CN112257695A - Method for generating confrontation network to generate vibration signal by using sparse constraint - Google Patents
Method for generating confrontation network to generate vibration signal by using sparse constraint Download PDFInfo
- Publication number
- CN112257695A CN112257695A CN202011523721.2A CN202011523721A CN112257695A CN 112257695 A CN112257695 A CN 112257695A CN 202011523721 A CN202011523721 A CN 202011523721A CN 112257695 A CN112257695 A CN 112257695A
- Authority
- CN
- China
- Prior art keywords
- sparse
- generator
- countermeasure network
- discriminator
- vibration signal
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 54
- 239000013598 vector Substances 0.000 claims abstract description 81
- 230000004913 activation Effects 0.000 claims abstract description 40
- 239000011159 matrix material Substances 0.000 claims abstract description 22
- 230000003213 activating effect Effects 0.000 claims abstract description 4
- 238000012549 training Methods 0.000 claims description 94
- 230000006870 function Effects 0.000 claims description 52
- 210000002569 neuron Anatomy 0.000 claims description 33
- 238000007781 pre-processing Methods 0.000 claims description 7
- 238000013459 approach Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 3
- 230000008569 process Effects 0.000 abstract description 11
- 238000012886 linear function Methods 0.000 abstract 1
- 238000001228 spectrum Methods 0.000 description 12
- 238000009826 distribution Methods 0.000 description 9
- 230000036541 health Effects 0.000 description 9
- 238000013528 artificial neural network Methods 0.000 description 7
- 238000010276 construction Methods 0.000 description 7
- 238000003745 diagnosis Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 7
- 108010076504 Protein Sorting Signals Proteins 0.000 description 6
- 238000001514 detection method Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 6
- 238000013135 deep learning Methods 0.000 description 5
- 238000005070 sampling Methods 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 230000003042 antagnostic effect Effects 0.000 description 3
- 230000007246 mechanism Effects 0.000 description 3
- 238000010606 normalization Methods 0.000 description 3
- 230000002411 adverse Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000007274 generation of a signal involved in cell-cell signaling Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000005096 rolling process Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013434 data augmentation Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000002955 isolation Methods 0.000 description 1
- 230000010355 oscillation Effects 0.000 description 1
- 239000000047 product Substances 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 238000009827 uniform distribution Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
-
- 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
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2136—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on sparsity criteria, e.g. with an overcomplete basis
-
- 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
- G06N3/088—Non-supervised learning, e.g. competitive learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Software Systems (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)
- Signal Processing (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses a method for generating a vibration signal by using a sparse constraint generation countermeasure network, which comprises the following steps: the generator of the trained sparse constraint generation countermeasure network transforms an input random noise vector into a sparse activation vector containing only a few non-zero values; activating the trained sparse constraint to generate column vectors of corresponding positions in a weight matrix of a hidden layer of a generator of the countermeasure network by using a small number of non-zero values contained in the sparse vectors, wherein the activated column vectors contain all key frequency components and combinations thereof in a real vibration signal; weighting the activated column vector by a non-zero activation value of a sparse vector to obtain a linear combination, wherein the linear combination comprises all key frequency components in the original vibration signal of the rotary machine to be generated; the linear combination is activated by a non-linear function, which is equivalent to an amplitude modulation process, to obtain the final generated sample.
Description
Technical Field
The invention relates to the technical field of rotary machine vibration signal generation, in particular to a sparse constraint generation countermeasure network implementation method of a rotary machine vibration signal.
Background
The rotary machine is an important component in industrial equipment and plays a key role in the working and running of the equipment, so that the running condition of the rotary machine can greatly influence the whole running condition of the equipment, and once the rotary machine breaks down, the whole fault of the equipment is easily caused, and the adverse effects of equipment halt, economic loss, personal safety damage and the like are caused. However, the rotary machine usually runs under severe environmental conditions such as high load, variable working conditions and the like, and is easy to degrade and lose efficacy, so that the rotary machine can carry out health management work such as fault detection, fault diagnosis, health assessment and the like, can effectively master the running state of the rotary machine, and can carry out alarm and isolation positioning when a fault occurs, thereby improving the running reliability of equipment. Since the operation of the rotary machine has obvious periodicity, the vibration signal caused by the method contains a large amount of information highly related to the operation condition and the fault condition of the rotary machine, and therefore methods such as fault detection and diagnosis for the rotary machine mostly depend on the acquired vibration signal. The traditional signal analysis and feature extraction method can be combined with the working and fault mechanisms of the rotary machine to carry out detection and diagnosis on the rotary machine, and a better effect is achieved. In recent years, with the development of deep learning technology, extensive attention and research have been drawn to end-to-end detection and diagnosis of rotary machine vibration signals based on a deep learning method. However, deep learning approaches typically rely on a large number of labeled data samples. In an actual scene, due to the high cost of data acquisition and the high risk of faulty operation of equipment, obtaining a large number of marked vibration signal samples is difficult to realize, and the implementation effect of the deep learning method is influenced by the problem. Therefore, the artificial data generation is carried out by using a data augmentation method, and a large number of synthetic samples are generated based on limited real samples to supplement a training set, so that the method becomes an effective means for solving the problem of data shortage.
The method is characterized in that a countermeasure network is generated and used as an unsupervised generation model, distribution can be learned from real data effectively, random noise vectors are converted into samples highly similar to the real data, and a related method is applied to the aspect of generating vibration signals of the rotary machine. The existing method mainly focuses on generation of vibration signal frequency spectrum and vibration signal characteristics by using a generation countermeasure network. The method comprises the steps of generating a vibration signal frequency spectrum, firstly, performing fast Fourier transform on an original rotary machine vibration signal to obtain a vibration signal frequency spectrum, using the vibration signal frequency spectrum as a real sample, training a generation countermeasure network to obtain a large number of generated frequency spectrum samples, and further serving health management work such as subsequent detection, diagnosis and health assessment of the rotary machine. And generating vibration signal characteristics, namely firstly extracting the characteristics of the vibration signals, including time domain characteristics, frequency domain characteristics and the like, forming characteristic vectors by using the extracted characteristics, and then training an impedance network by using the characteristic vectors as real samples to enable the network to obtain the capability of generating the vibration signal characteristics. However, neither spectral nor signature generation retains all the information of the original vibration signal. The existing method lacks the capability of stably generating the original time domain vibration signal of the rotating machine, and complex network structures are needed, so that the instability of training is further caused. Therefore, it is necessary to design a method capable of learning the distribution of the original time-domain vibration signal of the rotating machine without supervision by using the generation countermeasure network, and further generating a large number of vibration signal samples, so as to support the development and application of the detection and diagnosis method with high demand on data volume.
Disclosure of Invention
The invention aims to provide a method for generating a vibration signal by using a sparse constraint generation countermeasure network, which directly carries out unsupervised learning and mass generation on an original time domain vibration signal of a rotary machine.
The method for generating the vibration signal of the countermeasure network by using the sparse constraint comprises the following steps:
generating a countermeasure network by constructing and training a sparse constraint generation countermeasure network comprising a generator and a discriminator to obtain a trained sparse constraint generation countermeasure network, and inputting a random noise vector to the generator of the trained sparse constraint generation countermeasure network;
the generator of the trained sparse constraint generation countermeasure network transforms an input random noise vector into a sparse activation vector containing only a few non-zero values;
activating the trained sparse constraint to generate column vectors of corresponding positions in a weight matrix of a hidden layer of a generator of the countermeasure network by using a small number of non-zero values contained in the sparse vectors, wherein the activated column vectors contain all key frequency components and combinations thereof in a real vibration signal;
weighting the activated column vector by a non-zero activation value of a sparse vector to obtain a linear combination, wherein the linear combination comprises all key frequency components in the original vibration signal of the rotary machine to be generated;
the linear combination is activated by a nonlinear function, namely, the linear combination is subjected to amplitude modulation processing to obtain a final generated sample;
and the generator of the trained sparse constraint generation countermeasure network outputs the generated sample to obtain a generated sample which is highly similar to the original vibration signal.
Preferably, constructing and training a sparse constraint generating countermeasure network comprising a generator and a discriminator comprises:
constructing an input layer dimension and an output layer dimension ofwDimension of the hidden layer ofmThe sparse autoencoder of (1), wherein saidwAndmare all positive integers;
the dimension obtained by preprocessing the acquired vibration signal iswTraining the constructed sparse automatic encoder by using the vibration signal training sample to obtain a trained sparse automatic encoder;
constructing a sparse constraint generation countermeasure network comprising a generator and a discriminator by utilizing a trained sparse automatic encoder;
using dimensions ofwThe constructed sparse constraint generation countermeasure network is trained by the vibration signal training sample and the noise sample, and the sparse constraint generation countermeasure network capable of generating the rotary mechanical vibration signal by using the noise is obtained.
Preferably, constructing a sparse constraint generation countermeasure network comprising a generator and a discriminator using a trained sparse autoencoder comprises:
splitting the trained sparse automatic encoder, taking an input layer and a hidden layer as an encoder part, and taking the hidden layer and an output layer as a decoder part;
a classifier of a sparse constraint generation countermeasure network is obtained by accessing an output layer containing a neuron after a coder part obtained by splitting a trained sparse automatic coder;
the sparse constraint generation countermeasure network generator is obtained by accessing an input layer with dimension w (the number of neurons of the input layer is w) before splitting a decoder part obtained by a trained sparse automatic encoder.
Preferably, the utilization dimension iswTraining the constructed sparse constraint generation countermeasure network by using the vibration signal training sample and the noise sample comprises the following steps:
carrying out a plurality of times of iterative cycle training on the discriminator of the sparse constraint generation countermeasure network by using the vibration signal training sample and the noise sample to obtain a trained discriminator;
performing iterative loop training for a generator for generating a countermeasure network with sparse constraint by using the noise samples for a plurality of times;
during a plurality of times of iterative loop training of the generator, the trained discriminator discriminates the generated sample output by the generator and the signal training sample until the generated sample output by the generator approaches the signal training sample.
Preferably, performing a number of iterative loop trainings on the discriminator of the sparse constraint generation countermeasure network comprises:
calculating a loss value of the discriminator during a plurality of times of iterative cycle training of the discriminator of the sparse constraint generation countermeasure network by using the signal training sample and the noise sample;
and according to the calculated loss value of the discriminator, carrying out gradient descent updating on the network parameter of the discriminator until the loss value of the discriminator and the loss value of the generator reach a Nash equilibrium state.
Preferably, the parameters of the discriminator network include a weight matrix and a bias vector of the trained sparse automatic encoder and a weight matrix and a bias vector of the accessed output layer including one neuron.
Preferably, the discriminator hidden layer and the output layer activation function are both Sigmoid functions.
Preferably, performing several iterative loop trainings on the generator for generating the sparse constraint antagonistic network comprises:
calculating the generator loss value during several iterative loop trainings of the generator of the sparse constraint generation countermeasure network with the signal training samples and noise samples;
and according to the calculated generator loss value, performing gradient descent updating on the generator network parameter until the generator loss value and the discriminator loss value reach a Nash equilibrium state.
Preferably, the generator network parameters include a weight matrix and a bias vector of the trained sparse automatic encoder and a weight matrix and a bias vector of the input layer with dimension w.
Preferably, the hidden layer activation function of the generator is a Sigmoid function, and the output layer activation function is a tanh function.
Preferably, the dimension obtained after preprocessing the acquired vibration signal iswThe vibration signal training samples of (a) include:
normalizing the amplitude of the vibration signal sequence to be between-1 and 1;
the amplitude normalized vibration signal sequence is sliced into n vibration training signal samples of length w using a window of width w.
The beneficial technical effects of the invention comprise:
1. compared with the existing method capable of realizing the generation of the vibration signal frequency spectrum and the vibration signal characteristics, the method can directly perform unsupervised learning and mass generation on the original time domain vibration signals of the rotating machinery, and overcomes the inevitable information loss problem in the frequency spectrum generation and the characteristic generation;
2. compared with the existing method which can realize the generation of the time domain vibration signal of the rotary machine to a certain extent, the method can stably generate the vibration signal samples under different conditions without designing a complex neural network structure, carefully and balancedly generating a confrontation network training process and introducing a large amount of training skills, thereby effectively reducing the difficulty and instability of the generation;
3. the method can effectively expand the vibration signal sample set, reduce the adverse effect of methods with high requirements on data quantity, such as deep learning, caused by insufficient data quantity, and improve the model performance on tasks, such as fault diagnosis.
The present invention will be described in detail below with reference to specific embodiments thereof, which are illustrated in the accompanying drawings.
Drawings
FIG. 1 is a schematic diagram of a sparse constraint generation countermeasure network implementation of a rotating machine vibration signal of the present invention;
FIG. 2 is a schematic diagram of the sparse autoencoder of the present invention for decomposing and recomposing sparse constraints to generate a countermeasure network;
FIG. 3 is a schematic illustration of a bearing raw vibration signal in a normal state;
FIG. 4 is a schematic diagram of a vibration signal after preprocessing a bearing raw vibration signal in a normal state;
FIG. 5 is a schematic diagram of the sparse autoencoder training loss variation of the present invention;
FIG. 6 is a schematic diagram of sparse constraint generation versus network training loss variation of the present invention;
FIG. 7 is a schematic diagram showing the comparison of the normal state generated vibration signal and the normal state real vibration signal of the trained sparse constraint generation countermeasure network of the present invention and the frequency spectrum thereof;
FIG. 8 is a diagram showing the frequencies corresponding to corresponding neurons in the hidden layer of the trained sparse constraint generation countermeasure network generator of the present invention.
Detailed Description
FIG. 1 shows a sparse constraint generation countermeasure network implementation method of a rotary mechanical vibration signal, which comprises the following steps:
constructing an input layer dimension and an output layer dimension ofwDimension of the hidden layer ofmThe sparse autoencoder of (1), wherein saidwAndmare all positive integers;
the dimension obtained by preprocessing the acquired vibration signal iswTraining the constructed sparse automatic encoder by using the vibration signal training sample to obtain a trained sparse automatic encoder;
constructing a sparse constraint generation countermeasure network comprising a generator and a discriminator by utilizing a trained sparse automatic encoder;
using dimensions ofwThe constructed sparse constraint generation countermeasure network is trained by the vibration signal training sample and the noise sample, and the sparse constraint generation countermeasure network capable of generating the rotary mechanical vibration signal by using the noise is obtained.
Specifically, constructing a sparse constraint generation countermeasure network including a generator and a discriminator by using a trained sparse automatic encoder is shown in fig. 2, and includes:
splitting the trained sparse automatic encoder, taking an input layer and a hidden layer as an encoder part, and taking the hidden layer and an output layer as a decoder part, namely, obtaining the encoder part and the decoder part of the trained sparse automatic encoder by acquiring and/or copying codes corresponding to the input layer, the output layer and the hidden layer of the trained sparse automatic encoder;
a classifier of a sparse constraint generation countermeasure network is obtained by accessing an output layer containing a neuron after a coder part obtained by splitting a trained sparse automatic coder;
obtained by splitting a trained sparse automatic encoderBefore the decoder part of (2) an input layer of dimension w (the number of neurons in this input layer isw) And obtaining a generator for generating the countermeasure network by sparse constraint.
Specifically, using dimensions ofwTraining the constructed sparse constraint generation countermeasure network by using the vibration signal training sample and the noise sample comprises the following steps:
carrying out a plurality of times of iterative cycle training on the discriminator of the sparse constraint generation countermeasure network by using the vibration signal training sample and the noise sample to obtain a trained discriminator;
performing iterative loop training for a generator for generating a countermeasure network with sparse constraint by using the noise samples for a plurality of times;
during a plurality of times of iterative loop training of the generator, the trained discriminator discriminates the generated sample output by the generator and the signal training sample until the generated sample output by the generator approaches the signal training sample.
Specifically, performing a number of iterative loop trainings on the discriminator of the sparse constraint generation countermeasure network comprises:
calculating a loss value of the discriminator during a plurality of times of iterative cycle training of the discriminator of the sparse constraint generation countermeasure network by using the signal training sample and the noise sample;
according to the calculated arbiter loss value, gradient descent update is performed on the arbiter network parameter until the arbiter loss value and the generator loss value reach a nash equilibrium state, as shown in fig. 6.
The network parameters of the discriminator comprise a weight matrix and a bias vector of a trained sparse automatic encoder and a weight matrix and a bias vector of an accessed output layer containing a neuron, and activation functions of a hidden layer and an output layer of the discriminator are Sigmoid functions.
Specifically, performing a number of iterative loop trainings on the generator for generating the sparse constraint antagonistic network comprises:
calculating the generator loss value during several iterative loop trainings of the generator of the sparse constraint generation countermeasure network with the signal training samples and noise samples;
and according to the calculated generator loss value, performing gradient descent updating on the generator network parameter until the generator loss value and the discriminator loss value reach a Nash equilibrium state.
The generator network parameters comprise a weight matrix and a bias vector of a trained sparse automatic encoder and a weight matrix and a bias vector of an input layer with the dimension w, the hidden layer activation function of the generator is a Sigmoid function, and the output layer activation function of the generator is a tanh function.
The above-described method embodiment of the present invention comprises the steps of:
the method comprises the following steps: rotary machine vibration signal data preprocessing
The sensor is used for collecting vibration signals under the conditions that the rotary machine is in a certain working condition, a certain load and a certain health state. The vibration signal sequence obtained by acquisition is set asWhereinsRepresenting the total number of sample points. Firstly, the amplitude of the vibration signal sequence is normalized to be between-1 and 1, and the normalization formula is as follows
After normalization is complete, use the width ofwWindow of (2) vibrating the signal sequenceVIs divided intoHas a length ofwThe sample of (1), whereini]Represents no more thaniIs the largest integer of (a). Is provided withAfter pretreatment, the product is obtainednHas a length ofwTraining sample set of vibration signals of samples。
Step two: sparse autoencoder construction
Constructing an input layer dimension and an output layer dimension ofwDimension of the hidden layer ofmThe sparse autoencoder of (1).
The automatic encoder is a three-layer neural network comprising an input layer, a hidden layer and an output layer. For input vectorxThe purpose of the neural network is to learn an identity map, i.e.
WhereinIs a reconstructed sample of the output of the network,andrepresenting the weight matrix and the bias vector of the encoder and decoder, respectively.
The encoder part of the network converts the input samples into output vectors of the hidden layer, i.e.
Is the output vector of the hidden layer.Is a model parameter of the encoder, includingAnd。non-linear activation functions, usually Sigmoid functions, expressed as。
The decoder converts the hidden layer output vector into an output vector, i.e.
With root Mean Square Error (MSE) between input samples and output vector as a loss function for the automatic encoder, i.e.
The sparse automatic encoder introduces sparse constraint in a hidden layer of the automatic encoder, limits the activation condition of neurons of the hidden layer, and can more effectively encode input samples and extract features. Is set for the input sampleOf 1 atjThe activation value of hidden layer neuron isFor a batchnAn input sample, the average activation value of the neuron is
Sparse autoencoder expects the average activation value of hidden layer neurons to be kept at a low level,I.e. sparse parameters, a positive real number close to 0 is often taken. The degree of deviation between the actual average activation and the sparse parameter is measured using the Kullback-Leibler (KL) divergence
WhereinIs a hidden layermThe average activation value of each neuron constitutes an activation vector. When the actual average activation condition is higher, the KL divergence value is larger, so that the KL divergence is introduced into a loss function of the sparse automatic encoder and optimized together with the reconstruction error loss MSE (mean Square error), namely
For the loss of the sparse automatic encoder obtained by combining the MSE loss of the reconstruction error and the KL divergence loss of the sparse,is a hyper-parameter for controlling the strength of the sparse constraint penalty item.
Step three: sparse autoencoder training
And after the sparse automatic encoder is built, training the sparse automatic encoder by using the rotating mechanical vibration signal training sample set built in the step one. For one iteration (epoch), the training process is as follows. Let each pair of network parameters be updated once, and the number of samples used in a batch is batch _ size.
Step 302: let i = 1.
Step 303: when i + batch _ size-1 does not exceed the total number of samples n, the following steps are performed, otherwise, a jump is made to 308.
Step 306: updating parameters of sparse autoencoders using gradient descent algorithms, i.e.
Step 307: let i = i + batch _ size, jump back to step 303.
Step 308: the training of this epoch on the constructed sparse auto-encoder is completed.
And training the constructed sparse automatic encoder in a plurality of iterative cycles according to the preset epoch total number until the loss function is not reduced any more, and finishing the training process of the sparse automatic encoder.
Step four: sparse constraint generation countermeasure network construction
The generation countermeasure network is a neural network which is composed of a generator and a discriminator and is symmetrical to each other. The generator takes a random noise vector as input and outputs a generated sample, and the purpose is to enable the generated sample to be similar to a real sample as much as possible so as to confuse the discriminator to give an erroneous discrimination result; the discriminator takes the generated samples and the real samples as input, and outputs discrimination results, aiming at accurately discriminating whether the input samples come from the real data distribution or belong to the generated samples of the generator.
The loss function of the generator is as follows
WhereinIs a random noise vector that is a function of,is a priori noise distribution, and generally takes a uniform distribution or a gaussian distribution.GAndDrespectively representing the generator and the arbiter.
The loss function of the discriminator is as follows
In the training process, the generator and the discriminator update parameters in turn until a Nash equilibrium state is reached. The training objective function of the whole generation countermeasure network is as follows
As shown in fig. 2, the sparse automatic encoder obtained by training in step three is split into two parts: an encoder section and a decoder section. A network layer containing 1 neuron is accessed after the encoder part, the activation function of the network layer is a sigmoid function, and the network layer is used as a discriminator for generating a countermeasure network by sparse constraint; and a network layer with the neuron number being the same as the dimension of the vibration signal sample is accessed as an input layer before the decoder, and a generator for generating the countermeasure network by sparse constraint is formed. Therefore, the generator and the arbiter of the sparse generation countermeasure network are both a three-layer network structure of an input layer-a hidden layer-an output layer.
Due to the existence of the sparse constraint in step two and step three, the hidden layers of the generator and the arbiter still contain the sparse constraint at this time, so the method is called sparse constraint generation countermeasure network.
Step five: sparse constraint generation confrontation network training
And after the sparse constraint generation countermeasure network is constructed, training the sparse constraint generation countermeasure network by using the rotating machinery vibration signal training sample set obtained in the step one.
Due to the existence of sparse constraints in the hidden layers of the generator and the discriminator, compared with the basic generation countermeasure network, the sparse constraint generation countermeasure network objective function in the invention is different. Wherein the objective function of the generator is as follows
Wherein,an activation vector consisting of the average activation values of all neurons of the second layer (i.e. the hidden layer) of the generator,is a hyper-parameter for controlling the strength of the sparse constraint penalty term in the generator.
Similarly, the objective function of the discriminator is as follows
Wherein,an activation vector consisting of the average activation values of all neurons of the second layer (i.e. the hidden layer) of the arbiter,is a hyper-parameter for controlling the strength of the sparse constraint penalty term in the discriminator.
For one iteration (epoch), the training process is as follows. Setting each pair of network parameters to be updated once, and using a batch of samples with the quantity ofpAnd (4) respectively.
Step 501: order toq=0。
Step 502: when in useqWhen the number of times of training of the discriminator is less than the set number of times of training of the discriminator, the following steps are sequentially executed; otherwise, jump to step 506.
Step 503: in advance ofIn examining noise distributions, random samplingpA random noise vector; randomly sampling in the vibration signal sample set of the real rotating machinerypA true vibration signal sample.
Step 504: calculating the loss value of the discriminator in the formula (3) according to the noise sample and the real sampleAnd gradient descending updating is carried out on the network parameters of the discriminator:。
step 505: order toq=q+1, jump back to step 502.
Step 506: in the prior noise distribution, random samplingpA random noise vector.
Step 507: calculating the loss value of the generator as formula (2) according to the noise sampleAnd performing gradient descent updating on the generator network parameters:。
step 508: the training of the epoch to sparse constraint generation countermeasure network is completed.
And training the constructed sparse constraint generation countermeasure network for a plurality of iterative cycles according to the preset epoch total number until the generator and the discriminator reach balance, wherein the generator can generate a vivid vibration signal sample, and the training process of generating the sparse constraint generation countermeasure network is completed.
Principle for generating rotary mechanical vibration signal by sparse constraint generation countermeasure network
After the confrontation training stage in the step five is finished, the three-layer neural network serving as the generator can generate and output a vivid time domain vibration signal by taking the random noise vector as input. For a given noise vectorAfter entering the model, it is first converted into a sparse vector, as follows
In the formula,andis the weight matrix and bias vector for the first layer of the generator network.
Then, the sparse vector is obtainedIs converted into a generated vibration signal and is output by the generator as follows
In the formula,andis the weight matrix and bias vector of the second layer of the generator network. Then, the weight matrix is processedExpressed in the form of column vectors, sparse vectorsExpressed in terms of each element it contains, the resulting sample of the output can be expressed as
Since the activation function of the network is a monotonically increasing function, it affects only the amplitude of the output signal, not the frequency of the output signal. For rotating machine PHM methods based on vibration signals, frequency domain information is often the key information. Thus, the generated vibration signal output by the generator can be regarded as a weight matrix column vectorPlus the offset vector. The weight of the linear combination of the column vectors is the element value contained in the sparse vector. Due to the presence of the sparse constraint during training, there are a large number of values of 0 or values close to 0 in the sparse vector, so this linear combination can be approximated as a linear combination of a small number of column vectors, corresponding to the locations in the sparse vector activated by the input random noise vector. Therefore, by analyzing the condition of the column vector of the weight matrix of the hidden layer of the generator, the mechanism that the generator can generate the vivid time-domain vibration signal can be explained. The frequency components contained in the final output signal are a superposition of the frequency components contained in the weight matrix column vectors selected by the sparse vector activation elements.
DETAILED DESCRIPTION OF EMBODIMENT (S) OF INVENTION
In the embodiment of the invention, the validity of the method is verified by adopting a public data set provided by Kaiser Wechu University (Case Western Reserve University, CWRU).
The data set contains the ball bearing vibration signals collected by the accelerometer. The test bed for testing and collecting signals consists of a driving motor, a torque sensor/encoder, a dynamometer and a control circuit, and the accelerometer for collecting signals is connected with equipment in a magnetic attraction mode.
The load level is 1-hp and the sampling frequency of the vibration signal is 48 kHz. The data set collectively contains normal (N), inner ring failure (IR), rolling element failure (B), outer ring failure (OR), wherein the inner ring failure, rolling element failure and outer ring failure modes each contain three different failure sizes of 0.007, 0.014 and 0.021 inches. Thus, the data set contains a total of 10 different health states.
The specific methods for executing and generating the vibration signals under different health states are the same, so that the vibration signals under the normal state are taken as an example, and the implementation effects of the first step to the fifth step are shown; and in the sixth step, the generated results of all 10 health states are displayed.
Rotary machine vibration signal data preprocessing
Under normal conditions, the bearing raw vibration signal contains 10 seconds of sampled data, totaling 480000 points, as shown in FIG. 3.
First, the vibration signal sequence amplitude is normalized to between-1 and 1. After normalization is complete, the vibration signal sequence is sliced into 1500 samples of length 320 using a window of width 320. The partially normalized and sliced sample is shown in fig. 4.
Sparse autoencoder construction
A sparse autoencoder with input and output layer dimensions of 320 and hidden layer dimension of 160 is constructed. The hidden layer activation function uses a Sigmoid function, and the output layer activation function uses a tanh function. Hyper-parameters for controlling sparse constraint penalty term dynamics. The results of the construction are shown in the following table.
Sparse autoencoder training
Each pair of network parameters is updated once, and the number of the used batch of samples is 64. The number of epochs was 1000 trains. The training process loss variation is shown in fig. 5.
Sparse constraint generation countermeasure network construction and training
Splitting a sparse automatic encoder obtained by training into two parts: an encoder section and a decoder section. In the encoder partThen accessing a network layer containing 1 neuron, wherein the activation function is a sigmoid function, and the sigmoid function is used as a sparse constraint to generate a discriminator of the countermeasure network; a network layer with the neuron number being the same as the dimension of the vibration signal sample (320 dimensions) is accessed as an input layer before a decoder, and a generator for generating a countermeasure network by sparse constraint is formed. Therefore, the generator is a three-layer structure neural network of 320-160-320; the discriminator is a three-layer structure neural network of 320-160-1. The hidden layer activation function of the generator is a Sigmoid function, the output layer activation function is a tanh function, and the hidden layer sparse punishment hyperparameter(ii) a The activation function and the output layer function of the hidden layer of the discriminator are Sigmoid functions, and the sparse punishment of the hidden layer is hyperparametric. The random noise distribution is uniformly distributed with the value between-1 and 1.
The generator construction results are shown in the following table.
The discriminator construction results are shown in the following table.
Each pair of network parameters is updated once, and the number of the used batch of samples is 50. The number of epochs of challenge training was 2000. The training process loss variation is shown in fig. 6.
As can be seen from the loss variation trend of the generator and the discriminator, the antagonistic training process is relatively stable, and the situations of loss large-amplitude oscillation or loss divergence do not occur.
Principle for generating mechanical vibration signal by sparse constraint generation countermeasure network
First, the actual vibration signal of the bearing in the normal state, the frequency spectrum of the actual vibration signal, the generated vibration signal, and the frequency spectrum of the generated vibration signal are observed, as shown in fig. 7.
The generation result shows that the sparse constraint generation countermeasure network established by the invention can realize stable generation for the vibration signal under the bearing health state; in the frequency domain, the main frequency and the energy of the frequency spectrum of the generated signal and the real signal are basically consistent, and the model is proved to learn key frequency domain information from the time domain vibration signal. It can also be seen that for the healthy state, the vibration signal contains three major frequency components (the lowest critical frequency component, the middle critical frequency component, and the highest critical frequency component).
According to the weight analysis method, extracting a weight matrix of a hidden layer of the generator network and acquiring 160 column vectors in total; and respectively drawing an original curve of the column vector and a frequency spectrum curve after FFT. The column vectors represented by eight representative neurons were selected for presentation analysis, as shown in fig. 8. As can be seen, different neurons of the generator network learn signal features that contain key frequency information, and differ from each other. For example, a portion of neurons learn mainly a single key frequency component, such as neuron #26, neuron #18 (corresponding to the lowest key frequency component), neuron #135 (corresponding to the central key frequency component), neuron #114 (corresponding to the highest key frequency component); some neurons learn a combination of two key frequency components, such as neuron #72, neuron #155 (corresponding to the combination of the central and highest key frequency components); some of the neurons learn all the key frequency components, such as neuron #81 and neuron # 148. Therefore, the generation mechanism of the rotating mechanical vibration signal generation model for generating the countermeasure network based on the sparse constraint can be interpreted as the following process: (1) firstly, a random noise vector input into a generator network is transformed into a sparse activation vector containing only a few non-zero values; (2) activating column vectors at corresponding positions in a weight matrix of a hidden layer of a generator by using a small number of non-zero values contained in the sparse vector, wherein the activated column vectors contain all key frequency components and combinations thereof in a real vibration signal; (3) the method comprises the following steps that an activated column vector is weighted by a non-zero activation value of a sparse vector to obtain a linear combination, wherein the linear combination contains all key frequency components in original vibration signals of the rotary machine to be generated; (4) the linear combination is activated by a nonlinear function, namely, the final generated sample is obtained through amplitude modulation processing and is output by the generator, and the generated sample which is highly similar to the original vibration signal is obtained.
Although the present invention has been described in detail hereinabove, the present invention is not limited thereto, and various modifications can be made by those skilled in the art in light of the principle of the present invention. Thus, modifications made in accordance with the principles of the present invention should be understood to fall within the scope of the present invention.
Claims (10)
1. A method of generating a countermeasure network generated vibration signal with sparse constraints, comprising:
generating a countermeasure network by constructing and training a sparse constraint generation countermeasure network comprising a generator and a discriminator to obtain a trained sparse constraint generation countermeasure network, and inputting a random noise vector to the generator of the trained sparse constraint generation countermeasure network;
the generator of the trained sparse constraint generation countermeasure network transforms an input random noise vector into a sparse activation vector containing only a few non-zero values;
activating the trained sparse constraint to generate column vectors of corresponding positions in a weight matrix of a hidden layer of a generator of the countermeasure network by using a small number of non-zero values contained in the sparse vectors, wherein the activated column vectors contain all key frequency components and combinations thereof in a real vibration signal;
weighting the activated column vector by a non-zero activation value of a sparse vector to obtain a linear combination, wherein the linear combination comprises all key frequency components in the original vibration signal of the rotary machine to be generated;
the linear combination is activated by a nonlinear function, and amplitude modulation processing is carried out to obtain a final generated sample;
and the generator of the trained sparse constraint generation countermeasure network outputs the generated sample to obtain a generated sample which is highly similar to the original vibration signal.
2. The method of claim 1, wherein constructing and training a sparse constraint generating countermeasure network comprising a generator and a discriminator comprises:
constructing an input layer dimension and an output layer dimension ofwDimension of the hidden layer ofmThe sparse autoencoder of (1);
the dimension obtained by preprocessing the acquired vibration signal iswTraining the constructed sparse automatic encoder by using the vibration signal training sample to obtain a trained sparse automatic encoder;
constructing a sparse constraint generation countermeasure network comprising a generator and a discriminator by utilizing a trained sparse automatic encoder;
using dimensions ofwThe vibration signal training sample and the noise sample train the constructed sparse constraint generation countermeasure network to obtain the trained sparse constraint generation countermeasure network.
3. The method of claim 1 or 2, wherein constructing a sparse constraint generating countermeasure network comprising a generator and a discriminator using a trained sparse autoencoder comprises:
splitting the trained sparse automatic encoder, taking an input layer and a hidden layer as an encoder part, and taking the hidden layer and an output layer as a decoder part;
a classifier of a sparse constraint generation countermeasure network is obtained by accessing an output layer containing a neuron after a coder part obtained by splitting a trained sparse automatic coder;
by accessing the dimensionality before splitting the decoder part obtained by a trained sparse automatic encoderwThe sparse constraint is obtained to generate a generator of the countermeasure network.
4. The method of claim 3, wherein the utilization dimension iswTraining the constructed sparse constraint generation countermeasure network by using the vibration signal training sample and the noise sample comprises the following steps:
carrying out a plurality of times of iterative cycle training on the discriminator of the sparse constraint generation countermeasure network by using the vibration signal training sample and the noise sample to obtain a trained discriminator;
performing iterative loop training for a generator for generating a countermeasure network with sparse constraint by using the noise samples for a plurality of times;
during a plurality of times of iterative loop training of the generator, the trained discriminator discriminates the generated sample output by the generator and the signal training sample until the generated sample output by the generator approaches the signal training sample.
5. The method of claim 3, wherein iteratively training a discriminator of the sparse constraint generating countermeasure network a number of times comprises:
calculating a loss value of the discriminator during a plurality of times of iterative cycle training of the discriminator of the sparse constraint generation countermeasure network by using the signal training sample and the noise sample;
and according to the calculated loss value of the discriminator, carrying out gradient descent updating on the network parameter of the discriminator until the loss value of the discriminator and the loss value of the generator reach a Nash equilibrium state.
6. The method of claim 5, wherein the discriminator network parameters include weight matrices and bias vectors of the trained sparse auto-encoder and weight matrices and bias vectors of the accessed output layer containing one neuron.
7. The method of claim 6, wherein the discriminator hidden layer and output layer activation functions are both Sigmoid functions.
8. The method of claim 3, wherein iteratively training a generator of the sparse constraint generating countermeasure network a number of times comprises:
calculating the generator loss value during several iterative loop trainings of the generator of the sparse constraint generation countermeasure network with the signal training samples and noise samples;
and according to the calculated generator loss value, performing gradient descent updating on the generator network parameter until the generator loss value and the discriminator loss value reach a Nash equilibrium state.
9. The method of claim 8, wherein the generator network parameters include a weight matrix and a bias vector for the trained sparse auto-encoder and a weight matrix and a bias vector for an input layer of dimension w.
10. The method of claim 8, wherein the generator hidden layer activation function is a Sigmoid function and the output layer activation function is a tanh function.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011523721.2A CN112257695B (en) | 2020-12-22 | 2020-12-22 | Method for generating confrontation network to generate vibration signal by using sparse constraint |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011523721.2A CN112257695B (en) | 2020-12-22 | 2020-12-22 | Method for generating confrontation network to generate vibration signal by using sparse constraint |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112257695A true CN112257695A (en) | 2021-01-22 |
CN112257695B CN112257695B (en) | 2021-04-02 |
Family
ID=74225815
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011523721.2A Active CN112257695B (en) | 2020-12-22 | 2020-12-22 | Method for generating confrontation network to generate vibration signal by using sparse constraint |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112257695B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2022188650A1 (en) * | 2021-03-09 | 2022-09-15 | 华为技术有限公司 | Encoding method, decoding method, encoding device, and decoding device |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109635774A (en) * | 2018-12-21 | 2019-04-16 | 中山大学 | A kind of human face synthesizing method based on generation confrontation network |
US20190370969A1 (en) * | 2018-05-30 | 2019-12-05 | Siemens Healthcare Gmbh | Methods for generating synthetic training data and for training deep learning algorithms for tumor lesion characterization, method and system for tumor lesion characterization, computer program and electronically readable storage medium |
CN110674836A (en) * | 2019-08-06 | 2020-01-10 | 厦门大学 | Sparse countermeasure sample generation method based on generation network |
CN111711820A (en) * | 2020-05-13 | 2020-09-25 | 哈尔滨工业大学 | Crack image compression sampling method based on generation countermeasure network |
-
2020
- 2020-12-22 CN CN202011523721.2A patent/CN112257695B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190370969A1 (en) * | 2018-05-30 | 2019-12-05 | Siemens Healthcare Gmbh | Methods for generating synthetic training data and for training deep learning algorithms for tumor lesion characterization, method and system for tumor lesion characterization, computer program and electronically readable storage medium |
CN109635774A (en) * | 2018-12-21 | 2019-04-16 | 中山大学 | A kind of human face synthesizing method based on generation confrontation network |
CN110674836A (en) * | 2019-08-06 | 2020-01-10 | 厦门大学 | Sparse countermeasure sample generation method based on generation network |
CN111711820A (en) * | 2020-05-13 | 2020-09-25 | 哈尔滨工业大学 | Crack image compression sampling method based on generation countermeasure network |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2022188650A1 (en) * | 2021-03-09 | 2022-09-15 | 华为技术有限公司 | Encoding method, decoding method, encoding device, and decoding device |
Also Published As
Publication number | Publication date |
---|---|
CN112257695B (en) | 2021-04-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112257694B (en) | Sparse constraint generation countermeasure network implementation method for rotary mechanical vibration signals | |
Tao et al. | Bearing fault diagnosis method based on stacked autoencoder and softmax regression | |
Wu et al. | Induction machine fault detection using SOM-based RBF neural networks | |
CN105758644A (en) | Rolling bearing fault diagnosis method based on variation mode decomposition and permutation entropy | |
CN113935460A (en) | Intelligent diagnosis method for mechanical fault under class imbalance data set | |
CN110674892A (en) | Fault feature screening method based on weighted multi-feature fusion and SVM classification | |
Huang et al. | Application of wavelet neural networks on vibration fault diagnosis for wind turbine gearbox | |
CN106197999A (en) | A kind of planetary gear method for diagnosing faults | |
Bhadane et al. | Bearing fault identification and classification with convolutional neural network | |
Jiang et al. | A fault diagnostic method for induction motors based on feature incremental broad learning and singular value decomposition | |
CN110782041B (en) | Structural modal parameter identification method based on machine learning | |
Huang et al. | A Fault Diagnosis Approach for Rolling Bearing Based on Wavelet Packet Decomposition and GMM-HMM. | |
CN113256443B (en) | Nuclear power water pump guide bearing fault detection method, system, equipment and readable storage medium | |
Deng et al. | Rolling bearing fault diagnosis based on Deep Boltzmann machines | |
CN112257695B (en) | Method for generating confrontation network to generate vibration signal by using sparse constraint | |
CN115587290A (en) | Aero-engine fault diagnosis method based on variational self-coding generation countermeasure network | |
Cocconcelli et al. | Diagnostics of ball bearings in varying-speed motors by means of Artificial Neural Networks | |
CN113324758A (en) | Bearing fault diagnosis method based on semi-supervised countermeasure network | |
Rajeswari et al. | Bearing fault diagnosis using multiclass support vector machine with efficient feature selection methods | |
Sun et al. | Application of deep belief networks for precision mechanism quality inspection | |
Xiao et al. | Adaptive feature extraction based on Stacked Denoising Auto-encoders for asynchronous motor fault diagnosis | |
CN114136604B (en) | Rotary equipment fault diagnosis method and system based on improved sparse dictionary | |
CN116484258A (en) | Elevator traction machine bearing fault diagnosis method | |
Ren et al. | Fault diagnosis with imbalanced data based on auto-encoder | |
CN116028795A (en) | Wind turbine generator main bearing fault diagnosis method and related device |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |