CN112232252A - Optimal transport-based transmission chain unsupervised domain adaptive fault diagnosis method - Google Patents

Optimal transport-based transmission chain unsupervised domain adaptive fault diagnosis method Download PDF

Info

Publication number
CN112232252A
CN112232252A CN202011152085.7A CN202011152085A CN112232252A CN 112232252 A CN112232252 A CN 112232252A CN 202011152085 A CN202011152085 A CN 202011152085A CN 112232252 A CN112232252 A CN 112232252A
Authority
CN
China
Prior art keywords
domain
target domain
sample
fault
target
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
Application number
CN202011152085.7A
Other languages
Chinese (zh)
Other versions
CN112232252B (en
Inventor
刘朝华
蒋林博
王畅通
吴亮红
陈磊
张红强
李小花
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.)
Hunan University of Science and Technology
Original Assignee
Hunan University of Science and Technology
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 Hunan University of Science and Technology filed Critical Hunan University of Science and Technology
Priority to CN202011152085.7A priority Critical patent/CN112232252B/en
Publication of CN112232252A publication Critical patent/CN112232252A/en
Application granted granted Critical
Publication of CN112232252B publication Critical patent/CN112232252B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
    • 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)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Probability & Statistics with Applications (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Mathematical Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Signal Processing (AREA)
  • Algebra (AREA)
  • Computational Mathematics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Testing And Monitoring For Control Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a transmission chain unsupervised domain adaptive fault diagnosis method based on optimal transportation, which comprises the following steps of: building a transmission chain fault database; constructing an unsupervised domain adaptive feature extraction and classification model; establishing an unsupervised domain adaptive fault diagnosis model fusing optimal transportation; and (3) sending the target domain data into the trained feature extractor and softmax classifier to obtain a prediction label of the target domain data, and calculating the classification precision of the target domain data. The invention combines the optimal transport theory and the domain adaptation theory, aligns the distribution of the feature space and the label space by minimizing the distance between the features and the labels of the fault source domain and the target domain, solves the domain adaptation problem of the source domain and the target domain in fault diagnosis, and improves the precision of fault diagnosis.

Description

Optimal transport-based transmission chain unsupervised domain adaptive fault diagnosis method
Technical Field
The invention relates to a fault diagnosis method for a transmission chain, in particular to an unsupervised domain adaptive fault diagnosis method for the transmission chain based on optimal transportation.
Background
The transmission chain plays an important role in industrial application and is a component of many industrial systems, such as high-speed rails, airplanes, wind power systems and the like, but the transmission chain is prone to failure. The fault diagnosis of the transmission chain is important for ensuring the safe operation of the system, reducing the machine downtime and saving the maintenance cost. Therefore, the diagnosis of the failure of the drive train has received a great deal of attention, and a large number of failure diagnosis methods have been proposed.
Fault diagnosis methods can be broadly classified into model-based methods and data-driven methods. Early fault diagnosis is mainly based on a physical model, which can accurately describe the relation between a fault and a related industrial system; however, this fault diagnosis method has two main disadvantages: 1) highly dependent on a priori knowledge of the system; 2) disturbances in the industrial operation process and some assumptions about the system (e.g., the form of noise and the operating conditions of the system) may be inappropriate, which may lead to uncertainty and misdiagnosis. The data-driven fault diagnosis method directly analyzes the acquired data, and the technology of signal processing, machine learning and the like is used, so that the dependence on the prior knowledge of the system can be reduced, and the method is more suitable for modern industrial application; such as Hilbert transform and wavelet transform, are widely used for fault signature analysis. While these methods may achieve good performance on the task of fault diagnosis, they still face the challenge of automatically extracting fault signatures from the initial fault signal. Moreover, manual extraction of fault features is very time consuming.
In recent years, deep learning has been greatly developed for application in the field of fault diagnosis, such as a deep stack network and a convolutional neural network, which have good capability of automatically extracting features, and are used to solve the problem of fault diagnosis. Most existing methods are able to accurately detect faults, but the success of these methods is based on two assumptions: 1) a plurality of labeled training data; 2) the training data of the source domain and the test data of the target domain are distributed identically. When these two assumptions are violated, the performance of these algorithms may be significantly degraded. However, in practical industrial applications, such as wind power generation systems, these assumptions often cannot be established due to changes in the operating environment and instability in the load torque. Moreover, collecting data is very time consuming, so it is often difficult to collect a large amount of labeled training data. It is also generally recognized that models trained from raw data may be less reliable. Even if tag data is available under certain operating conditions, the data distribution may change with new operating conditions.
The domain adaptation is an effective method for solving the problem of data imbalance or data scarcity, can effectively relieve the influence of data distribution difference, and obviously improves the performance of the classifier. However, the existing fault diagnosis method based on the domain adaptation theory mainly has two disadvantages: 1) when the distribution of the source domain and the target domain is multi-modal, the probability measures used by these methods cannot represent the difference between the source domain and the target domain; 2) the difference between the source domain and the target domain is only considered in the high-dimensional feature space, and the difference in the label space is often ignored, resulting in insufficient adaptation. Optimal Transport (OT) theory, as a powerful tool, can calculate the distance between probability distributions. The OT distance can be calculated directly from the samples of the distribution without the need for density estimation or other non-parametric methods.
Disclosure of Invention
In order to solve the technical problems, the invention provides the optimal transport-based transmission chain unsupervised domain adaptive fault diagnosis method which is simple in algorithm and high in diagnosis precision.
The technical scheme for solving the problems is as follows: an optimal transportation-based transmission chain unsupervised domain adaptive fault diagnosis method comprises the following steps:
(1) building a transmission chain fault database: acquiring fault data of a source domain and a target domain of a transmission chain, wherein a labeled source domain fault sample set is
Figure BDA0002739833410000031
The corresponding label is
Figure BDA0002739833410000032
Wherein
Figure BDA0002739833410000033
For the m-th source domain fault sample,
Figure BDA0002739833410000034
is composed of
Figure BDA0002739833410000035
Corresponding labels, s represents a source domain, and m is the total number of source domain samples; unlabeled target domain fault sample set is
Figure BDA0002739833410000036
Wherein
Figure BDA0002739833410000037
An nth target domain fault sample is obtained, t represents a target domain, and n is the total number of target domain samples; the source domain fault sample set and the target domain fault sample set jointly form a transmission chain fault database;
(2) constructing an unsupervised domain adaptive feature extraction and classification model: an automatic encoder is used as a feature extractor f to automatically extract features of source domain and target domain fault samples, and a softmax classifier g is trained by using the source domain samples, so that a prediction label of the target domain fault samples is obtained, and unsupervised domain adaptive feature extraction and classification model construction of the target domain fault samples are realized;
(3) establishing an unsupervised domain adaptive fault diagnosis model fusing optimal transportation: processing fault samples of a source domain and a target domain through an unsupervised domain adaptive feature extraction and classification model to obtain sample features and corresponding labels; integrating the feature extractor f and the softmax classifier g into a target function of an optimal transport solver, calculating a transport planning matrix gamma through sample features and labels, aligning the distribution of a source domain and a target domain, and transporting source domain label information to the target domain; alternately optimizing the unsupervised domain adaptive feature extraction and classification model and the optimal transport solver, thereby obtaining an unsupervised domain adaptive fault diagnosis model fusing optimal transport;
(4) and (3) sending the target domain data into a trained feature extractor f and a softmax classifier g to obtain a prediction label of the target domain data, and calculating the classification precision of the target domain data.
In the optimal transportation-based unsupervised domain adaptive fault diagnosis method for the transmission chain, in the step (2), an automatic encoder is adopted as a feature extractor f, and the specific process of automatically extracting the features of the fault samples of the source domain and the target domain comprises the following steps:
the automatic encoder passes a function hθEncoding input data x into a representative feature y; this process is represented as:
y＝hθ(x)＝σ(Wx+b)
wherein θ represents a parameter of the encoding portion; w represents a weight matrix of the encoding process, b represents a bias vector of the encoding process; σ represents an activation function;
accordingly, the decoding part passes through a function h'θ′Reconstructing the representative features back into the input data:
x′＝h′θ′(y)＝σ′(W′y+b′)
wherein θ ' represents a parameter of the decoding part, x ' is reconstructed input data, W ' represents a weight matrix of the decoding process, b ' represents a bias vector of the decoding process, and σ ' represents an activation function of the decoding process;
the loss function of the entire autoencoder network is:
Figure BDA0002739833410000041
wherein x isρAnd x'ρRespectively represents the ρ -th input data and the input data reconstructed by the decoder, ρ ═ 1,2, …, I is the number of input data, | x'ρ-xρ||2Represents x'ρ-xρ2-norm of (d).
In the optimal transport-based drive chain unsupervised domain adaptive fault diagnosis method, in the step (2), the softmax classifier has the function of estimating the probability that each sample belongs to each class, and taking the class with the highest probability as the class of the sample;
given input data and corresponding labels { (x)(1),y(1)),(x(2),y(2)),…,(x(τ),y(τ)) In which x(τ)Denotes the τ th input data, y(τ)Denotes x(τ)Corresponding label, and y(τ)E {1,2, …, k }, where k represents the number of categories of the tag; then data x is input(τ)Probability of label of
Figure BDA0002739833410000043
Comprises the following steps:
Figure BDA0002739833410000042
wherein the content of the first and second substances,
Figure BDA00027398334100000514
denotes the parameters of the softmax classifier, T denotes the transpose of the vector,
Figure BDA0002739833410000051
is shown to all
Figure BDA0002739833410000052
Summing to realize the normalization of the probability value;
Figure BDA0002739833410000053
denotes the l parameter
Figure BDA00027398334100000515
Transposing;
cross entropy loss is used as a loss function for the softmax classifier
Figure BDA00027398334100000517
Figure BDA0002739833410000054
Where 1{ expression is true } -, 1{ expression is false } -, 0, and y(α)For alpha-th input data x(α)α ═ 1,2, …, τ, β denotes group β tags, β ═ 1,2, …, k;
Figure BDA0002739833410000055
denotes the beta parameter
Figure BDA00027398334100000516
The transposing of (1).
In the optimal transportation-based unsupervised domain adaptive fault diagnosis method for the transmission chain, in the step (3), an optimal transportation solver is used for calculating the distance between the probability distributions of a source domain and a target domain, and the domain adaptation problem is regarded as a special case of a discrete optimal transportation problem;
let the edge distribution of the source domain and target domain samples be mu respectivelysAnd mutAnd then the distance between the feature space and the label space of the source domain and the target domain is:
Figure BDA0002739833410000056
wherein C (i, j) represents the cost of moving the probability mass from the ith source domain sample feature and label to the jth target domain sample feature and label; c is a cost matrix, i is 1,2, …, m, j is 1,2, …, n;
Figure BDA0002739833410000057
features representing an ith source domain sample;
Figure BDA0002739833410000058
features representing a jth target domain sample; η represents a trade-off parameter;
Figure BDA0002739833410000059
indicating the L2 distance between the sample features of the computed source domain and the target domain;
Figure BDA00027398334100000510
representing a cross-entropy loss between the compute source domain and the target domain label;
since the target domain exemplar does not have a corresponding label,
Figure BDA00027398334100000511
cannot be used directly, so the source domain trained softmax classifier g generates labels for target domain sample features
Figure BDA00027398334100000512
Namely:
Figure BDA00027398334100000513
the objective function for solving the optimal transport plan is:
Figure BDA0002739833410000061
wherein, n ═ { gamma ∈ (R)+)m×n|γ1n＝μs,γ T1m＝μtDenotes the set of all transport plans γ; gamma (i, j) is the probability quality of the ith source domain sample feature and label transferring to the jth target domain sample feature and label; 1nAn n × 1 vector representing an element of 1; 1mAn mx 1 vector representing an element of 1; (R)+)m×nRepresenting a positive real matrix with dimension m × n;<·,·>expressing Frobenius dot product;
initial value gamma for transport planning0The method comprises the following steps:
Figure BDA0002739833410000062
wherein, C0A cost matrix for moving the probability mass from the source domain samples to the target domain samples, and
Figure BDA0002739833410000063
γ0(i, j) is the probability mass of the transfer of the ith source domain sample to the jth target domain sample.
In the above optimal transportation-based transmission chain unsupervised domain adaptive fault diagnosis method, in the step (3), the target loss function to be optimized by the whole model is as follows:
Figure BDA0002739833410000064
where f' denotes the decoder of the auto-encoder,
Figure BDA0002739833410000065
and
Figure BDA0002739833410000066
respectively representing source domain features
Figure BDA0002739833410000067
And
Figure BDA0002739833410000068
decoding into input data; l isfAnd LgRespectively representing the loss functions of the feature extractor f and the softmax classifier g;
alternately solving the objective loss function using a cooperative descent algorithm:
when the parameters of the feature extractor f and the softmax classifier g are fixed, the target loss function is written as:
Figure BDA0002739833410000071
the above formula is a standard linear programming problem, which is solved by a network simplex flow algorithm;
when the transport plan γ is fixed, the target loss function is written as:
Figure BDA0002739833410000072
the Adam algorithm is used to solve the above equation.
In the optimal transportation-based transmission chain unsupervised domain adaptive fault diagnosis method, in the step (4), the prediction accuracy accurve of the target domain sample label is calculated according to the following formula:
Figure BDA0002739833410000073
wherein the content of the first and second substances,
Figure BDA0002739833410000074
is a sample
Figure BDA0002739833410000075
The predicted label of the tag is used to predict,
Figure BDA0002739833410000076
is a corresponding true label; 1 {. is a binary function when
Figure BDA0002739833410000077
When the temperature of the water is higher than the set temperature,
Figure BDA0002739833410000078
is 1; when in use
Figure BDA0002739833410000079
When the temperature of the water is higher than the set temperature,
Figure BDA00027398334100000710
is 0.
The invention has the beneficial effects that:
1. according to the invention, an AE network is used as a feature extractor to extract representative features in original data, and a softmax classifier trained by source domain samples is used for classifying the target domain fault samples, so that unsupervised fault diagnosis of the target domain samples is realized.
2. According to the method, an optimal transmission theory is fused in the domain adaptation problem, the feature extractor and the softmax classifier trained by the source domain samples are used for the target domain samples by aligning the feature space and the label space of the source domain and the target domain, and the feature extraction capability and the fault diagnosis capability are improved.
Drawings
FIG. 1 is a flow chart of the present invention.
FIG. 2 is a schematic diagram of an automatic encoder according to the present invention.
FIG. 3 is a schematic diagram of the optimal transport problem of the present invention.
FIG. 4 is a bar graph of a comparative experiment of the present invention.
Detailed Description
The invention is further described below with reference to the figures and examples.
As shown in fig. 1, a method for diagnosing an unsupervised domain adaptive fault of a transmission chain based on optimal transportation includes the following steps:
(1) building a transmission chain fault database: acquiring fault data of a source domain and a target domain of a transmission chain, wherein a labeled source domain fault sample set is
Figure BDA0002739833410000081
The corresponding label is
Figure BDA0002739833410000082
Wherein
Figure BDA0002739833410000083
For the m-th source domain fault sample,
Figure BDA0002739833410000084
is composed of
Figure BDA0002739833410000085
Corresponding labels, s represents a source domain, and m is the total number of source domain samples; unlabeled target domain fault sample set is
Figure BDA0002739833410000086
Wherein
Figure BDA0002739833410000087
Is the nth target domain fault sample, t represents the target domain, and n is the total number of target domain samples.
(2) Constructing an unsupervised domain adaptive feature extraction and classification model: and (3) adopting an automatic encoder as a feature extractor f to automatically extract the features of the source domain fault samples and the target domain fault samples, and training a softmax classifier g by using the source domain samples, so as to obtain a prediction label of the target domain fault samples and realize unsupervised domain adaptive feature extraction and classification of the target domain fault samples.
The auto-encoder is used to learn more compact sample features with class discrimination, thereby speeding up the training process;
as shown in FIG. 2, the autoencoder passes a function hθInput data x is encoded as representative characteristics y. This process is represented as:
y＝hθ(x)＝σ(Wx+b)
where θ represents a parameter of the encoding portion. W denotes a weight matrix of the encoding process, and b denotes a bias vector of the encoding process. σ represents an activation function;
accordingly, the decoding part passes through a function h'θ′Reconstructing the representative features back into the input data:
x′＝h′θ′(y)＝σ′(W′y+b′)
where θ' represents a parameter of the decoding portion. x' is the reconstructed input data. W 'denotes a weight matrix of the decoding process, and b' denotes a bias vector of the decoding process. σ' represents an activation function of the decoding process;
the loss function of the entire autoencoder network is:
Figure BDA0002739833410000091
wherein x isρAnd x'ρThe input data ρ is the input data reconstructed by the decoder, and ρ is 1,2, …, and I are the number of input data. L x'ρ-xρ||2Represents x'ρ-xρ2-range ofCounting;
the softmax classifier is the most common algorithm to solve the multi-classification problem. The main function of Softmax is to estimate the probability that each sample belongs to each class and to take the class with the highest probability as the class of the sample;
given input data and corresponding labels { (x)(1),y(1)),(x(2),y(2)),…,(x(τ),y(τ))}. Wherein x is(τ)Denotes the τ th input data, y(τ)Denotes x(τ)Corresponding label, and y(τ)E {1,2, …, k }, k representing the number of categories of the tag. Then data x is input(τ)The label probability of (a) is:
Figure BDA0002739833410000092
wherein the content of the first and second substances,
Figure BDA0002739833410000093
denotes the parameters of the softmax classifier, T denotes the transpose of the vector,
Figure BDA0002739833410000101
is shown to all
Figure BDA0002739833410000102
And summing to realize the normalization of the probability value.
Figure BDA0002739833410000103
Denotes the l parameter
Figure BDA0002739833410000104
Transposing;
the cross entropy loss is used as a loss function for the softmax classifier:
Figure BDA0002739833410000105
where 1{ expression is true } -, 1, and 1{ expression is false } -, 0. y is(α)For alpha-th input data x(α)α ═ 1,2, …, τ. β represents a β -th class label, β ═ 1,2, …, k.
Figure BDA0002739833410000106
Denotes the beta parameter
Figure BDA0002739833410000107
The transposing of (1).
(3) Establishing an unsupervised domain adaptive fault diagnosis model fusing optimal transportation: processing fault samples of a source domain and a target domain through an unsupervised domain adaptive feature extraction and classification model to obtain sample features and corresponding labels; integrating the feature extractor f and the softmax classifier g into a target function of an optimal transport OT solver, calculating a transport planning matrix gamma through sample features and labels to align the distribution of a source domain and a target domain, and transporting label information of the source domain to the target domain; and alternately optimizing the unsupervised domain adaptive feature extraction and classification model and the OT solver, thereby obtaining the unsupervised domain adaptive fault diagnosis model fusing the optimal transportation.
The optimal transport OT is used to calculate the distance between the source and target domain probability distributions. The domain adaptation problem can be seen as a special case of the discrete OT problem;
as shown in FIG. 3, let the edge distributions of the source domain and target domain samples be μsAnd mutAnd then the distance between the feature space and the label space of the source domain and the target domain is:
Figure BDA0002739833410000108
where C (i, j) represents the cost of moving the probability mass from the ith source domain sample feature and label to the jth target domain sample feature and label. C is a cost matrix, i is 1,2, …, m, j is 1,2, …, n.
Figure BDA0002739833410000109
Representing the characteristics of the ith source domain sample.
Figure BDA00027398334100001010
Representing the characteristics of the jth target domain sample. η represents a trade-off parameter.
Figure BDA00027398334100001011
Indicating the L2 distance between the computed source and target domain sample features.
Figure BDA00027398334100001012
Representing a cross-entropy loss between the compute source domain and the target domain label;
since the target domain exemplar does not have a corresponding label,
Figure BDA0002739833410000111
cannot be used directly, so the source domain trained softmax classifier g generates labels for target domain sample features
Figure BDA0002739833410000112
Namely:
Figure BDA0002739833410000113
the objective function for solving the optimal transport plan is:
Figure BDA0002739833410000114
wherein, n ═ { gamma ∈ (R)+)m×n|γ1n＝μs,γ T1m＝μtRepresents the set of all transport plans γ. And gamma (i, j) is the probability quality of the transfer of the ith source domain sample characteristic and label to the jth target domain sample characteristic and label. 1nRepresenting an n × 1 vector with an element of 1, 1mRepresenting an m 1 vector with elements of 1. (R)+)m×nRepresenting a matrix of positive real numbers of dimension m n.<·,·>Expressing Frobenius dot product;
initial value gamma for transport planning0The method comprises the following steps:
Figure BDA0002739833410000115
wherein, C0A cost matrix for moving the probability mass from the source domain samples to the target domain samples, and
Figure BDA0002739833410000116
γ0(i, j) is the probability mass of the ith source domain sample being transferred to the jth target domain sample;
the objective loss function to be optimized for the entire model is:
Figure BDA0002739833410000117
where f' denotes the decoder of the auto-encoder,
Figure BDA0002739833410000118
and
Figure BDA0002739833410000119
respectively representing source domain features
Figure BDA0002739833410000121
And
Figure BDA0002739833410000122
decoded into input data. L isfAnd LgRespectively representing the loss functions of the feature extractor f and the softmax classifier g;
and alternately solving the target loss function by using a cooperative descent algorithm. When the parameters of the feature extractor f and the softmax classifier g are fixed, the target loss function can be written as:
Figure BDA0002739833410000123
the above formula is a standard linear programming problem, which can be solved by a network simplex flow algorithm;
when the transport plan γ is fixed, the target loss function can be written as:
Figure BDA0002739833410000124
the Adam algorithm is used to solve the above equation.
(4) And (3) sending the target domain data into a trained feature extractor f and a softmax classifier g to obtain a prediction label of the target domain data, and calculating the classification precision of the target domain data.
The prediction accuracy of the target domain sample label is calculated according to the following formula:
Figure BDA0002739833410000125
wherein the content of the first and second substances,
Figure BDA0002739833410000126
is a sample
Figure BDA0002739833410000127
The predicted label of the tag is used to predict,
Figure BDA0002739833410000128
is the corresponding true label. 1 {. is a binary function when
Figure BDA0002739833410000129
When the temperature of the water is higher than the set temperature,
Figure BDA00027398334100001210
is 1; when in use
Figure BDA00027398334100001211
When the temperature of the water is higher than the set temperature,
Figure BDA00027398334100001212
is 0.
In order to verify the effectiveness of the invention, the invention selects real bearing fault data in the transmission system for verification. The source domain samples are normal and fault samples under a load condition and are labeled. And the target domain samples are normal and fault samples in another state and have no label. The motor loads are four, namely 0hp, 1hp, 2hp and 3hp, and 6 cross domains are formed for experimental verification: 0-1hp, 0-2hp, 0-3hp, 1-2hp, 1-3hp and 2-3 hp. The failure diameters were 0.007 inches and 0.014 inches. And selecting methods such as a Support Vector Machine (SVM), K-nearest neighbor, softmax classifier, back propagation neural network (BP), migration component analysis (TCA), Joint Distribution Adaptation (JDA), correlation alignment method (CORAL) and the like for comparison experiment with the method. The experimental result is shown in fig. 4, the method can realize bearing fault diagnosis under various working conditions, and the precision is higher than that of other comparison methods.
In conclusion, the optimal transportation-based transmission chain unsupervised domain adaptive fault diagnosis method aligns the distribution of the source domain and the target domain by minimizing the distance between the fault feature space of the source domain and the target domain and the label space, better extracts the depth feature of the target domain sample, ensures that a label classifier trained by the source domain can be well used for target domain data, realizes unsupervised domain adaptive label prediction of the target domain, and improves the fault diagnosis precision.

Claims (6)

1. A transmission chain unsupervised domain adaptive fault diagnosis method based on optimal transportation is characterized by comprising the following steps: the method comprises the following steps:
(1) building a transmission chain fault database: acquiring fault data of a source domain and a target domain of a transmission chain, wherein a labeled source domain fault sample set is
Figure FDA0002739833400000011
The corresponding label is
Figure FDA0002739833400000012
Wherein
Figure FDA0002739833400000013
For the m-th source domain fault sample,
Figure FDA0002739833400000014
is composed of
Figure FDA0002739833400000015
Corresponding labels, s represents a source domain, and m is the total number of source domain samples; unlabeled target domain fault sample set is
Figure FDA0002739833400000016
Wherein
Figure FDA0002739833400000017
An nth target domain fault sample is obtained, t represents a target domain, and n is the total number of target domain samples; the source domain fault sample set and the target domain fault sample set jointly form a transmission chain fault database;
(2) constructing an unsupervised domain adaptive feature extraction and classification model: an automatic encoder is used as a feature extractor f to automatically extract features of source domain and target domain fault samples, and a softmax classifier g is trained by using the source domain samples, so that a prediction label of the target domain fault samples is obtained, and unsupervised domain adaptive feature extraction and classification model construction of the target domain fault samples are realized;
(3) establishing an unsupervised domain adaptive fault diagnosis model fusing optimal transportation: processing fault samples of a source domain and a target domain through an unsupervised domain adaptive feature extraction and classification model to obtain sample features and corresponding labels; integrating the feature extractor f and the softmax classifier g into a target function of an optimal transport solver, calculating a transport planning matrix gamma through sample features and labels, aligning the distribution of a source domain and a target domain, and transporting source domain label information to the target domain; alternately optimizing the unsupervised domain adaptive feature extraction and classification model and the optimal transport solver, thereby obtaining an unsupervised domain adaptive fault diagnosis model fusing optimal transport;
(4) and (3) sending the target domain data into a trained feature extractor f and a softmax classifier g to obtain a prediction label of the target domain data, and calculating the classification precision of the target domain data.
2. The optimal transport based drive chain unsupervised domain adaptive fault diagnosis method according to claim 1, characterized in that: in the step (2), an automatic encoder is used as a feature extractor f, and the specific process of automatically extracting the features of the source domain fault sample and the target domain fault sample comprises the following steps:
the automatic encoder passes a function hθEncoding input data x into a representative feature y; this process is represented as:
y＝hθ(x)＝σ(Wx+b)
wherein θ represents a parameter of the encoding portion; w represents a weight matrix of the encoding process, b represents a bias vector of the encoding process; σ represents an activation function;
accordingly, the decoding part passes through a function h'θ′Reconstructing the representative features back into the input data:
x′＝h′θ′(y)＝σ′(W′y+b′)
wherein θ ' represents a parameter of the decoding part, x ' is reconstructed input data, W ' represents a weight matrix of the decoding process, b ' represents a bias vector of the decoding process, and σ ' represents an activation function of the decoding process;
the loss function of the entire autoencoder network is:
Figure FDA0002739833400000021
wherein x isρAnd x'ρRespectively represents the ρ -th input data and the input data reconstructed by the decoder, ρ ═ 1,2, …, I is the number of input data, | x'ρ-xρ||2Represents x'ρ-xρ2-norm of (d).
3. The optimal transport based drive chain unsupervised domain adaptive fault diagnosis method according to claim 2, characterized in that: in the step (2), the softmax classifier has the function of estimating the probability of each sample belonging to each class and taking the class with the highest probability as the class of the sample;
given input data and corresponding labels { (x)(1),y(1)),(x(2),y(2)),…,(x(τ),y(τ)) In which x(τ)Denotes the τ th input data, y(τ)Denotes x(τ)Corresponding label, and y(τ)E {1,2, …, k }, where k represents the number of categories of the tag; then data x is input(τ)Probability of label of
Figure FDA0002739833400000022
Comprises the following steps:
Figure FDA0002739833400000031
wherein the content of the first and second substances,
Figure FDA00027398334000000312
denotes the parameters of the softmax classifier, T denotes the transpose of the vector,
Figure FDA0002739833400000032
is shown to all
Figure FDA0002739833400000033
Summing to realize the normalization of the probability value;
Figure FDA0002739833400000034
denotes the l parameter
Figure FDA00027398334000000313
Transposing;
cross entropy loss is used as a loss function for the softmax classifier
Figure FDA00027398334000000314
Figure FDA0002739833400000035
Where 1{ expression is true } -, 1{ expression is false } -, 0, and y(α)For alpha-th input data x(α)α ═ 1,2, …, τ, β denotes group β tags, β ═ 1,2, …, k;
Figure FDA0002739833400000036
denotes the beta parameter
Figure FDA00027398334000000315
The transposing of (1).
4. The optimal transport based drive chain unsupervised domain adaptive fault diagnosis method according to claim 3, characterized in that: in the step (3), the optimal transport solver is used for calculating the distance between the probability distributions of the source domain and the target domain, and the domain adaptation problem is regarded as a special case of the discrete optimal transport problem;
let the edge distribution of the source domain and target domain samples be mu respectivelysAnd mutAnd then the distance between the feature space and the label space of the source domain and the target domain is:
Figure FDA0002739833400000037
wherein C (i, j) represents the cost of moving the probability mass from the ith source domain sample feature and label to the jth target domain sample feature and label; c is a cost matrix, i is 1,2, …, m, j is 1,2, …, n;
Figure FDA0002739833400000038
features representing an ith source domain sample;
Figure FDA0002739833400000039
features representing a jth target domain sample; η represents a trade-off parameter;
Figure FDA00027398334000000310
indicating the L2 distance between the sample features of the computed source domain and the target domain;
Figure FDA00027398334000000311
representing a cross-entropy loss between the compute source domain and the target domain label;
since the target domain exemplar does not have a corresponding label,
Figure FDA0002739833400000041
cannot be used directly, so the source domain trained softmax classifier g generates labels for target domain sample features
Figure FDA0002739833400000042
Namely:
Figure FDA0002739833400000043
the objective function for solving the optimal transport plan is:
Figure FDA0002739833400000044
wherein the content of the first and second substances,
Figure FDA0002739833400000045
represents a set of all transport plans γ; gamma (i, j) is the probability quality of the ith source domain sample feature and label transferring to the jth target domain sample feature and label; 1nAn n × 1 vector representing an element of 1; 1mAn mx 1 vector representing an element of 1; (R)+)m×nRepresenting a positive real matrix with dimension m × n;<·,·>expressing Frobenius dot product;
initial value gamma for transport planning0The method comprises the following steps:
Figure FDA0002739833400000046
wherein, C0A cost matrix for moving the probability mass from the source domain samples to the target domain samples, and
Figure FDA0002739833400000047
γ0(i, j) is the probability mass of the transfer of the ith source domain sample to the jth target domain sample.
5. The optimal transport based drive chain unsupervised domain adaptive fault diagnosis method according to claim 4, characterized in that: in the step (3), the objective loss function to be optimized by the whole model is as follows:
Figure FDA0002739833400000048
where f' denotes the decoder of the auto-encoder,
Figure FDA0002739833400000051
and
Figure FDA0002739833400000052
respectively representing source domain features
Figure FDA0002739833400000053
And
Figure FDA0002739833400000054
decoding into input data; l isfAnd LgRespectively representing the loss functions of the feature extractor f and the softmax classifier g;
alternately solving the objective loss function using a cooperative descent algorithm:
when the parameters of the feature extractor f and the softmax classifier g are fixed, the target loss function is written as:
Figure FDA0002739833400000055
the above formula is a standard linear programming problem, which is solved by a network simplex flow algorithm;
when the transport plan γ is fixed, the target loss function is written as:
Figure FDA0002739833400000056
the Adam algorithm is used to solve the above equation.
6. The optimal transport based drive chain unsupervised domain adaptive fault diagnosis method according to claim 5, characterized in that: in the step (4), the prediction accuracy accurve of the target domain sample label is calculated according to the following formula:
Figure FDA0002739833400000057
wherein the content of the first and second substances,
Figure FDA0002739833400000058
is a sample
Figure FDA0002739833400000059
The predicted label of the tag is used to predict,
Figure FDA00027398334000000510
is a corresponding true label; 1 {. is a binary function when
Figure FDA00027398334000000511
When the temperature of the water is higher than the set temperature,
Figure FDA00027398334000000512
is 1; when in use
Figure FDA00027398334000000513
When the temperature of the water is higher than the set temperature,
Figure FDA00027398334000000514
is 0.
CN202011152085.7A 2020-10-23 2020-10-23 Transmission chain unsupervised domain adaptive fault diagnosis method based on optimal transportation Active CN112232252B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011152085.7A CN112232252B (en) 2020-10-23 2020-10-23 Transmission chain unsupervised domain adaptive fault diagnosis method based on optimal transportation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011152085.7A CN112232252B (en) 2020-10-23 2020-10-23 Transmission chain unsupervised domain adaptive fault diagnosis method based on optimal transportation

Publications (2)

Publication Number Publication Date
CN112232252A true CN112232252A (en) 2021-01-15
CN112232252B CN112232252B (en) 2023-12-01

Family

ID=74109601

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011152085.7A Active CN112232252B (en) 2020-10-23 2020-10-23 Transmission chain unsupervised domain adaptive fault diagnosis method based on optimal transportation

Country Status (1)

Country Link
CN (1) CN112232252B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112784920A (en) * 2021-02-03 2021-05-11 湖南科技大学 Cloud-side-end-coordinated dual-anti-domain self-adaptive fault diagnosis method for rotating part
CN112990259A (en) * 2021-02-04 2021-06-18 西交利物浦大学 Early fault diagnosis method of rotary mechanical bearing based on improved transfer learning
CN113822228A (en) * 2021-10-27 2021-12-21 南京大学 User expression recognition method and system based on continuous learning
WO2023044978A1 (en) * 2021-09-27 2023-03-30 苏州大学 Adversarial-flow-model-based unsupervised fault diagnosis method for mechanical device

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109376620A (en) * 2018-09-30 2019-02-22 华北电力大学 A kind of migration diagnostic method of gearbox of wind turbine failure
CN109580215A (en) * 2018-11-30 2019-04-05 湖南科技大学 A kind of wind-powered electricity generation driving unit fault diagnostic method generating confrontation network based on depth
CN109902393A (en) * 2019-03-01 2019-06-18 哈尔滨理工大学 Fault Diagnosis of Roller Bearings under a kind of variable working condition based on further feature and transfer learning
US20190325861A1 (en) * 2018-04-18 2019-10-24 Maneesh Kumar Singh Systems and Methods for Automatic Speech Recognition Using Domain Adaptation Techniques
CN110837850A (en) * 2019-10-23 2020-02-25 浙江大学 Unsupervised domain adaptation method based on counterstudy loss function

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190325861A1 (en) * 2018-04-18 2019-10-24 Maneesh Kumar Singh Systems and Methods for Automatic Speech Recognition Using Domain Adaptation Techniques
CN109376620A (en) * 2018-09-30 2019-02-22 华北电力大学 A kind of migration diagnostic method of gearbox of wind turbine failure
CN109580215A (en) * 2018-11-30 2019-04-05 湖南科技大学 A kind of wind-powered electricity generation driving unit fault diagnostic method generating confrontation network based on depth
CN109902393A (en) * 2019-03-01 2019-06-18 哈尔滨理工大学 Fault Diagnosis of Roller Bearings under a kind of variable working condition based on further feature and transfer learning
CN110837850A (en) * 2019-10-23 2020-02-25 浙江大学 Unsupervised domain adaptation method based on counterstudy loss function

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
金余丰;刘晓锋;姚美常;黄凤良;: "基于域对抗迁移的变工况滚动轴承故障诊断模型", 自动化仪表, no. 12, pages 55 - 60 *

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112784920A (en) * 2021-02-03 2021-05-11 湖南科技大学 Cloud-side-end-coordinated dual-anti-domain self-adaptive fault diagnosis method for rotating part
CN112784920B (en) * 2021-02-03 2024-02-13 湖南科技大学 Yun Bianduan coordinated rotating component reactance domain self-adaptive fault diagnosis method
CN112990259A (en) * 2021-02-04 2021-06-18 西交利物浦大学 Early fault diagnosis method of rotary mechanical bearing based on improved transfer learning
CN112990259B (en) * 2021-02-04 2023-12-26 西交利物浦大学 Early fault diagnosis method for rotary mechanical bearing based on improved transfer learning
WO2023044978A1 (en) * 2021-09-27 2023-03-30 苏州大学 Adversarial-flow-model-based unsupervised fault diagnosis method for mechanical device
CN113822228A (en) * 2021-10-27 2021-12-21 南京大学 User expression recognition method and system based on continuous learning
CN113822228B (en) * 2021-10-27 2024-03-22 南京大学 User expression recognition method and system based on continuous learning

Also Published As

Publication number Publication date
CN112232252B (en) 2023-12-01

Similar Documents

Publication Publication Date Title
CN112232252B (en) Transmission chain unsupervised domain adaptive fault diagnosis method based on optimal transportation
CN110209823B (en) Multi-label text classification method and system
Li et al. Confidence-based active learning
CN111860677B (en) Rolling bearing migration learning fault diagnosis method based on partial domain countermeasure
CN109492610B (en) Pedestrian re-identification method and device and readable storage medium
CN110705607A (en) Industry multi-label noise reduction method based on cyclic re-labeling self-service method
CN116304905B (en) Permanent magnet synchronous motor demagnetizing fault diagnosis method under multi-load working condition
CN116610803B (en) Industrial chain excellent enterprise information management method and system based on big data
CN112560948B (en) Fundus image classification method and imaging method under data deviation
CN116911929B (en) Advertisement service terminal and method based on big data
Dai et al. Hybrid deep model for human behavior understanding on industrial internet of video things
CN116910013A (en) System log anomaly detection method based on semantic flowsheet mining
CN114299326A (en) Small sample classification method based on conversion network and self-supervision
CN114780767A (en) Large-scale image retrieval method and system based on deep convolutional neural network
CN114943233A (en) Parallelization-improvement-based entity naming identification method for Transformer-CRF network
CN111177010A (en) Software defect severity identification method
CN112949097B (en) Bearing residual life prediction model and method based on deep migration learning
CN113239720B (en) Subway vehicle running gear fault diagnosis method based on deep migration learning
Li et al. Dual-source gramian angular field method and its application on fault diagnosis of drilling pump fluid end
CN116610846A (en) Breeding screening method and system for cattle breeding
CN109978013B (en) Deep clustering method for character action recognition
CN115098681A (en) Open service intention detection method based on supervised contrast learning
Brønstad Data-driven detection and identification of undesirable events in subsea oil wells
CN115795037B (en) Multi-label text classification method based on label perception
CN111158640B (en) One-to-many demand analysis and identification method based on deep learning

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