjected to the Fourier transform and the wavelet transform are polar coordinate coded. The data of the bearing vibration signal coded in the polar coordinates is converted from one-dimensional data into two-dimensional data based on the Gram angle difference field; the bearing vibration signal subjected to the wavelet transform is converted from one-dimensional data into two-dimensional data based on the Gram angle and field; the bearing vibration signal is converted from one-dimensional data into two-dimensional data based on the Markov transition field; and the three two-dimensional data obtained are spliced on a channel to serve as an input of the light-weight neural network.

[0011] Further, the process of the polar coordinate coding is represented as:

[0012]

[0013] wherein, φ i is a′ i the cosine after the polar coordinate coding; a′ i is the normalized acceleration value, A′ is a set of all normalized acceleration values, a′ i is the i-th element in the set A′; r i is the radius after the polar coordinate coding; t i is a time stamp, and N′ is a constant factor.

[0014] Further, the bearing vibration signal converted into two dimensions based on the Gram angle and field and the Gram angle difference field after polar coordinate coding includes:

[0015] The angle cosine after polar coordinate coding is converted into two-dimensional data according to the Gram angle field, and the element of the i-th row and the j-th column of the two-dimensional data is represented as:

[0016] GASF(i,j)=sin(Φ i -Φ j )

[0017] The angle cosine after polar coordinate coding is converted into two-dimensional data according to the Gram angle difference field, and the element of the i-th row and the j-th column of the two-dimensional data is represented as:

[0018] GADF(i,j)=cos(Φ i +Φ j )

[0019] Wherein, Φ i represents the angle cosine after polar coordinate coding based on the i-th normalized signal; Φ j represents the angle cosine after polar coordinate coding based on the j-th normalized signal.

[0020] Further, the bearing vibration signal converted into two dimensions based on the Markov transition field after polar coordinate coding includes: dividing the polar coordinate coded data into Q quantile boxes, constructing a QxQ Markov transition matrix, and the quantile boxes of the data on the time stamp i and the time stamp j are q i and q j The element of the i-th row and the j-th column in the Markov transition matrix represents the transition probability of q i →q j .

[0021] Further, before inputting the two-dimensional signal into the lightweight neural network, the channels of the image are standardized one by one using a Gaussian distribution with a mean of 0 and a standard deviation of 1.

[0022] Further, the lightweight neural network includes a sixteen-layer structure stack, the first layer is a 3x3 convolution layer, the second to twelfth layers are feature extraction layers, the thirteenth and fifteenth layers are both 1x1 convolution layers, the fourteenth layer is a pooling layer, and the sixteenth layer is a full connection layer.

[0023] Further, each feature extraction layer is composed of an Antisym module or an Antisym module and a split convolution; the data input into the feature extraction layer is sequentially processed by the Antisym module and the split convolution, and the output after superimposition of the processed data and the input of the feature extraction layer is taken as the output of the feature extraction layer; and when the step length of the feature extraction layer is 1, the feature extraction layer is composed of two cascaded Antisym modules, and when the step length of the feature extraction layer is 2, the feature extraction layer is composed of an Antisym module, a split convolution layer and an Antisym module.

[0024] Further, the Antisym module includes a forward branch and a reverse branch, the input data are extracted features by the forward branch and the reverse branch respectively, and the two branches are spliced together to be taken as the output of the Antisym module, and the process specifically includes the following processes:

[0025] If the two-dimensional data channel number of the data input into the lightweight neural network is C, the height is H and the width is W, the feature channel number of the output of the lightweight neural network is C', the height is H' and the width is W', the data are processed by 1*1 point convolution and 3*3 split convolution in the forward branch and the reverse branch, and the processing process of the forward branch is represented as:

[0026] Y1=Φ(Y′1);

[0027]

[0028] The processing process of the reverse branch is represented as:

[0029]

[0030] Y′2=Φ(X);

[0031] Wherein, Φ() represents a split convolution operation; X represents the data input into the lightweight neural network; PConv1(i) represents point convolution operation with the i-th convolution kernel in the forward branch; b1 represents a bias term; and PConv2(i) represents point convolution operation with the i-th convolution kernel in the reverse branch.

[0032] The vibration signal and machine vision-based rolling bearing fault diagnosis method based on dimension expansion provided by the application can effectively realize visualization of the vibration signal, provide RGB three channels for neural network learning, and thus accurately identify and diagnose the rolling bearing fault through a relatively small sample machine vision method. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 It is a schematic diagram of the lightweight neural network structure of the application;

[0034] Figure 2 Figure 1 is a schematic diagram of the Antisym bottleneck structure of the present application;

[0035] Figure 3 Figure 2 is a schematic diagram of the comparison of the precision of the present application and the prior art TOP-1 on the MiniImageNet dataset;

[0036] Figure 4 Figure 3 is a schematic diagram of the window sliding mode in the embodiment of the present application;

[0037] Figure 5 Figure 4 is a schematic diagram of the process of dimension expansion on the original data of the present application. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0039] The present application provides a bearing fault diagnosis method based on a lightweight neural network and dimension expansion, specifically comprising the following steps:

[0040] Obtain historical bearing vibration signals, and perform polar coordinate coding on the obtained bearing vibration signals after normalization;

[0041] Convert the bearing vibration signals after polar coordinate coding into two-dimensional bearing vibration signals based on the Gram angle field, the Gram angle difference field and the Markov transition field;

[0042] Construct a lightweight neural network, and train the neural network using the two-dimensional bearing vibration signals;

[0043] Convert the bearing vibration signals to be detected into two-dimensional bearing vibration signals, and input them into the trained lightweight neural network to obtain a diagnosis result.

[0044] Since the rolling bearing data belongs to a small sample, in the present application, a large visual data set ImageNet2012 or a larger data set is not used. Instead, a benchmark test is performed on a smaller data set MiniImageNet. It is hoped to explore the performance of the designed neural network on a smaller data set (in the case of a small sample), and the results are shown in Table 2; it can be seen that the AntisymNet can further reduce the parameter amount while ensuring the precision compared with the latest lightweight neural network.

[0045] Although deep learning can directly process one-dimensional data, current deep learning methods mainly process two-dimensional structured data, especially in the fields of computer vision (CV) and natural language processing (NLP). Moreover, the effect of directly using deep learning to classify one-dimensional signals is poor. Therefore, one-dimensional sequence data needs to be converted into two-dimensional (image) data for processing by a convolutional neural network.

[0046] Rolling bearing vibration signals belong to one-dimensional signals, although CNN has strong image recognition capabilities, it cannot directly recognize one-dimensional signals. In the present application, the time-frequency sequence of the vibration signal is encoded as an image to allow machine "vision" to recognize, classify and learn structures and patterns.

[0047] The present application adopts a window sliding method to pre-intercept the original vibration signal sample, as shown in Figure 4 In order to facilitate the calculation and recognition of a lightweight neural network, the window size is set to 2048.

[0048] Three preprocessing methods are used for the pre-intercepted original time-domain signal: 1. Fast Fourier transform, 2. Wavelet transform, 3. No processing.

[0049] The pre-processed signal is standardized (mean 0, standard deviation 1) using a Gaussian distribution to accelerate model convergence. The standardization process is represented as:

[0050]

[0051] Where output[channel] represents the standardized output signal; input[channel] represents the pre-processed signal, i.e. the signal that needs to be standardized; mean[channel] represents the mean of the signal; and std[channel] represents the standard deviation of the signal.

[0052] The pre-processed time-frequency signal is converted into a two-dimensional array using GASF, GADF and MTF, respectively, to obtain three single-channel images; finally, the three single-channel images are fused to obtain a two-dimensional image that maps one-dimensional signal information, as shown in Figure 5 .

[0053] The final processed image data is made into an image data set, and the designed AntisymNet lightweight neural network is used for training. The trained network is used for diagnosis and identification.

[0054] In actual working environment, the obtained rolling bearing fault characteristic signal is mostly non-stationary sequence signal, and cannot avoid containing a large degree of noise, which covers the original fault characteristic signal. Therefore, the pretreatment of the signal is very important. Since Fourier transform (FT) is for continuous signal, it is not suitable for discrete signal. The actually measured vibration signal is discrete, and the discrete Fourier transform (DFT) is generally used to perform Fourier transform on the discrete signal. Fourier forward transform and inverse transform include:

[0055]

[0056]

[0057] Wherein, DFT[x(n)] represents performing discrete Fourier transform on the input sequence x(n), and N represents the transform point number of Fourier transform; IDFT[X(k)] represents performing discrete Fourier inverse transform on the sequence X(k).

[0058] The discrete wavelet transform is similar to the discrete Fourier transform, and for N-point discrete signal (N=2 J ), there are:

[0059]

[0060] Wherein:

[0061]

[0062]

[0063] Wherein, j=0, 1, …, J-1, and k=0, 1, …, 2 j -1; f(x) represents data obtained after wavelet transform, represents a scale function of wavelet transform, represents a scale function of Haar wavelet; ψ j,k (x) represents a wavelet function of wavelet transform; represents a scale function of wavelet transform has been conjugated, represents that the wavelet function of wavelet transform has been conjugated, and if the scale function and the wavelet function are real functions, the conjugation can be omitted.

[0064] DFT is mathematically feasible, but computer calculation is relatively complex, especially in source-limited embedded devices. Therefore, the present application adopts fast Fourier transform which is more friendly to hardware devices, and discrete wavelet transform.

[0065] The time-frequency signal, after the above preprocessing, is encoded into a two-dimensional image through dimensional expansion to allow machine "vision" to recognize, classify, and learn structures and patterns.

[0066] The vibration signal of a rolling bearing is generally a one-dimensional time series, A={a1,a2,…,a…} n}, a n Let be the acceleration value of the nth sampling point; normalize the numerical values ​​of the one-dimensional vibration signal sequence so that all values ​​in sample set A are between [-1, 1], as shown in formula (1).

[0067]

[0068] Or it may be between [0,1], as shown in formula (2).

[0069]

[0070] In Cartesian coordinates, the horizontal axis of a one-dimensional vibration signal is generally time and the vertical axis is acceleration. The horizontal axis of the normalized vibration signal sequence is encoded into the radius in polar coordinates, and the vertical axis is encoded into the angle cosine in polar coordinates. The time series is then re-encoded in polar coordinates using formula (3).

[0071]

[0072] Where, a′ i Here, A' represents the normalized acceleration values, and r represents all normalized acceleration values. i φ represents the radius in the transformed polar coordinates. i The encoded angle cosine, t i It is a timestamp, and N′ is a constant factor to regularize the span of the polar coordinate system. The encoding mapping of formula (3) has two important properties. First, it is bijective because when When ∈[0,π] It is monotonic; secondly, given a time series, the proposed mapping produces one and only one result in a polar coordinate system with a unique inverse mapping.

[0073] After converting the scaled time series into a polar coordinate system, the angular perspective can be easily used to identify time correlations within different time intervals by considering the trigonometric sum / difference between points. Based on the Gramian Angular Summation Field (GASF), Gramian Angular Difference Field (GADF), and Markov Transition Field (MTF), as many data sets as possible are obtained through simple improvements and data processing for lightweight neural network training and recognition, as follows:

[0074] (1) Gramian Angular Sum / Difference Field

[0075] GAF provides a method to keep time dependence, when the position moves from the top left to the bottom right, the time increases. GAF contains time correlation because G(i,j||i-j| = k) represents the relative correlation of the sum / difference with respect to the direction of time interval k. The main diagonal line Gi,i is a special case when k = 0, which contains the original value / angle information. From the main diagonal line, the time series can be reconstructed using the high-level features learned by the deep neural network. GAF is divided into GASF and GADF;

[0076] GASF(i,j) represents the i-th row and j-th column of the Gramian Angular Sum Field,

[0077] GASF(i,j) = sin(Φ i -Φ j ) (4)

[0078] GADF(i,j) represents the i-th row and j-th column of the Gramian Angular Difference Field,

[0079] GADF(i,j) = cos(Φ i +Φ j ) (5)

[0080] (2) Markov Transition Field

[0081] By dividing the data (magnitude) into Q quantile bins, a QxQ Markov transition matrix (W) is constructed. The quantile bins containing data at time stamps i and j (time axis) are q i and q j (q ∈ [1, Q]). M ij in MTF represents the transition probability from q i to q j . That is, by considering the time position, the matrix W containing the transition probability on the amplitude axis is unfolded into the MTF matrix.

[0082]

[0083] In the embodiment, the data of the bearing vibration signal after polar coordinate coding is converted from one-dimensional data to two-dimensional data based on the Gram angle difference field; the bearing vibration signal after wavelet transform is converted from one-dimensional data to two-dimensional data based on the Gram angle and field; the bearing vibration signal is converted from one-dimensional data to two-dimensional data based on the Markov transition field; and the three obtained two-dimensional data are spliced on the channel to serve as the input of the lightweight neural network.

[0084] (3) Window smoothing

[0085] The original vibration signal sample is valued in a sliding window manner, the window size is set to 2048 for the convenience of calculation and identification of the lightweight neural network, and the converted picture size is set to 224x224. In order to obtain as many training samples as possible, the step length of window sliding is 400 (or an integer multiple of the sampling period).

[0086] (4) Image standardization

[0087] Either the Gram angle and difference field or the Markov transition field can be used for the conversion of one-dimensional vibration signals, and the specific method to be used is determined according to the actual situation. Experiments show that the use of MTF can achieve better accuracy. During neural network training, the converted image needs to be standardized again, and the Gaussian distribution is used to standardize each channel of the picture (mean value is 0 and standard deviation is 1) in order to accelerate the model convergence.

[0088] The specific formula (6) is as follows:

[0089]

[0090] (5) Training and use of neural network

[0091] The final processed image data is divided into a data set and a test set in a ratio of 8:2; the AntisymNet lightweight neural network is used for training. The trained network is used for diagnosis and identification.

[0092] For the vibration signal data with fault state labels, the two-dimensional image obtained by the above algorithm is pasted with its corresponding fault label. The designed Antisym Net lightweight neural network is selected for training. The training parameters are set, and the Antisym Net for visualizing the vibration signal is trained to realize the fault diagnosis of the rolling bearing.

[0093] Deep convolutional neural networks consist of a large number of convolutions, which leads to significant computational costs. With the increasing need to deploy neural networks on embedded devices, a series of methods for developing compact deep neural networks have been proposed in recent years, such as network pruning, low-order quantization, and knowledge distillation. Furthermore, redesigning convolutional modules holds great potential for building lightweight neural networks with fewer parameters and computations. Typically, for vibration signals, more advanced features are needed to improve classification and recognition accuracy. This invention designs an efficient convolutional module with a shallower and wider architecture. Figure 1 As shown, this invention refers to it as a lightweight neural network convolutional module (or Antisym module). Figure 1 Antisymmodule is divided into two branches: a forward branch and a reverse branch.

[0094] Given input data in practice Where C is the number of input channels, and H and W are the height and width of the input feature map. Assume the output feature map... Where C′ is the number of output channels, and H′ and W′ are the height and width of the output feature map. For efficient implementation, we use 1×1 pointwise convolutions for the two branches and set the size of the separable convolutions to 3×3. This is represented as:

[0095]

[0096] Y1=Φ(Y′1)

[0097] Where Y′1 represents the generated intrinsic features, i represents the i-th convolutional kernel, PConv1 represents the point convolution used in the forward branch, Φ represents the split convolution operation, and Y1 represents the output feature map of the mapping branch.

[0098] The reverse branch is represented as:

[0099] Y′2=Φ(X)

[0100]

[0101] Where Y′2 represents the intermediate features generated by the split convolution, PConv2 represents the point convolution used by the reverse branch, and Y2 represents the output feature map of the reverse branch.

[0102] Therefore, the output feature map can be represented as:

[0103]

[0104] in This represents the feature map stitching operation. Note that the number of point convolutions used in both branches is C′ / 2.

[0105] Leveraging the advantages of the Antisym module, an Antisym bottleneck (A-bneck) specifically designed for small CNNs was created, such as... Figure 2 As shown, this is similar to the basic residual block of ResNet. We borrowed the stacked design from Ghostbottleneck, where the first Antisym module acts as an extension layer, increasing the number of channels; the second Antisym module is used to reduce the number of channels. Similarly, a shortcut is used between the input and output of these two Antisym modules.

[0106] Figure 2 The left side of the middle section represents the structure of the feature extraction layer (Antisym bottleneck, or A-bneck for short) with a stride of 1. That is, when the stride is 1, the feature extraction layer consists of two cascaded Antisym modules, and the output of the last Antisym module is superimposed with the input of the first Antisym module as the output of the Antisym module with a stride of 1. Figure 2 The right side shows the structure of the Antisym module with a stride of 2. That is, with a stride of 2, the feature extraction layer is composed of an Antisym module, a split convolutional layer, and an Antisym module cascaded together. The output of the last Antisym module is superimposed with the input of the first Antisym module to serve as the output of the Antisym module with a stride of 2.

[0107] Based on the designed Antisym bottleneck, this embodiment constructs AntisymNet as shown in Table 1. In Table 1, the first column, Input, represents the shape change of the output feature map of each layer of AntisymNet; the second column, Operator, represents the block structure that the feature layer will experience each time and the size of the structure, for example, "Conv2d, 3×3" means passing through a 3×3 Conv2d; the third and fourth columns represent the number of channels after the inverse residual structure rises in the bneck and the number of channels of the feature layer when input to the bneck, respectively; the fifth column, SE, represents whether an attention mechanism is introduced in this layer, and the value represents the sensitivity of the attention mechanism; the seventh column, s, represents the stride used for each block structure.

[0108] Table 1 Network Architecture

[0109]

[0110] In this embodiment, the first layer of AntisymNet is a standard convolutional layer using a common 3×3 convolution with 16 filters. This is followed by 11 feature extraction layers, with the number of channels gradually increasing. Each layer consists of an Antisym bottleneck, with the Antisym module serving as a building block. Thanks to the advantages of the Antisym module designed in this paper, an excessively deep network architecture is not required. This not only reduces the computational load but also minimizes unnecessary computations during feedforward and backpropagation.

[0111] Table 2 Comparison of Parameter Quantity and Accuracy

[0112]

[0113] To demonstrate the advantages of the network designed in this embodiment, this embodiment compares the designed network with currently popular lightweight neural networks, using MiniImageNet as the dataset, and the results are shown in Table 2.

[0114] For clarity, this embodiment also uses a scatter plot for illustration, such as... Figure 3 All test results are typical.

[0115] The results show that larger FLOPs generally lead to higher accuracy, demonstrating its effectiveness. The AntisymNet provided in this embodiment consistently outperforms its competitors at various computational complexity levels, thanks to the superiority of the Antisym module.

[0116] To better illustrate the effects of this invention, the following examples further explain the proposed bearing fault diagnosis method based on dimensional expansion vibration signals and machine vision.

[0117] For bearing fault diagnosis, this example uses bearing data from Case Western Reserve University as the raw data, which consists of one-dimensional vibration data under different speeds or loads. Ten high-quality sample data sets were selected as test data to ensure the validity of the experiment. These include nine fault data sets and one normal data set, each with approximately 480,000 sample values. This paper uses a sliding window method to pre-truncate the raw vibration signal samples. To facilitate computation and recognition by the lightweight neural network, the window size is set to 2048.

[0118] The processing of the pre-captured sample signal consists of three steps: First, the signal is preprocessed by performing Fast Fourier Transform, Wavelet Transform, and no processing on the original signal respectively; then, the dimensions of these three preprocessed signals are expanded to obtain three single-channel images respectively; finally, the three single-channel images are fused to obtain a two-dimensional image that maps one-dimensional signal information.

[0119] The images converted from GAF or MTF are 224×224 in size. To obtain as many training samples as possible, the window sliding step size is set to 400, approximately one cycle. We ultimately generated 10,000 images as a bearing fault diagnosis dataset, with 1,000 images for each category. We divided the training and test sets in a 9:1 ratio.

[0120] For this case, the test results are shown in Table 3. As can be seen, our method achieves an accuracy rate of up to 99.79% in diagnosing ten types of faults.

[0121] Table 3 Fault Diagnosis Manifestations

[0122]

[0123] In the description of this invention, it should be understood that the terms "coaxial," "bottom," "one end," "top," "middle," "other end," "upper," "side," "top," "inner," "outer," "front," "center," "both ends," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0124] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "rotation," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0125] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A bearing fault diagnosis method based on lightweight neural networks and dimensionality expansion, characterized in that, Specifically, the following steps are included: Historical bearing vibration signals are acquired, and the acquired bearing vibration signals are normalized and then encoded in polar coordinates. Based on Gram angle sum field, Gram angle difference field and Markov transition field, the polar coordinate encoded signal is converted into a two-dimensional bearing vibration signal. A lightweight neural network was constructed and trained using two-dimensional bearing vibration signals. The lightweight neural network consists of a stacked sixteen-layer structure, with the first layer being... Convolutional layers, layers 2-12 are feature extraction layers, and layers 13 and 15 are... The system consists of convolutional layers, with the fourteenth layer being a pooling layer and the sixteenth layer being a fully connected layer. Each feature extraction layer is composed of an Antisym module, or an Antisym module and a separable convolution. The data input to the feature extraction layer is processed sequentially by the Antisym module and the separable convolution, and the output of the processed data is superimposed on the input of the feature extraction layer to serve as the output of the feature extraction layer. When the stride of the feature extraction layer is 1, the feature extraction layer consists of two cascaded Antisym modules; when the stride of the feature extraction layer is 2, the feature extraction layer consists of an Antisym module, a separable convolution layer, and an Antisym module cascaded. The Antisym module consists of a forward branch and a backward branch. The input data is processed through the forward and backward branches to extract features, and the two branches are concatenated as the output of the Antisym module. Includes the following processes: If the input data to the lightweight neural network has C two-dimensional data channels, H height, and W width, then the number of output feature channels of the lightweight neural network is... Gao Wei Width In forward and reverse branches, use Point convolution, Separating convolutions process the data; the forward branch's processing is represented as follows: ； ； The process of handling reverse branches is represented as follows: ； ； in, Indicates the separation convolution operation; This represents the data input to the lightweight neural network; This indicates that a point convolution operation is performed with the i-th convolution kernel in the forward branch; Indicates the bias term; This indicates that a point convolution operation is performed using the i-th convolution kernel in the reverse branch; The bearing vibration signal to be detected is converted into a two-dimensional bearing vibration signal and input into a trained lightweight neural network to obtain the diagnostic result.

2. The bearing fault diagnosis method based on lightweight neural network and dimensionality expansion according to claim 1, characterized in that, The acquired bearing vibration signal is subjected to Fourier transform or wavelet transform, and the bearing signal and the bearing vibration signal after Fourier transform or wavelet transform are encoded in polar coordinates. The data of the bearing vibration signal after polar coordinate encoding is converted from one-dimensional data to two-dimensional data based on Gram angle difference field; the bearing vibration signal after wavelet transform is converted from one-dimensional data to two-dimensional data based on Gram angle sum field; the bearing vibration signal is converted from one-dimensional data to two-dimensional data based on Markov transformation field; the three two-dimensional data are concatenated on the channel and used as input to the lightweight neural network.

3. The bearing fault diagnosis method based on lightweight neural networks and dimensionality expansion according to claim 1 or 2, characterized in that, The process of polar coordinate encoding is represented as follows: in, for Based on the angle cosine after polar coordinate encoding; The normalized acceleration value. The set of all normalized acceleration values. For set The i-th element; The radius is the result of polar coordinate encoding. It's a timestamp. It is a constant factor.

4. The bearing fault diagnosis method based on lightweight neural networks and dimensionality expansion according to claim 1 or 2, characterized in that, Based on Gram angles and field, and Gram angle difference field, the polar coordinate encoded bearing vibration signal is converted into a two-dimensional signal, including: Based on the Gram angle and field, the angle cosine of the polar coordinate encoding is converted into two-dimensional data. The element in the i-th row and j-th column of the two-dimensional data is represented as follows: Based on the Gram angle difference field, the angular cosine of the polar coordinate encoding is converted into two-dimensional data. The element in the i-th row and j-th column of the two-dimensional data is represented as follows: in, This represents the angle cosine of the i-th normalized signal after polar coordinate encoding; Let represent the angle cosine of the j-th normalized signal after polar coordinate encoding.

5. The bearing fault diagnosis method based on lightweight neural networks and dimensionality expansion according to claim 1 or 2, characterized in that, Converting bearing vibration signals from one-dimensional to two-dimensional data based on Markov transition fields includes: The bearing vibration signal data was divided into Q quantile bins to construct a... Markov transition matrix, timestamp and timestamp The quantile bins of the data are and The element in the i-th row and j-th column of the Markov transition matrix represents... The transition probability.

6. The bearing fault diagnosis method based on lightweight neural network and dimensionality expansion according to claim 1, characterized in that, Before inputting the two-dimensional signal into the lightweight neural network, the image channels are standardized one by one using a Gaussian distribution with a mean of 0 and a standard deviation of 1.

Images