CN113203566B - Motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN - Google Patents

Motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN Download PDF

Info

Publication number
CN113203566B
CN113203566B CN202110365267.0A CN202110365267A CN113203566B CN 113203566 B CN113203566 B CN 113203566B CN 202110365267 A CN202110365267 A CN 202110365267A CN 113203566 B CN113203566 B CN 113203566B
Authority
CN
China
Prior art keywords
data
training
neural network
data set
dimensional
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
CN202110365267.0A
Other languages
Chinese (zh)
Other versions
CN113203566A (en
Inventor
李宏亮
龙云瑶
周武能
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.)
Shanghai Tushan Intelligent Technology Co ltd
Original Assignee
Shanghai Tushan Intelligent Technology Co ltd
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 Shanghai Tushan Intelligent Technology Co ltd filed Critical Shanghai Tushan Intelligent Technology Co ltd
Priority to CN202110365267.0A priority Critical patent/CN113203566B/en
Publication of CN113203566A publication Critical patent/CN113203566A/en
Application granted granted Critical
Publication of CN113203566B publication Critical patent/CN113203566B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • 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
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent

Abstract

The invention relates to a motor bearing fault diagnosis method based on one-dimensional data enhanced CNN, which comprises the steps of dividing a preprocessed one-dimensional vibration signal into a training data set and a test data set, inputting the test data set into a data conversion module to be directly converted into two-dimensional image test data, and inputting the two-dimensional image test data into a trained convolutional neural network module to perform fault diagnosis test; the trained convolutional neural network module is obtained by training by taking a two-dimensional image data set as a training set; the two-dimensional image data set is obtained by integrating the enhanced data set and the training data set and then directly converting the input data set by the data conversion module; the enhanced data set is obtained by inputting a training data set into a one-dimensional data enhancement module and outputting; the one-dimensional data enhancement module comprises SAE improved by a BP neural network with a softmax classifier; the invention combines the one-dimensional data enhancement module and the CNN to be applied to the fault diagnosis problem, and aims to solve the fault diagnosis under the conditions of various fault types and insufficient sample data.

Description

Motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN
Technical Field
The invention belongs to the technical field of bearing production fault diagnosis, and relates to a motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN.
Background
With the rapid development of intelligent manufacturing, under the background of industrial big data, an intelligent fault diagnosis technology is gradually a hot problem in the industrial field as an important tool for guaranteeing the safe operation of mechanical equipment, and a large amount of related researches are shown. The rolling bearing of the motor is an important part of the rotating machinery, and the operation condition of the rolling bearing has direct influence on the production safety and the working efficiency of large-scale industrial equipment. The fault diagnosis of the rolling bearing is therefore one of the important points of research as an important means of avoiding the risks of industrial production.
The machine learning algorithm takes an important weight in the data-driven fault diagnosis algorithm by virtue of its advantages in automatically extracting data features. Traditional machine learning algorithms such as a support vector machine, principal component analysis, random forest and the like are all applied to the fault diagnosis problem one by one, and good effects are achieved in the aspects of feature extraction, fault classification and the like. However, researches find that the selection of important parameters in various machine learning algorithms is still largely limited by the experience of researchers, and the generalization capability of the model after parameter determination often does not meet the requirement.
And the successful application of deep learning models such as a stacked self encoder (SAE), a cyclic neural network, a generative countermeasure network and a convolutional neural network to the fault diagnosis problem alleviates the defects of a machine learning algorithm. The neural networks avoid one-sidedness and uncertainty of artificial feature extraction by deeply mining the feature information of fault diagnosis data, and the generalization capability of the model is greatly improved. However, the deep learning algorithm often requires a large amount of sufficient data to train the network, and when the training data is too little or the failure types are too many, the diagnosis accuracy is obviously affected. The CNN has great advantages for classification of two-dimensional images due to a special neural network structure, but firstly, bearing fault diagnosis data are mostly one-dimensional vibration data sets, and the CNN can be used for subsequent processing only by converting one-dimensional data into two-dimensional data; secondly, when the training data set has complex classification types and each type of data is insufficient, the classification accuracy of the CNN may have a significant drop. Therefore, there is no example of CNN-based optimization and application of the fault diagnosis model for the situation where the fault type is complicated and the collected data is insufficient. And the prior art has more obvious defects in solving the problems, such as: the fault diagnosis precision of an unoptimized single neural network is low, and the neural network which does not relate to pooling operation is more prone to serious overfitting problems during classification; in data enhancement, methods such as GAN are mostly used for a two-dimensional data set which is converted in advance, and due to a complex model structure and a large amount of convolution operation in a data enhancement stage, the fault diagnosis efficiency of the whole model is greatly influenced. Namely, the efficiency and accuracy cannot be balanced and optimized.
Therefore, it is very important to research a bearing fault diagnosis method that can achieve both efficiency and accuracy.
Disclosure of Invention
In order to solve the problems in the prior art, the invention designs a fault diagnosis method based on one-dimensional data enhancement and CNN, firstly, a BP neural network with a softmax classifier is used for carrying out optimization processing on the traditional SAE, and the optimized SAE is used for directly carrying out data enhancement on a one-dimensional vibration signal; then, a one-dimensional data enhancement module is combined with the CNN to provide a new model, and the new model is applied to the field of fault diagnosis. The invention can remarkably improve the fault diagnosis precision and simultaneously consider the diagnosis efficiency under the conditions of various fault types and insufficient sample data. The system can be applied to fault diagnosis of the motor bearing in an industrial big data environment.
In order to achieve the purpose, the invention adopts the following scheme:
a motor bearing fault diagnosis method based on one-dimensional data enhanced CNN comprises the steps of dividing a preprocessed one-dimensional vibration signal into a training data set and a test data set, inputting the test data set into a data conversion module to be directly converted into two-dimensional image test data, and inputting the two-dimensional image test data into a trained convolutional neural network module to perform fault diagnosis test; the fault diagnosis test means that the output layer of the trained convolutional neural network module outputs a one-dimensional vector with a classification result of 16 multiplied by 1, only one value of the one-dimensional vector is 1, the rest values are 0, and the position of 1 is the test result;
the one-dimensional vibration signal is provided with a fault diagnosis type label;
the trained convolutional neural network module is obtained by training by taking a two-dimensional image data set as a training set; the two-dimensional image data set is obtained by integrating the enhanced data set and the training data set and then directly converting the input data set by the data conversion module;
the enhancement data set is obtained by inputting a training data set into a one-dimensional data enhancement module and outputting the training data set; the one-dimensional data enhancement module includes an SAE modified with a BP neural network with a softmax classifier.
Namely, the motor bearing fault diagnosis method based on the CNN enhanced by the one-dimensional data comprises the following steps:
(1) after data preprocessing is carried out on the one-dimensional vibration signals, the one-dimensional vibration signals are divided into a training data set and a testing data set;
(2) inputting the training data set into a one-dimensional data enhancement module to output an enhancement data set (the enhancement data set is obtained by the output of a decoder of SAE);
(3) after the enhancement data set and the training data set are integrated, the input data are directly converted into a two-dimensional image data set by a data conversion module;
(4) inputting the two-dimensional image data set into a convolutional neural network module for training;
(5) and inputting the test data set into a data conversion module to be directly converted into two-dimensional image test data, and testing the fault diagnosis result of the trained convolutional neural network module.
As a preferred technical scheme:
the method for diagnosing the motor bearing fault based on the one-dimensional data enhanced CNN comprises the following steps:
(1) selecting the training data set; no more than 100 samples of the training data set;
(2) constructing three layers of SAE and three layers of BP neural networks with softmax classifiers, and initializing parameters of the SAE and the BP neural networks;
(3) the reconstruction error (i.e. the error between the output of the decoder and the input of the encoder) of SAE is calculated using a mean square error function, and the error function L of a single input 1 Comprises the following steps:
Figure BDA0003007043010000031
in the formula, W is a normal vector of the hyperplane and is initialized to a unit vector with the same dimension as the number of fault characteristic parameter data of each group of input data, b is an intercept of the hyperplane and is initialized to 0, x is decoder input data, y is encoder input data, h is output data w，b (x) For the decoder output, | | | · | |, is the euclidean distance;
for a set of inputs { (x) (1) ，y (1) )，(x (2) ，y (2) )，...，(x (m) ，y (m) ) The total error function is:
Figure BDA0003007043010000032
in the formula, M is the number of input data;
(4) selecting a cross entropy function to calculate the fault diagnosis category of the softmax layer of the BP neural network and the classification error L of the label in the training data set 2 The specific function is as follows:
Figure BDA0003007043010000033
wherein K is the number of classified categories; q. q of is Q is a classification parameter when the class s is the same as that of the ith sample data is Get 1, otherwise q is Taking 0; p is a radical of is A predicted probability of belonging to class s for the observation of the ith sample;
(5) the total error L in the optimized SAE training process is as follows:
L＝L 1 (W，b)+L 2 ；
(6) in each training process, in order to minimize L, updating the optimized network parameters of SAE through a back propagation algorithm, wherein the calculation formula is as follows:
Figure BDA0003007043010000034
in the formula (I), the compound is shown in the specification,
Figure BDA0003007043010000035
and
Figure BDA0003007043010000036
is the partial differential of L with respect to W and b, a is the learning rate of SAE;
(7) judging whether the preset training times are reached, if so, stopping training; otherwise, returning to the step (2), and obtaining the SAE after the training of the BP neural network with the softmax classifier is finished by taking the connection weight and the deviation of the neurons of each layer of neural network as the connection weight and the deviation of each neuron when the training is terminated and forming a data enhancement module by the SAE.
The motor bearing fault diagnosis method based on the one-dimensional data enhanced CNN comprises the following steps that the convolutional neural network module comprises a convolutional neural network model, and the convolutional neural network model comprises an input layer, two alternating convolutional layers and pooling layers, a full connection layer and an output layer.
The motor bearing fault diagnosis method based on the one-dimensional data enhanced CNN comprises the following training processes by taking the two-dimensional image data set as a training set: inputting a two-dimensional image data set into a convolutional neural network model, training a CNN model by taking a fault type label corresponding to an input two-dimensional image as theoretical output, and determining the connection weight and deviation of each neuron, wherein the method specifically comprises the following steps:
(1) initializing parameters of a CNN model;
(2) selecting the two-dimensional image dataset;
(3) inputting a two-dimensional image data set, training by using an Adam optimizer, and adjusting the weight and the deviation of each layer of neural network in each step of iteration process to enable the fault precision to be higher than that before the iterative computation; the fault precision calculation mode is as follows: the ratio of the number of data with the fault diagnosis type label given to the output data and the collected fault type label consistent with each other to the total sample data number after model classification;
(4) judging whether the training times reach a preset value, if so, stopping training; otherwise, returning to the step (3), and the connection weight and the deviation of the neurons of each layer of the neural network when the training is terminated are the connection weight and the deviation of each neuron.
When the convolutional neural network module operates, zero values automatically fill the convolution process, let epsilon be the output size, delta be the input size, eta be the span, gamma be the filter width, and the left zero padding number PL and the right zero padding number PR are calculated accordingly, the padding formula is as follows:
Figure BDA0003007043010000041
PT＝(ε-1)×η+γ-δ；
Figure BDA0003007043010000042
PR＝PT--PL；
in the formula, ceil () is an upward rounding function, floor () is a downward rounding function, and PT is the upper limit of the zero padding number.
If the initialized training step length is not matched with the size of the input image, namely the convolution kernel crosses the image edge in the process of performing convolution calculation by the step length line-by-line translation, a zero filling method is used to prevent the size loss problem of image conversion.
According to the motor bearing fault diagnosis method based on the one-dimensional data enhanced CNN, the process of direct conversion of the data conversion module is as follows:
(1) randomly intercepting a section with the length of N from a one-dimensional vibration signal 2 Sequentially filling the signal values into an n × n matrix line by line, namely filling each n signal values into one line of the matrix;
(2) standardizing the obtained matrix to 0-255, namely obtaining an n multiplied by n image within the range of the pixel value of the gray image; the formula for the conversion is as follows:
Figure BDA0003007043010000051
in the formula, D (e, f) is the value of the pixel point of the ith row and the fth column of the obtained picture; s (i) is the value of the ith sample point in the intercepted signal segment, max (·) is a maximum function, min (·) is a minimum function; round (·) is a rounding function that preserves the integer part according to the rounding principle.
The preprocessing refers to performing normalization processing on a data matrix formed by fault characteristic parameters of each signal point in the one-dimensional vibration signal, and the processing formula is as follows:
Figure BDA0003007043010000052
in the formula, x * Is normalized data, x is data before normalization, x min Is the minimum value of x, x max Is the maximum value of x.
The motor bearing fault diagnosis method based on the one-dimensional data enhanced CNN is characterized in that the fault diagnosis type labels comprise a ball fault (RF), an outer periphery fault (OF) and an inner periphery fault (IF).
In the above method for diagnosing a motor bearing fault based on a one-dimensional data enhanced CNN, initializing the parameters of the SAE and the BP neural network refers to: presetting the training times to be 800, initializing the learning rate of SAE to be 0.0001, and setting the number of training samples in each batch to be 50; the number of neurons of each layer of the three-layer SAE neural network is initialized to be 1024, 1000 and 1000, and the number of neurons of each layer of the three-layer BP neural network is 600, 300 and 16 respectively; initializing the weight of each layer of neural network as a unit vector with the same dimension as the input data dimension, wherein the deviation is 0; the activation function of the initialized hidden layer is a ReLU function, and the activation function of the output layer is a sigmoid function. The initialized content can be modified according to a specific application scenario, for example, the initialized preset training times are according to the size of a training data set, and when the number of single type fault diagnosis samples is 10-1000, a multiple of 100 can be generally selected from 100-3000.
As described above, in the motor bearing fault diagnosis method based on the one-dimensional data enhanced CNN, the initializing parameters of the CNN model refer to: the initial learning rate is 0.0005, the training step length is 1, the training times are 800, and the size of each batch of training sample sets is 256; initializing convolution layer convolution kernel size to be 5 x 5, pooling layer filter size to be 2 x 2, and number of full-connection layer neurons to be 2048; the weight for initializing each layer of neural network is a unit vector with the same dimension as the input data dimension, and the deviation is 0.
The principle of the invention is as follows:
the method comprises the steps of firstly optimizing SAE by using a BP neural network with a softmax classifier, then directly performing data enhancement on the collected one-dimensional vibration signals by using the optimized SAE as a one-dimensional data enhancement module, and finally converting an enhanced data set into two-dimensional data to be input into a CNN for fault diagnosis. Compared with the prior art, firstly, the optimized SAE can use supervised learning and unsupervised learning at the same time, and the data characteristics and data labels of the collected fault data are fully utilized, so that the decoder output of the trained SAE can better keep the characteristics of the original data set, the similarity of the enhanced data set and the original data set is improved, and the influence of the insufficient training data on the fault diagnosis result is creatively solved. Secondly, the obvious advantages of the convolutional neural network in image classification are fully exerted, the precision deficiency caused by the structural limitation of the neural network (namely no pooling layer is used for reducing the occurrence of overfitting problems) of the classification methods such as SAE and DBN which are applied more at present is made up, the influence of excessive convolution operation on the fault diagnosis efficiency of the existing two-dimensional data enhancement methods such as GAN is reduced, and particularly, the optimization effect is obvious under the conditions of complicated fault types and insufficient sample data. Finally, a direct conversion method is adopted in the data conversion process, so that the loss of data characteristics in the conversion process is avoided as much as possible, and compared with the existing methods such as wavelet packet conversion, Fourier transform and the like, the method simplifies the complex calculation process and improves the overall operation efficiency of the model.
Advantageous effects
(1) The invention relates to a motor bearing fault diagnosis method based on one-dimensional data enhancement CNN, which innovatively uses optimized SAE for one-dimensional data enhancement, and reserves the original characteristics of data to the greatest extent when data enhancement is carried out by using a special SAE neural network structure. The method makes full use of the advantages of the convolutional neural network in image processing in fault classification, does not involve redundant convolution operation, and can conveniently and uniformly process a large number of collected vibration signals. The method uses a simple and direct data conversion method, does not relate to any parameter selection requiring expert experience, almost retains all effective characteristics of the original data, and improves the intelligence and the high efficiency of fault diagnosis. The invention provides a good solution to the problems of the existing research;
(2) the invention discloses a motor bearing fault diagnosis method based on one-dimensional data enhanced CNN, which combines a one-dimensional data enhanced module and the CNN to be applied to the problem of fault diagnosis and aims to solve the problem of fault diagnosis under the conditions of various fault types and insufficient sample data. Most of the existing fault diagnosis algorithms aim at the conditions of sufficient data sets and simple fault types (less than or equal to five fault types), and the fault problems in the industrial background often cause the conditions of insufficient sample data and complex fault types due to high data collection cost, complex mechanical equipment and the like.
Drawings
FIG. 1 is a flow chart of a motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN;
FIG. 2 is a flow chart of a motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN;
FIG. 3 is a schematic structural diagram of a motor bearing fault diagnosis system based on one-dimensional data enhancement and CNN;
FIG. 4 is a schematic diagram of the structure of SAE after optimization using BP neural network with softmax classifier;
FIG. 5 is a flow chart for training an improved SAE;
FIG. 6 is a schematic diagram of a method of converting a one-dimensional dataset into a two-dimensional dataset;
fig. 7 is a structural diagram of the CNN model.
Detailed Description
The invention will be further illustrated with reference to specific embodiments. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the present invention. Further, it should be understood that various changes or modifications of the present invention may be made by those skilled in the art after reading the teaching of the present invention, and such equivalents may fall within the scope of the present invention as defined in the appended claims.
A motor bearing fault diagnosis method based on one-dimensional data enhanced CNN is disclosed, the flow chart is shown in figure 1 and figure 2, the structural schematic diagram of the diagnosis system is shown in figure 3, and the specific steps are as follows:
(1) preprocessing a one-dimensional vibration signal;
normalizing a data matrix formed by fault characteristic parameters of each signal point in the one-dimensional vibration signal, wherein the processing formula is as follows:
Figure BDA0003007043010000071
in the formula, x * Is normalized data, x is data before normalization, x min Is the minimum value of x, x max Is the maximum value of x;
(2) dividing the preprocessed one-dimensional vibration signal into a training data set and a testing data set; the one-dimensional vibration signal is provided with a fault diagnosis type label, and the fault diagnosis type label comprises a ball fault, an outer periphery fault and an inner periphery fault;
(3) constructing a one-dimensional data enhancement module:
(3.1) selecting the training data set; no more than 100 samples of the training data set;
(3.2) constructing three layers of SAE and three layers of BP neural networks with softmax classifiers, wherein the structure diagram is shown in FIG. 4, and initializing parameters of the SAE and the BP neural networks;
initializing parameters of the SAE and the BP neural network refers to: presetting the training times to be 800, initializing the learning rate of SAE to be 0.0001, and setting the number of training samples in each batch to be 50; the number of neurons of each layer of the three-layer SAE neural network is initialized to be 1024, 1000 and 1000, and the number of neurons of each layer of the three-layer BP neural network is 600, 300 and 16 respectively; initializing the weight of each layer of neural network as a unit vector with the same dimension as the input data dimension, wherein the deviation is 0; initializing an activation function of a hidden layer to be a ReLU function, and initializing an activation function of an output layer to be a sigmoid function;
(3.3) selecting a mean square error function to calculate the reconstruction error of SAE, and selecting a single-input error function L 1 Comprises the following steps:
Figure BDA0003007043010000072
in the formula, W is a normal vector of the hyperplane and is initialized to a unit vector with the same dimension as the number of fault characteristic parameter data of each group of input data, b is an intercept of the hyperplane and is initialized to 0, x is decoder input data, y is encoder input data, h is output data w，b (x) For the decoder output, | | | · | |, is the euclidean distance;
for a set of inputs { (x) (1) ，y (1) )，(x (2) ，y (2) )，...，(x (m) ，y (m) ) The total error function is:
Figure BDA0003007043010000073
in the formula, M is the number of input data;
(3.4) selecting a cross entropy function to calculate the fault diagnosis category of the softmax layer of the BP neural network and the classification error L of the label in the training data set 2 The specific function is as follows:
Figure BDA0003007043010000081
wherein K is the number of classified categories; q. q.s is Q is a classification parameter when the class s is the same as that of the ith sample data is Get 1, otherwise q is Taking 0; p is a radical of is A predicted probability for observing that the ith sample belongs to class s;
(3.5) the total error L in the optimized SAE training process is:
L＝L 1 (W，b)+L 2 ；
(3.6) in each training process, in order to minimize L, updating the optimized network parameters of SAE through a back propagation algorithm, and calculating the formula as follows:
Figure BDA0003007043010000082
in the formula (I), the compound is shown in the specification,
Figure BDA0003007043010000083
and
Figure BDA0003007043010000084
is the partial differential of L with respect to W and b, and α is the learning rate of SAE;
(3.7) judging whether the preset training times are reached, and if so, stopping training; otherwise, returning to the step (3.2), wherein the connection weight and the deviation of the neurons of each layer of the neural network at the end of training are the connection weight and the deviation of each neuron, obtaining the trained improved SAE of the BP neural network with the softmax classifier, and forming a data enhancement module by the SAE, wherein the SAE training flow is shown in FIG. 5;
(4) inputting a training data set into a one-dimensional data enhancement module to output an enhanced data set;
(5) after the enhancement data set and the training data set are integrated, the input data are directly converted into a two-dimensional image data set by a data conversion module;
the process of the direct conversion of the data conversion module is that a schematic diagram of a conversion method is shown in fig. 6:
(5.1) randomly intercepting a section with the length of N from the one-dimensional vibration signal 2 Sequentially filling the signal values into an n × n matrix line by line, namely filling each n signal values into one line of the matrix;
(5.2) carrying out standardization treatment on the obtained matrix, and normalizing to 0-255 to obtain an n multiplied by n image; the formula for the conversion is as follows:
Figure BDA0003007043010000085
in the formula, D (e, f) is the value of the pixel point of the ith row and the fth column of the obtained picture; (i) is the value of the ith sample point in the intercepted signal segment, max (-) is the maximum function, min (-) is the minimum function; round (·) is a rounding function that preserves the integer part according to the rounding principle;
(6) inputting the two-dimensional image data set into a convolutional neural network module for training to obtain a trained convolutional neural network module;
the convolutional neural network module comprises a convolutional neural network model, the convolutional neural network model comprises an input layer, two alternating convolutional layers and pooling layers, a full-link layer and an output layer, and the structure diagram is shown in FIG. 7;
when the convolutional neural network module runs, zero values automatically fill the convolution process, let epsilon be the output size, delta be the input size, eta be the span, gamma be the filter width, and the left zero padding PL and the right zero padding PR are calculated by the following filling formula:
Figure BDA0003007043010000091
PT＝(ε-1)×η+γ-δ；
Figure BDA0003007043010000092
PR＝PT-PL；
in the formula, ceil (·) is an upward rounding function, floor (·) is a downward rounding function, and PT is an upper limit of zero padding;
the process of training by taking the two-dimensional image data set as a training set comprises the following steps: inputting a two-dimensional image data set into a convolutional neural network model, training a CNN model by taking a fault type label corresponding to an input two-dimensional image as theoretical output, and determining the connection weight and deviation of each neuron, wherein the method specifically comprises the following steps:
(6.1) initializing parameters of the CNN model; the method specifically comprises the following steps: the initial learning rate is 0.0005, the training step length is 1, the training times are 800, and the size of each batch of training sample sets is 256; initializing convolution layer convolution kernel size to be 5 x 5, pooling layer filter size to be 2 x 2, and number of full-connection layer neurons to be 2048; initializing the weight of each layer of neural network as a unit vector with the same dimension as the input data dimension, wherein the deviation is 0;
(6.2) selecting the two-dimensional image data set;
(6.3) inputting a two-dimensional image data set, training by using an Adam optimizer, and adjusting the weight and the deviation of each layer of neural network in each step of iteration process to enable the fault precision to be higher than that before the iterative computation; the fault precision calculation mode is as follows: the ratio of the number of data with the fault diagnosis type label given to the output data and the collected fault type label consistent with each other to the total sample data number after model classification;
(6.4) judging whether the training times reach a preset value, if so, stopping training; otherwise, returning to the step (6.3), wherein the connection weight and the deviation of the neurons of each layer of neural network when the training is terminated are the connection weight and the deviation of each neuron;
(7) the test data set is input into a data conversion module to be directly converted into two-dimensional image test data, and the two-dimensional image test data is input into a trained convolutional neural network module to be subjected to fault diagnosis test; the fault diagnosis test means that the output layer of the trained convolutional neural network module outputs a one-dimensional vector with a classification result of 16 multiplied by 1, the one-dimensional vector has only one value of 1, the rest values are 0, and the position of 1 is the test result.
The specific case is as follows:
and selecting the public motor driving end bearing fault diagnosis data collected by a Kaiser-Si university motor experiment platform with the sampling frequency of 12kHz and the load of 0hp as experiment data. The data contained 16 types of failure diagnosis in total, among which 15 failure environments and one normal environment (NO). The failure types can be roughly classified into a rolling body failure (RF), an inner ring failure (IF), and an outer ring failure (OF). The position of the bearing outer ring which is in fault is relatively fixed, and the influence of the position difference on the vibration response of the motor bearing system is large. Therefore, when data are collected, fault signal acquisition points of the bearing outer ring are respectively placed at three different positions in the 3 o ' clock direction, the 6 o ' clock direction and the 12 o ' clock direction, outer ring faults occurring in the three directions are regarded as different fault types, each fault type has different damage diameters, and the different damage diameters are also regarded as different fault environments. Specifically, the rolling element failure and the inner ring failure contained damage diameters of 0.007, 0.014, 0.021, 0.028 inches, the 3 o ' clock outer raceway contained damage diameters of 0.007, 0.021 inches, the 6 o ' clock outer raceway contained damage diameters of 0.007, 0.014, 0.021 inches, and the 12 o ' clock outer raceway introduced damage diameters of 0.007, 0.021 inches. Each set of signals contained 1024 data points in the experiment. About 89 groups of signals are selected from the training samples in each fault environment, 354 groups of signals are selected from the samples in the normal environment, and 1682 groups of signals form a training sample data set. The test samples are randomly extracted to obtain 568 groups of signals under 16 fault diagnosis environments according to the principle of no repeated sampling. That is, the experimental data set selected according to the selection principle has the characteristics of various fault categories, less sample data and obviously more sample data under normal conditions than under fault conditions. The experimental result shows that although the training time of the model provided by the invention is 691 seconds, the training time is only prolonged by about 1/3 compared with that of a single neural network, the testing time is still very short, the training precision is about 0.06s as same as that of other neural network models, the training precision is 98.89%, the testing precision is 97.25%, and effective optimization is achieved, namely the fault diagnosis performance is improved on the premise of ensuring the operation efficiency by the model.

Claims (9)

1. A motor bearing fault diagnosis method based on one-dimensional data enhanced CNN is characterized by comprising the following steps: dividing the preprocessed one-dimensional vibration signal into a training data set and a test data set, inputting the test data set into a data conversion module to be directly converted into two-dimensional image test data, and inputting the two-dimensional image test data into a trained convolutional neural network module to perform fault diagnosis test; the fault diagnosis test means that the output layer of the trained convolutional neural network module outputs a one-dimensional vector with a classification result of 16 multiplied by 1, only one value of the one-dimensional vector is 1, the rest values are 0, and the position of 1 is the test result;
the one-dimensional vibration signal is provided with a fault diagnosis type label;
the trained convolutional neural network module is obtained by training by taking a two-dimensional image data set as a training set; the two-dimensional image data set is obtained by integrating an enhanced data set and a training data set and then directly converting the enhanced data set and the training data set by an input data conversion module;
the enhancement data set is obtained by inputting a training data set into a one-dimensional data enhancement module and outputting the training data set; the one-dimensional data enhancement module comprises SAE improved by a BP neural network with a softmax classifier;
the construction process of the one-dimensional data enhancement module comprises the following steps:
(1) selecting the training data set; no more than 100 samples of the training data set;
(2) constructing three layers of SAE and three layers of BP neural networks with softmax classifiers, and initializing parameters of the SAE and the BP neural networks;
(3) selecting a mean square error function to calculate the reconstruction error of SAE, and a single input error function L 1 Comprises the following steps:
Figure FDA0003710431970000011
in the formula, W is a normal vector of the hyperplane and is initialized to a unit vector with the same dimension as the number of fault characteristic parameter data of each group of input data, b is an intercept of the hyperplane and is initialized to 0, x is decoder input data, y is encoder input data, h is output data W，b (x) For the decoder output, | | | · | |, is the euclidean distance;
for a set of inputs { (x) (1) ，y (1) )，(x (2) ，y (2) )，...，(x (m) ，y (m) ) The total error function is:
Figure FDA0003710431970000012
in the formula, M is the number of input data;
(4) selecting a cross entropy function to calculate the fault diagnosis category of the softmax layer of the BP neural network and the classification error L of the label in the training data set 2 The specific function is as follows:
Figure FDA0003710431970000013
wherein K is the number of classified categories; q. q.s is Q is a classification parameter when the class s is the same as that of the ith sample data is Get 1, otherwise q is Taking 0; p is a radical of is A predicted probability for observing that the ith sample belongs to class s;
(5) the total error L in the optimized SAE training process is as follows:
L＝L 1 (W，b)+L 2 ；
(6) in each training process, in order to minimize L, updating the optimized network parameters of SAE through a back propagation algorithm, wherein the calculation formula is as follows:
Figure FDA0003710431970000021
in the formula (I), the compound is shown in the specification,
Figure FDA0003710431970000022
and
Figure FDA0003710431970000023
is the partial differential of L with respect to W and b, and α is the learning rate of SAE;
(7) judging whether the preset training times are reached, if so, stopping training; otherwise, returning to the step (2), wherein the connection weight and the deviation of the neurons of each layer of the neural network when the training is terminated are the connection weight and the deviation of each neuron, obtaining the SAE after the training of the BP neural network with the softmax classifier is finished, and forming a data enhancement module by the SAE.
2. The method as claimed in claim 1, wherein the convolutional neural network module comprises a convolutional neural network model, and the convolutional neural network model comprises an input layer, two alternating convolutional layers and pooling layers, a fully-connected layer and an output layer.
3. The method for diagnosing the motor bearing fault based on the CNN with the enhanced one-dimensional data of claim 2, wherein the process of training with the two-dimensional image data set as the training set comprises: inputting a two-dimensional image data set into a convolutional neural network model, training a CNN model by taking a fault type label corresponding to an input two-dimensional image as theoretical output, and determining the connection weight and deviation of each neuron, wherein the method specifically comprises the following steps:
(1) initializing parameters of a CNN model;
(2) selecting the two-dimensional image dataset;
(3) inputting a two-dimensional image data set, training by using an Adam optimizer, and adjusting the weight and the deviation of each layer of neural network in each step of iteration process to enable the fault precision to be higher than that before the iterative computation; the fault precision calculation mode is as follows: the ratio of the number of data with the fault diagnosis type label given to the output data and the collected fault type label consistent with each other to the total sample data number after model classification;
(4) judging whether the training times reach a preset value or not, and if so, stopping training; otherwise, returning to the step (3), and the connection weight and the deviation of the neurons of each layer of the neural network when the training is terminated are the connection weight and the deviation of each neuron.
4. The method as claimed in claim 1, wherein when the convolutional neural network module operates, zero values automatically fill the convolutional process, let epsilon be the output size, delta be the input size, eta be the span, gamma be the filter width, and left zero padding PL and right zero padding PR are calculated therefrom, and the padding formula is as follows:
Figure FDA0003710431970000031
PT＝(ε-1)×η+γ-δ；
Figure FDA0003710431970000032
PR＝PT-PL；
in the formula, ceil (. cndot.) is an upward rounding function, floor (. cndot.) is a downward rounding function, and PT is the upper limit of zero padding.
5. The method for diagnosing the motor bearing fault based on the one-dimensional data enhanced CNN as claimed in claim 1, wherein the process of direct conversion by the data conversion module is as follows:
(1) randomly intercepting a section with the length of N from a one-dimensional vibration signal 2 Sequentially filling the signal values into an n × n matrix line by line, namely filling each n signal values into one line of the matrix;
(2) standardizing the obtained matrix to 0-255 to obtain an n multiplied by n image; the formula for the conversion is as follows:
Figure FDA0003710431970000033
in the formula, D (e, f) is the value of the pixel point of the ith row and the fth column of the obtained picture; (i) is the value of the ith sample point in the intercepted signal segment, max (-) is the maximum function, min (-) is the minimum function; round (·) is a rounding function that preserves the integer part according to the rounding principle.
6. The method for diagnosing the motor bearing fault based on the one-dimensional data enhanced CNN as claimed in claim 1, wherein the preprocessing is to normalize a data matrix composed of fault characteristic parameters of each signal point in the one-dimensional vibration signal, and the processing formula is as follows:
Figure FDA0003710431970000034
in the formula, x * Is normalized data, x is data before normalization, x min Is the minimum value of x, x max Is the maximum value of x.
7. The method of claim 1, wherein the fault diagnosis type labels comprise ball fault, outer periphery fault and inner periphery fault.
8. The method as claimed in claim 1, wherein the initializing parameters of the SAE and the BP neural network refer to: presetting the training times to be 800, initializing the learning rate of SAE to be 0.0001, and setting the number of training samples in each batch to be 50; the number of neurons of each layer of the three-layer SAE neural network is initialized to be 1024, 1000 and 1000, and the number of neurons of each layer of the three-layer BP neural network is 600, 300 and 16 respectively; initializing the weight of each layer of neural network as a unit vector with the same dimension as the input data dimension, wherein the deviation is 0; the activation function of the initialized hidden layer is a ReLU function, and the activation function of the output layer is a sigmoid function.
9. The method for diagnosing the motor bearing fault based on the one-dimensional data enhanced CNN as claimed in claim 3, wherein the parameters for initializing the CNN model refer to: the initial learning rate is 0.0005, the training step length is 1, the training times are 800, and the size of each batch of training sample sets is 256; initializing convolution layer convolution kernel size to be 5 x 5, pooling layer filter size to be 2 x 2, and number of full-connection layer neurons to be 2048; the weight for initializing each layer of neural network is a unit vector with the same dimension as the input data dimension, and the deviation is 0.
CN202110365267.0A 2021-04-06 2021-04-06 Motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN Active CN113203566B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110365267.0A CN113203566B (en) 2021-04-06 2021-04-06 Motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110365267.0A CN113203566B (en) 2021-04-06 2021-04-06 Motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN

Publications (2)

Publication Number Publication Date
CN113203566A CN113203566A (en) 2021-08-03
CN113203566B true CN113203566B (en) 2022-09-16

Family

ID=77026161

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110365267.0A Active CN113203566B (en) 2021-04-06 2021-04-06 Motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN

Country Status (1)

Country Link
CN (1) CN113203566B (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113740667B (en) * 2021-08-30 2022-06-14 华北电力大学 Power grid fault diagnosis method integrating self-encoder and convolutional neural network
CN113836027B (en) * 2021-09-28 2023-07-21 重庆大学 Method for generating failure test case by using generation type network
CN113935351B (en) * 2021-11-22 2023-12-19 山西警察学院 Non-contact vibration frequency detection and forward and backward rotation identification system
CN114757239B (en) * 2022-06-15 2022-08-30 浙江大学 Fan fault migratable diagnosis method based on data enhancement and capsule neural network
CN115753103A (en) * 2022-08-30 2023-03-07 山东建筑大学 Fault diagnosis method and system based on standard self-learning data enhancement
CN116701912B (en) * 2023-06-14 2023-11-14 盐城工学院 Bearing fault diagnosis method and system based on one-dimensional convolutional neural network

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107526853B (en) * 2016-06-22 2018-09-21 北京航空航天大学 Rolling bearing fault mode identification method and device based on stacking convolutional network
CN108268940B (en) * 2017-01-04 2022-02-18 意法半导体股份有限公司 Tool for creating reconfigurable interconnect frameworks
CN109918752B (en) * 2019-02-26 2022-12-16 华南理工大学 Mechanical fault diagnosis method, equipment and medium based on migration convolution neural network
CN110334764B (en) * 2019-07-04 2022-03-04 西安电子科技大学 Rotary machine intelligent fault diagnosis method based on integrated depth self-encoder
EP4030346A4 (en) * 2019-09-13 2022-10-19 Mitsubishi Electric Corporation Information processing device and electronic apparatus equipped with same
CN110647867B (en) * 2019-10-09 2022-05-13 中国科学技术大学 Bearing fault diagnosis method and system based on self-adaptive anti-noise neural network
CN112396109A (en) * 2020-11-19 2021-02-23 天津大学 Motor bearing fault diagnosis method based on recursion graph and multilayer convolution neural network

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
Saban Öztürk.Stacked auto-encoder based tagging with deep features for content-based medical image retrieval.《Expert Systems with Applications》.2020, *

Also Published As

Publication number Publication date
CN113203566A (en) 2021-08-03

Similar Documents

Publication Publication Date Title
CN113203566B (en) Motor bearing fault diagnosis method based on one-dimensional data enhancement and CNN
CN112417954B (en) Bearing fault mode diagnosis method and system for small sample data set
Zhu et al. Acoustic signal-based fault detection of hydraulic piston pump using a particle swarm optimization enhancement CNN
Yin et al. Wasserstein generative adversarial network and convolutional neural network (WG-CNN) for bearing fault diagnosis
CN110657984B (en) Planetary gearbox fault diagnosis method based on reinforced capsule network
Liang et al. Rolling bearing fault diagnosis based on one-dimensional dilated convolution network with residual connection
CN109389171B (en) Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology
CN111753891B (en) Rolling bearing fault diagnosis method based on unsupervised feature learning
CN112766355B (en) Electroencephalogram signal emotion recognition method under label noise
CN111222457B (en) Detection method for identifying authenticity of video based on depth separable convolution
CN110929765A (en) Convolution self-coding fault monitoring method based on batch imaging
CN112883994A (en) Rotating machinery variable working condition fault diagnosis method with balanced distribution adaptation
CN113281029A (en) Rotating machinery fault diagnosis method and system based on multi-scale network structure
CN114358124B (en) New fault diagnosis method for rotary machinery based on deep countermeasure convolutional neural network
CN112766283B (en) Two-phase flow pattern identification method based on multi-scale convolution network
CN113923104A (en) Network fault diagnosis method, equipment and storage medium based on wavelet neural network
CN114169377A (en) G-MSCNN-based fault diagnosis method for rolling bearing in noisy environment
CN113177577A (en) Bearing fault diagnosis method based on improved convolutional neural network
CN115290326A (en) Rolling bearing fault intelligent diagnosis method
CN116628605A (en) Method and device for electricity stealing classification based on ResNet and DSCAttention mechanism
CN114491823A (en) Train bearing fault diagnosis method based on improved generation countermeasure network
CN114441173A (en) Rolling bearing fault diagnosis method based on improved depth residual shrinkage network
Ding et al. Machinery probabilistic few-shot prognostics considering prediction uncertainty
CN113758709A (en) Rolling bearing fault diagnosis method and system combining edge calculation and deep learning
CN117516939A (en) Bearing cross-working condition fault detection method and system based on improved EfficientNetV2

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