CN112683532B - Cross-working condition countermeasure diagnostic method for bearing - Google Patents
Cross-working condition countermeasure diagnostic method for bearing Download PDFInfo
- Publication number
- CN112683532B CN112683532B CN202011342399.3A CN202011342399A CN112683532B CN 112683532 B CN112683532 B CN 112683532B CN 202011342399 A CN202011342399 A CN 202011342399A CN 112683532 B CN112683532 B CN 112683532B
- Authority
- CN
- China
- Prior art keywords
- discriminator
- sample
- parameters
- loss function
- expression
- 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
Links
Images
Abstract
The invention discloses a cross-working condition countermeasure diagnostic method of a bearing, which comprises the steps of collecting vibration data of the bearing in an operating state, dividing the vibration data to generate signal samples, and dividing the samples into a test set for testing and a training set for training; constructing a training module which comprises a feature extractor for extracting signal features, a classifier for classifying bearing faults and a discriminator for distinguishing the working conditions and faults of the features; the training module is trained on the basis of a training set, wherein the feature extraction parameters and the classifier parameters are updated according to a loss function by using a BP (back propagation) method; fixing the characteristic extractor parameters, and updating discriminator parameters by using the loss function; fixing discriminator parameters, and updating the characteristic extractor parameters by using a countermeasure loss function; and constructing a test module based on the updated classifier and the updated feature extractor, and inputting the test set and/or the target domain working condition sample into the test module to carry out fault diagnosis.
Description
Technical Field
The invention belongs to the technical field of intelligent diagnosis of bearing faults, and particularly relates to a cross-working condition countermeasure diagnosis method for a bearing.
Background
The rolling bearing is an important key part in the rotating machinery, and whether the inside bearing is safe or not is directly related to whether the machinery can normally run or not. The traditional intelligent diagnosis relies on expert to extract features and complex signal processing, and the traditional method cannot be widely applied due to high dependence on expert knowledge.
The traditional artificial intelligence fault diagnosis method based on deep learning depends on a new working condition field sample to participate in training in the field of variable working conditions (the rotating speed and the load of a training set are different from the rotating speed or the load of a testing set), retraining is needed when the new working conditions are generated in actual operation, and a large amount of time and resources are wasted.
In the wheel set bearing fault diagnosis method based on the equal-weight local feature sparse filter network of the Western-Ann traffic university, firstly, vibration signals of a vibration bearing in different health states are utilized to establish a fault diagnosis model based on the equal-weight local feature sparse filter network, then the sparse filter network is trained, fault features are automatically extracted from the vibration signals by utilizing the trained sparse filter network, and finally, a Softmax classifier is trained based on the extracted fault features to perform fault diagnosis on the bearing. The method relies on expert knowledge and utilizes a complex signal processing method to extract screening characteristics, and a classifier is adopted to classify faults based on the extracted characteristics. The method needs manual feature extraction, and the quality of feature extraction particularly depends on the experience of experts and a signal processing method, so that certain obstruction is caused to the application of the technology.
In a bearing fault diagnosis method based on a deep countermeasure migration network of Shandong science and technology university, firstly, vibration signals of a bearing under different working conditions (different rotating speeds and loads) are obtained, a frequency spectrum signal is obtained through Fourier transform, data with a label under a certain working condition is used as source domain data, and data without labels under other working conditions is used as target domain data; two generators are used as a feature extraction network of a source domain and a target domain, Softmax cross entropy is used as a fault classifier, a discriminator is used as a domain discrimination network, a gradient inversion layer is added for domain discrimination training, and the performance of the feature extraction network is improved. Deep learning has great advantages in data mining and feature extraction, and can solve the problem of dependence on expert experience, but when the working condition changes, a semi-supervised method based on transfer learning is required to be adopted to be trained by combining new working condition label-free fault data. This method has a fatal disadvantage that when a new working condition is met, the training needs to be carried out again by combining the working condition data of the source domain again to adapt to the new working condition. However, this wastes a lot of time and resources, and it is difficult to obtain new condition fault data in some cases.
In the actual mechanical operation, the method adopted by the method has lower reliability and accuracy and cannot be quickly adapted to new working conditions. Especially under the environment of complex and changeable working conditions, the method for extracting the features needs to be continuously adjusted and the model needs to be continuously retrained. These shortcomings have led to the use of intelligent diagnostic methods in the field of bearing fault diagnosis.
The above information disclosed in this background section is only for enhancement of understanding of the background of the invention and therefore it may contain information that does not form the prior art that is already known in this country to a person of ordinary skill in the art.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a cross-working condition countermeasure diagnosis method for a bearing, which does not need to manually extract features to reduce the dependence on expert knowledge, and overcomes the defect of the traditional artificial intelligence semi-supervised variable working condition training, namely, data of a target field without labels are needed for training, and retraining is needed once the target field is changed. The method utilizes improved conditions to resist a training method, trains invariant features in the model learning field, and gets rid of the interference of working condition information on the model. Thus, it may perform well in the target domain.
The invention aims to realize the purpose through the following technical scheme, and the cross-working condition countermeasure diagnosis method of the bearing comprises the following steps:
the method comprises the steps that in the first step, vibration data of a rolling bearing under a plurality of different working conditions are collected, wherein the working conditions are divided into a source domain working condition and a target domain working condition; dividing the vibration data under the working condition of the source domain to generate a first signal sample, and taking out the signal sample according to a preset proportion to be used as a training set for training; dividing the vibration data under the working condition of the target domain to generate a second signal sample which is then used as a test set sample for testing;
in the second step, a training module is constructed, wherein the training module comprises a feature extractor for extracting signal features, a classifier for classifying bearing faults and a discriminator for distinguishing the working conditions and faults of the features, the feature extractor is provided with feature extraction parameters, the classifier is provided with classifier parameters, and the discriminator is provided with discriminator parameters;
in a third step, training the training module based on a training set, comprising:
updating the feature extraction parameters and the classifier parameters according to a loss function by using a BP method;
fixing the characteristic extractor parameters, and updating discriminator parameters by using the loss function;
fixing discriminator parameters, and updating the characteristic extractor parameters by using a countermeasure loss function;
and in the fourth step, a test module is constructed based on the updated classifier and the updated feature extractor, and a test set sample is input into the test module for fault diagnosis.
In the method, in the first step, the vibration data is divided to generate signal samples with the length of 1024 data points, and a test set and a training set for testing are randomly divided according to a preset proportion, wherein each signal sample from the test set and each signal sample from the training set are subjected to normalization processing,,are samples of the signal generated by the vibration data,is thatThe average value of (a) of (b),is thatStandard deviation of (2).
In the method, in the second step, the training set is a source domain data set with known fault labels and working condition information,Wherein the source domainData setIs provided withA sample,Class andthe working conditions of the individual source areas are set,the working condition serial number of the source domain;is the ith sample, which is from the ithThe working condition of the source region belongs toClass, which is labeled asTo aDiscriminator labelThe expression is as follows: ,for theClassifier (A)) The output is:,
whereinWherein, generationWatch sampleBelong to the firstSecond of the individual domainProbability of individual class, the ideal output of the discriminator required by the feature extractor is:
representing samples at ideal output of discriminatorBelong to the firstSecond of the individual domainProbability of individual class.
In the method, in the third step, the cross entropy is taken as a loss function, the BP method is used for updating the feature extraction parameters and the classifier parameters according to the loss function, the loss function is the cross entropy, and the expression is。
In the method, in the third step, the characteristic extractor parameter is fixed, and the discriminator parameter is updated by using the loss function, wherein the loss function expression is
In the method, in the third step, the loss function is based on mean square error and has the expression。
Wherein
,Is to beObtained by performing transformation with respect to,The method only contains the information of the fault category but not the information of the working condition category, in order to distinguish the fault category, the cross entropy is adopted as an optimization function, and the expression is as follows:
the entropy regularization of a single sample is defined as:whereinIs the number of the categories that the user is in,is that the sample isAnd (3) class probability, constructing an entropy regular loss function as follows:
in the method, in the first step, the bearing is a rolling bearing.
The invention has the beneficial effects that:
1) the deep learning is utilized, and the characteristics of strong data pattern deep mining capability and remarkable knowledge learning capability are achieved. Features can be automatically extracted from the original signal without using complex signal processing methods to rely on manual extraction of fault features. Therefore, compared with the traditional fault diagnosis method, the method is more efficient and has wider application range.
2) The problems that resources are wasted due to the fact that traditional artificial intelligence needs to use new working condition label-free data to participate in training in the variable working condition problem, and the new working condition data cannot be obtained in the training process in some cases are solved. The invention only uses a plurality of source domain working conditions, learns the invariant characteristics of the domain by a conditional countermeasure method and promotes the model to be in
The above description is only an overview of the technical solutions of the present invention, and in order to make the technical means of the present invention more clearly apparent, and to make the implementation of the content of the description possible for those skilled in the art, and to make the above and other objects, features and advantages of the present invention more obvious, the following description is given by way of example of the specific embodiments of the present invention.
Drawings
Various other advantages and benefits of the present invention will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the invention. It is obvious that the drawings described below are only some embodiments of the invention, and that for a person skilled in the art, other drawings can be derived from them without inventive effort. Also, like parts are designated by like reference numerals throughout the drawings.
In the drawings:
FIG. 1 is a schematic view of the overall structure of the present invention;
FIG. 2 is a schematic diagram of the training phase of the present invention;
FIG. 3 is a schematic diagram of the output of the feature extractor expectation discriminator of the present invention;
FIG. 4 is a diagram of the network architecture of the present invention;
fig. 5(a) to 5(d) are feature cluster maps of two working condition samples of a source domain according to the case-source domain of the present invention, wherein fig. 5(a) is an original cluster map, fig. 5(b) is a cluster map of a conditional penalty function based on mean square error (mse) and (c) is a cluster map of a conditional penalty function based on cross entropy and entropy of classes, and fig. 5(d) is a cluster map of a conditional penalty function based on cross entropy and cross entropy of domains;
fig. 6(a) to 6(d) are feature cluster maps of two working condition samples of the case two-source domain of the present invention, wherein fig. 6(a) is an original cluster map, fig. 6(b) is a cluster map based on a mean square error conditional countermeasures loss function, fig. 6(c) is a cluster map based on cross entropy and entropy of a class, and fig. 6(d) is a cluster map based on cross entropy and cross entropy of a domain;
7(a) to 7(d) are feature cluster maps of two working condition samples of a case three-source domain of the invention, FIG. 7(a) an original cluster map, FIG. 7(b) a cluster map based on a mean square error conditional adversity loss function, FIG. 7(c) a cluster map of a conditional adversity loss function based on cross entropy and entropy of classes, and FIG. 7(d) a cluster map of a conditional adversity loss function based on cross entropy and cross entropy of domains;
8(a) to 8(d) are feature cluster maps of two working condition samples of a case four-source domain of the invention, FIG. 8(a) an original cluster map, FIG. 8(b) a cluster map based on a mean square error conditional adversity loss function, FIG. 8(c) a cluster map of a conditional adversity loss function based on cross entropy and entropy of classes, and FIG. 8(d) a cluster map of a conditional adversity loss function based on cross entropy and cross entropy of domains;
FIGS. 9(a) to 9(c) are graphs of the value of the objective domain loss function of case one of the present invention as a function of the number of iteration steps-The line of-' is the conventional CNN--' is the conditional countermeasure adopted in the present invention; i is a method 1 based on a mean square error condition to confront a loss function; II, constructing a conditional adversity loss function based on the cross entropy and entropy of the class; III is constructing a conditional adversity loss function based on the cross entropy of the class and the cross entropy of the domain, FIG. 9(a) is a cluster map of the conditional adversity loss function based on the mean square error, FIG. 9(b) is a cluster map of the conditional adversity loss function based on the cross entropy and the entropy of the class, and FIG. 9(c) is a cluster map of the conditional adversity loss function based on the cross entropy and the entropy of the class;
fig. 10(a) to 10(c) are graphs of the target domain loss function values according to the iteration step number in case two of the present invention, where the line of "-" is the conventional CNN and "-" is the conditional countermeasure adopted by the present invention; i is a method 1 based on a mean square error condition to confront a loss function; II, constructing a conditional adversity loss function based on the cross entropy and entropy of the class; III is constructing a conditional adversity loss function based on the cross entropy of the class and the cross entropy of the domain, FIG. 10(a) is a cluster map of the conditional adversity loss function based on the mean square error, FIG. 10(b) is a cluster map of the conditional adversity loss function based on the cross entropy and entropy of the class, and FIG. 10(c) is a cluster map of the conditional adversity loss function based on the cross entropy and entropy of the class;
fig. 11(a) to 11(c) are graphs of the target domain loss function values according to the iteration step number for case three of the present invention, where the line of "-" is the conventional CNN and "-" is the conditional countermeasure adopted by the present invention; i is a method 1 based on a mean square error condition to confront a loss function; II, constructing a conditional adversity loss function based on the cross entropy and entropy of the class; III is constructing a conditional adversary loss function based on the cross entropy of the class and the cross entropy of the domain, FIG. 11(a) is a cluster map of the conditional adversary loss function based on the mean square error, FIG. 11(b) is a cluster map of the conditional adversary loss function based on the cross entropy and entropy of the class, and FIG. 11(c) is a cluster map of the conditional adversary loss function based on the cross entropy and entropy of the class;
fig. 12(a) to 12(c) are graphs of the target domain loss function value according to the iteration step number for case four of the present invention, the line of "-" is the conventional CNN, and "-" is the conditional countermeasure adopted by the present invention; i is a method 1 based on a mean square error condition to confront a loss function; II, constructing a conditional adversity loss function based on the cross entropy and entropy of the class; III is constructing a conditional adversary loss function based on the cross entropy of the class and the cross entropy of the domain, fig. 12(a) is a cluster map of the conditional adversary loss function based on the mean square error, fig. 12(b) is a cluster map of the conditional adversary loss function based on the cross entropy and entropy of the class, and fig. 12(c) is a cluster map of the conditional adversary loss function based on the cross entropy and entropy of the class.
The invention is further explained below with reference to the figures and examples.
Detailed Description
Specific embodiments of the present invention will be described in more detail below with reference to fig. 1 to 12 (c). While specific embodiments of the invention are shown in the drawings, it should be understood that the invention may be embodied in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.
It should be noted that certain terms are used throughout the description and claims to refer to particular components. As one skilled in the art will appreciate, various names may be used to refer to a component. This specification and claims do not intend to distinguish between components that differ in name but not function. In the following description and in the claims, the terms "include" and "comprise" are used in an open-ended fashion, and thus should be interpreted to mean "include, but not limited to. The description which follows is a preferred embodiment of the invention, but is made for the purpose of illustrating the general principles of the invention and not for the purpose of limiting the scope of the invention. The scope of the present invention is defined by the appended claims.
For the purpose of facilitating understanding of the embodiments of the present invention, the following description will be made by taking specific embodiments as examples with reference to the accompanying drawings, and the drawings are not to be construed as limiting the embodiments of the present invention.
The cross-working condition countermeasure diagnostic method of the bearing comprises the following steps:
in the first step, vibration data of the rolling bearing under a plurality of different working conditions are collected, and the working conditions are divided into source domain working conditions and target domain working conditions. Dividing the vibration data of the source domain working condition to generate signal samples, and taking the samples out of the source domain samples according to a certain proportion to serve as a training set for training; and the target domain vibration data is divided to generate signal samples which are used as test set samples of the test.
In the second step, a training module is constructed, wherein the training module comprises a feature extractor for extracting signal features, a classifier for classifying bearing faults and a discriminator for distinguishing the working conditions and faults of the features, the feature extractor is provided with feature extraction parameters, the classifier is provided with classifier parameters, and the discriminator is provided with discriminator parameters;
in the third step, the training module trains based on a training set, wherein the feature extraction parameters and the classifier parameters are updated according to a loss function by using a BP method (back propagation algorithm); fixing the characteristic extractor parameters, and updating discriminator parameters by using the loss function; fixing discriminator parameters, and updating the characteristic extractor parameters by using a countermeasure loss function; the step ensures that the influence of the working condition on the classification characteristic is considered as little as possible when the model extracts the characteristic through the invariant characteristic of the learning field,
and in the fourth step, a test module is constructed based on the updated classifier and the updated feature extractor, and a test set and/or a target domain working condition sample is input into the test module for fault diagnosis. This may test the network for fault diagnosis capability under unknown conditions.
In a preferred embodiment of the method, in a first step, the different operating conditions are divided into a source domain operating condition and a target domain operating condition. And (3) dividing the vibration data in the source domain working condition to generate signal samples with the length of 1024, and taking out the source domain working condition samples according to a certain proportion to be used as a training set for training. And dividing the vibration data of the target domain working condition into 1024-length signal samples, and dividing the signal samples into a test set.
In a preferred embodiment of the method, in a first step, the vibration data is divided into 1024-length signal samples, and the test set and the training set for testing are randomly divided according to a predetermined ratio, wherein each signal sample from the test set and the training set is normalized,,are samples of the signal generated from the raw vibration data,is thatThe average value of (a) of (b),is thatStandard deviation of (2).
In a preferred embodiment of the method, in the second step, the training set is a source domain data set of known fault labels and operating condition information,Wherein the source domain data setIs provided withA sample,Class andthe working conditions of the individual source areas are set,the working condition serial number of the source domain;is the ith sample, which is from the ithThe working condition of the source region belongs toClass, which is labeled asTo aDiscriminator labelThe expression is as follows:to aClassifier (A)) The output is:
whereinWherein, in the step (A),representative sampleBelong to the firstSecond of the individual domainProbability of individual class, the ideal output of the discriminator required by the feature extractor is:
andin a similar manner to the above-described embodiments,representing samples at ideal output of discriminatorBelong to the firstSecond of the individual domainProbability of individual class.
In a preferred embodiment of the method, in the third step, the first step uses cross entropy as a loss function, and updates the feature extraction parameter and the classifier parameter according to the loss function by using a BP method, wherein the cross entropy expression of the small step loss function is。
In a preferred embodiment of the method, in the third step, the second substep fixes the feature extractor parameters and updates only the discriminator parameters by using a cross entropy loss function, the expression of which is。
In a preferred embodiment of the method, the 3 rd step in the third step is a total of three loss functions.
In a preferred embodiment of said method, in the third step, the penalty-combating function is based on the mean square error, expressed as。
In a preferred embodiment of said method, in a third step, the cross entropy is used as a loss function for the samplesDefining new variables
Wherein,Is to beObtained by performing transformation with respect to,The method only contains information of categories but not information of working conditions, in order to realize category discrimination, cross entropy is adopted as an optimization function, and the expression is as follows:
the entropy of a single sample is regular as:whereinIs the number of the categories that the user is in,is that the sample isProbability of class, entropy loss function is:
in a preferred embodiment of said method, in the third step, the cross-entropy loss function is
in a preferred embodiment of the method, in the first step, the bearing is a rolling bearing.
In order to further understand the invention, the invention develops a bearing diagnosis model training method, and a traction motor bearing fault diagnosis model can be constructed by using the method. After the model is trained and trained based on a plurality of source domain working condition data, the generalization capability of the model in a rotating speed and load combined test set which does not appear in a training set can be improved.
The invention is generally divided into the following three modules of data acquisition and processing, a training module and a testing module:
1) the data acquisition and processing module mainly refers to the acquisition of data, the division of a data set and the normalization of a sample
2) The training module comprises three parts (1) a characteristic extraction moduleWith the parameter of): extracting deep-layer characteristics of signals by utilizing a plurality of working condition time-domain vibration signals acquired by a bearing fault prefabrication experiment; 2. classifier module (With the parameter of): using 1) general-purpose based on extracted deep featuresAnd (4) completing fault classification of the motor bearing. 3. Discriminator module (With the parameter of): based on the extracted features in 1), the discriminator must distinguish which condition the generated features come from and what kind of fault.
3) The test module comprises two parts, namely 2) a feature extraction module) And a classifier module (). During testing, the load and rotating speed combined condition data which do not appear in training are collected and used for testing
The first step of the present invention is the acquisition and processing of data. And (3) mounting a sensor at a proper position, and acquiring vibration data of the bearing in an operating state. The collected data is divided into 1024-length samples, and the source domain samples are taken out according to the proportion of 80% to be used as a training set for training. Each sample from the test set and training set is normalized by mean-std.
The second and third embodiments of the present invention will be described in more detail with reference to the accompanying drawings.
In the present invention, as shown in fig. 1, the overall structural diagram of the method is generally described. The method is described in detail in the following order.
The training phase is divided into three steps in total as shown in fig. 2.
The first step is to use the cross entropy as a loss function and update the characteristic extraction parameters and classifier parameters of the model according to the cross entropy loss function by using a BP method.
And secondly, fixing the parameters of the feature extractor, and updating the parameters of the discriminator by using a cross entropy loss function, so that the discriminator can distinguish which working condition and which fault type the sample comes from as far as possible.
Thirdly, the parameters of the discriminator are fixed and the parameters of the feature extractor are updated by utilizing the combination of the three loss functions of the invention, thereby achieving the effect of confusing the discriminator. By using the loss function combination of the invention, the parameters of the discriminator are fixed, and the parameters of the feature extractor are updated, thus achieving the effect of confusing the discriminator. One core of the invention is that the obfuscator is not a traditional full obfuscation, but a conditional obfuscation, i.e. a conditional countermeasure. The present invention requires that the feature extractor can make the discriminator unable to distinguish which condition the sample is from but can distinguish which type of sample the sample is from. By using the conditional countermeasure method, the interference of the working conditions on the feature extractor can be reduced, so that the feature extractor has the capability of extracting the domain-independent features. By utilizing the capability, the network test set and the diagnosis accuracy rate under unknown working conditions can be improved.
And in the testing stage, only the feature extractor and the classifier obtained by the countertraining are used, and the target domain working condition sample which does not appear in the training is input for fault diagnosis.
In one aspect of the invention, the construction of the conditional penalty function:
variables and problem descriptions are defined first to facilitate the following description.
Representing source domain data sets, in commonA sample, in totalClass toAll working conditions are common() The characteristics of each sample, source domain, are the label and working condition information of the known faultTherein is provided withAnd (4) operating conditions of each source domain.Represents the target domain dataset, but does not know the category information and does not participate in the training.Is from the firstA working condition belongs toClass one toA specimen labeled asFor discriminatorsThe expression is as follows:
whereinEach element in the formula represents a type of fault under one operating condition, i.e.Represents the firstSecond of the individual domainA class of the one or more classes,the calculation method of (A) is shown as the formula. The core idea of the invention is to let the discriminator distinguish which class a sample comes from but not which condition it comes from. The task for a discriminator is very simple, namely to distinguish which class and which domain the sample comes from, and the target output of the discriminator is shown by the equation:
for the feature extractor, the goal is to make the discriminator distinguish which class the sample comes from but not which condition it comes from, as shown in fig. 3, the target output expression is:
fig. 3 illustrates a sample of the first category for the first operating mode to explain how the core objectives of the present invention are achieved. In fig. 3, the circle represents a first operating condition, and the square represents a second operating condition. The goal of the generator requires that the sum of the first element of the circle box and the first element of the box (the first element representing a first type of failure) be 1. By summing to 1 it is ensured that the discriminator can distinguish which class the sample belongs to. However, in order for the discriminator to be unable to distinguish which case it belongs to, it is required that the first element of the circle and the first element of the square be both 0.5 (i.e., that is, that of the square)For the present example). To implement the above-mentioned patent core idea, the present invention is shown in fig. 2 at each iteration of the training phase.
The method comprises the following specific steps:
Similar to the traditional deep learning network, the invention uses a cross-entropy loss function in the first step (As shown) as an optimization objective to update parameters,At the time of optimizationThe cross entropy loss function is reduced as much as possible by using the BP algorithm.
As shown in the equation, the goal of the discriminator is to distinguish which condition and which class the sample comes from, and its nature remains a classification problem. Therefore, similar to step 1, the invention selects the cross entropy as the loss function (as shown in the formula), and fixes the cross entropyUpdate。
This step is the core of the present invention, and the main purpose of this step is to realize that the discriminator can distinguish which class the sample comes from but cannot distinguish which condition it comes from, and the core goal is formula. To achieve this goal, the present invention solves this problem with three loss functions.
The method comprises the following steps: it is straightforward to let the discriminator output resemble its target as much as possible, treat this problem as a fitting problem, and let the discriminator output fit its target output, the most common method in fitting being mean error square (MSE). Thus, the first conditional penalty of step 3 is (based on the mean square error conditional penalty function):
the method 2 comprises the following steps: the core idea of the invention is that the discriminator can distinguish between classes, but cannotAnd distinguishing the working conditions. The idea can be broken down into two parts, the first part distinguishes the category, and the second part cannot distinguish the working condition. To achieve class discrimination, cross entropy can be used as a loss function, for ease of description, for samplesDefining new variables
Wherein,Is to beObtained by performing transformation with respect to,Only the category information is included and the condition information is not included. In order to realize category discrimination, cross entropy is adopted as an optimization function, and the expression is as follows:
in order to realize the condition that the working condition cannot be distinguished, the entropy regulation is adopted to solve the problem, and for a single sample, the entropy regulation is as follows:
whereinIs the number of categories, isThe sample is the firstThe probability of a class. When entcopy is minimum, the sample class is determined. When the amount of the drug is the maximum,i.e. from which class the sample cannot be distinguished. Based on the above expression, entropy regularization is employed to solve this problem. Thus the second method is to construct a loss function (a conditional countermeasures loss function based on cross-entropy and entropy of classes) as
Thus the second method is to construct a loss function (a conditional countermeasures loss function based on cross-entropy and entropy of classes) as
WhereinConditional probability () Distinguishing from which condition the sample came from can be achieved by minimizing the equation. However, the aim of the invention is to achieve a discriminator that does not distinguish from which condition the sample came from, and therefore maximizes the optimization. Therefore, the expression of the loss function (cross entropy + conventional impedance function: cross-entropy + conventional adaptive method) of method 3 is:
according to the above explanation, the steps of the present invention are as follows:
2 fixed parametersMinimization, updating parametersTo makeCan distinguish which working condition and which type of fault sample comes from
3 fixed parametersSelecting any one from the formulas as an optimization parameter, and updating the parameterThrough conditional countermeasure trainingDomain invariant models can be learned, i.e. letsThe categories can be distinguished, but the conditions cannot be distinguished.
4 utilization of,And performing fault diagnosis, and inputting samples of the source domain working condition and the target domain working condition which does not appear in training for fault diagnosis.
In one embodiment of the present invention, the substrate is,
through a traction motor bearing acceleration prefabrication fault test, taking non-driving-end 6311 bearing data as an example, a three-way acceleration sensor is used for measuring time domain vibration data of three directions of the traction motor bearing from X, Y, ZAnd collecting the collected data. For convenient operation, selectingThe acceleration vibration signal time domain of the shaft is taken as a sample. In the example there are 6 faults in total, as shown in table 1. The source domain operating conditions and the target domain operating conditions are shown in table 2.
TABLE 1 failure types
TABLE 2 working condition table
Each sample length is 1024, normalized by mean-std (formula):
whereinIs the original signal of the input samples and,is thatThe average value of (a) of (b),is thatStandard deviation of (2).
The overall architecture of the network is shown in fig. 4, where the feature extractor framework refers to the structure of resnet18 with the addition of bottleeck, whose parameters are shown in table 3. The classifier and discriminator parameters are shown in tables 4 and 5. The hyper-parameters and the computer configuration are shown in tables 6 and 7.
TABLE 3 feature extractor
(a) Resnet architecture
(b) Bottleneck architecture
TABLE 4 classifier (C) architecture
TABLE 5 discriminator network architecture
In one aspect of the invention, the construction of the penalty function is conditioned during the training phase.
The computer configurations and hyper-parameters used in the training of the present invention are shown in tables 6 and 7.
TABLE 6 computer hardware and configuration
TABLE 7 Superparameter
In the training phase, each iteration step performs the following three steps.
The method comprises the steps of preprocessing data, dividing a vibration signal of a source domain working condition 1, a source domain working condition 2 and a target domain working condition into samples with the length of 1024, and marking the source domain working condition with a classifier label and a discriminator label. Target domain condition labeled classifier label
1, inputting the samples of the working condition 1 and the working condition 2 of the source domain into a characteristic extractor, and obtaining the prediction result of the samples by a classifier
The classifier obtains the prediction result of the sample. And optimizing the characteristic extractor and the classifier parameters by using a BP algorithm and adopting an adam optimizer to minimize cross entropy (formula).
2, inputting the samples of the working condition 1 and the working condition 2 of the source domain into a characteristic extractor, and obtaining which category the samples come from by a discriminator. Using the BP algorithm, only discriminator parameters are optimized using the adam optimizer to minimize cross entropy (formula). (this step is performed twice per iteration)
And 3, inputting the source domain working condition 1 and working condition 2 samples into a characteristic extractor together, and obtaining which class the samples come from by a discriminator. By utilizing a BP algorithm, an adam optimizer optimizes parameters of the feature extractor by adopting one of a conditional counterattack loss function (formula) based on mean square error, a conditional counterattack loss function (formula) based on cross entropy and entropy of a class and a conditional counterattack loss function (formula) based on cross entropy and cross entropy of a domain, so that the discriminator can distinguish the class of a sample, but cannot distinguish which working condition the fault comes from. In the third step of training, the invention provides three loss functions, namely a countermeasure loss function based on a mean square error condition:
construction of conditional adversarial loss function based on class cross entropy and entropy
Conditional penalty function based on mean square error
And 4, after one iteration is finished, inputting the target domain working condition test sample into the feature extractor and the classifier, and testing the performance of the network. See fig. 9-12.
TABLE 8 accuracy Table for four cases
(a) Case 1
(b) Case 2
(c) Case 3
(d) Case 4
Although the embodiments of the present invention have been described above with reference to the accompanying drawings, the present invention is not limited to the above-described embodiments and application fields, and the above-described embodiments are illustrative, instructive, and not restrictive. Those skilled in the art, having the benefit of this disclosure, may effect numerous modifications thereto without departing from the scope of the invention as defined by the appended claims.
Claims (6)
1. A cross-condition countermeasure diagnostic method for a bearing, the method comprising the steps of:
the method comprises the steps that in the first step, vibration data of a rolling bearing under a plurality of different working conditions are collected, wherein the working conditions are divided into a source domain working condition and a target domain working condition; dividing the vibration data under the working condition of the source domain to generate a first signal sample, and taking out the signal sample according to 80% of proportion to be used as a training set for training; and dividing the vibration data under the working condition of the target domain to generate a second signal sample which is then used as a test set sample for testing, dividing the vibration data to generate signal samples with the length of 1024 data points, and randomly dividing the test set for testing and a training set for training according to a preset proportion, wherein each signal sample from the test set and the training set is subjected to normalization processing, and the expression is as follows:
are samples of the signal generated by the vibration data,is thatThe average value of (a) of (b),is thatThe standard deviation of (a) is determined,representing the signal obtained after the normalization process;
in the second step, a training module is constructed, which comprises a feature extractor for extracting signal features, and a classification for classifying bearing faultsThe system comprises a characteristic extractor and a discriminator for distinguishing the working condition and the fault of the characteristic, wherein the characteristic extractor is provided with characteristic extractor parameters, the classifier is provided with classifier parameters, the discriminator is provided with discriminator parameters, and the training set is a source domain data set of known fault labels and working condition informationThe expression is as follows:
wherein the content of the first and second substances,
is the ith sample, which is from the ithIndividual source region working condition and belongs toClass, which is labeled as;
wherein the content of the first and second substances,,is a sampleBelong to the firstThe probability of a class;
wherein the content of the first and second substances,
wherein the content of the first and second substances,sample when the representative feature extractor requires the ideal output of the discriminatorProbability of belonging to class j;
in a third step, training the training module based on a training set, comprising:
updating the feature extractor parameters and the classifier parameters according to a first loss function by using a BP method;
fixing the characteristic extractor parameters, and updating discriminator parameters by using a second loss function;
fixing discriminator parameters, and updating the characteristic extractor parameters by using a countermeasure loss function;
and in the fourth step, a test module is constructed based on the updated classifier and the updated feature extractor, and a test set sample is input into the test module for fault diagnosis.
2. The method according to claim 1, wherein in the third step, the feature extractor parameters and the classifier parameters are updated by a BP method according to a first loss function using cross entropy as a loss function, and the first loss functionFor cross entropy, the expression is:
3. The method according to claim 1, wherein in a third step, the feature extractor parameters are fixed and the discriminator parameters are updated with a second loss function, the second loss functionNumber ofThe expression is as follows:
5. The method of claim 1, wherein in the third step, for the sampleDefining new variablesThe expression is as follows:
wherein;
wherein the content of the first and second substances,is to beAnd (3) transforming to obtain: relative to,Only the information of the fault category is contained in the test data, but the information of the working condition category is not contained;
in order to distinguish fault categories, cross entropy is adopted to define an optimization functionThe expression is as follows:
wherein N isnIndicating the nth source domain condition-N in totalnA sample is obtained;
the entropy of a single sample is regularly defined as the following expression:
wherein the content of the first and second substances,is the number of the categories that the user is in,is that the sample isThe probability of a class;
To achieve the desired output of the discriminator, the penalty function isThe expression is as follows:
6. the method of claim 5, wherein,
wherein the content of the first and second substances,
representing the actual output of the discriminator, the sampleBelong to the firstThe probability of a class;
then represents the actual output of the discriminator, the sampleBelong to the firstThe probability of a class;
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011342399.3A CN112683532B (en) | 2020-11-25 | 2020-11-25 | Cross-working condition countermeasure diagnostic method for bearing |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011342399.3A CN112683532B (en) | 2020-11-25 | 2020-11-25 | Cross-working condition countermeasure diagnostic method for bearing |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112683532A CN112683532A (en) | 2021-04-20 |
CN112683532B true CN112683532B (en) | 2022-04-05 |
Family
ID=75446825
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011342399.3A Active CN112683532B (en) | 2020-11-25 | 2020-11-25 | Cross-working condition countermeasure diagnostic method for bearing |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112683532B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114429152A (en) * | 2021-12-31 | 2022-05-03 | 苏州大学 | Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption |
Family Cites Families (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8014957B2 (en) * | 2005-12-15 | 2011-09-06 | Fred Hutchinson Cancer Research Center | Genes associated with progression and response in chronic myeloid leukemia and uses thereof |
US10825219B2 (en) * | 2018-03-22 | 2020-11-03 | Northeastern University | Segmentation guided image generation with adversarial networks |
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 |
CN110414383A (en) * | 2019-07-11 | 2019-11-05 | 华中科技大学 | Convolutional neural networks based on Wasserstein distance fight transfer learning method and its application |
CN111504644B (en) * | 2020-04-01 | 2021-04-09 | 西安交通大学 | Bearing fault diagnosis method for adaptive anti-domain based on embedding distinguishability condition |
CN111626345A (en) * | 2020-05-15 | 2020-09-04 | 北京航空航天大学 | Multi-stage deep convolution transfer learning fault diagnosis method between different bearing devices |
CN111738455B (en) * | 2020-06-02 | 2021-05-11 | 山东大学 | Fault diagnosis method and system based on integration domain self-adaptation |
CN111898634B (en) * | 2020-06-22 | 2022-08-16 | 西安交通大学 | Intelligent fault diagnosis method based on depth-to-reactance-domain self-adaption |
-
2020
- 2020-11-25 CN CN202011342399.3A patent/CN112683532B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN112683532A (en) | 2021-04-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109580215B (en) | Wind power transmission system fault diagnosis method based on deep generation countermeasure network | |
CN109102005B (en) | Small sample deep learning method based on shallow model knowledge migration | |
Li et al. | Fault diagnosis of rotating machinery based on combination of deep belief network and one-dimensional convolutional neural network | |
Lei et al. | Fault diagnosis of wind turbine based on Long Short-term memory networks | |
CN109918752B (en) | Mechanical fault diagnosis method, equipment and medium based on migration convolution neural network | |
CN110502991B (en) | Internal combustion engine health monitoring method and system based on random convolutional neural network structure | |
CN112161784B (en) | Mechanical fault diagnosis method based on multi-sensor information fusion migration network | |
CN206504869U (en) | A kind of rolling bearing fault diagnosis device | |
CN110006649B (en) | Bearing fault diagnosis method based on improved ant lion algorithm and support vector machine | |
CN106874957A (en) | A kind of Fault Diagnosis of Roller Bearings | |
CN112883994B (en) | Rotating machinery variable working condition fault diagnosis method with balanced distribution adaptation | |
CN115270956A (en) | Cross-equipment incremental bearing fault diagnosis method based on continuous learning | |
CN111753891B (en) | Rolling bearing fault diagnosis method based on unsupervised feature learning | |
CN109946080B (en) | Mechanical equipment health state identification method based on embedded circulation network | |
CN114358124B (en) | New fault diagnosis method for rotary machinery based on deep countermeasure convolutional neural network | |
Joshuva et al. | A machine learning approach for condition monitoring of wind turbine blade using autoregressive moving average (ARMA) features through vibration signals: a comparative study | |
CN115358259A (en) | Self-learning-based unsupervised cross-working-condition bearing fault diagnosis method | |
CN116451150A (en) | Equipment fault diagnosis method based on semi-supervised small sample | |
Zhang et al. | MS-TCN: A multiscale temporal convolutional network for fault diagnosis in industrial processes | |
CN114429152A (en) | Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption | |
CN112683532B (en) | Cross-working condition countermeasure diagnostic method for bearing | |
Yuan et al. | Fault diagnosis approach for rotating machinery based on feature importance ranking and selection | |
Xu et al. | An optimal method based on HOG-SVM for fault detection | |
Wei et al. | WSAFormer-DFFN: A model for rotating machinery fault diagnosis using 1D window-based multi-head self-attention and deep feature fusion network | |
Du et al. | Fault diagnosis of plunger pump in truck crane based on relevance vector machine with particle swarm optimization algorithm |
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 |