CN111275007B - Bearing fault diagnosis method and system based on multi-scale information fusion - Google Patents

Bearing fault diagnosis method and system based on multi-scale information fusion Download PDF

Info

Publication number
CN111275007B
CN111275007B CN202010112157.9A CN202010112157A CN111275007B CN 111275007 B CN111275007 B CN 111275007B CN 202010112157 A CN202010112157 A CN 202010112157A CN 111275007 B CN111275007 B CN 111275007B
Authority
CN
China
Prior art keywords
layer
module
input
output end
signal
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.)
Active
Application number
CN202010112157.9A
Other languages
Chinese (zh)
Other versions
CN111275007A (en
Inventor
李沂滨
王代超
贾磊
高辉
宋艳
张天泽
胡晓平
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shandong University
Original Assignee
Shandong University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Shandong University filed Critical Shandong University
Priority to CN202010112157.9A priority Critical patent/CN111275007B/en
Publication of CN111275007A publication Critical patent/CN111275007A/en
Application granted granted Critical
Publication of CN111275007B publication Critical patent/CN111275007B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G01M13/04—Bearings
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G01M13/04—Bearings
    • G01M13/045—Acoustic or vibration analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Signal Processing (AREA)
  • Acoustics & Sound (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

The invention discloses a bearing fault diagnosis method and system based on multi-scale information fusion, which comprises the following steps: acquiring a vibration signal and a torque signal of a bearing to be subjected to fault diagnosis; carrying out Fourier transform on the obtained vibration signal and the torque signal of the bearing subjected to fault diagnosis; and inputting the vibration signal and the torque signal of the bearing subjected to fault diagnosis obtained after Fourier transform into a multi-scale information fusion fault diagnosis model, and outputting the fault type of the bearing to be subjected to fault diagnosis. The network structure provided by the disclosure can effectively extract complementary fault characteristics in the bearing vibration signal and the torque signal, and improves the accuracy of fault diagnosis to a great extent.

Description

Bearing fault diagnosis method and system based on multi-scale information fusion
Technical Field
The disclosure relates to the technical field of bearing fault diagnosis, in particular to a bearing fault diagnosis method and system based on multi-scale information fusion.
Background
The statements in this section merely provide background information related to the present disclosure and may not constitute prior art.
With the continuous development of intelligent manufacturing, the intelligent degree of an industrial system is higher and higher, and meanwhile, the industrial system is more and more complex, and the loss caused by equipment damage is larger and larger. Early fault detection not only allows for troubleshooting before significant economic losses are incurred, but also avoids major safety hazards. However, due to the complexity and non-linearity of industrial systems, it is difficult to build accurate mathematical models. Due to the rapid development of information technology, a large amount of operation data is generated in the industrial system, which contains a large amount of valuable device status information. For complex systems with high integration, data-driven based fault diagnosis methods have proven to be more efficient than manual models based on manual experience.
The traditional fault diagnosis and state monitoring method depends on health indexes such as current imbalance, overvoltage and the like, but faults can not be accurately judged and positioned at the early stage of weak fault signals and short duration. In the past decades, a variety of intelligent fault diagnosis methods such as a Support Vector Machine (SVM), an Artificial Neural Network (ANN), and a decision tree have been developed. Although these intelligent methods have achieved some success in fault diagnosis, there are problems in use:
1) they need to be used in conjunction with feature extraction methods, and the selection of features will largely influence the final classification result. In addition, the feature extraction network and the classifier are designed separately, which consumes a lot of time and cannot realize global optimization.
2) Most of the methods belong to shallow structures, and the effective feature representation and the nonlinear mapping relation of a complex system are difficult to learn.
In recent years, with the development of Deep Learning (DL), more and more studies have shown that DL can learn Deep feature expressions and nonlinear mapping relationships due to its Deep structure. In addition, unlike the conventional fault diagnosis method, DL can adaptively extract fault features and perform global optimization. The deep neural network can solve the problems in the traditional machine learning due to the deep network architecture, and is widely applied to the field of fault diagnosis. Common Deep learning methods include Convolutional Neural Network (CNN), Auto-Encoder (AE), Deep Belief Network (DBN), and the like.
In the course of implementing the present disclosure, the inventors found that the following technical problems exist in the prior art:
although DL has good performance in feature extraction, it has two main problems, which are two technical problems to be solved by the present application:
1) most DL methods use only a single source input. Currently, many researches in the field of fault diagnosis are focused on a single signal source. However, in a complex industrial system, it is difficult to obtain complete fault information, and using a single signal source will result in incomplete fault features extracted from the signal, and poor generalization capability of the network.
2) The network structure is single. In order to improve the performance of fault diagnosis, many researches only focus on increasing the depth of a network, so that not only can multi-scale fault features in signals not be extracted, but also gradient disappearance can be caused, parameters are difficult to update, and great obstacles are brought to the training of the network.
Disclosure of Invention
In order to solve the defects of the prior art, the present disclosure provides a bearing fault diagnosis method and system based on multi-scale information fusion;
in a first aspect, the present disclosure provides a bearing fault diagnosis method based on multi-scale information fusion;
the bearing fault diagnosis method based on multi-scale information fusion comprises the following steps:
acquiring a vibration signal and a torque signal of a bearing to be subjected to fault diagnosis;
carrying out Fourier transform on the obtained vibration signal and the torque signal of the bearing subjected to fault diagnosis;
and inputting the vibration signal and the torque signal of the bearing subjected to fault diagnosis obtained after Fourier transform into a multi-scale information fusion fault diagnosis model, and outputting the fault type of the bearing to be subjected to fault diagnosis.
In a second aspect, the present disclosure also provides an electronic device comprising a memory and a processor, and computer instructions stored on the memory and executed on the processor, which when executed by the processor, perform the method of the first aspect.
In a third aspect, the present disclosure also provides a computer-readable storage medium for storing computer instructions which, when executed by a processor, perform the method of the first aspect.
Compared with the prior art, the beneficial effect of this disclosure is:
1. the proposed data fusion network is adopted to respectively extract fault characteristics for Fourier transformation of current and torque signals, and the fault characteristics are subjected to feature layer fusion and finally applied to final fault classification, and the result shows that the method further improves the accuracy of fault diagnosis of the vibration signals. It was therefore concluded that: complementary information aiming at bearing faults exists in the vibration signal and the torque signal, and the performance of bearing fault diagnosis can be effectively improved through the fusion of the two data.
2. The network structure provided by the disclosure can effectively extract complementary fault characteristics in the bearing vibration signal and the torque signal, and improves the accuracy of fault diagnosis to a great extent.
Drawings
The accompanying drawings, which are included to provide a further understanding of the disclosure, illustrate embodiments of the disclosure and together with the description serve to explain the disclosure and are not to limit the disclosure.
Fig. 1 is a convolution-pooling layer in the 1D CNN of the first embodiment;
FIG. 2 is a multi-scale information fusion fault diagnosis model based on a residual convolutional neural network according to a first embodiment;
fig. 3 is a diagram showing the construction of an inclusion unit of the first embodiment;
FIG. 4 is a view showing an internal structure of a fusion module according to the first embodiment;
Detailed Description
It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the disclosure. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments according to the present disclosure. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.
English explanation:
BN layer, representing the block Normalization layer, Bach-Normalization;
the method comprises the steps that firstly, a bearing fault diagnosis method based on multi-scale information fusion is provided;
the bearing fault diagnosis method based on multi-scale information fusion comprises the following steps:
s1: acquiring a vibration signal and a torque signal of a bearing to be subjected to fault diagnosis;
s2: carrying out Fourier transform on the obtained vibration signal and the torque signal of the bearing subjected to fault diagnosis;
s3: and inputting the vibration signal and the torque signal of the bearing subjected to fault diagnosis obtained after Fourier transform into a multi-scale information fusion fault diagnosis model, and outputting the fault type of the bearing to be subjected to fault diagnosis.
As one or more embodiments, the obtaining step of the multi-scale information fusion fault diagnosis model includes:
s31: constructing a neural network model;
s32: constructing a training set; the training set is vibration signals and torque signals of known bearing fault types after Fourier transformation;
s33: and inputting the training set into a neural network model for training, and obtaining the trained neural network model after the training is finished, namely the multi-scale information fusion fault diagnosis model.
As one or more embodiments, as shown in fig. 2, in S31, constructing a neural network model; the neural network model specifically includes:
two branches in parallel: a first branch and a second branch; the output end of the first branch and the output end of the second branch are both connected with the input end of the fusion module, and the output end of the fusion module is connected with the classifier;
the first branch, comprising: the device comprises a first input layer, a first ResidualUnit module, a first MaxPholing module, a second ResidualUnit module, a second MaxPholing module, a third ResidualUnit module, a third MaxPholing module and a first output layer which are connected in sequence;
the first input layer is used for inputting a vibration signal;
after an output end signal of the first ResidualUnit module passes through the first MaxPholing module, the output end signal of the first ResidualUnit module is used as an input signal of the second ResidualUnit module;
after an output end signal of the second ResidualUnit module passes through the second MaxPholing module, the output end signal of the second ResidualUnit module is used as an input signal of the third ResidualUnit module;
after an output end signal of the third ResidualUnit module passes through the third MaxPholing module, the output end signal is used as an input signal of the first output layer;
the first output layer is a BN + Relu function layer, the input end of the first output layer is connected with the output end of the third MaxPholing layer, and the output end of the first output layer is connected with the input end of the fusion module;
the second branch circuit includes: the second input layer, the fourth ResidualUnit module, the fourth Maxboosting module, the fifth ResidualUnit module, the fifth Maxboosting module, the sixth ResidualUnit module, the sixth Maxboosting module and the second output layer are connected in sequence;
the second input layer is used for inputting a torque signal;
after passing through the fourth MaxPholing module, an output end signal of the fourth ResidualUnit module is used as an input signal of the fifth ResidualUnit module;
after an output end signal of the fifth ResidualUnit module passes through the fifth Maxboosting module, the output end signal is used as an input signal of the sixth ResidualUnit module;
after an output end signal of the sixth ResidualUnit module passes through the sixth MaxPholing module, the output end signal is used as an input signal of the second output layer;
the second output layer is a BN + Relu function layer, the input end of the second output layer is connected with the output end of the sixth MaxPholing layer, and the output end of the second output layer is connected with the input end of the fusion module;
further, the first residaualunit module performs a first feature extraction assumption on the input vibration signal to obtain M1 features, and M1 features pass through a first MaxPooling layer to obtain M2 features;
the second Residual Unit module performs second feature extraction on the M2 features to obtain N1 features, and the N1 features pass through a second Max boosting layer to obtain N2 features;
the third ResidualUnit module carries out third-time feature extraction on the N2 features to obtain P1 features, and the P1 features pass through a third Max boosting layer to obtain P2 features;
p2 characteristics are obtained after the P2 characteristics pass through the first output layer, wherein M1, M2, N1, N2, P1 and P2 are positive integers.
Further, the fourth residaualunit module performs a first feature extraction assumption on the input torque signal to obtain X1 features, and X1 features pass through a fourth MaxPooling layer to obtain X2 features;
performing secondary feature extraction on the X2 features by using a fifth ResidualUnit module to obtain Y1 features, and obtaining Y2 features after the Y1 features pass through a fifth Max boosting layer;
a sixth ResidualUnit module carries out third-time feature extraction on Y2 features to obtain Z1 features, Z1 features pass through a sixth Max boosting layer to obtain Z2 features,
z2 characteristics are obtained after the Z2 characteristics pass through the second output layer, wherein X1, X2, Y1, Y2, Z1 and Z2 are positive integers.
Further, the fusion module performs feature fusion on the P2 features and the Z2 features to obtain the features to be classified of the bearing.
As one or more embodiments, the functions of the first, second, third, fourth, fifth and sixth Residual Unit modules are the same, and all the modules are used for extracting multi-scale features in a signal and solving the problem of gradient disappearance when the network depth increases through Residual connection.
Feature extraction relies on the inclusion Unit in the Residual Unit module, and the Residual linker now adds the inclusion input data to the Addition of the BN + Relu output data.
As one or more embodiments, the first, second, third, fourth, fifth, and sixth Residual Unit modules are identical in structure, and each include: the device comprises an inclusion unit, a BN + Relu unit and an Addition unit which are connected in sequence;
further, the inclusion unit is used for extracting multi-scale fault features in the signals. The inclusion unit comprises four one-dimensional convolutions (convolution kernel sizes are 1, 3, 5 and 7 respectively) with different sizes and a maximum pooling layer, and the convolution with the convolution kernel size of 1 is used for reserving original information in a signal. The method comprises the steps of extracting fault features of different scales in signals by adopting four convolution kernels of different sizes, connecting the fault features of the different scales through a concatenate layer, and using the fault features as output of an inclusion unit after dimension reduction through a Maxpoiling layer.
Further, the BN + Relu unit, that is, the block normalization and Relu activation function, is configured to pull back the distribution of any neuron input value of each layer of the neural network to a standard normal distribution with a mean value of 0 and a variance of 1, so that the activation input value falls in a region where the nonlinear function is sensitive to input, and even a small change in input will result in a large change in the loss function, and further the gradient becomes large, thereby avoiding a problem of gradient disappearance, and the gradient becomes large meaning that a learning convergence speed is fast, and a training speed is greatly increased.
Furthermore, the Addition unit is configured to perform Addition operation on corresponding elements of the two sets of input data to implement a residual error structure. The dimensions of the two inputs of the Addition unit must be the same, and the output dimension of the Addition unit after the Addition operation is the same as the input dimension.
As one or more embodiments, each Residual Unit module is followed by a max pooling layer MaxPooling, which is used to reduce feature dimensions and highlight useful information without information loss.
Further, as shown in fig. 3, the inclusion unit includes: a previous layer, a first convolution layer, a second convolution layer, a third convolution layer, a fourth convolution layer, a bicarbonate layer and a maxporoling layer;
the output end of the previous layer is connected with the input end of the first convolution layer, and the output end of the first convolution layer is connected with the bicarbonate layer; the convolution kernel size of the first convolution layer is 1;
the output end of the previous layer is connected with the input end of the second convolution layer, and the output end of the second convolution layer is connected with the concatenate layer; the convolution kernel size of the second convolution layer is 3;
the output end of the previous layer is connected with the input end of the third convolution layer, and the output end of the third convolution layer is connected with the concatenate layer; the convolution kernel size of the third convolution layer is 5;
the output end of the previous layer is connected with the input end of the fourth convolution layer, and the output end of the fourth convolution layer is connected with the concatenate layer; the convolution kernel size of the fourth convolution layer is 7;
the output end of the catenate layer is connected with the input end of the Maxpooling layer, the length of the pooling window of the Max Pooling is 4, and the step length is 4.
Further, the inclusion unit extracts features of different source data by respectively adopting convolution kernels with three scales of 1, 3, 5 and 7, the convolution with the convolution kernel size of 1 is used for retaining original information in the signal, and the maximum pooling layer is used for reducing dimensions.
Further, the previous layer refers to: the input layer of the inclusion unit.
Further, the previous layer is configured to: the input is the output data of the previous layer and provides input for three convolution operations and one maximum pooling-convolution operation after the previous layer.
Further, the concatenate layer refers to: and a feature connection layer.
Further, the concatenate layer is used for physically connecting the four scales of fault features extracted by the four convolution operations, combining the four scales of fault features into one, and using the one as an input of the maxporoling layer.
Further, the maxpololing layer refers to: the output layer of the inclusion unit.
Further, the maxporoling layer is used for matching the input and output dimensions of the inclusion unit, so that the input and output dimensions of the inclusion unit are consistent, otherwise, the Addition operation cannot be performed.
In one or more embodiments, the fusion module is a concatemate layer that physically connects features extracted from the vibration signal and the torque signal, as shown in fig. 4. The fusion module has the function of fusing the features extracted from different signals and is used for final fault classification, and the output end of the fusion module is connected with the input end of the softmax classification layer.
As one or more embodiments, the classifier is softmax.
The first embodiment solves the problem of multi-source data fusion strategy: the single data source is often difficult to cover the complete characteristics of the fault, and the diagnosis accuracy is not high due to the fact that the single data source is adopted for fault diagnosis, so that the first problem to be solved is the multi-source data fusion strategy problem.
Aiming at the problem, vibration and torque signal data of a Paderborn bearing data set are adopted to respectively verify the performance of the two data in fault diagnosis by adopting CNN, and the obtained diagnosis accuracy is not ideal;
then, Fourier transformation is carried out on the two data, the performance of fault diagnosis by adopting CNN is verified respectively again, the diagnosis accuracy of the obtained vibration signal is greatly improved, and the diagnosis accuracy of the torque signal is still not ideal;
and finally, respectively extracting fault characteristics for Fourier transformation of the current and torque signals by adopting the proposed data fusion network, carrying out feature layer fusion, and finally applying to final fault classification. It was therefore concluded that: complementary information aiming at bearing faults exists in the vibration signal and the torque signal, and the performance of bearing fault diagnosis can be effectively improved through the fusion of the two data.
The first embodiment also solves the problem of building a data fusion network model: in the fault diagnosis of mechanical equipment, a nonlinear relation is often formed between detection quantity and fault characteristics and between fault characteristics and a fault source, and the nonlinear relation and deep fault characteristics cannot be effectively extracted by simply increasing the network depth, and the phenomenon of gradient disappearance can be caused, so that the second problem to be solved is the problem of building a data fusion network model.
Aiming at the problem, a residual error network structure and an inclusion structure in GoogLeNet are adopted to optimize the CNN, extract effective characteristics and fuse the characteristics, and the method is applied to final fault classification. Network performance was verified using vibration and torque signal data from the Paderborn bearing data set. The result shows that the provided network structure can effectively extract the complementary fault characteristics in the bearing vibration signal and the torque signal, and the fault diagnosis accuracy is improved to a great extent.
Similar to other computational techniques, CNN's inspiration comes from image recognition mechanisms of the mammalian visual cortex. Unlike the global image processing method, the image it obtains from the retina is processed in a hierarchical and distributed manner. A group of nerve cells acts directly on the input to extract basic features, such as edge features. Convolution is a common spatial linear filtering method in image processing, and the three most important features using convolution kernel are: local/sparse connections, weight/parameter sharing, and translation invariance representation, which makes CNN require less pre-processing. Unlike other DNNs, CNNs use smaller convolution kernels to act on local regions of the input image to extract subtle and critical features.
The 1D CNN generally includes a convolutional layer, a pooling layer, and a fully-connected layer, and its working mechanism can be summarized as: the convolution kernel slides in the whole sequence by proper steps to extract local features, and the extracted feature value changes with different convolution kernel weight vectors in the convolution layer. The convolutional layer is always connected to the sub-sampling layer (e.g., max-firing) through a non-linear mapping function (e.g., ReLu), and the appropriate sub-sampling layer can effectively reduce the dimensionality of the input without losing information. And after connecting the plurality of convolutions and the pools, extracting effective characteristic vectors, and classifying results through a full connection layer. In the training process, all the weights such as convolution kernels of different convolution layers are updated by effective learning algorithms such as random gradient descent and the like. A convolution-pooling layer is shown in fig. 1, where the dashed box in the first row and the dashed box in the second row represent convolution and sub-sampling operations, respectively.
In the figure, the output of the first convolutional layer can be calculated as follows:
Figure BDA0002390397310000111
wherein N isl-1Representing the number of (l-1) layer pooling layer outputs,
Figure BDA0002390397310000112
a bias scalar representing the kth neuron in the l-th convolutional layer,
Figure BDA0002390397310000113
represents the weight of the kth neuron in the first convolutional layer,
Figure BDA0002390397310000114
represents the output of the i-th neuron in the (l-1) -th pooling layer, cov1D represents a one-dimensional convolution operation, and f (-) represents the activation function of the convolutional layer.
The output of the l-th pooling layer can be calculated as follows:
Sl＝ss(Yl) (2)
where ss denotes the down-sampling operation.
Firstly, respectively verifying the performance of vibration and torque signal data of a Paderborn bearing data set in fault diagnosis by adopting CNN (CNN);
secondly, Fourier transformation is carried out on the two signals, and the CNN is adopted again to verify the performance of fault diagnosis.
The Fourier transform method is used for analyzing signals by utilizing sine waveforms in signal components, can simplify complex convolution operation into product operation, and is an important algorithm in the field of digital signal processing. For a vector x containing n uniform sample points, its fourier transform is defined as:
Figure BDA0002390397310000121
wherein w ═ e-2πi/nIs one of n complex unit roots, i is an imaginary unit. For any x and y, the index numbers j and k are natural numbers between 0 and n-1, and the specific expansion of the Fourier is disclosed as follows:
Figure BDA0002390397310000122
finally, Fourier transformation of the vibration signal and the torque signal in the Paderborn bearing data set is input into the multi-scale information fusion fault diagnosis model provided by the scheme, and the model is shown in figure 2.
In the multi-scale information fusion fault diagnosis model, Fourier transformation of a vibration signal and a torque signal is respectively input into two networks from two channels, and feature extraction is synchronously performed.
The feature extraction network of each channel consists of three Residual units (as shown in fig. 2), each of which consists of an inclusion Unit, a block Normalization (back-Normalization) + Relu activation function layer, and an Addition layer. And simultaneously, combining the input of each inclusion unit and the output of the BN + Relu layer in an 'add' mode by adopting a residual connection mode, namely residual connection. A Maxpooling layer is connected behind each Residual Unit to realize dimension reduction.
In the inclusion unit (as shown in fig. 3), convolution kernels of 1, 3, 5 and 7 are respectively adopted to extract features of different source data, the convolution with the convolution kernel size of 1 is used for retaining information in an original signal, and the maxpoling layer is used for ensuring that the input dimension and the output dimension of the inclusion unit are consistent.
In a data fusion network architecture based on the inclusion and residual error structure, features extracted from different source data are fused. The fused features are finally used for fault diagnosis, and softmax is used as a classifier of the architecture and is defined as:
Figure BDA0002390397310000131
Figure BDA0002390397310000132
wherein z isiRepresenting the input of the softmax function, hjRepresents the output of the layer preceding softmax, WjiRepresenting the weight connecting the softmax layer and its previous layer.
When a normal network is trained by using a standard optimization algorithm (such as a gradient descent method), if residual error connection does not exist, research proves that training errors are reduced and then increased along with the deepening of the network. Theoretically, as the depth of the network increases, the training should be better and better, i.e., the deeper the depth of the network, the better. In practice, however, if there is no residual concatenation, deeper depth for a normal network will result in more training errors, meaning more difficult training with an optimization algorithm. And the residual error structure is connected in a crossing manner, so that the problems of gradient disappearance and gradient explosion are solved, and the good performance can be ensured while a deeper network is trained.
The invention mainly provides a multi-scale information fusion fault diagnosis method based on a deep residual convolution neural network, which adopts a residual network structure and an inclusion structure in GoogLeNet to optimize CNN, extracts effective characteristics and fuses, is applied to final fault classification, and mainly solves two technical problems: the method comprises the steps of multi-source data fusion strategy problem and data fusion network model building problem.
For the first problem, the vibration signal and the torque signal of the Paderborn bearing data set are used to verify the performance of the two signals in bearing fault diagnosis.
Firstly, the 1D CNN is adopted to verify the diagnosis performance of the two signals, and the experimental result shows that the fault diagnosis accuracy of the vibration signal is up to 81 percent, the fault diagnosis accuracy of the torque signal is up to 78 percent, and the diagnosis performance is not ideal.
Then, Fourier transformation is carried out on the two signals, and the 1D CNN is adopted to verify the diagnosis performance of the two transformed signals, and the experimental result shows that the accuracy of Fourier transformation fault diagnosis of the vibration signal reaches 95%, so that the accuracy is greatly improved; while the fourier transform of the torque signal has an 80% accuracy of fault diagnosis, an improvement.
Therefore, the fourier transform of the vibration signal and the torque signal is used as the input of the information fusion fault diagnosis network.
For the second problem, the Fourier transformation of the vibration signal and the torque signal is used as the input of the information fusion fault diagnosis network, the feature extraction is synchronously carried out, the fusion of deep features is carried out, and the classification for the fault is finally carried out. The experimental result shows that the classification accuracy after the features of the two signals are fused reaches 97%, and is improved by 2% on the basis of 95% of the diagnosis accuracy of the vibration signal.
The experimental results show that complementary information aiming at bearing faults exists in the vibration signals and the torque signals, and the fusion of the two data can effectively improve the performance of bearing fault diagnosis; meanwhile, the method for diagnosing the multi-scale information fusion fault based on the residual convolutional neural network is practical and effective.
In a second embodiment, the present invention further provides an electronic device, which includes a memory, a processor, and computer instructions stored in the memory and executed on the processor, where the computer instructions, when executed by the processor, implement the method in the first embodiment.
In a third embodiment, the present embodiment further provides a computer-readable storage medium for storing computer instructions, and the computer instructions, when executed by a processor, implement the method of the first embodiment.
The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure, and various modifications and changes may be made to the present disclosure by those skilled in the art. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present disclosure should be included in the protection scope of the present disclosure.

Claims (10)

1. The bearing fault diagnosis method based on multi-scale information fusion is characterized by comprising the following steps:
acquiring a vibration signal and a torque signal of a bearing to be subjected to fault diagnosis; complementary information for bearing faults exists in the vibration signal and the torque signal;
carrying out Fourier transform on the obtained vibration signal and the torque signal of the bearing subjected to fault diagnosis;
inputting a vibration signal and a torque signal of the bearing subjected to fault diagnosis obtained after Fourier transform into a multi-scale information fusion fault diagnosis model, and outputting a fault type of the bearing to be subjected to fault diagnosis;
the step of obtaining the multi-scale information fusion fault diagnosis model comprises the steps of constructing a neural network model, wherein the neural network model specifically comprises the following steps: two branches in parallel: a first branch and a second branch; the output end of the first branch and the output end of the second branch are both connected with the input end of the fusion module, and the output end of the fusion module is connected with the classifier; the fusion module fuses the features respectively extracted from the vibration signal and the torque signal obtained after Fourier transform;
the network structures of the first branch and the second branch are the same, the input layer of the first branch is used for inputting vibration signals, and the input layer of the second branch is used for inputting torque signals;
the first branch and/or the second branch respectively comprise 3 ResidualUnit modules and 3 MaxPholing modules, each ResidualUnit module is connected with the MaxPholing module, and the MaxPholing modules are used for reducing characteristic dimensions and highlighting useful information on the premise of not causing information loss of vibration signals and torque signals; the ResidualUnit module comprises an increment unit, a BN + Relu unit and an Addition unit which are connected in sequence; the Incep unit is used for extracting multi-scale fault features in the vibration signals and the torque signals; and the Addition unit is used for performing Addition operation on corresponding elements of the inclusion input end data and the BN + Relu output end data to realize a residual error structure.
2. The method of claim 1, wherein the multi-scale information fusion fault diagnosis model, the obtaining step comprises:
constructing a neural network model;
constructing a training set; the training set is vibration signals and torque signals of known bearing fault types after Fourier transformation;
and inputting the training set into a neural network model for training, and obtaining the trained neural network model after the training is finished, namely the multi-scale information fusion fault diagnosis model.
3. The method of claim 2, wherein a neural network model is constructed; the neural network model specifically includes:
the first branch, comprising: the device comprises a first input layer, a first ResidualUnit module, a first MaxPholing module, a second ResidualUnit module, a second MaxPholing module, a third ResidualUnit module, a third MaxPholing module and a first output layer which are connected in sequence;
the first input layer is used for inputting a vibration signal;
after an output end signal of the first ResidualUnit module passes through the first MaxPholing module, the output end signal of the first ResidualUnit module is used as an input signal of the second ResidualUnit module;
after an output end signal of the second ResidualUnit module passes through the second MaxPholing module, the output end signal of the second ResidualUnit module is used as an input signal of the third ResidualUnit module;
after an output end signal of the third ResidualUnit module passes through the third MaxPholing module, the output end signal is used as an input signal of the first output layer;
the first output layer is a BN + Relu function layer, the input end of the first output layer is connected with the output end of the third MaxPholing layer, and the output end of the first output layer is connected with the input end of the fusion module.
4. The method as set forth in claim 3,
the second branch circuit includes: the second input layer, the fourth ResidualUnit module, the fourth Maxboosting module, the fifth ResidualUnit module, the fifth Maxboosting module, the sixth ResidualUnit module, the sixth Maxboosting module and the second output layer are connected in sequence;
the second input layer is used for inputting a torque signal;
after passing through the fourth MaxPholing module, an output end signal of the fourth ResidualUnit module is used as an input signal of the fifth ResidualUnit module;
after an output end signal of the fifth ResidualUnit module passes through the fifth Maxboosting module, the output end signal is used as an input signal of the sixth ResidualUnit module;
after an output end signal of the sixth ResidualUnit module passes through the sixth MaxPholing module, the output end signal is used as an input signal of the second output layer;
the second output layer is a BN + Relu function layer, the input end of the second output layer is connected with the output end of the sixth Max machining layer, and the output end of the second output layer is connected with the input end of the fusion module.
5. The method as set forth in claim 3,
the first ResidualUnit module performs first feature extraction assumption on the input vibration signal to obtain M1 features, and M1 features pass through a first Max boosting layer to obtain M2 features;
the second Residual Unit module performs second feature extraction on the M2 features to obtain N1 features, and the N1 features pass through a second Max boosting layer to obtain N2 features;
the third ResidualUnit module carries out third-time feature extraction on the N2 features to obtain P1 features, and the P1 features pass through a third Max boosting layer to obtain P2 features;
p2 characteristics are obtained after the P2 characteristics pass through the first output layer, wherein M1, M2, N1, N2, P1 and P2 are positive integers.
6. The method as set forth in claim 3,
the first, second and third Residual Unit modules have the same structure and all comprise: the device comprises an inclusion unit, a BN + Relu unit and an Addition unit which are connected in sequence;
the BN + Relu unit, namely a block normalization and Relu activation function, is used for forcibly pulling back the distribution of any neuron input values of each layer of neural network to a standard normal distribution with a mean value of 0 and a variance of 1.
7. The method as set forth in claim 6, wherein,
the inclusion unit comprising: a previous layer, a first convolution layer, a second convolution layer, a third convolution layer, a fourth convolution layer, a bicarbonate layer and a maxporoling layer;
the output end of the previous layer is connected with the input end of the first convolution layer, and the output end of the first convolution layer is connected with the bicarbonate layer; the convolution kernel size of the first convolution layer is 1;
the output end of the previous layer is connected with the input end of the second convolution layer, and the output end of the second convolution layer is connected with the concatenate layer; the convolution kernel size of the second convolution layer is 3;
the output end of the previous layer is connected with the input end of the third convolution layer, and the output end of the third convolution layer is connected with the concatenate layer; the convolution kernel size of the third convolution layer is 5;
the output end of the previous layer is connected with the input end of the fourth convolution layer, and the output end of the fourth convolution layer is connected with the concatenate layer; the convolution kernel size of the fourth convolution layer is 7;
the output end of the bicarbonate layer is connected with the input end of the Maxpooling layer;
the Incep unit extracts features of different source data by respectively adopting convolution kernels with four scales of 1, 3, 5 and 7.
8. The method as set forth in claim 7, wherein,
the previous layer refers to: an input layer of the inclusion unit;
the previous layer is used for: the input is the output data of the front layer, and provides input for three convolution operations and one maximum pooling-convolution operation after the previous layer;
the bicarbonate layer refers to: a feature connection layer;
the concatenate layer is used for physically connecting the four-scale fault features extracted by the four convolution operations, combining the four-scale fault features into one and taking the combined fault feature as the input of the Maxpooling layer;
the Maxpooling layer refers to: an output layer of the inclusion unit;
the Maxpooling layer is used for matching the input dimension and the output dimension of the inclusion unit, so that the input dimension and the output dimension of the inclusion unit are consistent, and otherwise, the Addition operation cannot be performed.
9. An electronic device comprising a memory and a processor, and computer instructions stored on the memory and executable on the processor, the computer instructions when executed by the processor performing the method of any of claims 1-8.
10. A computer-readable storage medium storing computer instructions which, when executed by a processor, perform the method of any one of claims 1 to 8.
CN202010112157.9A 2020-02-24 2020-02-24 Bearing fault diagnosis method and system based on multi-scale information fusion Active CN111275007B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010112157.9A CN111275007B (en) 2020-02-24 2020-02-24 Bearing fault diagnosis method and system based on multi-scale information fusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010112157.9A CN111275007B (en) 2020-02-24 2020-02-24 Bearing fault diagnosis method and system based on multi-scale information fusion

Publications (2)

Publication Number Publication Date
CN111275007A CN111275007A (en) 2020-06-12
CN111275007B true CN111275007B (en) 2021-10-22

Family

ID=71002271

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010112157.9A Active CN111275007B (en) 2020-02-24 2020-02-24 Bearing fault diagnosis method and system based on multi-scale information fusion

Country Status (1)

Country Link
CN (1) CN111275007B (en)

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112067289A (en) * 2020-08-21 2020-12-11 天津电气科学研究院有限公司 Motor shaft and transmission shaft abnormal vibration early warning algorithm based on neural network
CN112016004B (en) * 2020-08-21 2023-03-31 重庆邮电大学 Multi-granularity information fusion-based job crime screening system and method
CN113158722A (en) * 2020-12-24 2021-07-23 哈尔滨理工大学 Rotary machine fault diagnosis method based on multi-scale deep neural network
CN112798280B (en) * 2021-02-05 2022-01-04 山东大学 Rolling bearing fault diagnosis method and system
CN112926641B (en) * 2021-02-22 2022-05-17 山东大学 Three-stage feature fusion rotating machine fault diagnosis method based on multi-mode data
CN112884093B (en) * 2021-04-29 2021-08-31 四川大学 Rotary machine fault diagnosis method and equipment based on DSCRN model and storage medium
CN113221996A (en) * 2021-04-30 2021-08-06 山东大学 Bearing fault diagnosis method and system based on multi-head attention mechanism
CN113155466B (en) * 2021-05-24 2024-04-23 合肥工业大学 Bearing fault visual vibration detection method and system
CN113295413B (en) * 2021-06-24 2022-10-25 北京交通大学 Traction motor bearing fault diagnosis method based on indirect signals
CN113236595B (en) * 2021-07-13 2021-09-28 湖南师范大学 Fan fault analysis method, device, equipment and readable storage medium
CN113780151B (en) * 2021-09-07 2022-05-20 山东大学 Bearing fault diagnosis method and system based on bilinear feature fusion
CN117932561B (en) * 2024-03-22 2024-06-11 绵阳师范学院 Method and system for analyzing torque data of coupler

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109779894A (en) * 2018-12-29 2019-05-21 西安交通大学 A kind of reciprocating compressor fault diagnosis system and method based on neural network algorithm
CN110595775A (en) * 2019-09-19 2019-12-20 电子科技大学 Rolling bearing fault diagnosis method based on multi-branch multi-scale convolutional neural network

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101432786B1 (en) * 2013-11-14 2014-09-23 엠앤디테크놀로지 주식회사 Fault diagnosis method of motor and system thereof
CN108921303A (en) * 2018-05-29 2018-11-30 青岛鹏海软件有限公司 The Fault diagnosis and forecast maintaining method of industrial motor

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109779894A (en) * 2018-12-29 2019-05-21 西安交通大学 A kind of reciprocating compressor fault diagnosis system and method based on neural network algorithm
CN110595775A (en) * 2019-09-19 2019-12-20 电子科技大学 Rolling bearing fault diagnosis method based on multi-branch multi-scale convolutional neural network

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Gear Pitting Fault Diagnosis Using Integrated CNN and GRU Network with Both Vibration and Acoustic Emission Signals;Xueyi Li 等;《applied sciences》;20190222;第9卷(第4期);正文第2-4部分 *
Xueyi Li 等.Gear Pitting Fault Diagnosis Using Integrated CNN and GRU Network with Both Vibration and Acoustic Emission Signals.《applied sciences》.2019,第9卷(第4期),正文第2-4部分. *
融合多传感器的故障诊断方法研究;王裕峰;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20200115;第2020年卷(第01期);正文第34页第一段-第39页图4-7 *

Also Published As

Publication number Publication date
CN111275007A (en) 2020-06-12

Similar Documents

Publication Publication Date Title
CN111275007B (en) Bearing fault diagnosis method and system based on multi-scale information fusion
CN112926641B (en) Three-stage feature fusion rotating machine fault diagnosis method based on multi-mode data
CN111275618A (en) Depth map super-resolution reconstruction network construction method based on double-branch perception
CN114398961A (en) Visual question-answering method based on multi-mode depth feature fusion and model thereof
Guan et al. Autoattend: Automated attention representation search
WO2024040828A1 (en) Method and device for fusion and classification of remote sensing hyperspectral image and laser radar image
CN113157919B (en) Sentence text aspect-level emotion classification method and sentence text aspect-level emotion classification system
CN107491782A (en) Utilize the image classification method for a small amount of training data of semantic space information
CN111353583B (en) Deep learning network based on group convolution characteristic topological space and training method thereof
CN114742985A (en) Hyperspectral feature extraction method and device and storage medium
CN115188440A (en) Intelligent matching method for similar medical records
CN114170657A (en) Facial emotion recognition method integrating attention mechanism and high-order feature representation
CN111612046B (en) Feature pyramid graph convolution neural network and application thereof in 3D point cloud classification
Li et al. Knowing where to look? Analysis on attention of visual question answering system
CN109583406B (en) Facial expression recognition method based on feature attention mechanism
CN111553424A (en) CGAN-based image data balancing and classifying method
CN110543567A (en) Chinese text emotion classification method based on A-GCNN network and ACELM algorithm
CN114494284B (en) Scene analysis model and method based on explicit supervision area relation
CN115620064A (en) Point cloud down-sampling classification method and system based on convolutional neural network
CN113837263B (en) Gesture image classification method based on feature fusion attention module and feature selection
CN113283530B (en) Image classification system based on cascade characteristic blocks
Srinivas et al. Convolution neural networks for binary classification
Liu et al. A deep neural network pruning method based on gradient L1-norm
CN113537325A (en) Deep learning method for image classification based on logic of extracting high-low-level features
CN113239678A (en) Multi-angle attention feature matching method and system for answer selection

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
GR01 Patent grant
GR01 Patent grant