CN116026569A - Mechanical equipment unsupervised migration intelligent fault diagnosis method for releasing source data - Google Patents
Mechanical equipment unsupervised migration intelligent fault diagnosis method for releasing source data Download PDFInfo
- Publication number
- CN116026569A CN116026569A CN202211643639.2A CN202211643639A CN116026569A CN 116026569 A CN116026569 A CN 116026569A CN 202211643639 A CN202211643639 A CN 202211643639A CN 116026569 A CN116026569 A CN 116026569A
- Authority
- CN
- China
- Prior art keywords
- migration
- model
- source
- data
- fault diagnosis
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Testing And Monitoring For Control Systems (AREA)
Abstract
The invention provides an unsupervised migration intelligent fault diagnosis method for mechanical equipment releasing source data, which comprises the following steps: s1, collecting equipment vibration monitoring data under different working conditions; s2, dividing the source field training sample set into a target field testing sample set; s3, data standardization is carried out; s4, establishing an unsupervised migration fault diagnosis model for releasing source data; s5, vibration monitoring data of the mechanical equipment are collected and input into an unsupervised migration fault diagnosis model, and mechanical health state identification is achieved. The method introduces the maximization of the nuclear norm to further encourage the legibility and the diversity of the prediction output, and interacts with a self-training mechanism to improve the migration performance of the source model to the target domain, so that the health status recognition of the unlabeled target sample under the condition of releasing the source data is realized. The invention considers the cost of expensive data storage and transmission and privacy protection in the actual industrial scene, so that the unsupervised migration has the cross-domain diagnosis capability and simultaneously releases the requirement on source domain data in the model migration process.
Description
Technical Field
The application relates to the technical field of mechanical fault diagnosis, in particular to an unsupervised migration intelligent fault diagnosis method for mechanical equipment releasing source data.
Background
Rotary machines are becoming increasingly more of a major support in industrial manufacturing, with increased concerns regarding safety and reliability. However, due to long-term operation, degradation and damage to critical components inevitably occurs. Therefore, fault diagnosis techniques have been widely studied in the academia and are becoming an indispensable technique in the modern industry.
With the vigorous development of the Internet, the Internet of things and industrial big data, the deep learning model lifts the tide of the intelligent diagnosis method, and attractive achievement is achieved. The technology breaks through the conventional thought of a diagnosis method based on a model or a signal analysis, and treats the diagnosis problem as an end-to-end health mode identification problem. While the performance of these models is attractive, the premise is to satisfy independent co-distribution assumptions of the data. There is always an inevitable difference in the distribution of data characteristics of the training and testing domains due to the variation of the operating state and the uncertainty of the monitoring environment. Thus, a serious performance degradation will be observed, since the above-mentioned preconditions are not satisfied. In other words, it is difficult to apply the trained model directly to new tasks with domain distribution differences.
In response to the above problems, diagnostic techniques based on unsupervised migration have emerged in recent years. The method aims at reducing domain bias and aligning marked source data and unmarked target data in a common representation space so that models trained on source domain can be generalized to target domain. One common strategy is to measure and reduce the differences in feature distribution with different metrics, another popular model is to use generative countermeasure network concepts to learn inter-domain migratable knowledge.
However, while the above approaches have met with some success, they must work under stringent conditions that keep the source data available throughout the training process. The following two conditional limitations make unsupervised migration diagnostics unsuitable in many real-world scenarios. First, one key reason is that the source data is not always accessible due to privacy protection or intellectual property issues. Another key reason is long-term monitoring, particularly high frequency sampling, which means huge storage and loading costs. Therefore, it is necessary to study migration models that do not require source data, and this approach can accomplish the objective domain diagnostic tasks without accessing the source data.
Disclosure of Invention
In order to overcome the above-mentioned deficiencies of the prior art, it is an object of the present invention to provide a Source-free transfer diagnosis (SFTD) Source-data-releasing migration diagnostic network for cross-domain fault diagnosis of a mechanical device, which is capable of performing fault diagnosis without relying on Source data in combination with the migration diagnostic network. SFTD includes two phases, a source model generation phase and a source model migration phase, respectively. In the generation stage of the source model, the source field data with the labels is taken as input, and the smooth cross entropy of the labels is designed as a loss function, so that the source field diagnosis information is fully mined, and meanwhile, the problems of over fitting and weak adaptation of the traditional cross entropy are restrained; in the source model migration stage, the source model migration is realized by only using unlabeled target field data. The invention firstly provides a new self-training mechanism to fully explore the application of the label-free data, and simultaneously considers the cost of expensive data storage and transmission and privacy protection in the actual industrial scene, so that the unsupervised migration has the cross-domain diagnosis capability and simultaneously reduces the requirement on the source field data in the model migration process.
Specifically, in order to achieve the above object, the solution adopted by the present invention is:
an intelligent fault diagnosis method for unsupervised migration of mechanical equipment releasing source data comprises the following steps:
s1, aiming at rotary machines with various health states, collecting equipment vibration monitoring data under different working conditions, and constructing a sample data space and a corresponding label thereof through the monitoring data;
s2, dividing the equipment vibration monitoring data with different working conditions in the step S1 into a source field training sample set and a target field testing sample set;
s3, carrying out data standardization on sample data of the source field training sample set and the target field testing sample set obtained in the step S2 to obtain a standardized source field training sample set and a standardized target field testing sample set;
s4, establishing an unsupervised migration fault diagnosis model for releasing source data, wherein the unsupervised migration fault diagnosis model comprises the following substeps:
s41, constructing an unsupervised migration diagnosis model, wherein the unsupervised migration diagnosis model f comprises a feature encoding module
And classification module->
Wherein d is the feature dimension after encoding, and K is the total category number; the feature coding module comprises a multi-layer convolution module and a full connection layer, and the classification module comprises a full connection layer and a nonlinear activation layer;
s42, generating a source model, generating the source model by using a supervised learning paradigm, and introducing a label smooth cross entropy L s ls As a loss function for source model training;
s43, generating a pseudo tag by using a class prototype, wherein the method specifically comprises the following substeps:
s431, calculating a class prototype of the target field by using the following formula:
in sigma k (. Cndot.) represents the kth element, X, in the softmax output t Representing a target domain space;
representing a learned target model in a previous iteration, the target model being obtained by initializing a source model;
s432, obtaining an initial pseudo tag through distance measurement:
in the method, in the process of the invention,
is an initial pseudo tag;
s433, updating the class prototype by using the initial pseudo tag, optimizing the initial pseudo tag, and obtaining the optimized pseudo tag
The specific process is as follows:
wherein, when the parameter is true, the indicator function ζ (·) is equal to 1;
is an initial pseudo tag; x is x t Representing a target domain sample; k represents the predicted class of the target domain sample; />
A classifier in the target model;
s434, introducing normalized symmetrical cross entropy, and reducing noise label interference of pseudo labels:
the formula of normalized symmetric cross entropy is:
the right side represents normalized cross entropy and inverse representation thereof, and the normalized cross entropy is specifically:
where p is the predictive probability, i.e., p=σ (f (x));
representing a pseudo tag;
s435, forming the following self-training target loss function based on the normalized symmetrical cross entropy in the step S434:
s44, utilizing the F norm maximization of the output matrix to provide reliable target output prediction discrimination capability, and adding a kernel norm maximization loss function;
s45, performing model migration training on the unsupervised migration fault diagnosis model by using the standardized target field training sample set obtained in the step S3, fixing network parameters of a classification module, optimizing a model feature extraction module by using a random gradient descent method in combination with the self-training target loss function in the step S43 and the kernel norm maximization loss function in the step S44, iterating until the model converges, and completing source model migration;
s5, vibration monitoring data of the mechanical equipment are collected and input into an unsupervised migration fault diagnosis model, and mechanical health state identification is achieved.
Preferably, the loss function in step S42 is:
q' k ＝(1-α)q k +α(1/K)
wherein x is s For source field samples, y s Corresponding tags for source domain samples; σ (·) represents the softmax function,
a kth element in the softmax output representing the K-dimensional non-normalized probability vector; if and only if y s When=k, q k 1, otherwise, q k Is 0; alpha is a smoothing parameter, and the default value is 0.1; f (f) s :X s →Y s Representing a source model, f t :X t →Y t For the learned objective function, by reasoning +.>
Preferably, the output matrix F norm maximization expression in step S44 is:
wherein P= [ P ] ij ] B×K A softmax output matrix representing a batch process of size B.
Preferably, the kernel norm maximization loss function in step S44 is:
L nnm ＝-||P|| * /B
in the formula, a convex hull of matrix P rank is its nuclear norm P * 。
Preferably, the feature encoding module in step S41 includes four convolution layers, a Batch-normalization layer, a nonlinear activation function, a pooling layer, a full-connection layer, and a Dropout layer; the classification module comprises a full connection layer and a nonlinear activation layer.
Preferably, source field D is assumed in step S2 s Comprising n s Each having a markSign sample
Wherein the method comprises the steps of
Target area D t Comprising n t Label-free sample->
Wherein->
Preferably, the formula for data normalization in step S3 is specifically:
wherein: x is x i Is the i-th data sample; mu (mu) i Is x i Average value of (2); sigma (sigma) i Is x i Standard deviation of (2); x is x i (j) Is x i Is the j-th element of (c).
Preferably, the mechanical device in step S1 comprises a multistage acceleration transmission bearing failure simulation test bench.
Preferably, the sampling frequency of the vibration data of the multistage acceleration transmission is 10kHz.
Compared with the prior art, the invention has the beneficial effects that:
1) According to the invention, an original time domain vibration signal is used as data input, and the relevant information of the health state decision of the mechanical key component is adaptively extracted through the diagnosis network model, so that an end-to-end mechanical fault diagnosis model is established, the requirements on human resources, expert experience and the like in the diagnosis decision process are reduced, and the efficiency of automatic diagnosis and the accuracy of a diagnosis result are ensured;
2) The invention realizes the complex mechanical health state identification of the target field sample under the condition of no label and different field data distribution, completes the migration of the labeled source field data to the unlabeled target field data, makes up the defect of the conventional deep learning diagnosis model, and improves the fault identification precision;
3) The invention considers the availability of most of the deep migration fault diagnosis methods depending on source data, and aims at the technical defect that the traditional method cannot be realized in a plurality of actual industrial scenes due to the cost of expensive data storage, transmission and privacy protection.
4) The invention utilizes pseudo tag learning strategy oriented to class prototypes and standardized symmetrical cross entropy series deployment to promote robust training; and meanwhile, the nuclear norm maximization is introduced to further improve the legibility and diversity of the prediction output, and the nuclear norm maximization interacts with a self-training mechanism to improve the migration performance of the source model to the target field, so that the health state identification of the unlabeled target sample under the condition of no source data is realized.
Drawings
Fig. 1 is a flowchart of an intelligent fault diagnosis method for mechanical equipment based on SFTD according to an embodiment of the present invention;
FIG. 2 is a flowchart of a training step of an unsupervised migration fault diagnosis model for releasing source data according to an embodiment of the present invention;
FIG. 3 is a network structure and parameter setting of a feature encoding module of the SFTD in an embodiment of the invention;
fig. 4 is a network structure and parameter setting of a classification module of SFTD according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of a multi-stage acceleration transmission experimental apparatus in the present embodiment of the invention;
the partial reference numerals in the figures are as follows:
1-a drive motor; a 2-coupling; 3-planetary gear box; 4-fixed-axis gear box; 5-tachometer; 6-bearing seats; 7-a vibration sensor; 8-magnetic powder brake.
Detailed Description
Hereinafter, embodiments of the present invention will be described with reference to the drawings.
The invention provides an intelligent fault diagnosis method for unsupervised migration of mechanical equipment for releasing source data, which is shown in fig. 1-4 and specifically comprises the following steps:
s1, aiming at rotary machines with various health states, collecting equipment vibration monitoring data under different working conditions, so that the collected data contain different health state information, and then constructing a sample data space and a corresponding label thereof through the collected monitoring data.
S2, dividing the equipment monitoring data under different working conditions in the step S1 into two groups, namely a source field training sample set and a target field testing sample set. Source field training sample set D s Comprising n s Each labeled sample
Wherein the method comprises the steps of
Target area test sample set D t Comprising n t Label-free sample->
Wherein->
The source field training sample set is a labeled data set, the target field testing sample set is a non-labeled data set, and the data sets can be cross-validated under different working conditions.
S3, carrying out data standardization on sample data of the source field training sample set and the target field testing sample set obtained in the step S2 to obtain a standardized source field training sample set and a standardized target field testing sample set, wherein a specific standardization formula is as follows:
wherein x is i Is the i-th data sample; mu (mu) i Is x i Average value of (2); sigma (sigma) i Is x i Standard deviation of (2); x is x i (j) Represents x i Is the j-th element of (c).
S4, constructing an unsupervised migration fault diagnosis model (Source-freetransfer diagnosis, SFTD) for releasing Source data, primarily training an SFTD feature encoding module formed by a convolutional neural network by using a standardized Source field training sample set, and an SFTD classifying module formed by a full-connection layer and a nonlinear activation layer, performing model migration training on the SFTD feature encoding module only based on a standardized target field test sample set, and completing migration from Source field knowledge to target field knowledge, wherein the method comprises the following specific steps:
s41, constructing an SFTD framework, wherein the framework adopts a One-dimensional convolutional neural network (One-dimensional Convolutional NeuralNetwork, 1D-CNN) to construct an integral network framework, and comprises a feature coding module and a classification module, wherein the feature coding module is used for carrying out feature learning on source domain data and target domain data so as to finish the migration of knowledge in different domains, the two domains comprise a source domain and a target domain, and comprises four convolutional layers, a Batch-standardized layer, a nonlinear activation function, a pooling layer, a full-connection layer, a Batch-standardized layer and a Dropout layer. Fig. 3 shows a network structure and parameter settings of the feature encoding module of SFTD, where: "Maps" represents the number of convolution kernels for the current convolution layer; "Pad" means zero-padding operation, aimed at keeping the feature dimensions unchanged before and after the convolution operation; "BN" means batch normalization; "ReLU" means a linear rectifying unit activation function; "Maxpool:2×1" means a maximum pooling operation with a pooling area size of 2×1; "Stride 2" means that the slip step of the pooling operation is 2; "Linear" represents a fully attached layer; "Drop" means a Drop out operation. The classification module comprises a full connection layer and a nonlinear activation layer, and the network structure and parameter setting of the classification module of the SFTD are shown in fig. 4.
S42, generating a source model: in the source model generation stage, in order to enable the network model to accurately classify the source domain training sample set, a standard supervised learning paradigm may be used to generate the source model based on given source data and its corresponding real labels. Meanwhile, in order to inhibit the problems of overfitting and weak adaptation existing in the traditional cross entropy, a label smooth cross entropy L is introduced s ls As a loss function for source model training, the expression of the loss function is as follows:
wherein x is s For source field samples, y s Corresponding tags for source domain samples; σ (·) represents the softmax function,
a kth element in the softmax output representing the K-dimensional non-normalized probability vector; if and only if y s When=k, q k 1, otherwise, q k Is 0; alpha is a smoothing parameter and the default value is 0.1.f (f) s :X s →Y s Representing a source model, f t :X t →Y t For the learned objective function, by reasoning +.>
f is the proposed diagnostic model. And (3) performing preliminary training on the diagnosis model by using the standardized source field training sample set obtained in the step (S3), optimizing the model by using a random gradient descent method, iterating until the model converges, and completing the generation of the source model in the first stage.
S43, after the source model is generated, the source data is inaccessible. During the source model migration process, the classification module is fixed and only the feature encoding module is updated. Because of the lack of source data, the present invention proposes a simple and competitive passive domain data model migration strategy using a self-training mechanism if a generalized solution using explicit matching of source domain and target domain distributions is not feasible. Unlike conventional strategies, first, computational class prototypes are utilized, enabling more reliable pseudo tags to be generated. Class prototypes for the target domain are calculated as follows:
in sigma k (. Cndot.) represents the kth element, X, in the softmax output t Representing a target domain space;
representing the target model that has been learned in the last iteration, the model being initialized by the source model.
The pseudo tag is then obtained by distance measurement:
in the method, in the process of the invention,
is an initial pseudo tag.
After this, the class prototype is updated again using the initial label, and the pseudo-label is optimized with the updated class prototype to calculate a more accurate pseudo-label
The process is as follows:
wherein when the parameter is true, the indicator function ζ (·) is equal to 1;
is an initial pseudo tag; x is x t Representing a target domain sample; k represents the predicted class of the target domain sample; />
Is a classifier in the object model.
Neither pseudo tag generation algorithm can completely avoid the interference of noise tags, which would lead to under-learning and over-fitting problems of standard cross entropy. To improve noise tolerance, the method of the present invention introduces a variant, normalized Symmetric Cross Entropy (NSCE), for robust self-training instead of the original CE. The formula for NSCE is:
where the right side represents the Normalized Cross Entropy (NCE) and its inverse. The Normalized Cross Entropy (NCE) is calculated as follows:
where p is the predictive probability, i.e., p=σ (f (x));
representing a pseudo tag.
Then, based on the normalized cross entropy, a robust self-training objective function is formed as follows:
s44, in order to enhance the resolving power of the model to different categories, the output matrix F norm maximization is utilized to provide the reliable prediction discrimination capability of the target output, and the output matrix F norm maximization is expressed as:
wherein P= [ P ] ij ] B×K A softmax output matrix representing a batch process of size B.
Given that incorrect pseudo tags can compromise the diversity of predictions, thereby impeding migration of the model, constraints are established that maximize matrix rank (P). Directly computing the matrix rank is an NP-hard and non-convex problem, however, if I P I F And (2) the convex hull of the matrix P rank is equal to or less than 1, and the kernel norm P * Maximizing the kernel norms thus can guarantee both discriminatory and diversity of the target output. A kernel-norm maximization loss function is deployed in the model in the form of:
L nnm ＝-||P|| * /B (10)
s45, performing model migration training on the model by using the standardized training sample set of the target field obtained in the step S3, fixing network parameters of the classification module, integrating the loss functions in the step S43 and the step S44, optimizing the model feature extraction module by using a random gradient descent method, and iterating until the model converges to finish source model migration in the second stage.
S5, vibration monitoring data of the mechanical equipment are collected and input into an unsupervised migration network for releasing source data, and then mechanical health state identification is achieved.
DETAILED DESCRIPTION OF EMBODIMENT (S) OF INVENTION
The working principle of the invention is further described below in connection with specific embodiments. The mechanical device in this embodiment is a rolling bearing. The sampling frequency for collecting vibration data of the rolling bearing is 10kHz.
In the example verification process, a diagnosis Method (Method 1) based on source field data only, a migration diagnosis Method (Method 2) based on maximum mean difference, a migration diagnosis Method (Method 3) based on related rearrangement, an anti-migration diagnosis Method (Method 4) based on a domain discriminator, and a source hypothesis migration diagnosis Method (Method 5) are selected and compared with the Method of the invention to verify the effectiveness of the Method of the invention. Where Method1 is a typical representation of a deep learning diagnostic model that is trained only from labeled source domain data and then applied directly to the target domain. In order to ensure the effectiveness of comparison, the network framework of the Method1 is connected in series with the feature coding module and the classification module of the SFTD provided by the invention. Methods 2 and 3 are typical representatives of two migration methods based on distance measures. The Method2 adopts the maximum mean value difference constraint distribution difference of multi-Gaussian kernels; and the Method3 is to utilize the second-order correlation statistic difference to realize the distribution rearrangement of the features. The Method4 is an anti-migration diagnosis Method based on a domain discriminator, the feature coding module and the classification module are the same as the SFTD of the invention, a domain discriminating module consisting of three full-connection layers is added on the basis, and the feature coding module output is taken as input to perform an anti-game with the feature coding module. Method5 is an excellent unsupervised migration Method of releasing source data, and has an impressive result in the field of image classification, so that it is applied to the field of migration diagnosis as a comparison Method of the present invention.
The invention is applied to a bearing fault diagnosis embodiment of a multistage acceleration transmission for verification.
In particular to bearing fault diagnosis of a multistage acceleration transmission device, the invention adopts bearing experimental data in a multistage transmission system to carry out embodiment verification, a schematic diagram of a laboratory table is shown in fig. 5, and the laboratory table comprises a driving motor 1, a coupler 2, a planetary gear box 3, a fixed shaft gear box 4, a tachometer 5, a bearing pedestal 6, a vibration sensor 7 and a magnetic powder brake 8. The power is provided by the driving motor 1, the rising speed of the planetary gear box 3 and the fixed shaft gear box 4 reaches the magnetic powder brake 8 through the bearing test base, and the magnetic powder brake 8 can apply different loads to the system. The present invention uses vibration data with a sampling frequency of 10kHz for analysis. The health state of the rolling bearing includes 7 kinds of normal state, inner ring crack failure, inner ring wear failure, inner ring crack failure, outer ring wear failure and cage crack failure. In the experimental process, each bearing operates under four different working conditions (the rotating speed is about 1200rpm, the loads of the four working conditions are respectively 0.7Nm, 1.2Nm, 2Nm and 5 Nm), an acceleration sensor arranged at the top of a bearing seat is adopted for data acquisition, and the sampling frequency is 10kHz. For each bearing, 1000 samples were collected in each operating state, each sample being 1024 in size. To simulate a real noisy industrial environment, noise is mixed into the measured data to reduce the signal-to-noise ratio of the data. Four fields, labeled A1, A2, A3 and A4, are constructed according to the working condition differences, with different data distributions between the four fields. Thus, 12 migration diagnostic tasks are arranged in four areas, as shown in table 1. Wherein A1→A2 represents using the data sample obtained under the working condition A1 as the source domain, using the data collected under the working condition A2 as the target domain, and the target domain data does not contain the tag information. In the training stage of the model, 70% of samples are randomly selected from the two fields to perform model training, and after the training is completed, the remaining samples in the target field are subjected to test analysis.
Table 1 multistage driveline bearing migration diagnostic tasks
The method of the present invention and various comparison methods were performed based on the data set, and the obtained diagnostic results are shown in table 2. In the present table, "SF" indicates passive setting, "a1→a2" indicates that data of A1 is a source field and A2 is a target field. The best results for each task are shown in bold for visual display. Looking at these results, we can conclude that: (1) Under the condition that no source data is used for model migration, the average accuracy of 96.89% is still achieved in 12 tasks, and the method is superior to all existing comparison methods. (2) Of course, method1 performs poorly in all diagnostic tasks, as it cannot do anything about domain drift. In contrast, the four migration-based methods can solve the data distribution differences and obtain better diagnostic results. (3) By evaluating the existing similar Method for releasing source data, the Method disclosed by the invention can be found to obtain remarkable performance improvement and has better model migration capability. Although the diagnostic effect of Method5 is not as good as that of the anti-migration diagnostic Method (Method 4), it is clearly superior to methods 2 and 3. This also shows that the result of Method5 is quite competitive, since it does not require source field data during migration, but the result is still far from that of this patent, and therefore the diagnostic effect of the Method of this patent is excellent.
Table 2 classification accuracy and standard deviation (%)
The mechanical fault recognition rate of the invention is superior to that of other several existing intelligent diagnosis methods based on a deep learning model by comparing with the diagnosis Method (Method 1) based on source field data only, the migration diagnosis Method (Method 2) based on the maximum mean difference, the migration diagnosis Method (Method 3) based on the relevant rearrangement, the anti-migration diagnosis Method (Method 4) based on the domain discriminator and the source hypothesis migration diagnosis Method (Method 5) recognition results.
The passive migration diagnosis method for the rotary machinery provided by the invention is different from the traditional unsupervised migration diagnosis method, and the SFTD does not need source data in the migration process, so that more practical diagnosis scenes only available by a source model can be met. In SFTD, we introduce pseudo tag generation algorithm and normalized system cross entropy loss to achieve more robust self-training, while deploying target output optimization oriented to kernel norm maximization to improve the legibility and diversity of predictions. Finally, on the premise of releasing source field data, the cross-domain complex mechanical health state identification under the conditions of multiple operation conditions, no-label target samples and different field distribution differences is realized. The invention makes up the defect of the conventional depth migration diagnosis model, reduces the requirement on source field data in the diagnosis model migration process, improves the accuracy of fault identification, ensures the accurate fault identification, and simultaneously reduces the cost of data storage and transmission and privacy protection in the actual engineering scene.
The above examples are only illustrative of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention, and various modifications and improvements made by those skilled in the art to the technical solution of the present invention should fall within the scope of protection defined by the claims of the present invention without departing from the spirit of the present invention.
Claims (9)
1. An intelligent fault diagnosis method for unsupervised migration of mechanical equipment for releasing source data is characterized in that: which comprises the following steps:
s1, aiming at rotary machines with various health states, collecting equipment vibration monitoring data under different working conditions, and constructing a sample data space and a corresponding label thereof through the monitoring data;
s2, dividing the equipment vibration monitoring data with different working conditions in the step S1 into a source field training sample set and a target field testing sample set;
s3, carrying out data standardization on sample data of the source field training sample set and the target field testing sample set obtained in the step S2 to obtain a standardized source field training sample set and a standardized target field testing sample set;
s4, establishing an unsupervised migration fault diagnosis model for releasing source data, wherein the unsupervised migration fault diagnosis model comprises the following substeps:
s41, an unsupervised migration diagnosis model is built, wherein the unsupervised migration diagnosis model f comprises a feature encoding module g:
and a classification module h:>
wherein d is the feature dimension after encoding, and K is the total category number; the feature coding module comprises a multi-layer convolution module and a full connection layer, and the classification module comprises a full connection layer and a nonlinear activation layer;
s42, generating a source model, generating the source model by using a supervised learning paradigm, and introducing a label smooth cross entropy
As a loss function for source model training;
s43, generating a pseudo tag by using a class prototype, wherein the method specifically comprises the following substeps:
s431, calculating a class prototype of the target field by using the following formula:
in sigma k (. Cndot.) represents the kth element, X, in the softmax output t Representing target domain space；
Representing a learned target model in a previous iteration, the target model being obtained by initializing a source model;
s432, obtaining an initial pseudo tag through distance measurement:
in the method, in the process of the invention,
is an initial pseudo tag;
s433, updating the class prototype by using the initial pseudo tag, optimizing the initial pseudo tag, and obtaining the optimized pseudo tag
The specific process is as follows:
wherein, when the parameter is true, the indicator function ζ (·) is equal to 1;
is an initial pseudo tag; x is x t Representing a target domain sample; k represents the predicted class of the target domain sample; />
A classifier in the target model;
s434, introducing normalized symmetrical cross entropy, and reducing noise label interference of pseudo labels:
the formula of normalized symmetric cross entropy is:
the right side represents normalized cross entropy and inverse representation thereof, and the normalized cross entropy is specifically:
where p is the predictive probability, i.e., p=σ (f (x));
representing a pseudo tag;
s435, forming the following self-training target loss function based on the normalized symmetrical cross entropy in the step S434:
s44, utilizing the F norm maximization of the output matrix to provide reliable target output prediction discrimination capability, and adding a kernel norm maximization loss function;
s45, performing model migration training on the unsupervised migration fault diagnosis model by using the standardized target field training sample set obtained in the step S3, fixing network parameters of a classification module, optimizing a model feature extraction module by using a random gradient descent method in combination with the self-training target loss function in the step S43 and the kernel norm maximization loss function in the step S44, iterating until the model converges, and completing source model migration;
s5, vibration monitoring data of the mechanical equipment are collected and input into an unsupervised migration fault diagnosis model, and mechanical health state identification is achieved.
2. The method for intelligent fault diagnosis of unsupervised migration of mechanical equipment releasing source data according to claim 1, wherein the method comprises the following steps: the loss function in step S42 is:
q' k ＝(1-α)q k +α(1/K)
wherein x is s For source field samples, y s Corresponding tags for source domain samples; σ (·) represents the softmax function,
a kth element in the softmax output representing the K-dimensional non-normalized probability vector; if and only if y s When=k, q k 1, otherwise, q k Is 0; alpha is a smoothing parameter, and the default value is 0.1; f (f) s :X s →Y s Representing a source model, f t :X t →Y t For the learned objective function, by reasoning +.>
3. The method for intelligent fault diagnosis of unsupervised migration of mechanical equipment releasing source data according to claim 1, wherein the method comprises the following steps: the output matrix F-norm maximization expression in step S44 is:
wherein P= [ P ] ij ] B×K A softmax output matrix representing a batch process of size B.
4. The method for intelligent fault diagnosis of unsupervised migration of mechanical equipment releasing source data according to claim 3, wherein the method comprises the following steps: the kernel-norm maximization loss function in step S44 is:
L nnm ＝-||P|| * /B
in the formula, a convex hull of matrix P rank is its nuclear norm P * 。
5. The method for intelligent fault diagnosis of unsupervised migration of mechanical equipment releasing source data according to claim 1, wherein the method comprises the following steps: the feature coding module in the step S41 comprises four convolution layers, a Batch-standardized layer, a nonlinear activation function, a pooling layer, a full-connection layer and a Dropout layer; the classification module comprises a full connection layer and a nonlinear activation layer.
6. The method for intelligent fault diagnosis of unsupervised migration of mechanical equipment releasing source data according to claim 1, wherein the method comprises the following steps: assume source field D in step S2 s Comprising n s Each labeled sample
Wherein->
Target area D t Comprising n t Label-free sample->
Wherein->
7. The method for intelligent fault diagnosis of unsupervised migration of mechanical equipment releasing source data according to claim 1, wherein the method comprises the following steps: the formula for data normalization in step S3 specifically includes:
wherein: x is x i Is the i-th data sample; mu (mu) i Is x i Average value of (2); sigma (sigma) i Is x i Standard deviation of (2); x is x i (j) Is x i Is the j-th element of (c).
8. The method for intelligent fault diagnosis of unsupervised migration of mechanical equipment releasing source data according to claim 1, wherein the method comprises the following steps: the mechanical equipment in step S1 comprises a multistage acceleration transmission bearing failure simulation test bed.
9. The method for intelligent fault diagnosis of unsupervised migration of mechanical equipment releasing source data according to claim 8, wherein the method comprises the steps of: the sampling frequency of the vibration data of the multistage acceleration transmission device is 10kHz.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211643639.2A CN116026569A (en) | 2022-12-20 | 2022-12-20 | Mechanical equipment unsupervised migration intelligent fault diagnosis method for releasing source data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211643639.2A CN116026569A (en) | 2022-12-20 | 2022-12-20 | Mechanical equipment unsupervised migration intelligent fault diagnosis method for releasing source data |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116026569A true CN116026569A (en) | 2023-04-28 |
Family
ID=86075199
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211643639.2A Pending CN116026569A (en) | 2022-12-20 | 2022-12-20 | Mechanical equipment unsupervised migration intelligent fault diagnosis method for releasing source data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116026569A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116502072A (en) * | 2023-06-27 | 2023-07-28 | 北京理工大学 | Robust fault diagnosis method for key components of wind generating set under complex variable working conditions |
| CN117909853A (en) * | 2024-03-19 | 2024-04-19 | 合肥通用机械研究院有限公司 | Intelligent monitoring method and system for equipment damage based on mechanism and working condition big data |
-
2022
- 2022-12-20 CN CN202211643639.2A patent/CN116026569A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116502072A (en) * | 2023-06-27 | 2023-07-28 | 北京理工大学 | Robust fault diagnosis method for key components of wind generating set under complex variable working conditions |
| CN116502072B (en) * | 2023-06-27 | 2023-09-08 | 北京理工大学 | Robust fault diagnosis method for key components of wind generating set under complex variable working conditions |
| CN117909853A (en) * | 2024-03-19 | 2024-04-19 | 合肥通用机械研究院有限公司 | Intelligent monitoring method and system for equipment damage based on mechanism and working condition big data |
| CN117909853B (en) * | 2024-03-19 | 2024-05-31 | 合肥通用机械研究院有限公司 | Intelligent monitoring method and system for equipment damage based on mechanism and working condition big data |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Li et al. | Intelligent fault diagnosis of rolling bearings under imbalanced data conditions using attention-based deep learning method | |
| CN106124212B (en) | Fault Diagnosis of Roller Bearings based on sparse coding device and support vector machines | |
| CN116026569A (en) | Mechanical equipment unsupervised migration intelligent fault diagnosis method for releasing source data | |
| CN112763214B (en) | Rolling bearing fault diagnosis method based on multi-label zero-sample learning | |
| CN110657984B (en) | Planetary gearbox fault diagnosis method based on reinforced capsule network | |
| CN113673346B (en) | Motor vibration data processing and state identification method based on multiscale SE-Resnet | |
| Gao et al. | ASM1D-GAN: An intelligent fault diagnosis method based on assembled 1D convolutional neural network and generative adversarial networks | |
| CN112633339A (en) | Bearing fault intelligent diagnosis method, bearing fault intelligent diagnosis system, computer equipment and medium | |
| Liang et al. | Unsupervised fault diagnosis of wind turbine bearing via a deep residual deformable convolution network based on subdomain adaptation under time-varying speeds | |
| Chen et al. | An anti-noise fault diagnosis approach for rolling bearings based on multiscale CNN-LSTM and a deep residual learning model | |
| CN111459144A (en) | Airplane flight control system fault prediction method based on deep cycle neural network | |
| You et al. | An efficient lightweight neural network using BiLSTM-SCN-CBAM with PCA-ICEEMDAN for diagnosing rolling bearing faults | |
| CN115508073B (en) | Prototype adaptive mechanical equipment fault diagnosis method based on multi-scale attention | |
| CN116704431A (en) | On-line monitoring system and method for water pollution | |
| Wang et al. | Recent advancement of deep learning applications to machine condition monitoring part 1: a critical review | |
| CN116702076A (en) | Small sample migration learning fault diagnosis method, system, computer and storage medium based on CNN feature fusion | |
| CN116223020A (en) | Intelligent fault diagnosis method for gearbox based on Li-MDCAN | |
| CN116108346A (en) | Bearing increment fault diagnosis life learning method based on generated feature replay | |
| Ge et al. | A deep condition feature learning approach for rotating machinery based on MMSDE and optimized SAEs | |
| Gao et al. | Semi-supervised small sample fault diagnosis under a wide range of speed variation conditions based on uncertainty analysis | |
| Li et al. | Source-free domain adaptation framework for fault diagnosis of rotation machinery under data privacy | |
| Man et al. | Bearing remaining useful life prediction based on AdCNN and CWGAN under few samples | |
| Jiang et al. | An orbit-based encoder–forecaster deep learning method for condition monitoring of large turbomachines | |
| Wang et al. | Distilling the knowledge of multiscale densely connected deep networks in mechanical intelligent diagnosis | |
| CN113884300A (en) | Rolling bearing fault diagnosis method for deep anti-migration 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 |