CN112101085B - Intelligent fault diagnosis method based on importance weighted domain antagonism self-adaptation - Google Patents

Intelligent fault diagnosis method based on importance weighted domain antagonism self-adaptation Download PDF

Info

Publication number
CN112101085B
CN112101085B CN202010711903.6A CN202010711903A CN112101085B CN 112101085 B CN112101085 B CN 112101085B CN 202010711903 A CN202010711903 A CN 202010711903A CN 112101085 B CN112101085 B CN 112101085B
Authority
CN
China
Prior art keywords
domain
network
importance
classifier
source
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.)
Active
Application number
CN202010711903.6A
Other languages
Chinese (zh)
Other versions
CN112101085A (en
Inventor
王宇
孙晓杰
訾艳阳
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.)
Xian Jiaotong University
Original Assignee
Xian Jiaotong University
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 Xian Jiaotong University filed Critical Xian Jiaotong University
Priority to CN202010711903.6A priority Critical patent/CN112101085B/en
Publication of CN112101085A publication Critical patent/CN112101085A/en
Application granted granted Critical
Publication of CN112101085B publication Critical patent/CN112101085B/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
    • G06F18/21—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
    • 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
    • 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
    • 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

Abstract

An adaptive intelligent fault diagnosis method based on importance weighted domain antagonism collects vibration signals of a rotary machine under different working conditions, and divides the signals by adopting a moving time window for data sets under different working conditions respectively; constructing a domain category identification network, and outputting importance weights of the source domain samples in confrontation training; extracting discriminative features in the data set; combining the feature extractor and the domain discriminator to construct an importance weighted domain adaptive network; and training the network model by adopting a training strategy of the countermeasure network until the model converges, and identifying the bearing health state of the target domain data set lacking the fault label by using the trained class classifier. The invention completes the migration of diagnosis knowledge by diagnosing the fault of the working condition with insufficient data information by means of the working condition with abundant data information, constructs a deep learning network, overcomes the dependence on expert knowledge in the traditional diagnosis method, and provides an effective tool for reducing the cost of the future intelligent fault diagnosis system.

Description

Intelligent fault diagnosis method based on importance weighted domain antagonism self-adaptation
Technical Field
The invention relates to a rolling bearing state evaluation method, in particular to an importance weighted domain based adaptive intelligent fault diagnosis method.
Background
The rolling bearing is used as an important rotating part in mechanical equipment and is widely applied to the fields of aerospace, engineering machinery, ship equipment, hydraulic engineering and the like. The health and performance of the rolling bearing directly affect the safety and reliability of the mechanical equipment. However, rolling bearings are vulnerable components of mechanical equipment, and failure of a bearing may result in shutdown of the entire mechanical system, causing an unexpected economic loss. Therefore, the research on the health state detection and fault diagnosis technology of the bearing is of great significance.
By using the sensor to collect the vibration signal of the bearing to analyze in the operation process of the bearing, the state of the monitored equipment can be judged. With the development of the intelligent diagnosis method, fault data under certain working conditions are difficult to obtain and are not enough to support the establishment of a diagnosis model, and the intelligent fault diagnosis method based on transfer learning is widely researched. The traditional intelligent diagnosis method based on the transfer learning can realize that the diagnosis knowledge is transferred from the working condition with large data volume and easy acquisition to the working condition with incomplete data volume, so that the generalization performance of the mechanical equipment diagnosis system is improved. Although many achievements have been made in the state detection of the intelligent diagnostic method based on the transfer learning, there are still many places to be ignored. The traditional intelligent method based on transfer learning is established on the premise of certain assumptions: i.e. a destination domain with insufficient data size needs to have all the types of failures present in a source domain with sufficient data size, i.e. the label space of both is the same. However, this has a great limitation in practical industrial applications, and it is difficult to ensure that the types of failures present in all source domains occur in the target domain. It would therefore be of great interest if source domain knowledge could be transmitted without having to wait for every possible failure to occur in the target domain.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide an importance weighting domain-based adaptive intelligent fault diagnosis method, which overcomes the defects of the traditional transfer learning-based intelligent diagnosis method in industrial application and overcomes the generation of negative transfer caused by the fact that a target domain training set is a data set which does not contain complete fault categories.
In order to realize the purpose, the invention is realized by the following technical scheme:
an adaptive intelligent fault diagnosis method based on importance weighted domain antagonism comprises the following steps,
(1) collecting vibration signals of a rotary machine under two different working conditions, respectively dividing the vibration signals under the two working conditions by adopting a moving time window to obtain data sets under the two working conditions, and dividing the data sets under the two working conditions into a source domain data set and a target domain data set;
(2) constructing a domain category identification network consisting of a feature extractor and a domain classifier, training the domain category identification network by adopting a source domain data set and a target domain data set to obtain the importance weight of a source domain sample in the source domain data set, and obtaining the importance weight of a normalized source domain sample according to the importance weight;
(3) forming a fault mode identification network by the feature extractor and the category label classifier;
(4) forming an importance weighted domain impedance adaptive network by the feature extractor, the importance weight of the normalized source domain sample and the domain discriminator;
(5) and sequentially iterating the training domain classifier, the fault mode identification network and the importance weighting domain antagonistic adaptive network by adopting an antagonistic game strategy until all network parameters are converged, and then identifying the health state of the rotating machine in the target domain data set by using the class label classifier.
The invention is further improved in that the specific process of the step (1) is as follows:
1) using a sensor to collect vibration signals of each fault type of a rolling bearing under two different working conditions;
2) selecting an optimal advancing step length, and segmenting the vibration signal by adopting a moving time window; and respectively obtaining data sets under two working conditions, wherein the working condition data set with complete fault types and fault labels is set as a source domain data set, and the data set with incomplete fault types and missing fault labels is set as a target domain data set.
In a further development of the invention, in step (2), a feature extractor and domain classifier D is constructed0The specific process of the formed domain category identification network is as follows:
1) constructing a feature extractor for the multilayer one-dimensional convolution neural network layer;
2) connecting the domain classifier with the feature extractor through a full connection layer to form a domain category identification network; identifying the network according to the domain type to obtain a domain type label prediction result;
3) calculating a domain class identification network loss function according to the result of the domain class label prediction, and iteratively training a domain classifier until the domain classifier is stable;
4) and obtaining the importance weight of the source domain sample by training a stable domain category identification network, and obtaining the importance weight of the normalized source domain sample according to the importance weight of the source domain sample.
The further improvement of the invention is that the specific process of the step 2) is as follows: training a data set from a source domain
Figure BDA0002596876160000031
Training data set with target domain
Figure BDA0002596876160000032
Wherein the content of the first and second substances,
Figure BDA0002596876160000033
a sample point of a source domain;
Figure BDA0002596876160000034
a label for a source domain sample point; n issThe number of the source domain sample points is; y is the label space of the source domain sample;
Figure BDA0002596876160000035
a sample point that is a target domain;
Figure BDA0002596876160000036
a label for a target domain sample point, the target domain label being absent at the time of training; n istThe number of the target domain sample points is; y issubIs the label space of the target domain sample, YsubA subset of label space that is a source domain sample; obtaining the characteristics of the vibration signal through a characteristic extractor:
Figure BDA0002596876160000037
in formula 1-1:
Figure BDA0002596876160000039
is a function of the network output, where xs、xt、θh、Hs、HtRespectively representing a source domain sample input by a network, an input target domain sample, network parameters of a feature extractor, source domain features output after the source domain sample passes through the feature extractor, and target domain features output after the target domain sample passes through the feature extractor;
passing through domain classifier D for source domain features and target domain features0And obtaining the result of the domain category label prediction:
Figure BDA0002596876160000038
in formulas 1-2: sigmoid (·) function maps inputs between 0, 1; p0Predicting a probability vector for the domain class label of the normalized network sample; thetad0Representative Domain classifier D0The network parameter of (2).
A further improvement of the present invention is that, in step 3), the domain class identification network loss function is as follows:
Figure BDA0002596876160000041
in formulas 1-3: j (D)0,Hs,Ht) A loss function for a domain class classifier;
Figure BDA0002596876160000042
the function is to obtain the mean value based on the source domain samples;
Figure BDA0002596876160000043
to find a mean value based on samples of the target domain, P0A probability vector is predicted for the domain class label of the normalized network sample.
A further improvement of the present invention is that, in step 4), the importance weights of the source domain samples are:
Figure BDA0002596876160000044
in formulae 1-4: dzSigmoid (. cndot.) function maps inputs to between 0, 1, H, which is the importance weight of the source domain samplessIs the source domain feature, theta, output after the source domain sample passes through the feature extractord0Representative Domain classifier D0The network parameter of (2);
the importance weights of the normalized source domain samples are:
Figure BDA0002596876160000045
in formulas 1-5: w (z) is the importance weight of the normalized source domain samples.
The invention has the further improvement that the specific process of the step (3) is as follows:
the category label classifier is connected with the feature extractor through a full connection layer, and carries out pattern recognition classification on the discriminant features extracted by the feature extractor to obtain a fault type prediction result;
and secondly, calculating a failure mode recognition network loss function according to the failure type prediction result to form a failure mode recognition network.
The invention is further improved in that the specific process of the step (4) is as follows:
a) obtaining the domain invariant features of the source domain data set and the target domain data set in a high-dimensional space through a feature extractor;
b) according to the domain invariant characteristics of the source domain data set and the target domain data set in the high-dimensional space, passing through a domain discriminator D1Performing domain category identification and classification to obtain a domain category label prediction result;
c) and calculating a loss function of the importance weighted-reactance self-adaptive network according to the domain class label prediction result and the importance weight of the normalized source domain sample, and completing the construction of the importance weighted-domain-reactance self-adaptive network.
The further improvement of the invention is that the specific steps of adopting a countermeasure game strategy to sequentially iterate the training domain classifier, the fault mode identification network and the importance weighting domain countermeasure adaptive network in the step (5) until all network parameters are converged are as follows:
1) in k iterations, the domain classifier D is minimized using the following equation0Loss function, updating domain classifier D by back propagation algorithm0Network parameter θ ofd0Until stable;
Figure BDA0002596876160000051
Figure BDA0002596876160000052
in the formulae 1 to 11,. alpha.1Is the learning rate;
Figure BDA0002596876160000053
as domain classifier D0A gradient of (a);
2) training the domain discriminant D against the loss function of the adaptive network by minimizing the importance weighting1Then training a class label classifier through a loss function of a minimum fault mode recognition network, and finally training a feature extractor through a loss function of the minimum label classifier and a maximum importance weighted impedance adaptive network loss function;
3) and repeating the iteration steps 1) and 2) until all network parameters are converged.
Compared with the prior art, the invention has the following beneficial effects:
the method uses the domain-impedance self-adaptive network, has the characteristic of automatic extraction of domain-invariant features, and reduces the feature distribution difference caused by working condition change; the local domain self-adaption is realized by using the weighted domain countermeasure based on the source domain samples, and the negative migration caused by incomplete target domain label space is overcome; the deep convolutional neural network is used for extracting the signal characteristics, the deep fault characteristics can be extracted, and the accuracy of establishing a fault diagnosis model is higher. The method can complete the migration of the mechanical equipment state recognition model under different working conditions, has the characteristics of low cost, high efficiency, practicality and the like, is suitable for identifying the health state of the rotary mechanical system bearing on site in real time, provides a reliable and convenient tool for the intelligent diagnosis method based on the migration learning, and has important field significance and wide application prospect.
Furthermore, the signal discriminant features are extracted in a self-adaptive mode through the one-dimensional convolutional neural network, and the dependence of feature extraction on expert knowledge in the traditional machine learning is overcome.
Drawings
Fig. 1 is a flowchart of the adaptive rolling bearing intelligent fault diagnosis based on importance weighted domain impedance under different working conditions.
Fig. 2 is a schematic diagram of a domain class identification network structure.
Fig. 3 is a schematic diagram of an adaptive network with importance weighted domain antagonism.
Fig. 4 is a simplified diagram of the procedure of the countermeasure domain adaptation method.
FIG. 5 shows the result of the confusion matrix comparing the outputs of different comparison methods. Wherein (a) is the result of training on the target domain dataset using only the source domain dataset; (b) the result that the DANN method is used and the target domain data set has no fault species missing exists; (c) the method is a result of using the DANN method and the target domain data set has fault type missing; (d) the same result as in (c) for the target domain dataset is achieved using the proposed method.
Detailed Description
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.
The present invention will now be described in further detail with reference to specific examples, which are intended to be illustrative, but not limiting, of the invention.
Referring to fig. 1, an intelligent fault diagnosis method based on importance weighted domain antagonism adaptation includes the following steps,
(1) acquiring vibration signals of a rotary machine under two working conditions by using a sensor, respectively dividing the vibration signals under the two working conditions by adopting a moving time window to obtain data sets under the two well-divided working conditions, and dividing the data sets under the two working conditions into a source domain data set with complete fault types and fault labels and a target domain data set with incomplete fault types and fault labels missing;
the specific process of the step (1) is as follows:
1) using a sensor to collect vibration signals of each fault type of a rolling bearing under two different working conditions;
2) selecting an optimal advancing step length, and segmenting the vibration signal by adopting a moving time window; respectively obtaining data sets under two working conditions, wherein the working condition data set with complete fault types and fault labels is set as a source domain data set
Figure BDA0002596876160000071
Setting the data set with incomplete fault type and missing fault label as the target domain data set
Figure BDA0002596876160000072
Wherein the content of the first and second substances,
Figure BDA0002596876160000073
a sample point of a source domain;
Figure BDA0002596876160000074
a label for a source domain sample point; n issThe number of the source domain sample points is; y is the label space of the source domain sample;
Figure BDA0002596876160000075
a sample point that is a target domain;
Figure BDA0002596876160000076
a label for a target domain sample point, the target domain label being absent at the time of training; n istThe number of the target domain sample points is; y issubIs the label space of the target domain sample, YsubA subset of the label space of the source domain samples.
(2) Construct pass feature extractor and domain classifier D0Forming a domain type identification network, forming the domain type identification network by adopting the source domain data set and the target domain data set obtained in the step (1) and through a training feature extractor and a domain discriminator, obtaining the importance weight of the source domain sample, and obtaining the importance weight of the normalized source domain sample according to the importance weight of the source domain sample;
in the step (2), a feature extractor and a domain classifier D are constructed0The specific process of forming the domain category identification network is as follows:
1) the characteristic extractor is built by a plurality of layers of one-dimensional convolution neural network layers and can adaptively extract the characteristics of the vibration signals in the data set of the source domain and the target domain;
specifically, the specific process of the step 1) is as follows:
a) training data from source domain
Figure BDA0002596876160000077
Training data with target domain
Figure BDA0002596876160000078
Obtaining the characteristics of the vibration signal through a characteristic extractor:
Figure BDA0002596876160000079
in the formula 1-1:
Figure BDA00025968761600000710
is a function of the network output, where xs、xt、θh、Hs、HtRespectively representing source domain samples of network input, target domain samples of input, feature extractorAnd the network parameters comprise source domain characteristics output after the source domain samples pass through the characteristic extractor and target domain characteristics output after the target domain samples pass through the characteristic extractor.
2) Domain classifier D0Connecting with the feature extractor through a full connection layer to form a domain category identification network; and identifying the network according to the domain type to obtain the result of the domain type label prediction.
Specifically, the specific process of the step 2) is as follows:
source domain and target domain features pass through domain classifier D0And obtaining the result of the domain category label prediction:
Figure BDA0002596876160000081
in formulas 1-2: sigmoid (·) function maps inputs between 0, 1; p0Predicting probability vectors for domain class labels of the normalized network samples, wherein the domain labels of the samples are set to be 1 from the source domain data set and 0 from the target domain data set; theta.theta.d0Representative Domain classifier D0The network parameter of (2).
3) Iteratively training a domain classifier D by computing a domain class recognition network loss function0Until stable;
specifically, in step 3), the specific process of calculating the domain class identification network loss function is as follows:
calculating a domain type identification network loss function according to the result of the domain type label prediction;
Figure BDA0002596876160000082
in formulas 1-3: j (D)0,Hs,Ht) A loss function for a domain class classifier;
Figure BDA0002596876160000083
the function is to obtain the mean value based on the source domain samples;
Figure BDA0002596876160000084
to find the mean based on the target domain samples.
4) And obtaining the importance weight of the source domain sample by training a stable domain category identification network, and obtaining the importance weight of the normalized source domain sample according to the importance weight of the source domain sample.
Specifically, the specific process of step 4) is as follows:
the source domain sample is trained to obtain the importance weight of the source domain sample through a stable domain category identification network:
Figure BDA0002596876160000085
in formulas 1 to 4: dzIs the importance weight of the source domain samples. When the output result of the sample is close to 1, the sample comes from the outlier class of the source domain (fault class not existing in the target domain), and almost no target domain sample is similar to the outlier class, so the weight of the sample in the next confrontation learning should be reduced; when its output is close to 0.5, indicating that the sample is from a common class (fault class present in the target domain), the weight of the sample should be maximized, so:
Figure BDA0002596876160000091
in formulas 1-5: w (z) is the importance weight of the normalized source domain sample.
(3) Forming a fault mode identification network by the feature extractor and the class label classifier in the step (2), extracting discriminative features in the source domain data set by the feature extractor, and identifying various fault states under the working condition corresponding to the source domain data set by the class label classifier according to the discriminative features;
the fault identification network in the step (3) is formed by the following processes:
connecting a label classifier with the feature extractor in the step (2) through a full connection layer, and performing pattern recognition classification on the discriminant features extracted by the feature extractor in the step (2) to obtain a fault type prediction result; the failure type prediction result is as follows;
Figure BDA0002596876160000092
in formulae 1-6: the softmax (.) function maps the input to a probability distribution that sums to 1;
Figure BDA0002596876160000093
outputting the function for the network; y issPredicting an output probability vector, θ, for a fault type of a normalized network source domain samplecNetwork parameters representing a class label classifier.
And secondly, calculating a failure mode recognition network loss function according to the failure type prediction result to form a failure mode recognition network.
Step two, the network loss function of fault mode identification is as follows:
Figure BDA0002596876160000094
in formulae 1-7:
Figure BDA0002596876160000095
a loss function of the source domain samples passing through the pattern recognition network is adopted, and B is a batch size (batch-size) of each iteration process; y islabelA true fault type label vector for the source domain sample;
(4) feature extractor, normalized source domain sample importance weights and domain discriminator D1Forming an importance weighting domain adaptive network, and extracting domain invariant features of the source domain data set and the target domain data set according to the importance weighting domain adaptive network;
specifically, in step (4), the importance weight of the feature extractor and the normalized source domain sample in step (2) is combined with a domain discriminator D1The specific process of forming the importance weighting domain versus adaptive network is as follows:
a) obtaining the domain invariant features of the source domain data set and the target domain data set in a high-dimensional space by adopting the source domain data set and the target domain data set and passing through a feature extractor;
b) according to the domain invariant characteristics of the source domain data set and the target domain data set in the high-dimensional space, passing through a domain discriminator D1And performing domain category identification and classification to obtain a domain category label prediction result.
The specific process of the step b) is as follows: the source domain sample characteristics H after passing through the characteristic extractorsAnd target domain sample feature HtInput field discriminator D1Outputting a domain category label prediction result of the sample;
Figure BDA0002596876160000101
in formulae 1-8: p1Predicting probability vectors for domain class labels of the normalized network samples, wherein the domain labels of the samples are set to be 1 from the source domain data set and 0 from the target domain data set; thetad1Representative domain discriminator D1The network parameter of (2).
c) And calculating a loss function of the importance weighted-reactance self-adaptive network according to the domain class label prediction result and the importance weight of the normalized source domain sample, and completing the construction of the importance weighted-domain-reactance self-adaptive network.
The invention further improves that the concrete process of the step c) is as follows: the specific process of calculating the importance weighted impedance adaptive network's loss function is as follows:
Figure BDA0002596876160000102
in formulas 1 to 9: j. the design is a squareweight(D1,Hs,Ht) A loss function expressed as an importance weighted impedance adaptive network;
(5) successively and iteratively training the domain classifier D in the step (2) by adopting a countermeasure game strategy0The failure mode identification network in the step (3) and the importance weighted domain countermeasure adaptive network in the step (4) untilAll network parameters are converged, the trained feature extractor can extract discriminant features and domain invariant features, and finally the class label classifier is used for identifying the bearing health state in the target domain data set lacking fault labels, so that the migration of the diagnosis knowledge from the source domain to the target domain is completed.
Wherein, the classifier D in the step (2) is sequentially and iteratively trained by adopting a confrontation game strategy in the step (5)0The failure mode identification network in the step (3) and the importance weighted domain antagonistic adaptive network in the step (4) until all network parameters are converged comprise the following specific steps:
1) in k iterations, the domain classifier D is minimized0The loss function, as shown in equations 1-10, updates the domain classifier D through a back propagation algorithm0Network parameter θ ofd0Until stable, as shown in equations 1-11;
Figure BDA0002596876160000111
Figure BDA0002596876160000112
in the formulae 1 to 11,. alpha.1Is the learning rate;
Figure BDA0002596876160000113
as domain classifier D0A gradient of (a);
2) training the fault mode recognition network in the step (3) and the importance weighted domain impedance adaptive network in the step (4) in a combined manner; the specific process is as follows:
as shown in equations 1-12, the domain arbiter D is trained by minimizing the loss function of the importance weighted adversary adaptive network1Guaranteed domain discriminator D1The method comprises the steps of having the function of identifying domain class labels, training a class label classifier by minimizing a loss function of a fault mode identification network to enable the class label classifier to identify fault types under working conditions, and finally minimizing a loss function of the class label classifier and maximizing the loss functionThe importance weighted impedance adaptive network loss function trains a feature extractor, and guides the feature extracted by the feature extractor to extract not only domain discriminant features but also domain invariant features; network parameter theta of domain discriminatord1The updating process of (2) is shown in equations 1-13; network parameter θ of class label classifiercThe updating process of (2) is shown in equations 1-14; network parameter theta of feature extractorhThe updating process of (2) is shown in equations 1-15;
Figure BDA0002596876160000114
Figure BDA0002596876160000115
Figure BDA0002596876160000116
θh←θhβJweight(D1,Hs,Ht) (1-15)
in the above formula: alpha is alpha2Is the learning rate; beta is a weight coefficient;
Figure BDA0002596876160000121
is domain discriminator D1A gradient of (a);
Figure BDA0002596876160000122
a gradient of a class classifier;
Figure BDA0002596876160000123
is the gradient of the feature extractor;
3) and repeating the process of the step 1) and the step 2) until all network parameters are converged.
The following gives a specific application example process, and at the same time, the effectiveness of the invention in engineering application is verified.
The method for identifying the health state of the rolling bearing of the rotary machine under different working conditions by using the importance weighted domain impedance self-adaption method is implemented according to the following specific steps:
the experiment is carried out on a SQ mechanical fault comprehensive simulation experiment table, bearing damage simulation is respectively carried out under multiple groups of rotating speeds, and the SQ mechanical fault comprehensive simulation experiment table has 5 types of bearings including healthy bearings, mild bearings in inner rings, moderate bearings in inner rings, mild bearings in outer rings and moderate bearings in outer rings. The experiment was performed using a data acquisition instrument CoCo80 for vibration signal acquisition with a sampling frequency of 11.52 kHz.
By utilizing the method, the method for identifying the health state of the bearing in the target domain comprises the following steps:
(1) data set partitioning:
firstly, acquiring sensor vibration data under three rotation speed working conditions of 300rpm, 480rpm and 660rpm, wherein the acquisition time is 120s, and the sampling frequency is 11.52 khz. Firstly, carrying out FFT (fast Fourier transform) on vibration data to obtain frequency domain signals, secondly, carrying out sample division on the vibration data by adopting a moving time window in order to expand the vibration data, selecting a step size of 1024 with an overlap rate of 80% to divide each fault frequency domain signal, finally setting a source domain data set as a full fault type and labeled data set, wherein the total number of samples is 5 x 1000, and the total number of samples is 3 x 1000 when a target domain data set lacks two fault types, namely a light fault type in the outer circle and a medium fault type in the outer circle and is not labeled. The test set was a full-scale labeled target domain dataset with a total of 3 x 1000 samples, and all experiments were repeated 5 times to avoid chance and specificity.
(2) Domain class identification network construction
In the present invention, as shown in FIG. 2, the designed domain category identification network is composed of a feature extractor and a domain classifier D0The device comprises a feature extractor, a data processing module and a data processing module, wherein the feature extractor comprises three one-dimensional convolution layers, three pooling layers and an expansion layer; domain classifier D0From two fully-connected layers, the structural parameters were obtained from multiple experiments. In this study, the output activation function for each convolutional layer is Relu, domain classifier D0The activation function of the last layer is Sigmoid.
The three pooling layers play a role in reducing the feature dimension and characterizing the translational invariance. The pooling operation makes the neural network learned features robust.
After the last pooling layer, an unwind layer is used as a transition between the fully-connected layer and the convolutional layer. To avoid overfitting, dropout and l are used2Two methods of regularization.
(3) Failure mode identification network construction
In the invention, the designed fault mode identification network consists of the feature extractor and the class label classifier in the step (2), wherein the class label classifier consists of a full connection layer and a softmax layer, and the structural parameters are obtained by a plurality of tests.
(4) Construction of importance weighted domain antagonistic adaptive network
In the present invention, as shown in FIG. 3, the designed importance weighted domain adaptive network is formed by the feature extractor domain discriminator D in step (2)1Composition, domain discriminator D1And domain classifier D0The same network structure is adopted, the domain invariant feature is obtained through the countermeasure game strategy in the training process, and the domain countermeasure self-adaption process is shown in the figure 4.
(5) Distributed iterative training process
Firstly, the minimum domain class recognition network loss function is used for iterative training 5 times, and the domain classifier D is updated0Network parameter θ ofd0The data set is stable, so that the data set can be accurately identified from a source domain or a target domain, and a prediction result of a source domain sample is output; next, the feature extractor, the tag classifier, and the domain discriminator D are iteratively trained 5 times1Until the network reaches convergence temporarily in the confrontation iteration, repeating the two processes to make each network parameter reach convergence finally; and finally, inputting the target domain test data set into the network, and identifying the bearing state of the target domain test data set.
In order to verify the effectiveness of the invention, a CNN network with the same network structure as that of the experiment and without domain impedance adaptive extraction is selected, a target domain training set contains all fault types of training using the DANN network, the target domain training set is the same as that of the experiment and is compared with the result of the training using the DANN network, a plurality of groups of migration tasks are tested, taking the example of migration from 480rpm to 600rpm, and the confusion matrix of the output results of the network is shown as (a), (b), (c) and (d) in fig. 5. The result shows that the importance weighted domain impedance adaptive intelligent fault diagnosis method provided by the invention is superior to other traditional methods, and not only can extract discriminant characteristics to complete the detection of a fault mode, but also can extract domain invariant characteristics and avoid the generation of negative migration, thereby realizing the migration of fault diagnosis knowledge.
The invention carries out fault diagnosis on the working condition with insufficient data information by means of the working condition with abundant data information, completes the migration of diagnosis knowledge, constructs a deep learning network, overcomes the dependence on expert knowledge in the traditional diagnosis method, is a novel fault diagnosis method based on local self-adaptation, and provides an effective tool for reducing the cost of the future intelligent fault diagnosis system.

Claims (8)

1. An adaptive intelligent fault diagnosis method based on importance weighted domain antagonism is characterized by comprising the following steps,
(1) collecting vibration signals of a rotary machine under two different working conditions, respectively dividing the vibration signals under the two working conditions by adopting a moving time window to obtain data sets under the two working conditions, and dividing the data sets under the two working conditions into a source domain data set and a target domain data set;
(2) constructing a domain category identification network consisting of a feature extractor and a domain classifier, training the domain category identification network by adopting a source domain data set and a target domain data set to obtain the importance weight of a source domain sample in the source domain data set, and obtaining the importance weight of a normalized source domain sample according to the importance weight;
(3) forming a fault mode identification network by the feature extractor and the category label classifier;
(4) forming an importance weighted domain impedance adaptive network by the feature extractor, the importance weight of the normalized source domain sample and the domain discriminator; the specific process is as follows:
a) obtaining the domain invariant features of the source domain data set and the target domain data set in a high-dimensional space through a feature extractor;
b) according to the domain invariant characteristics of the source domain data set and the target domain data set in the high-dimensional space, passing through a domain discriminator D1Performing domain category identification and classification to obtain a domain category label prediction result;
c) calculating a loss function of the importance weighted-reactance adaptive network according to the domain category label prediction result and the importance weight of the normalized source domain sample, and completing the construction of the importance weighted-domain-reactance adaptive network;
(5) and sequentially iterating the training domain classifier, the fault mode recognition network and the importance weighting domain adaptive network by adopting an antagonistic game strategy until all network parameters are converged, and then recognizing the health state of the rotating machine in the target domain data set by using the class label classifier.
2. The adaptive intelligent fault diagnosis method based on importance weighted domain antagonism according to claim 1, wherein the specific process of step (1) is as follows:
1) using a sensor to collect vibration signals of each fault type of a rolling bearing under two different working conditions;
2) selecting an optimal advancing step length, and segmenting the vibration signal by adopting a moving time window; and respectively obtaining data sets under two working conditions, wherein the working condition data set with complete fault types and fault labels is set as a source domain data set, and the data set with incomplete fault types and missing fault labels is set as a target domain data set.
3. The intelligent fault diagnosis method based on importance weighted domain impedance adaptation of claim 1, wherein in step (2), the structure is composed of a feature extractor and a domain classifier D0The specific process of the formed domain category identification network is as follows:
1) building a feature extractor for the multilayer one-dimensional convolution neural network layer;
2) connecting the domain classifier with the feature extractor through a full connection layer to form a domain category identification network; identifying the network according to the domain type to obtain a domain type label prediction result;
3) calculating a domain class identification network loss function according to the result of the domain class label prediction, and iteratively training a domain classifier until the domain classifier is stable;
4) and obtaining the importance weight of the source domain sample by training a stable domain category identification network, and obtaining the importance weight of the normalized source domain sample according to the importance weight of the source domain sample.
4. The adaptive intelligent fault diagnosis method based on the importance weighted domain antagonism according to claim 3, wherein the specific process of the step 2) is as follows: training a data set from a source domain
Figure FDA0003551957230000021
Training data set with target domain
Figure FDA0003551957230000022
Wherein the content of the first and second substances,
Figure FDA0003551957230000023
a sample point of a source domain;
Figure FDA0003551957230000024
a label for a source domain sample point; n issThe number of the source domain sample points is; y is the label space of the source domain sample;
Figure FDA0003551957230000025
a sample point that is a target domain;
Figure FDA0003551957230000026
a label for a target domain sample point, the target domain label being absent at the time of training; n istThe number of the target domain sample points is; y issubIs the label space of the target domain sample, YsubA subset of label space that is a source domain sample; obtaining the characteristics of the vibration signal through a characteristic extractor:
Figure FDA0003551957230000027
in formula 1-1:
Figure FDA0003551957230000028
is a function of the network output, where xs、xt、θh、Hs、HtRespectively representing a source domain sample input by a network, an input target domain sample, network parameters of a feature extractor, source domain features output after the source domain sample passes through the feature extractor, and target domain features output after the target domain sample passes through the feature extractor;
the source domain feature and the target domain feature pass through a domain classifier D0And obtaining the result of the domain category label prediction:
Figure FDA0003551957230000031
in formulas 1-2: sigmoid (·) function maps the input to between 0, 1; p0Predicting a probability vector for the domain class label of the normalized network sample; theta.theta.d0Representative Domain classifier D0The network parameter of (2).
5. The intelligent fault diagnosis method based on importance weighted domain impedance adaptation of claim 3, wherein in step 3), the domain class identification network loss function is as follows:
Figure FDA0003551957230000032
in formulas 1-3: j (D)0,Hs,Ht) A loss function for a domain class classifier;
Figure FDA0003551957230000033
function as a solutionA mean based on source domain samples;
Figure FDA0003551957230000034
to find a mean value based on samples of the target domain, P0Predicting probability vectors for domain class labels of normalized network samples, HsRepresenting the source domain features, H, output by the source domain samples after passing through the feature extractortAnd the target domain sample represents the target domain characteristics output after passing through the characteristic extractor.
6. The adaptive intelligent fault diagnosis method based on importance weighted domain antagonism according to claim 3, wherein in step 4), the importance weights of the source domain samples are as follows:
Figure FDA0003551957230000035
in formulae 1-4: dzSigmoid (. cndot.) function maps inputs to between 0, 1, H, which is the importance weight of the source domain samplessIs the source domain feature, theta, output after the source domain sample passes through the feature extractord0Representative Domain classifier D0The network parameter of (2);
the importance weights of the normalized source domain samples are:
Figure FDA0003551957230000036
in formulas 1-5: w (z) is the normalized source domain sample importance weight,
Figure FDA0003551957230000037
to find a mean value based on the samples of the target domain, DzIs the importance weight of the source domain samples.
7. The adaptive intelligent fault diagnosis method based on the importance weighted domain antagonism according to claim 1, wherein the specific process of step (3) is as follows:
the category label classifier is connected with the feature extractor through a full connection layer, and carries out pattern recognition classification on the discriminant features extracted by the feature extractor to obtain a fault type prediction result;
and secondly, calculating a failure mode recognition network loss function according to the failure type prediction result to form a failure mode recognition network.
8. The intelligent fault diagnosis method based on significance weighting domain antagonism self-adaptation as claimed in claim 1, wherein in step (5) adopting an antagonism game strategy to successively iterate the training domain classifier, the fault pattern recognition network and the significance weighting domain antagonism self-adaptation network until all network parameters are converged comprises the following specific steps:
1) in k iterations, the domain classifier D is minimized using the following equation0Loss function, updating domain classifier D by back propagation algorithm0Network parameter θ ofd0Until stable;
Figure FDA0003551957230000041
Figure FDA0003551957230000042
in the formulae 1 to 11,. alpha.1Is the learning rate;
Figure FDA0003551957230000043
as domain classifier D0A gradient of (a);
2) training the domain discriminant D against the loss function of the adaptive network by minimizing the importance weighting1Then training a class label classifier through a loss function of a minimum fault mode recognition network, and finally training a feature extractor through a loss function of the minimum label classifier and a maximum importance weighted impedance adaptive network loss function;
3) and repeating the iteration steps 1) and 2) until all network parameters are converged.
CN202010711903.6A 2020-07-22 2020-07-22 Intelligent fault diagnosis method based on importance weighted domain antagonism self-adaptation Active CN112101085B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010711903.6A CN112101085B (en) 2020-07-22 2020-07-22 Intelligent fault diagnosis method based on importance weighted domain antagonism self-adaptation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010711903.6A CN112101085B (en) 2020-07-22 2020-07-22 Intelligent fault diagnosis method based on importance weighted domain antagonism self-adaptation

Publications (2)

Publication Number Publication Date
CN112101085A CN112101085A (en) 2020-12-18
CN112101085B true CN112101085B (en) 2022-05-20

Family

ID=73749834

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010711903.6A Active CN112101085B (en) 2020-07-22 2020-07-22 Intelligent fault diagnosis method based on importance weighted domain antagonism self-adaptation

Country Status (1)

Country Link
CN (1) CN112101085B (en)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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
CN113159126B (en) * 2021-03-22 2022-08-12 河海大学 Industrial Internet of things equipment fault diagnosis method based on universal domain adaptation
CN113204280B (en) * 2021-05-08 2023-09-26 山东英信计算机技术有限公司 Method, system, equipment and medium for diagnosing power failure
CN113343855B (en) * 2021-06-09 2022-09-16 西南交通大学 Rolling bearing fault diagnosis system and method based on guide type sub-field self-adaption
CN113591625B (en) * 2021-07-16 2022-12-09 西安交通大学 Gearbox fault diagnosis method based on unbalanced sample importance weighting
CN114021285B (en) * 2021-11-17 2024-04-12 上海大学 Rotary machine fault diagnosis method based on mutual local countermeasure migration learning
CN114580284B (en) * 2022-03-07 2023-04-07 重庆大学 Method and system for diagnosing variable working condition fault of rotary machine
CN117272032B (en) * 2023-11-22 2024-02-13 青岛埃恩斯信息技术科技有限公司 Air compressor fault diagnosis method and device based on vibration detection

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017070657A1 (en) * 2015-10-23 2017-04-27 John Cameron Methods and systems for generating a state of being construct
CN110849627A (en) * 2019-11-27 2020-02-28 哈尔滨理工大学 Width migration learning network and rolling bearing fault diagnosis method based on same
CN111060318A (en) * 2020-01-09 2020-04-24 山东科技大学 Bearing fault diagnosis method based on deep countermeasure migration network
US10664550B1 (en) * 2017-02-07 2020-05-26 DemandJump, Inc. Technologies for web domain ranking with temporal effects within competitive network environments
CN111275024A (en) * 2020-03-22 2020-06-12 东北电力大学 Fault diagnosis method for rolling bearing of variable-working-condition fan driven by limited data containing noise label

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110619342B (en) * 2018-06-20 2023-02-03 鲁东大学 Rotary machine fault diagnosis method based on deep migration learning
CN110674849B (en) * 2019-09-02 2021-06-18 昆明理工大学 Cross-domain emotion classification method based on multi-source domain integrated migration

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017070657A1 (en) * 2015-10-23 2017-04-27 John Cameron Methods and systems for generating a state of being construct
US10664550B1 (en) * 2017-02-07 2020-05-26 DemandJump, Inc. Technologies for web domain ranking with temporal effects within competitive network environments
CN110849627A (en) * 2019-11-27 2020-02-28 哈尔滨理工大学 Width migration learning network and rolling bearing fault diagnosis method based on same
CN111060318A (en) * 2020-01-09 2020-04-24 山东科技大学 Bearing fault diagnosis method based on deep countermeasure migration network
CN111275024A (en) * 2020-03-22 2020-06-12 东北电力大学 Fault diagnosis method for rolling bearing of variable-working-condition fan driven by limited data containing noise label

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
An improved correlation-based anomaly detection approach for condition monitoring data of industrial equipment;Shisheng Zhong 等;《2016 IEEE International Conference on Prognostics and Health Management (ICPHM)》;20160815;全文 *
LiftingNet: A Novel Deep Learning Network With Layerwise Feature Learning From Noisy Mechanical Data for Fault Classification;Jun Pan 等;《IEEE Transactions on Industrial Electronics》;20180630;第65卷(第6期);全文 *
一种用于轴承故障诊断的迁移学习模型;张根保 等;《吉林大学学报（工学版）》;20190902;全文 *
域自适应学习研究进展;刘建伟 等;《自动化学报》;20140831;第40卷(第8期);全文 *

Also Published As

Publication number Publication date
CN112101085A (en) 2020-12-18

Similar Documents

Publication Publication Date Title
CN112101085B (en) Intelligent fault diagnosis method based on importance weighted domain antagonism self-adaptation
CN111898634B (en) Intelligent fault diagnosis method based on depth-to-reactance-domain self-adaption
CN110361176B (en) Intelligent fault diagnosis method based on multitask feature sharing neural network
CN109635677B (en) Compound fault diagnosis method and device based on multi-label classification convolutional neural network
CN112161784B (en) Mechanical fault diagnosis method based on multi-sensor information fusion migration network
CN111898095A (en) Deep migration learning intelligent fault diagnosis method and device, storage medium and equipment
CN111967294A (en) Unsupervised domain self-adaptive pedestrian re-identification method
Li et al. Multiscale dynamic fusion prototypical cluster network for fault diagnosis of planetary gearbox under few labeled samples
CN110657984A (en) Planetary gearbox fault diagnosis method based on reinforced capsule network
CN113469219B (en) Rotary machine fault diagnosis method under complex working condition based on element transfer learning
CN111612051A (en) Weak supervision target detection method based on graph convolution neural network
CN114004252A (en) Bearing fault diagnosis method, device and equipment
CN111753891A (en) Rolling bearing fault diagnosis method based on unsupervised feature learning
CN111595541A (en) Multi-dimensional structure damage identification method based on mass vibration transmissibility data convolutional neural network processing
CN116026593A (en) Cross-working-condition rolling bearing fault targeted migration diagnosis method and system
CN116451150A (en) Equipment fault diagnosis method based on semi-supervised small sample
CN116502135A (en) Bearing fault diagnosis method and system
CN115435892A (en) Intelligent fault diagnosis method for complex electromechanical equipment
CN113537152A (en) Flow field state fault detection method based on deep neural network
CN114429152A (en) Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption
CN115859077A (en) Multi-feature fusion motor small sample fault diagnosis method under variable working conditions
CN113076920B (en) Intelligent fault diagnosis method based on asymmetric domain confrontation self-adaptive model
CN114942140A (en) Rolling bearing fault diagnosis method based on multi-input parallel graph convolution neural network
CN110779722B (en) Rolling bearing fault diagnosis method based on encoder signal local weighting
CN116911377A (en) Radiation source individual identification method, equipment and medium based on transfer 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