CN114386526A - Combined convolution neural network diagnosis method for rotary machine fault - Google Patents

Combined convolution neural network diagnosis method for rotary machine fault Download PDF

Info

Publication number
CN114386526A
CN114386526A CN202210054695.6A CN202210054695A CN114386526A CN 114386526 A CN114386526 A CN 114386526A CN 202210054695 A CN202210054695 A CN 202210054695A CN 114386526 A CN114386526 A CN 114386526A
Authority
CN
China
Prior art keywords
dimensional
layer
convolution
model
neural network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210054695.6A
Other languages
Chinese (zh)
Inventor
杜文辽
王宏超
李川
胡鹏杰
侯绪坤
巩晓赟
赵峰
谢贵重
孟凡念
郭志强
王良文
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.)
Zhengzhou University of Light Industry
Original Assignee
Zhengzhou University of Light Industry
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 Zhengzhou University of Light Industry filed Critical Zhengzhou University of Light Industry
Priority to CN202210054695.6A priority Critical patent/CN114386526A/en
Publication of CN114386526A publication Critical patent/CN114386526A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/2433—Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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

Landscapes

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

Abstract

The invention provides a combined convolution neural network diagnosis method for rotary machine faults, which comprises the steps of data acquisition, data preprocessing, 1D-2D JCNN model construction, model training, verification, diagnosis and the like. The method specifically aims at the vibration signals acquired in different states, multi-scale feature vectors of the signals are acquired in a self-adaptive mode through one-dimensional convolution, the feature vectors are constructed into two-dimensional vectors, and the two-dimensional vectors are used as the input of a two-dimensional convolution neural network. According to the invention, when a 1D-2D JCNN model is constructed, the two-dimensional structural expression of a one-dimensional convolutional neural network self-adaptive construction signal and the strong characteristic learning capacity of the two-dimensional convolutional neural network are fully utilized, the two convolutional neural networks with different structures are unified into an integral frame, and a combined convolutional neural network model for rotary machine fault diagnosis is developed.

Description

Combined convolution neural network diagnosis method for rotary machine fault
Technical Field
The invention relates to a combined convolution neural network diagnosis method for faults of rotary machinery, and belongs to the technical field of intelligent fault diagnosis of rotary machinery.
Background
The rotating machinery is widely applied to various industries and develops towards the direction of precision and intellectualization. If mechanical equipment fails, abnormal shutdown of the equipment is caused, which not only causes great property loss, but also may endanger life safety of people. The study of fault diagnosis of rotary machines has been the focus of research by researchers. Due to the fact that a large amount of state information is carried by vibration signals of equipment, a large amount of research results are obtained by using the vibration signals of the equipment and a fault diagnosis technology based on data driving. Generally, the diagnosis technology needs several steps of information acquisition, data preprocessing, feature extraction, feature dimension reduction, pattern recognition and the like, but the feature extraction and the feature dimension reduction usually need a large amount of manual experience, and the traditional pattern recognition technology is shallow learning, so that the obtained diagnosis precision is not ideal. In recent years, a deep learning method draws attention of a large number of scholars, a convolutional neural network is one of the most effective deep learning models, has the characteristics of weight sharing, local perception, multi-core participation, automatic feature extraction and the like, has great success in the fields of pattern recognition, image voice recognition and the like, and currently, a large number of expert scholars apply the convolutional neural network to the field of fault diagnosis.
CNN is mainly used for processing two-dimensional image problems initially, and for one-dimensional vibration signals, the one-dimensional signals need to be converted into two-dimensional vectors through some transformation, and common conversion methods include: short-time fourier transform (STFT), Continuous Wavelet Transform (CWT), Wavelet Packet Transform (WPT), and the like. The methods obtain better diagnosis results, particularly a convolutional neural network model based on wavelet transformation, the time-frequency analysis capability of the wavelet transformation is fully utilized, but the methods seriously depend on manual experience, different wavelet mother functions are selected, and the diagnosis results are often greatly different.
In recent years, deep learning has been successfully applied in a variety of fields because it is possible to directly learn features of different states from collected signals, and thus, good classification performance is obtained. However, the one-dimensional convolutional neural network is used for directly processing the original signal, and because of the limitation of the information utilization capacity, the one-dimensional convolutional neural network is only used as an intermediate step for establishing a diagnostic model, or further optimization is needed, and the manual experience intervention cannot be avoided.
Disclosure of Invention
In view of the shortcomings of the prior art, the invention aims to provide a joint convolutional neural network diagnosis method for rotating machinery faults.
In order to achieve the purpose, the invention adopts the technical scheme that:
a joint convolution neural network diagnosis method for rotary machine faults comprises the following steps:
step 1, data acquisition:
setting sampling frequency to complete the vibration signal acquisition of the diagnosis object in each state;
step 2, data preprocessing:
carrying out normalization processing on the acquired data, and dividing the data into a training set, a verification set and a test set;
step 3, constructing a 1D-2D JCNN model:
the 1D-2D JCNN model consists of a 1D convolution layer, a 2D convolution layer, a pooling layer and a full-connection layer; firstly, constructing a one-dimensional convolutional neural network, wherein the one-dimensional convolutional neural network only has one convolutional layer, and the size, the step length and the number of n convolutional kernels are set so that the length of a generated characteristic graph is n; inputting a one-dimensional vibration signal into the one-dimensional convolution neural network, performing convolution operation on the one-dimensional signal to generate n characteristic graphs, overlapping the n characteristic graphs together to generate an n multiplied by n two-dimensional matrix, and taking the matrix as the input of the two-dimensional convolution neural network; initializing connection weight values and bias parameters of each layer of the model;
step 4, model training:
selecting cross entropy as a loss function, inputting a training set and a verification set into a constructed 1D-2D JCNN model, starting training on the network model by using a back propagation algorithm, updating the weight by using a gradient descent method, calculating the gradient of each layer by using an error back propagation method according to an error chain propagation rule, and stopping training and storing the trained model when the condition of terminating model training is met;
step 5, verification:
verifying the model obtained in the step 4 by using a verification set sample; if the diagnosis precision does not meet the requirement, the process of the step 2-4 is executed again until the diagnosis precision meets the requirement, and then the training is completed to obtain the final model parameters;
step 6, diagnosis:
and inputting the test set into the trained model to obtain a diagnosis accuracy test result of the model.
The specific method for constructing the 1D-2D JCNN model in the step 3 comprises the following steps:
a. constructing a 1D-2D joint convolution neural network structure: the network consists of a 1D convolutional layer, a 2D convolutional layer, a pooling layer and a full-connection layer, wherein the model structure parameters of the 1D-2D JCNN are 16 layers of the network structure, and the network structure comprises 1 input layer, 1 one-dimensional convolutional layer, 4 two-dimensional convolutional layers, 4 BN layers, 4 pooling layers and 2 Dense layers; the convolutional layer activation function adopts a ReLU function, and a BN layer and a pooling layer are closely arranged behind each two-dimensional convolutional layer; the Dense _1 layer is a full connection layer, the Dense _2 layer is an output layer, and a classifier function adopts a Softmax classifier;
b. forward calculation: input X (X) to joint convolutional network0) Firstly, one-dimensional convolution is carried out, and the output of the one-dimensional convolution is a one-dimensional convolution characteristic diagram;
Figure BDA0003475970840000031
wherein the content of the first and second substances,
Figure BDA0003475970840000032
the ith convolution kernel representing the ith layer,
Figure BDA0003475970840000033
for the jth local input of the l-1 th layer,
Figure BDA0003475970840000034
to output, Conv1D (mush) is a one-dimensional convolution calculation,
Figure BDA0003475970840000035
bias ith for layer l, area calculated for convolution for N, and activation function for F (& ltSUB & gt); the ReLU function is chosen here as the activation function, expressed as:
F(a)＝max{0,a}
assuming that n convolution kernels are in total in the one-dimensional convolution layer, the result obtained by performing one-dimensional convolution on the p-th sample in the original signal X is Xij，xijThe result of the jth local part of the sample p passing through the ith convolution kernel is shown, and the characteristic graph obtained by the sample p passing through the ith convolution kernel is fi；
Superposing the obtained n characteristic maps to construct a two-dimensional characteristic map, wherein the specific construction process comprises the steps of reconstructing n 1 multiplied by n matrixes obtained by one-dimensional convolution for each sample into an n multiplied by n two-dimensional matrix, and the matrix element x of the two-dimensional matrixi,j＝xijUsing a matrix FpExpressed as:
Figure BDA0003475970840000036
taking a two-dimensional picture constructed by one-dimensional convolution as the input of a two-dimensional convolution network to carry out two-dimensional convolution operation, wherein the output of the neuron is as follows:
Figure BDA0003475970840000037
where k represents a convolution kernel, N is a convolution calculation region,
Figure BDA0003475970840000038
for the input of the first layer of convolution,
Figure BDA0003475970840000039
conv2D (mush) as the two-dimensional convolution calculation, b, for the convolution outputlFor bias, F (& lt) is the activation function, and likewise, is chosen to be the ReLU function.
The specific process of model training in step 4 is as follows: in the network training process, the weight value is updated by adopting a gradient descent method, the gradient of each layer is calculated by adopting an error L back propagation method according to the chain propagation rule of errors, and in the two-dimensional convolution part, the error of the convolution layer is assumed to be
Figure BDA00034759708400000310
The error of a hidden layer thereon
Figure BDA00034759708400000311
And need to update parameters
Figure BDA00034759708400000312
Comprises the following steps:
Figure BDA00034759708400000313
Figure BDA0003475970840000041
wherein the content of the first and second substances,
Figure BDA0003475970840000042
in order to correct the error of the convolution layer,
Figure BDA0003475970840000043
in order to be the error of the previous hidden layer,
Figure BDA0003475970840000044
in order to convolve the layer weights,
Figure BDA0003475970840000045
in order to obtain the weight of the previous hidden layer,
Figure BDA0003475970840000046
for convolutional layer output, L is the initial error.
When the error is propagated to the one-dimensional-two-dimensional connecting portion in the reverse direction, the error transmitted from the two-dimensional convolution layer is set
Figure BDA0003475970840000047
Comprises the following steps:
Figure BDA0003475970840000048
then its one-dimensional convolution kernel gradient update procedure can be formulated as:
Figure BDA0003475970840000049
wherein the content of the first and second substances,
Figure BDA00034759708400000410
error of convolution kernel, w1+iAs a result of the original value of the value,
Figure BDA00034759708400000411
the error value transmitted by the two-dimensional convolution layer, L is the initial error.
English abbreviation of Chinese word and phrase: one-dimensional convolutional neural network: 1DCNN, two-dimensional convolutional neural network: 2 DCNN.
The invention has the beneficial effects that:
the invention provides a rotary machine fault diagnosis method based on a one-dimensional-two-dimensional combined convolution neural network, which is more specific to vibration signals acquired in different states, utilizes one-dimensional convolution to adaptively acquire multi-scale feature vectors of the signals, constructs the feature vectors into two-dimensional vectors, and uses the two-dimensional vectors as the input of the two-dimensional convolution neural network. According to the invention, when a 1D-2D JCNN model is constructed, the two-dimensional structural expression of a one-dimensional convolutional neural network self-adaptive construction signal and the strong characteristic learning capacity of the two-dimensional convolutional neural network are fully utilized, the two convolutional neural networks with different structures are unified into an integral frame, and a combined convolutional neural network model for rotary machine fault diagnosis is developed.
The method takes the cross entropy error function as a loss function, optimizes the filter parameters of the one-dimensional and two-dimensional combined network model by using an error back propagation algorithm, and obtains a final fault diagnosis model, so that the model obtains better diagnosis performance and good data adaptivity, and can be used for diagnosing various rotary machine faults. The invention establishes a training method of model parameters of a 1D-2D joint convolution neural network in model training. A mechanism for transmitting model training errors from the two-dimensional convolutional neural network to the one-dimensional convolutional neural network is deduced, seamless connection between the two-dimensional convolutional neural network and the one-dimensional convolutional neural network is achieved, and updating of weights and bias parameters of all layers of the whole model is achieved by a gradient descent method.
Drawings
Fig. 1 is a flow chart of the fault diagnosis of the present invention.
Fig. 2 is a waveform diagram of a vibration signal of a bearing fault of class 10 of an experimental data sample in an application example of the invention.
The method comprises the following steps of A, a normal state, B, a slight inner ring fault, C, a slight rolling body fault, d, a slight outer ring fault, e, a medium inner ring fault, f, a medium rolling body fault, g, a medium outer ring fault, h, a serious inner ring fault, i, a serious rolling body fault and j, a serious outer ring fault; the upper graph is a one-dimensional signal graph, and the lower graph is a two-dimensional gray scale graph.
Fig. 3 is a diagram showing the result of fault diagnosis of 10 training tests of bearing data in the application example of the present invention.
Wherein, Times represents the Times, and Accuracy represents the Accuracy.
Detailed Description
The following examples are provided to further illustrate the embodiments of the present invention, and the embodiments and specific procedures of the present invention are given on the premise of the technical solution of the present invention, but the scope of the present invention is not limited to the following examples.
A joint convolution neural network diagnosis method for rotary machine faults comprises the following steps:
step 1, data acquisition:
and finishing the acquisition of vibration signals of the diagnosis object in each state according to a preset sampling frequency.
Step 2, data preprocessing:
and carrying out normalization processing on the acquired data, and dividing the data into a training set, a verification set and a test set according to a set proportion.
Step 3, constructing a 1D-2D JCNN model:
the 1D-2D JCNN model consists of a 1D convolution layer, a 2D convolution layer, a pooling layer and a full-connection layer; firstly, a one-dimensional convolutional neural network is constructed, wherein the one-dimensional convolutional neural network only has one convolutional layer, and proper convolutional kernel size, step length (note: the length of the generated feature map is n) and the number of n convolutional kernels are set. Inputting a one-dimensional vibration signal into the one-dimensional convolution neural network, performing convolution operation on the one-dimensional signal to generate n characteristic maps, superposing the n characteristic maps together to generate an n multiplied by n two-dimensional matrix, and taking the matrix as the input of the two-dimensional convolution neural network. And initializing the connection weight and bias parameters of each layer of the model.
The specific method for constructing the 1D-2D JCNN model comprises the following steps:
a. and constructing a 1D-2D joint convolution neural network structure. The network is composed of a 1D convolutional layer, a 2D convolutional layer, a pooling layer and a full-connection layer, model structure parameters of the 1D-2D JCNN are 16 layers of the network structure, and the network structure comprises 1 input layer, 1 one-dimensional convolutional layer, 4 two-dimensional convolutional layers, 4 BN layers, 4 pooling layers and 2 Dense layers. The convolutional layer activation function adopts a ReLU function, and each two-dimensional convolutional layer is followed by a BN layer and a pooling layer. The Dense _1 layer is a full connection layer, the Dense _2 layer is an output layer, and the classifier function adopts a Softmax classifier.
b. And (4) forward calculation. Input X (X) to joint convolutional network0) First, after one-dimensional convolution, the output is a one-dimensional convolution characteristic diagram.
Figure BDA0003475970840000061
Wherein the content of the first and second substances,
Figure BDA0003475970840000062
the ith convolution kernel representing the ith layer,
Figure BDA0003475970840000063
for the jth local input of the l-1 th layer,
Figure BDA0003475970840000064
to output, Conv1D (mush) is a one-dimensional convolution calculation,
Figure BDA0003475970840000065
bias ith for layer l, area calculated for convolution for N, and activation function for F (& lt). The ReLU function is chosen here as the activation function, expressed as:
F(a)＝max{0,a}
assuming that n convolution kernels are in total in the one-dimensional convolution layer, the result obtained by performing one-dimensional convolution on the p-th sample in the original signal X is Xij，xijThe result of the jth local part of the sample p passing through the ith convolution kernel is shown, and the characteristic graph obtained by the sample p passing through the ith convolution kernel is fi。
Superposing the obtained n characteristic maps to construct a two-dimensional characteristic map, wherein the specific construction process comprises the steps of reconstructing n 1 multiplied by n matrixes obtained by one-dimensional convolution of each sample into oneTwo-dimensional matrix of n x n, the matrix elements x of whichi,j＝xijUsing a matrix FpExpressed as:
Figure BDA0003475970840000066
taking a two-dimensional picture of the one-dimensional convolution structure as the input of a two-dimensional convolution network to carry out two-dimensional convolution operation, wherein the output of the neuron is
Figure BDA0003475970840000067
Where k represents a convolution kernel, N is a convolution calculation region,
Figure BDA0003475970840000068
for the input of the first layer of convolution,
Figure BDA0003475970840000069
conv2D (mush) as the two-dimensional convolution calculation, b, for the convolution outputlFor bias, F (& lt) is the activation function, and likewise, is chosen to be the ReLU function.
And 4, model training:
selecting cross entropy as a loss function, inputting a training set and a verification set into a constructed 1D-2D JCNN model, starting training on the network model by using a back propagation algorithm, updating the weight by using a gradient descent method, calculating the gradient of each layer by using an error back propagation method according to an error chain propagation rule, and stopping training and storing the trained model when the condition of terminating the model training is met.
The method has the advantages that a mechanism for transmitting model training errors from the 2D convolutional neural network to the 1D convolutional neural network is deduced, seamless connection of the 2D convolutional neural network and the 1D convolutional neural network is achieved, and updating of weights and bias parameters of all layers of the whole model is achieved by using a gradient descent method.
The specific process is as follows: in the network training process, the weight value is updated by adopting a gradient descent method, and errors are adopted according to the chain propagation rule of the errorsCalculating the gradient of each layer by the difference L back propagation method, and in the two-dimensional convolution part, assuming the error of the convolution layer as
Figure BDA0003475970840000071
The error of a hidden layer thereon
Figure BDA0003475970840000072
And need to update parameters
Figure BDA0003475970840000073
Comprises the following steps:
Figure BDA0003475970840000074
Figure BDA0003475970840000075
wherein the content of the first and second substances,
Figure BDA0003475970840000076
in order to correct the error of the convolution layer,
Figure BDA0003475970840000077
in order to be the error of the previous hidden layer,
Figure BDA0003475970840000078
in order to convolve the layer weights,
Figure BDA0003475970840000079
in order to obtain the weight of the previous hidden layer,
Figure BDA00034759708400000710
for convolutional layer output, L is the initial error.
When the error is propagated to the one-dimensional-two-dimensional connecting portion in the reverse direction, the error transmitted from the two-dimensional convolution layer is set
Figure BDA00034759708400000711
Comprises the following steps:
Figure BDA00034759708400000712
then its one-dimensional convolution kernel gradient update procedure can be formulated as:
Figure BDA00034759708400000713
wherein the content of the first and second substances,
Figure BDA00034759708400000714
error of convolution kernel, w1+iAs a result of the original value of the value,
Figure BDA00034759708400000715
the error value transmitted by the two-dimensional convolution layer, L is the initial error.
Step 5, verification:
verifying the deep hybrid convolutional neural network model obtained in the step 4 by using a verification set sample; if the diagnosis precision does not meet the requirement, the process of the step 2-4 is executed again until the diagnosis precision meets the requirement, and then the training is completed to obtain the final model parameters;
and 6, diagnosis:
and inputting the test set into the trained model to obtain a diagnosis accuracy test result of the model.
Application example:
a fault diagnosis method for a rotary machine based on a one-dimensional-two-dimensional combined convolutional neural network is disclosed, and the fault diagnosis process is shown in figure 1.
The experimental data set adopts a published data set CWRU OF American West university OF storage, the bearing to be tested consisting OF a motor, a coupler and a load motor is a motor driving end bearing, the bearing supports a rotating shaft OF the motor, the damage OF the bearing is single-point damage simulated by electric spark machining, the state OF the bearing can be divided into normal (N), inner ring fault (IF), rolling Body Fault (BF) and outer ring fault (OF), the damage degree (fault diameter) according to the fault is different, each fault type can be divided into mild (fault diameter is 0.18mm), moderate (fault diameter is 0.36mm) and severe (fault diameter is 0.53mm) according to different degrees OF the damage degree (fault diameter) OF the fault. The motor operates at approximately 1800 rpm under four loads of 0, 1, 2, and 3. The data acquisition was performed at both 12kHz and 48kHz sampling frequencies, and the data acquired at the 48kHz sampling frequency was used herein as experimental data.
According to the different loads, the experimental data are constructed into 4 data sets, the data set A, B, C, D corresponds to the loads of 0hp, 1hp, 2hp and 3hp, the data set comprises all 4 loads, each data set has 10 health states (as shown in fig. 2, a-j), 10000 data samples, each sample comprises 2048 data, and the data are processed by the following steps as shown in fig. 2:
step 1: 10000 collected data samples are subjected to normalization processing, and the data are divided into a training set, a verification set and a test set according to the proportion of (6:1: 3).
Step 2: the JCNN model is constructed and comprises a 1D convolutional layer, a 2D convolutional layer, a pooling layer and a full-connection layer, and the model structure parameters of the 1D-2D JCNN are 16 layers of a network structure, and comprise 1 input layer, 1 one-dimensional convolutional layer, 4 two-dimensional convolutional layers, 4 BN layers, 4 pooling layers and 2 Dense layers. The convolutional layer activation function adopts a ReLU function, and each two-dimensional convolutional layer is followed by a BN layer and a pooling layer. The Dense _1 layer is a full connection layer, the Dense _2 layer is an output layer, and the classifier function adopts Softmax. The multi-scale feature vector of the signal is obtained in a self-adaptive mode through one-dimensional convolution, and the obtained feature vector is constructed into a two-dimensional vector.
Input X (X) to joint convolutional network0) Firstly, one-dimensional convolution is carried out, and the output is a one-dimensional convolution characteristic diagram
Figure BDA0003475970840000091
Wherein the content of the first and second substances,
Figure BDA0003475970840000092
i-th volume representing the l-th layerThe number of the kernels is accumulated,
Figure BDA0003475970840000093
for the jth local input of the l-1 th layer,
Figure BDA0003475970840000094
to output, Conv1D (mush) is a one-dimensional convolution calculation,
Figure BDA0003475970840000095
bias ith for layer l, area calculated for convolution for N, and activation function for F (& lt). The ReLU function is chosen here as the activation function, expressed as: f (a) ═ max {0, a }
Setting n convolution kernels in the one-dimensional convolution layer, wherein m samples are in the training set, and the result obtained by performing one-dimensional convolution on the p-th sample in the original signal X is Xij，xijThe result of the jth local part of the sample p passing through the ith convolution kernel is shown, and the characteristic graph obtained by the sample p passing through the ith convolution kernel is fi。
Superposing the obtained n characteristic maps to form a two-dimensional characteristic map, wherein the matrix element x of the two-dimensional characteristic mapi,j＝xijUsing a matrix FpExpressed as:
Figure BDA0003475970840000096
taking a two-dimensional picture of the one-dimensional convolution structure as the input of a two-dimensional convolution network to carry out two-dimensional convolution operation, wherein the output of the neuron is
Figure BDA0003475970840000097
Where k represents a convolution kernel, N is a convolution calculation region,
Figure BDA0003475970840000098
for the input of the first layer of convolution,
Figure BDA0003475970840000099
conv2D (mush) as the two-dimensional convolution calculation, b, for the convolution outputlFor bias, F (& lt) is the activation function, and likewise, is chosen to be the ReLU function.
And step 3: and initializing and setting the connection weight w and the bias parameter b of each layer of the model.
And 4, inputting the test set and the verification set into the constructed JCNN model, starting training the network model by using a back propagation algorithm, and stopping training and storing the trained model when the model training termination condition is met.
In the network training process, the weight value is updated by adopting a gradient descent method, the gradient of each layer is calculated by adopting an error back propagation method according to the chain propagation rule of errors, and in the two-dimensional convolution part, the error of the convolution layer is assumed to be
Figure BDA00034759708400000910
The error of a hidden layer thereon
Figure BDA00034759708400000911
And need to update parameters
Figure BDA00034759708400000912
Comprises the following steps:
Figure BDA00034759708400000913
Figure BDA0003475970840000101
wherein the content of the first and second substances,
Figure BDA0003475970840000102
in order to correct the error of the convolution layer,
Figure BDA0003475970840000103
in order to be the error of the previous hidden layer,
Figure BDA0003475970840000104
in order to convolve the layer weights,
Figure BDA0003475970840000105
in order to obtain the weight of the previous hidden layer,
Figure BDA0003475970840000106
for convolutional layer output, L is the initial error.
When the error is propagated to the one-dimensional-two-dimensional connecting portion in the reverse direction, the error transmitted from the two-dimensional convolution layer is set
Figure BDA0003475970840000107
Comprises the following steps:
Figure BDA0003475970840000108
then its one-dimensional convolution kernel gradient update procedure can be formulated as:
Figure BDA0003475970840000109
wherein the content of the first and second substances,
Figure BDA00034759708400001010
error of convolution kernel, w1+iAs a result of the original value of the value,
Figure BDA00034759708400001011
the error value transmitted by the two-dimensional convolution layer, L is the initial error.
And 5, testing the trained model by using the test sample for 10 times, wherein the obtained diagnosis precision of the training and testing faults is shown in figure 3.
And 6, diagnosing the actual sample by using the obtained model to obtain a diagnosis result.
In the model training process, a training method of the model parameters of the 1D-2D joint convolution neural network is established. A mechanism for transmitting model training errors from the 2D convolutional neural network to the 1D convolutional neural network is deduced, seamless connection of the 2D convolutional neural network and the 1D convolutional neural network is achieved, and updating of weights and bias parameters of all layers of the whole model is achieved by a gradient descent method.
The method comprises the steps of utilizing one-dimensional convolution to adaptively obtain multi-scale feature vectors of signals aiming at vibration signals acquired in different states, constructing the feature vectors into two-dimensional vectors, and using the two-dimensional vectors as the input of a two-dimensional convolution neural network. And selecting the cross entropy as a loss function, and optimizing the filter parameters of the one-dimensional-two-dimensional combined network model by using an error back propagation algorithm to obtain a final fault diagnosis model. Better diagnostic performance is obtained.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention.

Claims (3)

1. A joint convolution neural network diagnosis method for rotary machine faults is characterized by comprising the following steps:
step 1, data acquisition:
setting sampling frequency to complete the vibration signal acquisition of the diagnosis object in each state;
step 2, data preprocessing:
carrying out normalization processing on the acquired data, and dividing the data into a training set, a verification set and a test set;
step 3, constructing a 1D-2D JCNN model:
the 1D-2D JCNN model consists of a 1D convolution layer, a 2D convolution layer, a pooling layer and a full-connection layer; firstly, constructing a one-dimensional convolutional neural network, wherein the one-dimensional convolutional neural network only has one convolutional layer, and the size, the step length and the number of n convolutional kernels are set so that the length of a generated characteristic graph is n; inputting a one-dimensional vibration signal into the one-dimensional convolution neural network, performing convolution operation on the one-dimensional signal to generate n characteristic graphs, overlapping the n characteristic graphs together to generate an n multiplied by n two-dimensional matrix, and taking the matrix as the input of the two-dimensional convolution neural network; initializing connection weight values and bias parameters of each layer of the model;
step 4, model training:
selecting cross entropy as a loss function, inputting a training set and a verification set into a constructed 1D-2D JCNN model, starting training on the network model by using a back propagation algorithm, updating the weight by using a gradient descent method, calculating the gradient of each layer by using an error back propagation method according to an error chain propagation rule, and stopping training and storing the trained model when the condition of terminating model training is met;
step 5, verification:
verifying the model obtained in the step 4 by using a verification set sample; if the diagnosis precision does not meet the requirement, the process of the step 2-4 is executed again until the diagnosis precision meets the requirement, and then the training is completed to obtain the final model parameters;
step 6, diagnosis:
and inputting the test set into the trained model to obtain a diagnosis accuracy test result of the model.
2. The method for diagnosing the joint convolutional neural network of the rotating machine fault as claimed in claim 1, wherein the specific method for constructing the 1D-2D JCNN model in the step 3 is as follows:
a. constructing a 1D-2D joint convolution neural network structure: the network consists of a 1D convolutional layer, a 2D convolutional layer, a pooling layer and a full-connection layer, wherein the model structure parameters of the 1D-2D JCNN are 16 layers of the network structure, and the network structure comprises 1 input layer, 1 one-dimensional convolutional layer, 4 two-dimensional convolutional layers, 4 BN layers, 4 pooling layers and 2 Dense layers; the convolutional layer activation function adopts a ReLU function, and a BN layer and a pooling layer are closely arranged behind each two-dimensional convolutional layer; the Dense _1 layer is a full connection layer, the Dense _2 layer is an output layer, and a classifier function adopts a Softmax classifier;
b. forward calculation: input X (X) to joint convolutional network0) Firstly, one-dimensional convolution is carried out, and the output of the one-dimensional convolution is a one-dimensional convolution characteristic diagram;
Figure FDA0003475970830000021
wherein the content of the first and second substances,
Figure FDA0003475970830000022
the ith convolution kernel representing the ith layer,
Figure FDA0003475970830000023
for the jth local input of the l-1 th layer,
Figure FDA0003475970830000024
to output, Conv1D (mush) is a one-dimensional convolution calculation,
Figure FDA0003475970830000025
bias ith for layer l, area calculated for convolution for N, and activation function for F (& ltSUB & gt); the ReLU function is chosen here as the activation function, expressed as:
F(a)＝max{0,a}
assuming that n convolution kernels are in total in the one-dimensional convolution layer, the result obtained by performing one-dimensional convolution on the p-th sample in the original signal X is Xij，xijThe result of the jth local part of the sample p passing through the ith convolution kernel is shown, and the characteristic graph obtained by the sample p passing through the ith convolution kernel is fi；
Superposing the obtained n characteristic maps to construct a two-dimensional characteristic map, wherein the specific construction process comprises the steps of reconstructing n 1 multiplied by n matrixes obtained by one-dimensional convolution for each sample into an n multiplied by n two-dimensional matrix, and the matrix element x of the two-dimensional matrixi,j＝xijUsing a matrix FpExpressed as:
Figure FDA0003475970830000026
taking a two-dimensional picture constructed by one-dimensional convolution as the input of a two-dimensional convolution network to carry out two-dimensional convolution operation, wherein the output of the neuron is as follows:
Figure FDA0003475970830000027
where k represents a convolution kernel, N is a convolution calculation region,
Figure FDA0003475970830000028
for the input of the first layer of convolution,
Figure FDA0003475970830000029
conv2D (mush) as the two-dimensional convolution calculation, b, for the convolution outputlFor bias, F (& lt) is the activation function, and likewise, is chosen to be the ReLU function.
3. The method for diagnosing the convolutional neural network of the rotary machine related to the fault of claim 1, wherein the specific process of the model training in the step 4 is as follows: in the network training process, the weight value is updated by adopting a gradient descent method, the gradient of each layer is calculated by adopting an error L back propagation method according to the chain propagation rule of errors, and in the two-dimensional convolution part, the error of the convolution layer is assumed to be
Figure FDA00034759708300000210
The error of a hidden layer thereon
Figure FDA00034759708300000211
And need to update parameters
Figure FDA0003475970830000031
Comprises the following steps:
Figure FDA0003475970830000032
Figure FDA0003475970830000033
wherein the content of the first and second substances,
Figure FDA0003475970830000034
in order to correct the error of the convolution layer,
Figure FDA0003475970830000035
in order to be the error of the previous hidden layer,
Figure FDA0003475970830000036
in order to convolve the layer weights,
Figure FDA0003475970830000037
in order to obtain the weight of the previous hidden layer,
Figure FDA0003475970830000038
for convolutional layer output, L is the initial error.
When the error is propagated to the one-dimensional-two-dimensional connecting portion in the reverse direction, the error transmitted from the two-dimensional convolution layer is set
Figure FDA0003475970830000039
Comprises the following steps:
Figure FDA00034759708300000310
then its one-dimensional convolution kernel gradient update procedure can be formulated as:
Figure FDA00034759708300000311
wherein the content of the first and second substances,
Figure FDA00034759708300000312
error of convolution kernel, w1+iAs a result of the original value of the value,
Figure FDA00034759708300000313
the error value transmitted by the two-dimensional convolution layer, L is the initial error.
CN202210054695.6A 2022-01-18 2022-01-18 Combined convolution neural network diagnosis method for rotary machine fault Pending CN114386526A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210054695.6A CN114386526A (en) 2022-01-18 2022-01-18 Combined convolution neural network diagnosis method for rotary machine fault

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210054695.6A CN114386526A (en) 2022-01-18 2022-01-18 Combined convolution neural network diagnosis method for rotary machine fault

Publications (1)

Publication Number Publication Date
CN114386526A true CN114386526A (en) 2022-04-22

Family

ID=81203009

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210054695.6A Pending CN114386526A (en) 2022-01-18 2022-01-18 Combined convolution neural network diagnosis method for rotary machine fault

Country Status (1)

Country Link
CN (1) CN114386526A (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114970605A (en) * 2022-05-06 2022-08-30 大连理工大学 Multi-mode feature fusion neural network refrigeration equipment fault diagnosis method
CN115017945A (en) * 2022-05-24 2022-09-06 南京林业大学 Mechanical fault diagnosis method and system based on enhanced convolutional neural network
CN116090523A (en) * 2023-04-10 2023-05-09 中国航空油料集团有限公司 Rotary machine edge side fault diagnosis method and system based on network structure compression
CN117171544A (en) * 2023-09-06 2023-12-05 哈尔滨理工大学 Motor vibration fault diagnosis method based on multichannel fusion convolutional neural network
CN117454756A (en) * 2023-10-23 2024-01-26 广州航海学院 Microstrip antenna modeling method and device, electronic equipment and medium
CN117540190A (en) * 2023-11-22 2024-02-09 广东粤电科试验检测技术有限公司 Transformer fault diagnosis method and system based on CNN self-feature fusion

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114970605A (en) * 2022-05-06 2022-08-30 大连理工大学 Multi-mode feature fusion neural network refrigeration equipment fault diagnosis method
CN114970605B (en) * 2022-05-06 2024-06-14 大连理工大学 Refrigerating equipment fault diagnosis method of multi-mode feature fusion neural network
CN115017945A (en) * 2022-05-24 2022-09-06 南京林业大学 Mechanical fault diagnosis method and system based on enhanced convolutional neural network
CN116090523A (en) * 2023-04-10 2023-05-09 中国航空油料集团有限公司 Rotary machine edge side fault diagnosis method and system based on network structure compression
CN117171544A (en) * 2023-09-06 2023-12-05 哈尔滨理工大学 Motor vibration fault diagnosis method based on multichannel fusion convolutional neural network
CN117171544B (en) * 2023-09-06 2024-04-12 哈尔滨理工大学 Motor vibration fault diagnosis method based on multichannel fusion convolutional neural network
CN117454756A (en) * 2023-10-23 2024-01-26 广州航海学院 Microstrip antenna modeling method and device, electronic equipment and medium
CN117540190A (en) * 2023-11-22 2024-02-09 广东粤电科试验检测技术有限公司 Transformer fault diagnosis method and system based on CNN self-feature fusion

Similar Documents

Publication Publication Date Title
CN114386526A (en) Combined convolution neural network diagnosis method for rotary machine fault
Zhu et al. Intelligent fault diagnosis of hydraulic piston pump combining improved LeNet-5 and PSO hyperparameter optimization
US20210270244A1 (en) Method and system for fault diagnosis of gearbox of wind turbine generator
WO2020156348A1 (en) Structural damage identification method based on ensemble empirical mode decomposition and convolution neural network
CN112418013B (en) Complex working condition bearing fault diagnosis method based on meta-learning under small sample
CN112417954B (en) Bearing fault mode diagnosis method and system for small sample data set
CN111523509B (en) Equipment fault diagnosis and health monitoring method integrating physical and depth expression characteristics
CN110657984B (en) Planetary gearbox fault diagnosis method based on reinforced capsule network
CN106096562B (en) Gearbox of wind turbine method for diagnosing faults based on vibration signal blind sources separation and sparse component analysis
CN111458148A (en) CBAM-based convolutional neural network rolling bearing fault diagnosis method
CN114048769A (en) Multi-source multi-domain information entropy fusion and model self-optimization method for bearing fault diagnosis
CN113255882A (en) Bearing fault diagnosis method based on improved convolution capsule network
CN111810124A (en) Oil pumping well fault diagnosis method based on characteristic re-calibration residual convolution neural network model
Zhu et al. Research on deep feature learning and condition recognition method for bearing vibration
CN116340859A (en) Marine wind turbine generator gearbox fault diagnosis method based on vibration signals under noise background
CN114648044B (en) Vibration signal diagnosis and analysis method based on EEMD and depth domain countermeasure network
CN115290326A (en) Rolling bearing fault intelligent diagnosis method
CN114462446A (en) Rolling bearing fault diagnosis method based on vibration signal and electronic equipment
CN116754231A (en) Method for rapidly diagnosing faults of rolling bearing based on RegNet-SES
CN109696906A (en) Underwater robot propeller method for diagnosing faults based on small echo amendment Bayes's convolution energy
CN112160877A (en) Fan bearing fault diagnosis method based on SFA and CNN
CN117009814A (en) Engine fault diagnosis method and system based on multi-source information fusion
Kong et al. Untrained compound fault diagnosis for planetary gearbox based on adaptive learning VMD and DSSECNN
Jiang et al. Rolling bearing fault diagnosis based on convolutional capsule network
CN112347917B (en) Gas turbine fault diagnosis method, system, equipment and storage medium

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