CN114676733A - Fault diagnosis method for complex supply and delivery mechanism based on sparse self-coding assisted classification generation type countermeasure network - Google Patents

Fault diagnosis method for complex supply and delivery mechanism based on sparse self-coding assisted classification generation type countermeasure network Download PDF

Info

Publication number
CN114676733A
CN114676733A CN202210354689.2A CN202210354689A CN114676733A CN 114676733 A CN114676733 A CN 114676733A CN 202210354689 A CN202210354689 A CN 202210354689A CN 114676733 A CN114676733 A CN 114676733A
Authority
CN
China
Prior art keywords
sparse
discriminator
network
loss
image
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210354689.2A
Other languages
Chinese (zh)
Inventor
闫啸家
梁伟阁
张钢
佘博
王旋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Naval University of Engineering PLA
Original Assignee
Naval University of Engineering PLA
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Naval University of Engineering PLA filed Critical Naval University of Engineering PLA
Publication of CN114676733A publication Critical patent/CN114676733A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G06F2218/04—Denoising
    • G06F2218/06—Denoising by applying a scale-space analysis, e.g. using wavelet analysis
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01H—MEASUREMENT OF MECHANICAL VIBRATIONS OR ULTRASONIC, SONIC OR INFRASONIC WAVES
    • G01H17/00—Measuring mechanical vibrations or ultrasonic, sonic or infrasonic waves, not provided for in the preceding groups
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Signal Processing (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a fault diagnosis method for a complex supply and delivery mechanism based on a sparse self-coding assisted classification generation type countermeasure network, which applies an ACGANs framework to the field of fault diagnosis, and a trained discriminator can directly identify the truth of an input time-frequency characteristic diagram and the type of a fault state, and does not need to add a model for training and identifying after mixing generated data and real data. From the aspect of increasing image feature constraint on hidden variables, original data are coded by pre-training an SE, noise is added to coding vectors through standard normal distribution to jointly form hidden variables containing real image feature information, the hidden variables and the categories are jointly input into a generator to strengthen the capability of the hidden variables for representing the features related to the categories to which the images belong, the range of the generator for learning the feature space of the real samples is reduced, the image category features can be effectively maintained in the stages of model training and recognition, and therefore the performance of the discriminator is further improved.

Description

Fault diagnosis method for complex supply and delivery mechanism based on sparse self-coding assisted classification generation type countermeasure network
Technical Field
The invention belongs to the technical field of fault diagnosis, and particularly relates to a fault diagnosis method for a complex supply and delivery mechanism based on a sparse self-coding auxiliary classification generation type countermeasure network.
Background
The complex feeding mechanism is an important component of special equipment, the reliability and stability of the complex feeding mechanism directly influence the exertion of the functions of the special equipment, for example, the artillery feeding mechanism directly influences the motion state of ammunition, the motion process is accompanied by violent impact, vibration and the like, and the mechanical environment is very complex. Therefore, the development of fault diagnosis of the complex supply and delivery mechanism has important significance in the aspects of improving the working efficiency and reliability, reducing the maintenance time and cost and the like.
The complex mechanical environment and the unique mechanical structure enable the vibration signal of the supply and transmission mechanism to have the characteristics of nonlinearity, non-stationarity and the like. The time-frequency analysis method can reflect the energy intensity change of the signal in two dimensions of time and frequency, effectively describes the fine fault characteristics of the signal, and is one of important methods for analyzing and processing non-stationary signals. Conventional Short-time Fourier transform (STFT)) The one-dimensional fault vibration signal can be transformed into a two-dimensional matrix, so that a characteristic diagram containing time domain and frequency domain information is obtained. However, the size and shape of the STFT time window are relatively fixed, and cannot satisfy the requirements of both time resolution and frequency resolution. The Continuous Wavelet Transform (CWT) with continuous scale transform can overcome the defect of STFT, and the sampling step length of different frequencies can be adjusted by adopting a wavelet base similar to a characteristic distribution waveform, so that the frequency resolution is higher at the low frequency of a signal, and the time resolution is higher at the high frequency. Therefore, the CWT can finely depict the local morphology of the signal in the time-frequency characteristic diagram, has good time-frequency window characteristic and local resolution capability, and is a powerful tool for processing the non-stable signal abrupt change part[5]。
Therefore, the intelligent algorithm based on the wavelet time-frequency diagram and the deep learning is widely applied to feature extraction and fault diagnosis of complex mechanical systems. Compared with the traditional method, the deep learning can adaptively extract the fault characteristics of the time-frequency graph, and end-to-end fault diagnosis is realized. The compressed wavelet time-frequency graph is input to the convolutional neural network by the Yuan-Jian and the like, so that the fault type of the rolling bearing can be effectively identified; cheng et al propose a data-driven rotating machine fault diagnosis method based on CWT and a local binary convolution neural network model, and tests show that the method has more stable and reliable prediction accuracy. However, in fault diagnosis of complex delivery mechanisms, there are two difficulties in applying deep learning: sample data is insufficient: due to the complexity of a mechanical structure and a working mechanism of a supply and delivery mechanism, sufficient samples are generally difficult to obtain, and the failure diagnosis effect of the model is poor due to the fact that a deep learning model is easy to over-fit under the condition of small samples; sample data is not equalized: in order to adapt to the working and using environment of special equipment, a supply and delivery mechanism must be in a normal operation state for a long time, and key components are frequently replaced within a preset service life range, so that the number of fault samples is very deficient, the proportion of positive samples and negative samples in actual data sets is very different, the classification result of the model is biased to a majority of classes, and the identification performance of the minority of classes is poor.
The conventional method for solving the above problems is to perform data enhancement on training data based on geometric transformation, noise disturbance, gray-scale transformation, and the like. Although the method is simple to operate, the generated data distribution is too single, the data imbalance problem is not solved essentially, and the model is easy to over-fit on the target data set under the condition of a small sample. Goodfellow equals 2014 and proposes Generation of Antagonistic Networks (GANs), and mutual antagonism and learning of a generator and a discriminator are used to achieve 'Nash equilibrium', so that data distribution generated by the generator from simple hidden variables is close to real data distribution, and the problem of insufficient sample data is solved. However, the GANs network is only applied to the field of computer graphics, and cannot classify and identify the label data. Until Augustus equals 2016, an assistant classification generation type countermeasure network (ACGANs) is proposed, creatively considers that a discriminator has the capability of identifying class labels, and divides a loss function into true and false discrimination loss and sample classification loss, so that a generator can generate corresponding samples according to the labels, and the discriminator can realize multi-classification function while performing two classifications. With the progress of research, many scholars optimize ACGANs in order to obtain better classification effect. The Suspanish and the like adopt a stack shrinkage automatic coding network as a discriminator of the ACGANs, so that the fault diagnosis performance of the network is improved; the Nivian et al introduces a pooling layer in the discriminator to accelerate the calculation speed of the discriminator in multi-classification tasks and prevent over-fitting of the model; and introducing a weight coefficient into the decision probability of the ACGANs loss function by the Michelson et al, and selecting a high-quality generated sample according to the weight so as to further optimize and identify the network. In these ACGANs frameworks, image information can be learned from labels, but in the case that sample data features are not obvious, image distribution features cannot be learned effectively, and the interpretability of input hidden variables of a generator to generated data is not considered.
Disclosure of Invention
The invention aims to further improve the characteristic learning capability of ACGANs and better solve the problem of fault identification of a complex supply and delivery mechanism under the condition of unbalanced small samples.
The technical scheme adopted by the invention is as follows:
a fault diagnosis method for a complex supply and delivery mechanism based on a sparse self-coding auxiliary classification generation type countermeasure network comprises the following steps:
step 1: collecting vibration signals of a supply and transmission mechanism, calculating a wavelet time-frequency diagram, and dividing the wavelet time-frequency diagram into a training sample and a test sample;
step 2: inputting the training sample into a sparse self-encoder, and encoding to generate an implicit variable zr；
And step 3: combining class information c and hidden variables zrForming a joint feature, inputting the generator, obtaining a generated image
Figure BDA0003582402610000031
And 4, step 4: real image
Figure BDA0003582402610000032
And generating an image
Figure BDA0003582402610000033
Inputting the samples into a discriminator together to realize discrimination of authenticity and category of the samples, calculating a network loss value of the discriminator based on a loss function of the samples, and updating network parameters by using an Adam optimizer; fixing the parameters of the discriminator, and adjusting the parameters of the sparse self-encoder and the generator by using a loss function;
and 5: continuously repeating the steps 2-4 based on a counterstudy mode until the network is converged, and finishing training of the sparse self-coding auxiliary classification generation type counternetwork;
step 6: and taking out the discriminator as a fault identification network of the supply and delivery mechanism, inputting the test sample into the discriminator to realize fault diagnosis, and outputting a diagnosis result.
Preferably, in step 3, an image is generated
Figure BDA0003582402610000041
Can be expressed as:
Figure BDA0003582402610000042
in the formula (I), the compound is shown in the specification,
Figure BDA0003582402610000043
representing an original image by input
Figure BDA0003582402610000044
And a generated image obtained by the category label c; f. ofφ(. h) represents a mapping function of the sparse autoencoder, phi being a set of encoding parameters of the sparse autoencoder; the category information c is represented by a label one-hot code of the original image; generator fθ(. The) decoding parameter set is theta by combining the class information c and the hidden variable zrForm a combined feature (z)rC) to produce a generated image of a particular class
Figure BDA0003582402610000045
Then the real image is displayed
Figure BDA0003582402610000046
And generating an image
Figure BDA0003582402610000047
And inputting the input image into a discriminator together to realize the discrimination of the truth and the category of the input image.
Preferably, in step 4, the loss function of the discriminator is:
Figure BDA0003582402610000048
in the formula, the loss L is discriminatedSAnd a classification loss LCThe expression of (c) is as follows:
Figure BDA0003582402610000049
in the formula, Pdata(x) True distribution of the original image; q. q.sφ(zr| x) represents a coding mapping distribution;
Figure BDA00035824026100000410
a set of parameters for the arbiter;
Figure BDA00035824026100000411
and
Figure BDA00035824026100000412
respectively outputting the true and false of the input sample and the probability of the class to which the input sample belongs by the discriminator; loss (-) is the classification loss between the sample class and the real class of the D prediction, and a cross entropy loss function is adopted.
Preferably, the loss function of the sparse self-encoder and generator is:
Figure BDA00035824026100000413
in the formula, alpha is a coefficient for adjusting the size of the auxiliary classification generation type antagonistic network loss, and alpha is set to be 50% in the training process;
Figure BDA0003582402610000051
and representing a loss function of the sparse self-encoder, wherein the specific expression is as follows:
LSAE(x，z)＝lAE(x，z)+βΩ(ρ)
in the formula, x represents
Figure BDA0003582402610000052
z represents
Figure BDA0003582402610000053
Beta is the weight coefficient of the sparse penalty term; the expression of the sparse penalty term Ω is:
Figure BDA0003582402610000054
in the formulaD is the number of hidden layer neurons; rho is a sparse parameter; when rhojWhen the value is rho, the value of the sparse penalty term omega is zero; when rhojWhen not equal to ρ, Ω increases with the degree of deviation; controlling the average activation degree rho j of the network by limiting the sparse parameter rho;
second, lAEThe expression of (x, z) is as follows:
Figure BDA0003582402610000055
in the formula, thetazFor decoding a parameter set, the expression:
Figure BDA0003582402610000056
wherein the content of the first and second substances,
Figure BDA0003582402610000057
is a decoder weight matrix; bzIs a decoder offset vector; thetahFor a coding parameter set, the expression is:
Figure BDA0003582402610000058
wherein the content of the first and second substances,
Figure BDA0003582402610000059
as a weight matrix of the encoder, bhIs the encoder offset vector.
Preferably, the average degree of activation ρjThe expression of (a) is as follows:
Figure BDA00035824026100000510
in the formula, xiIs the ith input sample;
Figure BDA00035824026100000511
is the weight matrix of the jth neuron; b is a mixture ofjIs the bias vector for the jth neuron.
The invention has the beneficial effects that:
in the invention, an SAE structure is added in an ACGANs frame as an extractor of image characteristics, original data is coded, noise is added to a coding vector through standard normal distribution to jointly form a hidden variable containing real image characteristic information, and the hidden variable and the belonged category information are jointly input into a generator to strengthen the capability of the hidden variable for representing the characteristics related to the belonged category of the image, so that the image category characteristics can be effectively maintained in the stages of model training and identification, and the performance of a discriminator is improved. Therefore, SAE-ACGANs can capture the internal distribution of the wavelet time-frequency diagram, optimize the characteristic learning capacity of the model, improve the quality of generated data and the fault identification precision, effectively improve the influence of most types of classification preference, and finally realize the fault identification of a complex supply and delivery mechanism under the condition of small unbalanced samples. The experimental research result shows that: the SAE-ACGANs framework can fully learn the internal distribution and depth characteristics of the input samples, and compared with the original ACGANs framework, the performance of the discriminator is improved, and the improvement of the convergence speed, the training precision and the stability of the model is realized; compared with the traditional non-equilibrium data processing algorithm, the model effectively improves the influence of most types of classification preference, and greatly improves the identification capability of few types of fault samples.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the embodiments or the prior art descriptions will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.
FIG. 1 is a block diagram of the structures of GANs and ACGANs; (a) GANs; (b) ACGANs;
FIG. 2 is a schematic diagram of a fault diagnosis method based on wavelet time-frequency diagrams and SAE-ACGANs;
FIG. 3 is a flow chart of a fault diagnosis method based on wavelet time-frequency diagrams and SAE-ACGANs;
FIG. 4 is a time domain waveform diagram of a vibration signal;
FIG. 5 is a waveform diagram of the RC unit period signal in time-frequency domain; (a) a unit period time domain waveform; (b) unit period frequency domain waveform;
FIG. 6 is a time-frequency diagram of unit period signals in different states; (a) NM state signal time-frequency diagram; (b) an RC state signal time-frequency diagram; (c) SW state signal time-frequency diagram;
FIG. 7 is a PCA visualization contrast graph of extracted features; (a) visualizing the PCA of the original time-frequency graph characteristics; (b) extracting feature PCA visualization by SAE-ACGANs;
FIG. 8 is a comparison of SAE-ACGANs versus ACGANs performance; (a) generating a loss comparison map; (b) a classification loss comparison graph; (c) and (5) classifying the precision comparison graph.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention.
1. Sparse self-encoder
A single layer Automatic Encoder (AE) is a symmetric neural network composed of an encoder and a decoder. And performing cyclic training by using an encoder function and a decoder function through a back propagation algorithm, and extracting and expressing the characteristics of the input layer x by using the hidden layer h while realizing the reconstruction of the input layer x by using the output layer z.
For the original sample data set
Figure BDA0003582402610000071
AE will input layer xmMapping to an implicit layer hmThe expression of (a) is:
Figure BDA0003582402610000072
in the formula: thetahIs a set of encoding parameters;
Figure BDA0003582402610000073
as a weight matrix of the encoder, bhBiasing the vector for the encoder; f is an activation function, the invention adopts a nonlinear Sigmoid activation function, and the mathematical expression is：
Figure BDA0003582402610000074
Decoder will imply layer hmThe inverse output being input layer xmIs expressed by the reconstruction ofmThe expression is as follows:
Figure BDA0003582402610000075
in the formula: thetazIs a decoding parameter set;
Figure BDA0003582402610000076
is a decoder weight matrix; bzIs the decoder offset vector.
Learning encoding and decoding parameter sets { theta ] through back propagation trainingh，θzSuch that an input layer x is inputmAnd reconstructed output layer zmLoss function L betweenAE(x, z) minimize:
Figure BDA0003582402610000081
in order to improve the anti-noise capability of coding learning and realize deep mining of input data characteristics, the sparse autoencoder introduces an additional sparsity constraint in a loss function of AE so as to control the activation degree of hidden neurons. For the jth neuron of the hidden layer, the average output activation degree rhojComprises the following steps:
Figure BDA0003582402610000082
in the formula, xiIs the ith input sample;
Figure BDA0003582402610000083
is the weight matrix of the jth neuron; bjTo the jth godA biased vector of the element.
The KL divergence can be used to measure the difference between two asymmetric probability distributions, thus introducing the KL divergence as a penalty to constrain ρjTherefore, the expression of the sparse penalty term Ω is:
Figure BDA0003582402610000084
in the formula, D is the number of hidden layer neurons; rho is a sparse parameter; when rhojWhen the value is rho, the value of the sparse penalty term omega is zero; when rhojWhen not equal to ρ, Ω increases with the degree of deviation; the average activation ρ of the network is controlled by limiting the sparse parameter ρjTherefore, a sparse penalty mode of the network is realized.
In summary, the loss function l of SAESAEIs defined as:
LSAE(x，z)＝lAE(x，z)+βΩ(ρ) (7)
in the formula, β is a weight coefficient of the sparse penalty term.
In the training process of SAE, through back propagation of the gradient, the loss function is minimized, so that the weight matrix and the offset vector in the network are updated:
Figure BDA0003582402610000091
in the formula: ε represents the learning rate.
If the dimensionality of the hidden layer is smaller than that of the input layer, good hidden layer sparse representation of the input data can be obtained after training is completed, and the input data can effectively learn complex intrinsic features in the input data after being encoded.
2. Auxiliary classified generation type countermeasure network (ACGANs)
The GANs are composed of a generator G and a discriminator D, the structural framework is shown in figure 1(a), and the core idea of the GANs is based on double-person zero-sum game in game theory. Inputting random hidden variables z into G to yield a true data distribution P as much as possibledataOfA sample G (z) is formed, and G is used for hiding the distribution space P of the random variable z as much as possiblezMapping to the real data distribution space PdataResulting in a generated sample g (z), and D is used to determine whether the input sample is a real sample x or a generated sample g (z).
In the training process of GANs, D is trained by maximizing the difference in the data distribution of real samples x and generated samples g (z), and the generator is trained by minimizing this difference. Adopting a mechanism based on antagonism to continuously optimize the performance of G and D, and finally achieving Nash equilibrium, namely generating a sample G (z) to successfully mislead D, thereby realizing P pair by GdataAn approximate estimate of (c). Therefore, the loss function for GANs is:
Figure BDA0003582402610000092
in the formula: e [. cndot. ] is the expectation of the corresponding distribution.
ACGANs are improved on the basis of GANs, and the structure of the ACGANs is shown in figure 1 (b). The improvement of ACGANs is mainly two: inputting a hidden variable z and a class label c into G together, so as to guide G to generate a class condition sample G (c, z); secondly, a softmax classifier is added in the D, so that the probability distribution of the class labels can be calculated while judging whether the input sample is real data. Therefore, the loss function of ACGANs is represented by the discriminant loss LSAnd a classification loss LCThe method comprises the following steps:
Figure BDA0003582402610000101
in the formula: l isSLoss function for judging data true or false; l isCA loss function that is a probability distribution of data class labels; dS(x) Determining the probability that x is a true sample for D; dS(G (c, z)) judging the probability that the generated sample G (c, z) is a true sample for D; l isD(cx| x) is the loss of classification of D on x, cxIs the true category of x; l isD(c | G (c, z)) is the classification loss of D to the generating sample G (c, z), and c is the label of G (c, z).
D should distinguish the generated sample and the real sample as much as possible and effectively classify the samples; g should mimic the intrinsic distribution of real samples as much as possible and be classified as effectively.
The loss function for D is therefore as follows:
Figure BDA0003582402610000102
the loss function for G is shown below:
Figure BDA0003582402610000103
the embodiment provides a diagnosis method for generating the confrontation network fault based on sparse coding assistance all the time based on the analysis.
3.1 principle of the method
The principle of the fault diagnosis method of the supply and delivery mechanism based on the wavelet time-frequency diagram and SAE-ACGANs is shown in figure 2, and the method is divided into a discriminant training stage and a testing stage. Wherein the SAE-ACGANs framework is composed of a sparse encoder SE, a generator G and a discriminator D, and the decoder of SAE is designated as the generator of ACGANs to generate images
Figure BDA0003582402610000104
Can be expressed as:
Figure BDA0003582402610000105
in the formula:
Figure BDA0003582402610000106
representing an original image by input
Figure BDA0003582402610000107
And a generated image obtained by the category label c; f. ofφ(. represents a mapping function of a sparse autoencoder, phi is a set of encoding parameters of the sparse autoencoder(ii) a The category information c is represented by a label one-hot code of the original image; generator fθ(. The) decoding parameter set is theta by combining the class information c and the hidden variable zrForm a combined feature (z)rC) to produce a generated image of a particular class
Figure BDA0003582402610000111
Then the real image is displayed
Figure BDA0003582402610000112
And generating an image
Figure BDA0003582402610000113
And inputting the images into a discriminator together to realize discrimination of true and false and category of the input images.
As can be seen from the above, the SE will convert the raw data
Figure BDA0003582402610000114
Sparse coding as implicit variable zrFully expressing the characteristics of the original image; the decoder of the SAE is designated as G of ACGANs to fully optimize the objective function of the SAE. Thus, the complete objective function of SAE-ACGANs is:
Figure BDA0003582402610000115
in the formula:
Figure BDA0003582402610000116
the loss function of the SAE module is represented, and a specific expression is shown as a formula (7); alpha is a coefficient for adjusting the loss of ACGANs, and alpha is set to be 50% in the training process; j. the design is a squareACGANsThe loss function representing the ACGANs module is determined by the discrimination loss LSAnd a classification loss LCConsists of the following components:
Figure BDA0003582402610000117
in the formula, Pdata(x) True distribution of the original image; q. q ofφ(zr| x) represents a code mapping distribution;
Figure BDA0003582402610000118
a set of parameters for the arbiter;
Figure BDA0003582402610000119
and
Figure BDA00035824026100001110
the probability outputs of the discriminator to the true and false input samples and the belonged categories are respectively. loss (-) is the classification loss between the sample class and the true class of the D prediction, here taken as a cross entropy loss function.
Thus, the loss function for SE and G is:
Figure BDA00035824026100001111
the loss function for D is:
Figure BDA0003582402610000121
3.2 Process flow
The implementation flow of the complex supply and transportation mechanism fault diagnosis method based on the wavelet time-frequency diagram and SAE-ACGANs is shown in FIG. 3, and the specific steps are as follows:
step 1: collecting vibration signals of a supply and transmission mechanism, calculating a wavelet time-frequency diagram, and dividing the wavelet time-frequency diagram into a training sample and a test sample;
step 2: inputting the training sample into a sparse self-encoder, and encoding to generate a hidden variable zr；
And step 3: combining class information c and hidden variables zrForming a joint feature, inputting the generator, obtaining a generated image
Figure BDA0003582402610000122
And 4, step 4: will real image
Figure BDA0003582402610000123
And generating an image
Figure BDA0003582402610000124
Inputting the samples into a discriminator together to realize discrimination of authenticity and category of the samples, calculating a network loss value of the discriminator based on a loss function of the samples, and updating network parameters by using an Adam optimizer; fixing the parameters of the discriminator, and adjusting the parameters of the sparse self-encoder and the generator by using a loss function;
and 5: continuously repeating the steps 2-4 based on a counterstudy mode until the network is converged, and finishing training of the sparse self-coding auxiliary classification generation type counternetwork;
step 6: and taking out the discriminator as a fault identification network of the supply and delivery mechanism, inputting the test sample into the discriminator to realize fault diagnosis, and outputting a diagnosis result.
4. Test verification
4.1 Experimental design and Signal acquisition
The test platform structure of a certain type of supply and delivery mechanism basically comprises a power device, a test platform frame body, a control device, a swing mechanism and a test system. The test platform adopts a movable hydraulic station as power, after the manual backseat oil cylinder returns to the right position, the trigger releases the tail iron block to move backwards quickly, and the tail iron block stores energy for the swing mechanism in the re-feeding process.
The afterbody iron plate uses the gyro wheel to move backward as supporting under the slide drives for the simulation mechanism process of renaming, and later swing mechanism drives the swing arm and swings the horizontal direction by vertical direction, is used for the simulation mechanism swing process. The running states of the sliding plate and the roller directly influence whether the tail iron block can move in place or not in the working process, and further influence whether the swing arm can normally swing in place or not, so that the sliding plate and the roller are key parts for determining that the feeding mechanism can normally work. Since the slide and the roller cannot be directly provided with sensors, the test was carried out by arranging 6 vibration acceleration sensors, numbered a 1-a 6, near the swing machine of the bench device and near the press machine located above the roller. The sensor type is an ICP acceleration sensor, the sampling frequency is 10kHz, and a 32-channel LMS signal acquisition system is adopted. Repeated tests show that the data collected at the measuring point A1 right above the roller can effectively reflect the running state of the feeding mechanism, so that the vibration signal at the measuring point A1 is selected as an analysis object.
In order to simulate the faults of key parts of a supply and delivery mechanism, the sliding plate and the roller are respectively manually processed to form abrasion and crack damage, the maximum damage size is shown in table 1, and the sliding plate and the roller, the abraded sliding plate and the roller with cracks under a normal state can form three operation states: state 1-Normal state (Normal, NM), all critical parts are not damaged; state 2-Skateboard Wear (SW), only one Skateboard has long belt-like Wear; state 3-Roller Crack (RC), only one Roller has multiple cracks.
TABLE 1 Critical part maximum damage size
Tab.1 Maximum damage size of key components
Figure BDA0003582402610000131
The control device is adjusted to a cycle working mode, and vibration acceleration signals are respectively collected under the three states, and the time domain waveforms of the vibration acceleration signals are shown in fig. 4. The test runs 11 groups in cycles in NM state, each group consisting of 20 actions, and 8 groups in cycles in RC and SW state, each group consisting of 20 actions, so the total number of samples collected is 540.
4.2 Signal analysis and parameter selection
As can be seen from fig. 4, the vibration acceleration signal measured by the feeding mechanism has large impact vibration, and it is difficult to effectively distinguish the types of faults based on only the time domain waveform of the vibration signal. In addition, the test platform adopts an automatic control system to control the running time of each action, and the time of one complete unit cycle period is about 6.8 s. The time domain waveform and the frequency domain waveform of the vibration acceleration signal of the unit period in the RC state are extracted as shown in fig. 5. This cycle mainly includes four actions: energy storage, reciprocating, swinging and swinging. As can be seen from fig. 5(b), the vibration acceleration signal of the feeding mechanism belongs to a typical non-stationary signal, and it is difficult for the conventional spectrum analysis to extract the characteristics of the mechanism fault category.
CWT is carried out on the unit period vibration signals in the three states, cmor3-3 wavelets are selected to serve as wavelet basis functions of the CWT, and an obtained wavelet time-frequency diagram is shown in fig. 6.
As can be seen from fig. 6, the energy of the vibration signal of the feeding mechanism is distributed in almost all frequency bands, and the energy in the middle and low frequency bands is higher than the energy in the high frequency band. Within one unit period, the signal has obvious energy fluctuation along with the change of time, and the transient characteristic of the signal is obvious. Comparing the time-frequency diagrams of NM, RC and SW states shows that although the energy of the fault state signal is slightly higher than that of the normal state, the fault state of the supply and transmission mechanism cannot be accurately judged according to the difference in the signal transient process.
When an SAE-ACGANs method is adopted for research, the related network hyper-parameters mainly comprise iteration times, batch size, network structure and the like, and improper hyper-parameters can cause insufficient fault identification generalization capability to cause model overfitting. Comprehensively considering the size and sample size of the input time-frequency diagram, and adopting a single-factor analysis method[15]The parameters that maximize the accuracy of the model fault diagnosis are analyzed and selected as shown in table 2.
TABLE 2 model parameter Table
Tab.2 Model parameters table
Figure BDA0003582402610000151
4.3 analysis of diagnostic results
4.3.1 feature visualization analysis
In order to verify the effectiveness of the SAE-ACGANs method in extracting depth features from small input samples, Principal Component Analysis (PCA) is used to extract two principal components from the small input samples, and the two principal components are visually compared with the two principal components extracted from the time-frequency diagram features, so that a scatter diagram is formed as shown in fig. 7. As can be seen from the graph, the time-frequency diagram features of the original signals are wide in distribution range and mutually overlapped in a cross mode, the internal distribution of the input samples is fully learned through SAE-ACGANs, the depth features of different fault types can be effectively distinguished, the clustering performance is better, and the effectiveness of SAE-ACGANs in extracting the depth features is verified.
4.3.2 diagnostic accuracy analysis
In order to simulate the small sample condition of non-equilibrium, 40 NM samples, 10 RC samples and 10 SW samples are randomly selected from 540 collected test samples as training data, and 40 NM samples, RC samples and SW samples are respectively selected as test data. The model parameters were set according to table 2 for testing and compared to the original ACGANs framework using the same parameters and network structure, and the model performance was recorded, with the results shown in fig. 8.
Fig. 8 shows the ability of SAE-ACGANs and original ACGANs to learn the true sample intrinsic distribution and predict the input sample class, respectively, the generator loss shows the probability loss that the generated sample can fool the arbiter, and the arbiter loss shows the loss of whether the input sample is a generated sample. As can be seen from fig. 8(a) and (b), SAE-ACGANs adds a sparse encoder to enhance the capability of latent variable characterization of features related to the class to which the image belongs, and the generator and the encoder form a self-encoder, which can learn the feature distribution of real samples more quickly (shown by arrows in fig. 8 (a)); with the increase of the number of iterations, the generation and classification loss of the ACGANs is close to that of the SAE-ACGANs, but the variation amplitude is relatively large and cannot be stabilized, so that the performance fluctuation is large when the fault diagnosis is performed by the discriminator.
Fig. 8(c) is a comparison graph of classification accuracy of 2 frames on the test set, and large amplitude fluctuation (indicated by arrows in fig. 8 (c)) of ACGANs occurs at positions where the number of iterations is close to 160 and 350, and the instability of ACGANs performance is also reflected. In addition, in the later training period, the depth features of real samples are learned by the generator, and the classification accuracy of SAE-ACGANs can be slightly improved.
In conclusion, the SAE-ACGANs framework performs image feature constraint on the input hidden variables of the generator, the model can realize convergence at a higher speed, and the classification precision and the stability are better than those of the original ACGANs framework.
4.3.3 sample imbalance analysis
Supply and delivery mechanism fault diagnosis problem oriented to unbalanced data set needs to be more focusedFor the recognition performance of a few classes of samples, it is therefore necessary to calculate the precision P of the prediction classiAnd recall ratio RiThe calculation formula is shown in formula 18 and formula 19.
Figure BDA0003582402610000161
Figure BDA0003582402610000171
Wherein L is the number of categories of the data set; n isiiFor the number of class i samples correctly predicted as class i, nijIs the number of class j samples that were mispredicted as class i.
The precision ratio is a measure of classification accuracy, the recall ratio is a measure of classification comprehensiveness, and the performance evaluation standard for measuring and processing the unbalanced data set model needs to consider the comprehensive performance of the precision ratio and the recall ratio, so that the recognition effect of a plurality of classes and a few classes is considered. Therefore, in order to comprehensively and effectively evaluate the classification performance of the model under the condition of class imbalance, the classification accuracy rate lambda is selectedaccF1 metric λF1And G-mean index lambdaG-meanAs evaluation indexes of the comprehensive classification performance of the model, the calculation formulas are shown as formulas 20 to 22.
Figure BDA0003582402610000172
Figure BDA0003582402610000173
Figure BDA0003582402610000174
To further verify the superiority of SAE-ACGANs framework in processing unbalanced data, random over-sampling (ROS), random under-sampling (RUS) and synthesis of a few classes of over-sampling (Sy) are selectedNTHETIC MINITORY OVERSAMPLING TECHNIQUE, SMOTE) AND ADAPTIVE SAMPLE SYNTHESIS (ADASPYN)[19]Four traditional algorithms for processing non-equalized samples are used for comparative analysis. The above methods all ensure that data is generated which makes the class distribution in the training set completely balanced, i.e. the ratio of the number of samples in NM, RC and SW states in the training set is 1:1: 1. Then, CWT is carried out on the balanced data set, a time-frequency graph is used as a training set to train CNNs with the same network structure and super-parameter setting as the discriminant, and then a test set is input into the CWT, wherein the obtained fault diagnosis result is shown in Table 3.
TABLE 3 comparison of fault diagnosis results under different algorithms
Tab.3 Comparison table of fault diagnosis results under different algorithms
Figure BDA0003582402610000181
As can be seen from table 3, compared to the original data set, the diagnostic performance of the model after being optimized by the five algorithms is improved. Compared with the RUS algorithm and the CNN model, the ROS algorithm is slightly improved in diagnostic performance, and the fact that the model overfitting has a larger influence on the diagnostic performance due to oversampling compared with information loss caused by undersampling is shown. The SMOTE and the ADASYN algorithm thereof can artificially synthesize a new sample according to a few samples, so that the problem of sample imbalance is solved to a certain extent, and the possibility of overlapping samples of different classes is increased. Lambda of SAE-ACGANS algorithmacc、λF1、λG-meanCompared with the traditional unbalanced data processing algorithm, the method has the advantages that indexes are improved by 49.17% compared with original data, 0.3604 and 0.7416, and compared with the traditional unbalanced data processing algorithm, the method provided by the text can learn the actual distribution characteristics of fault samples, effectively solves the problem of unbalanced distribution of training set sample categories, improves the comprehensive classification capability of a discriminator, and reduces the failure rate of supply and output mechanism faults.
Conclusion
Aiming at the problems of few fault samples and serious imbalance among categories in the intelligent fault diagnosis of a complex supply and delivery mechanism, the invention provides a fault diagnosis method of the supply and delivery mechanism based on a wavelet time-frequency diagram and SAE-ACGANs, and test verification is carried out, wherein the result shows that:
(1) the SAE-ACGANs can fully learn the inherent distribution of input samples, effectively distinguish the depth characteristics of different fault types and have better aggregation performance.
(2) Compared with the original ACGANs framework, the SAE-ACGANs adds an SE structure in the ACGANs to form SAE together with a generator, so that image feature constraint is applied to an input hidden variable, the performance of the discriminator is further improved, and the convergence speed, the training precision and the stability of the model are improved.
(3) Compared with traditional unbalanced data processing algorithms such as ROS and RUS, the SAE-ACGANs framework can effectively improve the classification preference problem of the diagnosis model for a plurality of types of samples, and the identification capability of the model for a few types of fault samples is greatly improved.
The above description is only for the purpose of illustrating the technical solutions of the present invention and not for the purpose of limiting the same, and other modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention should be covered within the scope of the claims of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims (5)

1. A fault diagnosis method for a complex supply and delivery mechanism based on a sparse self-coding auxiliary classification generation type countermeasure network is characterized by comprising the following steps:
step 1: collecting vibration signals of a supply and transmission mechanism, calculating a wavelet time-frequency diagram, and dividing the wavelet time-frequency diagram into a training sample and a test sample;
step 2: inputting the training sample into a sparse self-encoder, and encoding to generate a hidden variable zr；
And step 3: combining class information c and hidden variables zrForming a joint feature, inputting the generator, obtaining a generated image
Figure FDA0003582402600000011
And 4, step 4: real image
Figure FDA0003582402600000012
And generating an image
Figure FDA0003582402600000013
Inputting the samples into a discriminator together to realize discrimination of authenticity and category of the samples, calculating a network loss value of the discriminator based on a loss function of the samples, and updating network parameters by using an Adam optimizer; fixing the parameters of the discriminator, and adjusting the parameters of the sparse self-encoder and the generator by using a loss function;
and 5: continuously repeating the steps 2-4 based on a mode of confrontation learning until the network is converged, and finishing training of the sparse self-coding auxiliary classification generation type confrontation network;
step 6: and taking out the discriminator as a fault identification network of the supply and delivery mechanism, inputting the test sample into the discriminator to realize fault diagnosis, and outputting a diagnosis result.
2. The method for diagnosing the faults of the complex supply and delivery mechanism based on the sparse self-coding assisted classification generation type countermeasure network as claimed in claim 1, wherein in step 3, an image is generated
Figure FDA0003582402600000014
Can be expressed as:
Figure FDA0003582402600000015
in the formula (I), the compound is shown in the specification,
Figure FDA0003582402600000016
representing an original image by input
Figure FDA0003582402600000017
And a generated image obtained by the category label c; f. ofφ(. h) represents a mapping function of the sparse autoencoder, phi being a set of encoding parameters of the sparse autoencoder; the category information c is coded by the label one-hot of the original imageCode representation; generator fθ(. The) decoding parameter set is theta by combining the class information c and the hidden variable zrForm a combined feature (z)rC) to produce a generated image of a particular class
Figure FDA0003582402600000018
Then the real image is displayed
Figure FDA0003582402600000019
And generating an image
Figure FDA00035824026000000110
And inputting the input image into a discriminator together to realize the discrimination of the truth and the category of the input image.
3. The method for diagnosing the fault of the complex supply and delivery mechanism based on the sparse self-coding assisted classification generation type countermeasure network as claimed in claim 1, wherein in step 4, the loss function of the discriminator is as follows:
Figure FDA0003582402600000021
in the formula, the loss L is discriminatedSAnd a classification loss LCThe expression of (a) is as follows:
Figure FDA0003582402600000022
in the formula, Pdata(x) The real distribution of the original image; q. q.sφ(zr| x) represents a code mapping distribution;
Figure FDA0003582402600000023
a set of parameters for the arbiter;
Figure FDA0003582402600000024
and
Figure FDA0003582402600000025
and respectively outputting the classification loss between the sample class and the real class with loss (-) predicted as D to the probability of the input sample true and false and the class to which the discriminator belongs, and adopting a cross entropy loss function.
4. The fault diagnosis method for the complex supply and delivery mechanism based on the sparse self-coding assisted classification generation type countermeasure network as claimed in claim 3, wherein the loss functions of the sparse self-encoder and the generator are as follows:
Figure FDA00035824026000000210
in the formula, alpha is a coefficient for adjusting the size of the auxiliary classification generation type antagonistic network loss, and alpha is set to be 50% in the training process;
Figure FDA0003582402600000026
the loss function of the sparse self-encoder is represented by the following specific expression:
LSAE(x，z)＝lAE(x，z)+βΩ(ρ)
in the formula, x represents
Figure FDA0003582402600000027
z represents
Figure FDA0003582402600000028
Beta is the weight coefficient of the sparse penalty term; the expression of the sparse penalty term Ω is:
Figure FDA0003582402600000029
in the formula, D is the number of hidden layer neurons; rho is a sparse parameter; when rhojWhen the value is rho, the value of the sparse penalty term omega is zero; when ρjWhen not equal to rho, omega can be deviated along withThe degree of separation increases; the average activation ρ of the network is controlled by limiting the sparse parameter ρj；
Secondly, lAEThe expression of (x, z) is as follows:
Figure FDA0003582402600000031
in the formula, thetazFor decoding a parameter set, the expression:
Figure FDA0003582402600000032
wherein the content of the first and second substances,
Figure FDA0003582402600000033
is a decoder weight matrix; bzIs a decoder offset vector; thetahFor a coding parameter set, the expression is:
Figure FDA0003582402600000034
wherein the content of the first and second substances,
Figure FDA0003582402600000035
as a weight matrix of the encoder, bhIs the encoder offset vector.
5. The method for diagnosing the fault of the complex supply and delivery mechanism based on the sparse self-coding assisted classification generation type countermeasure network as claimed in claim 4, wherein the average activation degree p isjThe expression of (c) is as follows:
Figure FDA0003582402600000036
in the formula, xiIs the ith input sample;
Figure FDA0003582402600000037
is the weight matrix of the jth neuron; bjIs the jth nerveThe offset vector of the element.
CN202210354689.2A 2022-01-17 2022-04-06 Fault diagnosis method for complex supply and delivery mechanism based on sparse self-coding assisted classification generation type countermeasure network Pending CN114676733A (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210048801 2022-01-17
CN202210048801X 2022-01-17

Publications (1)

Publication Number Publication Date
CN114676733A true CN114676733A (en) 2022-06-28

Family

ID=82077246

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210354689.2A Pending CN114676733A (en) 2022-01-17 2022-04-06 Fault diagnosis method for complex supply and delivery mechanism based on sparse self-coding assisted classification generation type countermeasure network

Country Status (1)

Country Link
CN (1) CN114676733A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116152890A (en) * 2022-12-28 2023-05-23 北京融威众邦电子技术有限公司 Medical fee self-service payment system
CN116340833A (en) * 2023-05-25 2023-06-27 中国人民解放军海军工程大学 Fault diagnosis method based on countermeasure migration network in improved field

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116152890A (en) * 2022-12-28 2023-05-23 北京融威众邦电子技术有限公司 Medical fee self-service payment system
CN116152890B (en) * 2022-12-28 2024-01-26 北京融威众邦电子技术有限公司 Medical fee self-service payment system
CN116340833A (en) * 2023-05-25 2023-06-27 中国人民解放军海军工程大学 Fault diagnosis method based on countermeasure migration network in improved field
CN116340833B (en) * 2023-05-25 2023-10-13 中国人民解放军海军工程大学 Fault diagnosis method based on countermeasure migration network in improved field

Similar Documents

Publication Publication Date Title
Ravikumar et al. Gearbox fault diagnosis based on Multi-Scale deep residual learning and stacked LSTM model
CN112784881B (en) Network abnormal flow detection method, model and system
Pan et al. Deep feature generating network: A new method for intelligent fault detection of mechanical systems under class imbalance
CN112101426B (en) Unsupervised learning image anomaly detection method based on self-encoder
CN110657984B (en) Planetary gearbox fault diagnosis method based on reinforced capsule network
CN112001270B (en) Ground radar automatic target classification and identification method based on one-dimensional convolutional neural network
CN103728551B (en) A kind of analog-circuit fault diagnosis method based on cascade integrated classifier
CN114676733A (en) Fault diagnosis method for complex supply and delivery mechanism based on sparse self-coding assisted classification generation type countermeasure network
Yu et al. PCWGAN-GP: A new method for imbalanced fault diagnosis of machines
Hou et al. Contrastive-weighted self-supervised model for long-tailed data classification with vision transformer augmented
CN113923104B (en) Network fault diagnosis method, equipment and storage medium based on wavelet neural network
Ma et al. An unsupervised domain adaptation approach with enhanced transferability and discriminability for bearing fault diagnosis under few-shot samples
CN109239669B (en) Self-evolution radar target detection algorithm based on deep learning
US20220261655A1 (en) Real-time prediction method for engine emission
Wu et al. A transformer-based approach for novel fault detection and fault classification/diagnosis in manufacturing: A rotary system application
CN115774851B (en) Method and system for detecting internal defects of crankshaft based on hierarchical knowledge distillation
CN110766060B (en) Time series similarity calculation method, system and medium based on deep learning
CN115859142A (en) Small sample rolling bearing fault diagnosis method based on convolution transformer generation countermeasure network
CN116754230A (en) Bearing abnormality detection and fault diagnosis method based on deep convolution generation countermeasure network
CN111191823A (en) Production logistics prediction method based on deep learning
Man et al. Bearing remaining useful life prediction based on AdCNN and CWGAN under few samples
Rofik et al. The Optimization of Credit Scoring Model Using Stacking Ensemble Learning and Oversampling Techniques
CN116152154A (en) Wave soldering quality detection method based on feature clustering and ensemble learning
Reshma et al. A survey on speech emotion recognition
CN113592028A (en) Method and system for identifying logging fluid by using multi-expert classification committee machine

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