CN114298096A - Bearing health state identification method and system based on stacked self-coding neural network - Google Patents
Bearing health state identification method and system based on stacked self-coding neural network Download PDFInfo
- Publication number
- CN114298096A CN114298096A CN202111579017.3A CN202111579017A CN114298096A CN 114298096 A CN114298096 A CN 114298096A CN 202111579017 A CN202111579017 A CN 202111579017A CN 114298096 A CN114298096 A CN 114298096A
- Authority
- CN
- China
- Prior art keywords
- bearing
- self
- coding
- vibration signal
- deep
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 64
- 230000036541 health Effects 0.000 title claims abstract description 63
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 58
- 238000005457 optimization Methods 0.000 claims abstract description 49
- 230000006870 function Effects 0.000 claims abstract description 44
- 210000002569 neuron Anatomy 0.000 claims abstract description 42
- 238000012549 training Methods 0.000 claims abstract description 40
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 24
- 238000007619 statistical method Methods 0.000 claims abstract description 21
- 239000013589 supplement Substances 0.000 claims abstract description 20
- 238000004140 cleaning Methods 0.000 claims abstract description 18
- 239000002245 particle Substances 0.000 claims abstract description 18
- 238000000605 extraction Methods 0.000 claims description 19
- 230000004913 activation Effects 0.000 claims description 16
- 238000012360 testing method Methods 0.000 claims description 13
- 238000007781 pre-processing Methods 0.000 claims description 11
- 230000003862 health status Effects 0.000 claims description 10
- 238000004364 calculation method Methods 0.000 claims description 9
- 230000000694 effects Effects 0.000 claims description 8
- 238000011156 evaluation Methods 0.000 claims description 8
- 230000002159 abnormal effect Effects 0.000 claims description 7
- 238000010276 construction Methods 0.000 claims description 7
- 238000011478 gradient descent method Methods 0.000 claims description 6
- 238000010835 comparative analysis Methods 0.000 claims description 4
- 239000006185 dispersion Substances 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 3
- 238000001994 activation Methods 0.000 description 12
- 230000005540 biological transmission Effects 0.000 description 4
- 238000003066 decision tree Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000005096 rolling process Methods 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000001502 supplementing effect Effects 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a bearing health state identification method and system based on a stacked self-coding neural network, belongs to the technical field of characteristic identification based on the neural network, and aims to solve the technical problem of accurately and timely identifying the engine bearing health state. The method comprises the following steps: acquiring bearing vibration signals of various damages, and carrying out data cleaning and integrity supplement on the bearing vibration signals; extracting characteristics by a time domain-frequency domain statistical analysis method; constructing a deep stacking self-coding network by adding a classifier in a basic stacking self-coding network and adding a sparse penalty item in a neuron, adding an additional penalty item in a corresponding loss function, training the network under the sparse penalty constraint condition, and performing hyper-parameter optimization through a particle swarm optimization algorithm; and carrying out unsupervised learning on the shallow feature of the vibration signal to be identified through the final depth stacking self-coding network to obtain the health state of the bearing.
Description
Technical Field
The invention relates to the technical field of feature recognition based on a neural network, in particular to a bearing health state recognition method and system based on a stacked self-coding neural network.
Background
In an engine, the main rotary motion mode is from a crankshaft connecting rod structure and transmission components such as gears, belt wheels, chain wheels and the like. The gas pressure in the engine cylinder is converted into crankshaft torsional moment through the connecting rod for output, so that the reciprocating motion of the piston can be changed into rotary motion by the crankshaft connecting rod mechanism, and in addition, the rotary motion of some accessories of the engine such as a camshaft and a gas distribution mechanism is required to be driven. A gear transmission set and other transmission devices in the engine are used as crankshaft power transmission intermediate parts, and various power outputs are realized. The crankshaft of the engine is usually of a large mass and runs under a high load for a long time, so that a main shaft bearing and a crankshaft bearing of the engine are very easy to wear, age and even irreversibly damage, so that the engine is abnormally vibrated during working, and further serious faults and even halt of a machine are caused. Similarly, when the gear set runs frequently at high speed, the rolling bearing working between the gear shaft and the engine fixing body is in rapid wear, the health state and performance thereof gradually decline, when the damage exceeds the critical point, serious bearing failure is caused, the brought violent vibration, impact component and abnormal friction can influence the related rotary subsystem of the engine and even the normal work of the whole machine, if the bearing component with the failure is not detected and replaced in time, the performance of the engine can be accelerated and degraded, and the service life of the engine is greatly influenced. Through the simple analysis, the method for identifying the health state of the engine bearing is researched, and an effective bearing state monitoring means is provided.
According to the relevant research data, when a certain part of the bearing is damaged, impact pulse force is excited on the surface of the damaged point, so that impact vibration is caused. According to different damage degrees, the impact strength is different, and the impact strength can be reflected by the impact component of the bearing vibration signal. When the bearing is in a normal working stage, the vibration signal is relatively stable, the vibration characteristic can be basically kept unchanged, if the bearing is damaged, the vibration characteristic can be changed, the damages of different parts and different degrees can be different on the vibration characteristic, and the specific difference can be expressed on the vibration signal of the bearing.
How to accurately and timely identify the health state of an engine bearing is a technical problem to be solved.
Disclosure of Invention
The invention aims to overcome the defects and provides a bearing health state identification method and system based on a stacked self-coding neural network to solve the technical problem of accurately and timely identifying the engine bearing health state.
In a first aspect, the bearing health status identification method based on the stacked self-coding neural network comprises the following steps:
acquiring bearing vibration signals of various damages and corresponding bearing health states, and performing data cleaning and integrity supplement on the bearing vibration signals to obtain initial historical data;
extracting the characteristics of the initial historical data through a time domain-frequency domain statistical analysis method to obtain initial vibration signal shallow layer characteristics, and constructing a characteristic label data set based on the initial vibration signal shallow layer characteristics and the corresponding bearing health state;
constructing a deep stacking self-coding network by adding a classifier in a basic stacking self-coding network and adding a sparse penalty item in a neuron, adding an additional penalty item in a corresponding loss function, training the deep stacking self-coding neural network under the sparse penalty constraint condition based on the feature label data set, and performing hyper-parameter optimization on the deep stacking self-coding neural network through a particle swarm optimization algorithm to obtain a final deep stacking self-coding network;
acquiring a bearing vibration signal to be identified, and performing data cleaning and integrity supplement on the bearing vibration signal to obtain a preprocessed vibration signal;
performing feature extraction on the preprocessed vibration signal through a time domain-frequency domain statistical analysis method to obtain shallow features of the vibration signal to be recognized;
and carrying out unsupervised learning on the shallow feature of the vibration signal to be identified through the final depth stacking self-coding network to obtain the health state of the bearing.
Preferably, when the bearing vibration signal is subjected to data cleaning, invalid and abnormal data are removed;
and for the bearing vibration signal with the missing data value, performing integrity supplement by a linear interpolation method.
Preferably, the deep stacking self-coding network comprises an input layer, a hidden layer, an identification layer and an output layer, wherein a sparse penalty term is added to neurons of the hidden layer and used for performing feature extraction on shallow features of the vibration signal to obtain deep features of the vibration signal, and the identification layer is configured with a classifier and used for classifying, identifying and outputting the health state of the deep features of the vibration signal.
Preferably, the KL divergence is selected as a sparse penalty term, and the calculation formula of the KL divergence is as follows:
wherein s is the total number of hidden layer neurons, and the average activation amount of hidden layer node j is rhojρ, where ρ is a sparsity parameter;
calculating a loss function of the depth stacking self-coding network according to the mean square error;
training the deep stacking self-coding neural network under sparse penalty constraint conditions, and adding L to prevent the deep stacking self-coding neural network from being over-fitted when carrying out hyper-parameter optimization on the deep stacking self-coding neural network through a particle swarm optimization algorithm2The regularization term constitutes a structural risk function, and the mathematical expression of the objective function is as follows:
wherein, λ represents regular term attenuation coefficient, and after adding KL divergence as sparse penalty term, the objective function is represented as:
wherein, beta is a sparse penalty term coefficient, and W and b respectively represent neuron connection weight and bias;
and obtaining sparse penalty term expression by calculating average activation amount, and solving an objective function by an optimization problem, thereby obtaining sparse expression of the shallow feature of the vibration signal.
Preferably, by LsparseSolving two sparse parameters W and b in an objective function by the optimization problem of (W and b), and carrying out L by a BP algorithmsparse(W, b) optimization problem;
meanwhile, through a batch training method, a gradient descent method is adopted to update the weight value in each iteration, and the iteration steps are as follows:
where ξ is the learning rate.
Preferably, the classifier is a soft-max classifier;
the hyper-parameters include dropout ratios of the input and hidden layers, sparse parameters, and the number of neurons in each hidden layer.
Preferably, the feature label data set is divided into two data sets which are respectively a training set and a testing set, the deep stacking self-coding neural network is trained under the sparse penalty constraint condition based on the training set, and the super-parameter optimization is carried out on the deep stacking self-coding neural network through a particle swarm optimization algorithm, so that a final deep stacking self-coding network is obtained;
selecting the shallow layer characteristic of the initial vibration signal in the test set as input, performing unsupervised learning through the trained deep stacking self-coding network, taking the output bearing health state as a health state evaluation value, and performing comparative analysis on the health state evaluation value and the health state of the bearing in the test set to evaluate the final deep stacking self-coding network.
In a second aspect, the system for identifying the health status of a bearing based on a stacked self-coding neural network of the present invention identifies the health status of a bearing by the method for identifying the health status of a bearing based on a stacked self-coding neural network according to any one of the first aspect, and the system includes:
the data acquisition module is used for acquiring bearing vibration signals of various damages and corresponding bearing health states and acquiring bearing vibration signals to be identified;
the data preprocessing module is used for carrying out data cleaning and integrity supplement on the bearing vibration signals of various damages to obtain initial historical data; the vibration signal processing device is used for carrying out data cleaning and integrity supplement on the bearing vibration signal to be identified to obtain a vibration signal after pretreatment;
the shallow feature extraction module is used for extracting features of the initial historical data through a time domain-frequency domain statistical analysis method to obtain initial vibration signal shallow features; the vibration signal preprocessing module is used for extracting the characteristics of the preprocessed vibration signal through a time domain-frequency domain statistical analysis method to obtain the shallow layer characteristics of the vibration signal to be recognized;
the network construction training module is used for constructing a deep stacking self-coding network by adding a classifier in a basic stacking self-coding network and adding a sparse penalty item in a neuron, adding an additional penalty item in a corresponding loss function, training the deep stacking self-coding neural network under a sparse penalty constraint condition based on the feature label data set, and performing hyper-parameter optimization on the deep stacking self-coding neural network through a particle swarm optimization algorithm to obtain a final deep stacking self-coding network;
and the classification identification module is used for carrying out unsupervised learning on the shallow feature of the vibration signal to be identified through the final depth stacking self-coding network to obtain the health state of the bearing.
Preferably, the deep stacking self-coding network comprises an input layer, a hidden layer, an identification layer and an output layer, wherein a sparse penalty term is added to neurons of the hidden layer and used for performing feature extraction on shallow features of the vibration signal to obtain deep features of the vibration signal, and the identification layer is provided with a classifier and used for classifying, identifying and outputting a health state on the deep features of the vibration signal;
the sparse penalty term is KL divergence, and the calculation formula of the KL divergence is as follows:
where s is the total number of hidden layer neurons; the nature of KL divergence determines the final sparse penalty term when rhojWhen ρ is, the KL dispersion value is 0, if ρjGradually deviating from rho, the divergence value of KL is gradually increased;
the average activation amount of the hidden layer node j is rhojRho is a sparsity parameter which represents the average expected activity of neurons in the hidden layer, the theoretical value is 0-1, and the smaller the value is, the smaller the number of activated neurons in the hidden layer is;
calculating a loss function of the depth stacking self-coding network according to the mean square error;
training the deep stacking self-coding neural network under sparse penalty constraint conditions, and adding L to prevent the deep stacking self-coding neural network from being over-fitted when carrying out hyper-parameter optimization on the deep stacking self-coding neural network through a particle swarm optimization algorithm2The regularization term constitutes a structural risk function, and the mathematical expression of the objective function is as follows:
wherein, λ represents regular term attenuation coefficient, and after adding KL divergence as sparse penalty term, the objective function is represented as:
wherein, beta is a sparse penalty term coefficient, and W and b respectively represent neuron connection weight and bias;
obtaining sparse penalty term expression by calculating average activation amount and obtaining sparse penalty term expression by LsparseSolving two sparse parameters W and b in an objective function by the optimization problem of (W and b), and carrying out L by a BP algorithmsparse(W, b) optimization problem;
meanwhile, through a batch training method, a gradient descent method is adopted to update the weight value in each iteration, and the iteration steps are as follows:
where ξ is the learning rate;
the classifier is a soft-max classifier;
the network construction training module is used for dividing the feature label data set into two data sets which are respectively a training set and a testing set, training the deep stacking self-coding neural network under the sparse penalty constraint condition based on the training set, and performing hyper-parameter optimization on the deep stacking self-coding neural network through a particle swarm optimization algorithm to obtain a final deep stacking self-coding network;
selecting the shallow layer characteristic of the initial vibration signal in the test set as input, performing unsupervised learning through the trained deep stacking self-coding network, taking the output bearing health state as a health state evaluation value, and performing comparative analysis on the health state evaluation value and the health state of the bearing in the test set to evaluate the final deep stacking self-coding network.
Preferably, the preprocessing module is used for removing invalid and abnormal data when the data of the bearing vibration signal is cleaned;
and for the bearing vibration signal with the missing data value, performing integrity supplement by a linear interpolation method.
The bearing health state identification method and system based on the stacked self-coding neural network have the following advantages:
the method comprises the steps that data are mined by using a signal time domain-frequency domain statistical analysis method and a deep learning method based on an engine bearing vibration signal, and the health state of the engine bearing is accurately identified;
firstly, preprocessing data of a bearing signal of an engine, supplementing missing data by using an interpolation method, extracting shallow features of the bearing signal based on time domain and frequency domain statistical analysis, and ensuring the accuracy of a final recognition result;
thirdly, the method constructs a depth stacking self-encoder by using a basic self-encoder, converts the extracted shallow features into depth features which are used as the final input of a classifier, and ensures the accuracy of the final recognition result;
adding a sparse penalty term in a basic self-encoder, learning sparse data characteristics under the sparse penalty constraint condition, and avoiding output and complete copy input, so that useful characteristics are learned in high-dimensional data;
and (V) optimizing the dropout proportion, the sparse parameters and the number of neurons of each hidden layer of the input layer and the hidden layer by using a particle swarm optimization algorithm, and optimizing the topological structure of the depth network and the complexity of the model to the maximum extent, so that the calculation efficiency of the model is improved.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed for the embodiments or the prior art descriptions will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on the drawings without creative efforts.
The invention is further described below with reference to the accompanying drawings.
Fig. 1 is a block flow diagram of a bearing health status identification method based on a stacked self-coding neural network according to embodiment 1.
Detailed Description
The present invention is further described in the following with reference to the drawings and the specific embodiments so that those skilled in the art can better understand the present invention and can implement the present invention, but the embodiments are not to be construed as limiting the present invention, and the embodiments and the technical features of the embodiments can be combined with each other without conflict.
The embodiment of the invention provides a bearing health state identification method and system based on a stacked self-coding neural network, which are used for solving the technical problem of accurately and timely identifying the health state of an engine bearing.
Example 1:
the invention relates to a bearing health state identification method based on a stacked self-coding neural network, which comprises the following steps of:
s100, obtaining vibration signals of various damaged bearings and corresponding bearing health states, and performing data cleaning and integrity supplement on the vibration signals of the bearings to obtain initial historical data;
s200, extracting characteristics of initial historical data through a time domain-frequency domain statistical analysis method to obtain initial vibration signal shallow layer characteristics, and constructing a characteristic label data set based on the initial vibration signal shallow layer characteristics and the corresponding bearing health state;
s300, constructing a deep stacking self-coding network by adding a classifier in a basic stacking self-coding network and adding a sparse penalty item in a neuron, adding an additional penalty item in a corresponding loss function, training the deep stacking self-coding neural network under a sparse penalty constraint condition based on a feature tag data set, and performing hyper-parameter optimization on the deep stacking self-coding neural network through a particle swarm optimization algorithm to obtain a final deep stacking self-coding network;
s400, obtaining a bearing vibration signal to be identified, and performing data cleaning and integrity supplement on the bearing vibration signal to obtain a preprocessed vibration signal;
s500, extracting the characteristics of the preprocessed vibration signal through a time domain-frequency domain statistical analysis method to obtain the shallow layer characteristics of the vibration signal to be recognized;
s600, carrying out unsupervised learning on the shallow feature of the vibration signal to be recognized through the final depth stacking self-coding network to obtain the health state of the bearing.
In the embodiment, historical vibration signals of various damages are firstly acquired, and the health states of corresponding bearings are marked so as to train a subsequently constructed deep stacking self-coding neural network. The depth stacking self-coding neural network is used for carrying out depth feature extraction on the extracted shallow features to obtain the depth features of the vibration signals, and identifying and classifying the depth features of the vibration signals to obtain the health state.
In step S100 and step S400, for each type of obtained lost bearing vibration signal and the bearing vibration signal to be identified obtained in real time, the integrity of the initial vibration signal data is analyzed, invalid and abnormal data therein are removed, and the missing data value is supplemented by a linear interpolation method.
In step S200 and step S500, for the bearing vibration signal after data cleaning and integrity supplement, a preliminary signal feature extraction is performed by using a time domain statistical analysis method and a frequency domain statistical analysis method, and the extracted features constitute a shallow feature data set.
And step S300 and step S400 are matched to construct and train an optimized deep stacking self-coding network.
A deep stacked self-coding network, also known as a stacked self-encoder, is composed by stacking a plurality of elementary self-encoders. The bearing vibration signal shallow layer characteristics extracted by adopting a time domain and frequency domain statistical analysis method are pre-trained by means of unsupervised learning characteristics of a stacked self-encoder. And extracting deep-level features with high abstract expression capability from the shallow-level features. In this embodiment, the deep stacking self-coding network is improved on the basis of the basic stacking self-coding network, and includes an input layer, a hidden layer, an identification layer, and an output layer after improvement, the input layer and the output layer are the original basic stacking self-coding network, a sparse penalty item is added to a neuron of the hidden layer, and is used to perform feature extraction on a shallow feature of a vibration signal to obtain a deep feature of the vibration signal, and a classifier is configured in the added identification layer and is used to classify, identify, and output a healthy state on the deep feature of the vibration signal. The classifier can be selected from a multi-classification support vector machine, a classification decision tree, a k neighbor classification method, a soft-max classifier and the like, and the soft-max classifier is selected in the embodiment.
The basic self-encoder is an under-complete self-encoder, and the sparse penalty term is added to the neurons of the basic under-complete self-encoder, so that the sparse self-encoder is obtained, the sparse encoder can avoid complete copy input, and useful characteristics are learned from original high-dimensional data. Suppose a given set of signal data sets x1,x2,...,xi,...,xn}，(xi∈Rn) The specific method for adding sparse constraint is to limit the number of hidden layer neuron activations and control the average activation amount of hidden layer nodes j to be rho in the whole signal data setjRho is a sparsity parameter which represents the average expected activity of neurons in the hidden layer, the theoretical value is 0-1, and the smaller the value is, the smaller the number of activated neurons in the hidden layer is. In practical applications, to ensure the learning effect of the network, ρ is usually close to 0, for example, ρ is 0.05, which means that the average value of the activated outputs of hidden nodes is 0.05. To achieve such sparse effect, an extra penalty term is usually added to the loss function of the self-encoder to constrain ρjWithout deviating from the parameter p. In the invention, KL three-degree is selected as a sparse penalty term, and the calculation formula of the KL divergence is as follows:
where s is the total number of hidden layer neurons. The nature of the KL divergence determines the final penalty, apparently when ρjWhere ρ is, the KL dispersion value is 0, e.g.Fruit rhojGradually deviating from ρ, the divergence value of KL gradually increases. The loss function of the self-encoder is calculated according to the mean square error, and L is added to prevent the overfitting of the network2The regularization term constitutes a structural risk function, and the mathematical expression of the objective function is as follows:
where λ represents the regular term attenuation coefficient. After adding KL divergence as a sparse penalty term, the objective function is expressed as:
wherein, beta is a sparse penalty term coefficient, and W and b respectively represent neuron connection weight and bias. As known from the loss function expression, the key for sparse self-coding is W and b of the autumn sparse optimization, which can be converted into LsparseSolving the two parameters by the optimal value problem of (W, b), wherein the optimization problem can be realized by a BP algorithm, a batch training method is adopted, a gradient descent method is adopted for updating the weight value in each iteration, and the iteration steps are as follows:
where ξ is the learning rate.
According to the steps, sparse penalty expression is obtained by calculating average activation amount, and then sparse cost function is solved through an optimization problem, so that sparse expression of shallow input features is obtained.
Feature extraction for stacked self-encoders generally involves two stages: pre-training and fine-tuning. In the pre-training stage, the dropout proportion of the input layer and the hidden layer has important influence on the network training time and the reconstruction error; in the fine tuning stage, the sparse parameter value determines the average activation value of all nodes of the hidden layer and the updating speed of the network parameter; in addition, there is the number of neurons in each hidden layer that is the most critical network parameter. In the invention, the particle swarm optimization algorithm is used for optimizing the hyper-parameters of the 3 types of deep neural networks, and the purpose is to optimize the topological structure of the network and the complexity of the model to the maximum extent and improve the calculation efficiency of the model.
After the shallow features are converted into the deep features through the deep stacking self-encoder, the bearing health state recognition is carried out by using the features, the classification method used in the embodiment is soft-max, a soft-max classifier is adopted to train on a deep feature data set to obtain a bearing health state recognition model, then the model is used for recognizing a test sample, the mapping relation between a signal sample and the bearing damage state is obtained, and the rolling bearing damage health state recognition is completed.
Example 2:
the invention relates to a bearing health state recognition system based on a stacked self-coding neural network, which comprises a data acquisition module, a data preprocessing module, a shallow feature extraction module, a network construction training module and a classification recognition module, wherein the data acquisition module is used for acquiring bearing vibration signals of various damages and corresponding bearing health states and acquiring bearing vibration signals to be recognized; the data preprocessing module is used for carrying out data cleaning and integrity supplement on the bearing vibration signals with various damages to obtain initial historical data; the vibration signal processing device is used for carrying out data cleaning and integrity supplement on the bearing vibration signal to be identified to obtain a vibration signal after pretreatment; the shallow feature extraction module is used for extracting features of the initial historical data through a time domain-frequency domain statistical analysis method to obtain initial vibration signal shallow features; the method comprises the steps of preprocessing vibration signals, performing feature extraction on the preprocessed vibration signals through a time domain-frequency domain statistical analysis method to obtain shallow features of the vibration signals to be recognized; the network construction training module is used for constructing a deep stacking self-coding network by adding a classifier in a basic stacking self-coding network and adding a sparse penalty item in a neuron, adding an additional penalty item in a corresponding loss function, training the deep stacking self-coding network under a sparse penalty constraint condition based on the feature tag data set, and performing hyper-parameter optimization on the deep stacking self-coding network through a particle swarm optimization algorithm to obtain a final deep stacking self-coding network; and the classification identification module is used for carrying out unsupervised learning on the shallow feature of the vibration signal to be identified through the final depth stacking self-coding network to obtain the health state of the bearing.
In this embodiment, the preprocessing module analyzes the integrity of the initial vibration signal data for the obtained various lost bearing vibration signals and the bearing vibration signals to be identified obtained in real time, removes invalid and abnormal data therein, and supplements the missing data values by using a linear interpolation method.
And the shallow feature extraction module is used for carrying out primary signal feature extraction on the bearing vibration signals subjected to data cleaning and integrity supplement by using a time domain statistical analysis method and a frequency domain statistical analysis method, and forming a shallow feature data set by the extracted features.
The network construction training module constructs and trains an optimized deep stacked self-coding network, which is also called a stacked self-encoder and is formed by stacking a plurality of basic self-encoders. The bearing vibration signal shallow layer characteristics extracted by adopting a time domain and frequency domain statistical analysis method are pre-trained by means of unsupervised learning characteristics of a stacked self-encoder. And extracting deep-level features with high abstract expression capability from the shallow-level features. In this embodiment, the deep stacking self-coding network is improved on the basis of the basic stacking self-coding network, and includes an input layer, a hidden layer, an identification layer, and an output layer after improvement, the input layer and the output layer are the original basic stacking self-coding network, a sparse penalty item is added to a neuron of the hidden layer, and is used to perform feature extraction on a shallow feature of a vibration signal to obtain a deep feature of the vibration signal, and a classifier is configured in the added identification layer and is used to classify, identify, and output a healthy state on the deep feature of the vibration signal. The classifier can be selected from a multi-classification support vector machine, a classification decision tree, a k neighbor classification method, a soft-max classifier and the like, and the soft-max classifier is selected in the embodiment.
The basic self-encoder is an under-complete self-encoder, and the sparse penalty term is added to the neurons of the basic under-complete self-encoder, so that the sparse self-encoder is obtained, the sparse encoder can avoid complete copy input, and useful characteristics are learned from original high-dimensional data. Suppose a given set of signal data sets x1,x2,...,xi,...,xn}，(xi∈Rn) The specific method for adding sparse constraint is to limit the number of hidden layer neuron activations and control the average activation amount of hidden layer nodes j to be rho in the whole signal data setjRho is a sparsity parameter which represents the average expected activity of neurons in the hidden layer, the theoretical value is 0-1, and the smaller the value is, the smaller the number of activated neurons in the hidden layer is. In practical applications, to ensure the learning effect of the network, ρ is usually close to 0, for example, ρ is 0.05, which means that the average value of the activated outputs of hidden nodes is 0.05. To achieve such sparse effect, an extra penalty term is usually added to the loss function of the self-encoder to constrain ρjWithout deviating from the parameter p. In the invention, KL three-degree is selected as a sparse penalty term, and the calculation formula of the KL divergence is as follows:
where s is the total number of hidden layer neurons. The nature of the KL divergence determines the final penalty, apparently when ρjWhen ρ is, the KL dispersion value is 0, if ρjGradually deviating from ρ, the divergence value of KL gradually increases. The loss function of the self-encoder is calculated according to the mean square error, and L is added to prevent the overfitting of the network2The regularization term constitutes a structural risk function, and the mathematical expression of the objective function is as follows:
where λ represents the regular term attenuation coefficient. After adding KL divergence as a sparse penalty term, the objective function is expressed as:
wherein, beta is a sparse penalty term coefficient, and W and b respectively represent neuron connection weight and bias. As known from the loss function expression, the key for sparse self-coding is W and b of the autumn sparse optimization, which can be converted into LsparseSolving the two parameters by the optimal value problem of (W, b), wherein the optimization problem can be realized by a BP algorithm, a batch training method is adopted, a gradient descent method is adopted for updating the weight value in each iteration, and the iteration steps are as follows:
where ξ is the learning rate.
According to the steps, sparse penalty expression is obtained by calculating average activation amount, and then sparse cost function is solved through an optimization problem, so that sparse expression of shallow input features is obtained.
Feature extraction for stacked self-encoders generally involves two stages: pre-training and fine-tuning. In the pre-training stage, the dropout proportion of the input layer and the hidden layer has important influence on the network training time and the reconstruction error; in the fine tuning stage, the sparse parameter value determines the average activation value of all nodes of the hidden layer and the updating speed of the network parameter; in addition, there is the number of neurons in each hidden layer that is the most critical network parameter. In the embodiment, the particle swarm optimization algorithm is used for optimizing the hyper-parameters of the 3 types of deep neural networks, so that the topological structure of the network and the complexity of the model are optimized to the maximum extent, and the calculation efficiency of the model is improved.
The system of the embodiment can perform the health status identification on the engine bearing to be identified by executing the bearing health status identification method based on the stacked self-coding neural network disclosed in embodiment 1.
While the invention has been shown and described in detail in the drawings and in the preferred embodiments, it is not intended to limit the invention to the embodiments disclosed, and it will be apparent to those skilled in the art that many more embodiments of the invention are possible that combine the features of the different embodiments described above and still fall within the scope of the invention.
Claims (10)
1. The bearing health state identification method based on the stacked self-coding neural network is characterized by comprising the following steps of:
acquiring bearing vibration signals of various damages and corresponding bearing health states, and performing data cleaning and integrity supplement on the bearing vibration signals to obtain initial historical data;
extracting the characteristics of the initial historical data through a time domain-frequency domain statistical analysis method to obtain initial vibration signal shallow layer characteristics, and constructing a characteristic label data set based on the initial vibration signal shallow layer characteristics and the corresponding bearing health state;
constructing a deep stacking self-coding network by adding a classifier in a basic stacking self-coding network and adding a sparse penalty item in a neuron, adding an additional penalty item in a corresponding loss function, training the deep stacking self-coding neural network under the sparse penalty constraint condition based on the feature label data set, and performing hyper-parameter optimization on the deep stacking self-coding neural network through a particle swarm optimization algorithm to obtain a final deep stacking self-coding network;
acquiring a bearing vibration signal to be identified, and performing data cleaning and integrity supplement on the bearing vibration signal to obtain a preprocessed vibration signal;
performing feature extraction on the preprocessed vibration signal through a time domain-frequency domain statistical analysis method to obtain shallow features of the vibration signal to be recognized;
and carrying out unsupervised learning on the shallow feature of the vibration signal to be identified through the final depth stacking self-coding network to obtain the health state of the bearing.
2. The method for identifying the health state of the bearing based on the stacked self-coding neural network as claimed in claim 1, wherein invalid and abnormal data are removed when data cleaning is carried out on the vibration signal of the bearing;
and for the bearing vibration signal with the missing data value, performing integrity supplement by a linear interpolation method.
3. The method for identifying the health status of the bearing based on the stacked self-coding neural network as claimed in claim 1, wherein the deep stacked self-coding neural network comprises an input layer, an implied layer, an identification layer and an output layer, wherein a sparse penalty term is added to neurons of the implied layer for performing feature extraction on shallow features of the vibration signal to obtain deep features of the vibration signal, and the identification layer is configured with a classifier for classifying the deep features of the vibration signal to identify the health status of the bearing.
4. The method for identifying the health state of the bearing based on the stacked self-coding neural network according to claim 3, wherein KL divergence is selected as a sparse penalty term, and the calculation formula of the KL divergence is as follows:
wherein s is the total number of hidden layer neurons, and the average activation amount of hidden layer node j is rhojρ, where ρ is a sparsity parameter;
calculating a loss function of the depth stacking self-coding network according to the mean square error;
training the deep-stacked self-coding neural network under sparse penalty constraint condition and passing through particle swarmWhen the optimization algorithm carries out hyper-parameter optimization on the depth stacking self-coding neural network, in order to prevent the depth stacking self-coding neural network from being over-fitted, L is added2The regularization term constitutes a structural risk function, and the mathematical expression of the objective function is as follows:
wherein, λ represents regular term attenuation coefficient, and after adding KL divergence as sparse penalty term, the objective function is represented as:
wherein, beta is a sparse penalty term coefficient, and W and b respectively represent neuron connection weight and bias;
and obtaining sparse penalty term expression by calculating average activation amount, and solving an objective function by an optimization problem, thereby obtaining sparse expression of the shallow feature of the vibration signal.
5. The method of claim 4, wherein the L is a function of the state of health of the bearingsparseSolving two sparse parameters W and b in an objective function by the optimization problem of (W and b), and carrying out L by a BP algorithmsparse(W, b) optimization problem;
meanwhile, through a batch training method, a gradient descent method is adopted to update the weight value in each iteration, and the iteration steps are as follows:
where ξ is the learning rate.
6. The method of claim 1, wherein the classifier is a soft-max classifier;
the hyper-parameters include dropout ratios of the input and hidden layers, sparse parameters, and the number of neurons in each hidden layer.
7. The method for identifying the health state of the bearing based on the stacked self-coding neural network according to any one of claims 1 to 6, wherein the feature label data set is divided into two data sets, namely a training set and a testing set, the deep stacked self-coding neural network is trained under the condition of sparse penalty constraint based on the training set, and the final deep stacked self-coding network is obtained by performing hyper-parametric optimization on the deep stacked self-coding neural network through a particle swarm optimization algorithm;
selecting the shallow layer characteristic of the initial vibration signal in the test set as input, performing unsupervised learning through the trained deep stacking self-coding network, taking the output bearing health state as a health state evaluation value, and performing comparative analysis on the health state evaluation value and the health state of the bearing in the test set to evaluate the final deep stacking self-coding network.
8. The system for identifying the health state of a bearing based on a stacked self-coding neural network is characterized in that the health state of the bearing is identified by the method for identifying the health state of the bearing based on the stacked self-coding neural network according to any one of claims 1 to 7, and the system comprises:
the data acquisition module is used for acquiring bearing vibration signals of various damages and corresponding bearing health states and acquiring bearing vibration signals to be identified;
the data preprocessing module is used for carrying out data cleaning and integrity supplement on the bearing vibration signals of various damages to obtain initial historical data; the vibration signal processing device is used for carrying out data cleaning and integrity supplement on the bearing vibration signal to be identified to obtain a vibration signal after pretreatment;
the shallow feature extraction module is used for extracting features of the initial historical data through a time domain-frequency domain statistical analysis method to obtain initial vibration signal shallow features; the vibration signal preprocessing module is used for extracting the characteristics of the preprocessed vibration signal through a time domain-frequency domain statistical analysis method to obtain the shallow layer characteristics of the vibration signal to be recognized;
the network construction training module is used for constructing a deep stacking self-coding network by adding a classifier in a basic stacking self-coding network and adding a sparse penalty item in a neuron, adding an additional penalty item in a corresponding loss function, training the deep stacking self-coding neural network under a sparse penalty constraint condition based on the feature label data set, and performing hyper-parameter optimization on the deep stacking self-coding neural network through a particle swarm optimization algorithm to obtain a final deep stacking self-coding network;
and the classification identification module is used for carrying out unsupervised learning on the shallow feature of the vibration signal to be identified through the final depth stacking self-coding network to obtain the health state of the bearing.
9. The system according to claim 8, wherein the deep stacked self-coding network comprises an input layer, an implied layer, a recognition layer and an output layer, the implied layer neuron is added with a sparse penalty term and is used for performing feature extraction on the shallow features of the vibration signal to obtain the deep features of the vibration signal, and the recognition layer is configured with a classifier and is used for classifying the deep features of the vibration signal to recognize the output health status;
the sparse penalty term is KL divergence, and the calculation formula of the KL divergence is as follows:
where s is the total number of hidden layer neurons; the nature of KL divergence determines the final sparse penalty term when rhojWhen ρ is, the KL dispersion value is 0, if ρjGradually deviating from rho, the divergence value of KL is gradually increased;
the average activation amount of the hidden layer node j is rhojRho is a sparsity parameter which represents the average expected activity of neurons in the hidden layer, the theoretical value is 0-1, and the smaller the value is, the smaller the number of activated neurons in the hidden layer is;
calculating a loss function of the depth stacking self-coding network according to the mean square error;
training the deep stacking self-coding neural network under sparse penalty constraint conditions, and adding L to prevent the deep stacking self-coding neural network from being over-fitted when carrying out hyper-parameter optimization on the deep stacking self-coding neural network through a particle swarm optimization algorithm2The regularization term constitutes a structural risk function, and the mathematical expression of the objective function is as follows:
wherein, λ represents regular term attenuation coefficient, and after adding KL divergence as sparse penalty term, the objective function is represented as:
wherein, beta is a sparse penalty term coefficient, and W and b respectively represent neuron connection weight and bias;
obtaining sparse penalty term expression by calculating average activation amount and obtaining sparse penalty term expression by LsparseSolving two sparse parameters W and b in an objective function by the optimization problem of (W and b), and carrying out L by a BP algorithmsparse(W, b) optimization problem;
meanwhile, through a batch training method, a gradient descent method is adopted to update the weight value in each iteration, and the iteration steps are as follows:
where ξ is the learning rate;
the classifier is a soft-max classifier;
the network construction training module is used for dividing the feature label data set into two data sets which are respectively a training set and a testing set, training the deep stacking self-coding neural network under the sparse penalty constraint condition based on the training set, and performing hyper-parameter optimization on the deep stacking self-coding neural network through a particle swarm optimization algorithm to obtain a final deep stacking self-coding network;
selecting the shallow layer characteristic of the initial vibration signal in the test set as input, performing unsupervised learning through the trained deep stacking self-coding network, taking the output bearing health state as a health state evaluation value, and performing comparative analysis on the health state evaluation value and the health state of the bearing in the test set to evaluate the final deep stacking self-coding network.
10. The system according to claim 8 or 9, wherein the preprocessing module is configured to remove invalid and abnormal data when performing data cleaning on the bearing vibration signal;
and for the bearing vibration signal with the missing data value, performing integrity supplement by a linear interpolation method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111579017.3A CN114298096A (en) | 2021-12-22 | 2021-12-22 | Bearing health state identification method and system based on stacked self-coding neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111579017.3A CN114298096A (en) | 2021-12-22 | 2021-12-22 | Bearing health state identification method and system based on stacked self-coding neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114298096A true CN114298096A (en) | 2022-04-08 |
Family
ID=80969934
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111579017.3A Pending CN114298096A (en) | 2021-12-22 | 2021-12-22 | Bearing health state identification method and system based on stacked self-coding neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114298096A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114819108A (en) * | 2022-06-22 | 2022-07-29 | 中国电力科学研究院有限公司 | Fault identification method and device for comprehensive energy system |
-
2021
- 2021-12-22 CN CN202111579017.3A patent/CN114298096A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114819108A (en) * | 2022-06-22 | 2022-07-29 | 中国电力科学研究院有限公司 | Fault identification method and device for comprehensive energy system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109918752B (en) | Mechanical fault diagnosis method, equipment and medium based on migration convolution neural network | |
| Saufi et al. | Gearbox fault diagnosis using a deep learning model with limited data sample | |
| CN112417954B (en) | Bearing fault mode diagnosis method and system for small sample data set | |
| Saravanan et al. | Incipient gear box fault diagnosis using discrete wavelet transform (DWT) for feature extraction and classification using artificial neural network (ANN) | |
| Guo et al. | Multifeatures fusion and nonlinear dimension reduction for intelligent bearing condition monitoring | |
| Saravanan et al. | Fault diagnosis of spur bevel gear box using artificial neural network (ANN), and proximal support vector machine (PSVM) | |
| Pacheco et al. | A statistical comparison of neuroclassifiers and feature selection methods for gearbox fault diagnosis under realistic conditions | |
| Li et al. | Gear pitting fault diagnosis using raw acoustic emission signal based on deep learning | |
| CN113505655B (en) | Intelligent bearing fault diagnosis method for digital twin system | |
| Gecgel et al. | Gearbox fault diagnostics using deep learning with simulated data | |
| CN114778112A (en) | Audio identification and fault diagnosis method for mechanical fault of wind turbine generator system | |
| CN114358124B (en) | New fault diagnosis method for rotary machinery based on deep countermeasure convolutional neural network | |
| Wang et al. | A sequence-to-sequence model with attention and monotonicity loss for tool wear monitoring and prediction | |
| de Paula Monteiro et al. | A hybrid prototype selection-based deep learning approach for anomaly detection in industrial machines | |
| CN110737976A (en) | mechanical equipment health assessment method based on multi-dimensional information fusion | |
| CN115290326A (en) | Rolling bearing fault intelligent diagnosis method | |
| CN116842364A (en) | Rolling bearing fault method for asynchronous motor | |
| Wang et al. | A factor analysis based transfer learning method for gearbox diagnosis under various operating conditions | |
| CN114298096A (en) | Bearing health state identification method and system based on stacked self-coding neural network | |
| CN115712817A (en) | Fault diagnosis method of industrial motor pump based on convolutional neural network | |
| Fu et al. | An improved deep convolutional neural network with multiscale convolution kernels for fault diagnosis of rolling bearing | |
| BENKEDJOUH et al. | Deep Learning for Fault Diagnosis based on short-time Fourier transform | |
| Yu et al. | Fault feature of gearbox vibration signals based on morphological filter dynamic convolution autoencoder | |
| Natarajan | Vibration signal analysis using histogram features and support vector machine for gear box fault diagnosis | |
| Bhavaraju et al. | A comparative study on bearings faults classification by artificial neural networks and self-organizing maps using wavelets |
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 |