CN114492534A - Construction method and application of cross-size motor bearing fault diagnosis model - Google Patents

Construction method and application of cross-size motor bearing fault diagnosis model Download PDF

Info

Publication number
CN114492534A
CN114492534A CN202210113864.9A CN202210113864A CN114492534A CN 114492534 A CN114492534 A CN 114492534A CN 202210113864 A CN202210113864 A CN 202210113864A CN 114492534 A CN114492534 A CN 114492534A
Authority
CN
China
Prior art keywords
network
bearing
data
cross
timegan
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210113864.9A
Other languages
Chinese (zh)
Inventor
刘红奇
吴义轩
全政宇
朱秋凝
李斌
彭芳瑜
毛新勇
贺松平
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huazhong University of Science and Technology
Original Assignee
Huazhong University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Huazhong University of Science and Technology filed Critical Huazhong University of Science and Technology
Priority to CN202210113864.9A priority Critical patent/CN114492534A/en
Publication of CN114492534A publication Critical patent/CN114492534A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G06F2218/04—Denoising
    • G06F2218/06—Denoising by applying a scale-space analysis, e.g. using wavelet analysis
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G01M13/04—Bearings
    • G01M13/045—Acoustic or vibration analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Signal Processing (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Acoustics & Sound (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

The invention discloses a construction method and application of a cross-size motor bearing fault diagnosis model, which belong to the technical field of fault diagnosis, construct a model based on a TimeGAN network and a DAAN network, capture potential complex dynamic characteristics of acceleration cross-time through the TimeGAN network, simultaneously generate new time series data of the acceleration, expand samples of cross-size fault bearings, migrate original fault data by using the DAAN network, and align the distribution of sample data of bearings with different sizes according to the bearings with known normal and fault sample data and the normal sample data of the bearings with different sizes; the established model is trained end to end, so that the technical problem of low accuracy of motor bearing fault diagnosis caused by difficulty in acquiring a large amount of labeled high-quality data in the prior art is solved.

Description

Construction method and application of cross-size motor bearing fault diagnosis model
Technical Field
The invention belongs to the technical field of fault diagnosis, and particularly relates to a construction method and application of a cross-size motor bearing fault diagnosis model.
Background
The bearing is one of the most basic components of the motor, and serves as a connecting rotating part and a supporting part to bear certain load. Therefore, the bearing is inevitably affected by a series of actions such as mechanical stress during normal use, and is worn and corroded during use over a long period of time. According to incomplete statistics, 40% of failures of the motor are caused by rolling bearing failures. If a fault occurs, such as failure to detect the fault in time or failure to effectively troubleshoot and repair the potential fault, further damage to the mechanical equipment and influence industrial production may occur, and even the life safety of workers may be endangered, thereby causing great loss.
At present, motor bearing mechanical health state detection represented by a deep learning method is gradually becoming a mainstream of a fault diagnosis method. Compared with the traditional model, the deep learning model has deeper network layers and strong nonlinear computing capability, can better approximate complex function relationship, and has more successful application in the field of fault diagnosis. However, the success of deep learning fault diagnosis depends on a large amount of labeled high-quality data, and at present, a large amount of manpower and time are consumed for data acquisition of fault bearing samples, so that the monitoring of abnormal data samples is insufficient for the requirement of deep learning. For motor bearing samples of different sizes, data of a failed bearing sample can be changed, and the original data is difficult to be suitable for monitoring training of a new failed bearing.
Disclosure of Invention
Aiming at the defects or the improvement requirements of the prior art, the invention provides a construction method and application of a cross-size motor bearing fault diagnosis model, which are used for solving the technical problem of low motor bearing fault diagnosis accuracy rate caused by difficulty in acquiring a large amount of labeled high-quality data in the prior art.
Generally, by the above technical solution conceived by the present invention, the following beneficial effects can be obtained:
in order to achieve the above object, in a first aspect, the present invention provides a method for constructing a cross-dimension motor bearing fault diagnosis model, including the following steps:
s1, building a cross-size motor bearing fault diagnosis model; wherein, stride size motor bearing fault diagnosis model includes: the system comprises a cascaded cross-size data generation module, a feature extraction module and a DAAN network; the cross-size data generation module comprises a first TimeGAN network, a second TimeGAN network, a third TimeGAN network and a fourth TimeGAN network which are connected in parallel; the feature extraction module comprises a first feature extraction unit and a second feature extraction unit which are connected in parallel; the output ends of the first TimeGAN network and the third TimeGAN network are respectively connected with the input end of the first characteristic extraction unit; the output ends of the second TimeGAN network and the fourth TimeGAN network are respectively connected with the input end of the second characteristic extraction unit; the DAAN network is used for carrying out transfer learning on the input bearing data characteristics so as to predict the health state of the bearing data characteristics;
s2, obtaining a training sample set; wherein the training samples in the training sample set include bearing data of two different sizes, denoted as (N)A,FA,NB)；NAThe vibration acceleration of the bearing A in a normal state and a working condition corresponding to the vibration acceleration are obtained; fAThe vibration acceleration and the corresponding working condition of the bearing A in the fault state are obtained; n is a radical ofBThe vibration acceleration and the corresponding working condition of the bearing B in the normal state are obtained; the working conditions of all the training samples are the same;
s3, training sample (N) in training sample setA,FA,NB) Inputting the fault diagnosis model into a cross-size motor bearing for training:
respectively adding NAVia a first TimeGAN network, NBInputting into the cross-size motor bearing fault diagnosis model via the second TimeGAN network to calibrate the first TimeGAN network, the second TimeGAN network and the DAThe AN network is trained, and the distance dis between the features output by the first feature extraction unit and the second feature extraction unit at the moment is calculated12；
F is to beAInputting the signals into a cross-size motor bearing fault diagnosis model through a third TimeGAN network and a fourth TimeGAN network respectively, and calculating the distance dis between the features output by the first feature extraction unit and the second feature extraction unit at the moment34(ii) a In minimizing dis34、dis12And dis34And dis12Training a third TimeGAN network, a fourth TimeGAN network and a DAAN network on the premise of the distance difference;
and S4, repeating the step S3 until all training samples in the training sample set are input into the cross-size motor bearing fault diagnosis model for training.
Further preferably, the method for acquiring the vibration acceleration in the training data set includes: and after data interception is carried out on the collected original vibration acceleration signals, angular domain resampling is carried out.
Further preferably, the method for acquiring the vibration acceleration in the training data set includes: and after data interception is carried out on the acquired original vibration acceleration signals, noise reduction processing and angular domain resampling are carried out in sequence.
Further preferably, the health status of the bearing comprises: a normal state and a fault state; wherein the fault states include an outer ring fault, an inner ring fault, a rolling element fault and a retainer fault.
Further preferably, the first TimeGAN network, the second TimeGAN network and the third TimeGAN network are all configured to generate bearing data similar to the input bearing data, in consideration of the timing correlation;
the fourth TimeGAN network is used for generating bearing data in the fault state of the bearing B based on the difference between the bearing data in the normal state of the bearing a and the bearing data in the fault state of the bearing B and considering the time sequence correlation.
Further preferably, the DAAN network comprises: the system comprises a feature extractor, a label classifier, a global area discriminator and a local area discriminator; the label classifier, the whole local area discriminator and the local area discriminator are connected in parallel and are respectively connected with the output end of the feature extractor; the local domain discriminator comprises C local domain discriminators which are mutually connected in parallel; c is the health state category number of the bearing;
in the training process of the model:
the bearing data characteristics of the bearing A input by the characteristic extraction module are used as source domain data, and the bearing data characteristics of the bearing B are used as target domain data;
the characteristic extractor is used for respectively extracting the domain invariant characteristics of the source domain data and the target domain data, which are related to the bearing health state category;
the label classifier is used for mapping the domain invariant features of the source domain data into corresponding health state classes and measuring the difference between the health state classes and the real health state classes;
the global area discriminator is used for aligning the edge distribution between the domain invariant features of the source domain data and the target domain data;
the local domain discriminator is used for respectively aligning the condition distribution between the domain invariant features of the source domain data and the target domain data under each type of health state;
during the application of the model:
the characteristic extractor is used for extracting the domain invariant characteristics of the bearing data to be diagnosed, which are related to the bearing health state category;
the label classifier is used for mapping the domain invariant features of the bearing data to be diagnosed into corresponding health status categories.
In a second aspect, a cross-dimension motor bearing fault diagnosis method includes: inputting bearing data to be diagnosed into a DAAN network in a cross-size motor bearing fault diagnosis model established by the method for establishing the cross-size motor bearing fault diagnosis model provided by the first aspect of the invention, thereby obtaining the health state of the bearing data to be diagnosed.
In a third aspect, the invention provides a cross-size motor bearing fault diagnosis method, which comprises the following steps: a memory storing a computer program and a processor executing the computer program to perform the cross-dimension motor bearing fault diagnosis method provided by the second aspect of the present invention.
In a fourth aspect, the present invention provides a machine-readable storage medium storing machine-executable instructions which, when invoked and executed by a processor, cause the processor to carry out the method of constructing a cross-dimensional motor bearing fault diagnosis model provided in the first aspect of the invention and/or the method of cross-dimensional motor bearing fault diagnosis provided in the second aspect of the invention.
Generally, by the above technical solution conceived by the present invention, the following beneficial effects can be obtained:
1. the invention provides a method for constructing a cross-size motor bearing fault diagnosis model, which constructs a model based on a TimeGAN network and a DAAN network, captures potential complex characteristics (namely potential complex dynamic characteristics of acceleration across time) of acceleration signals changing along with time through the TimeGAN network, simultaneously generates new time sequence data of acceleration, expands samples of cross-size fault bearings, uses the DAAN network to carry out anti-migration on original fault data, measures the distances among normal samples of the bearings with different sizes according to the bearings with known normal and fault sample data and in combination with normal sample data of the bearings with different sizes, and carries out end-to-end training on the constructed model by restricting the generated fault samples of the cross-working condition data to meet the same distance among corresponding fault classes, thereby solving the fault diagnosis problem of the new bearings under the condition that only the normal samples have no fault sample, the accuracy of motor bearing fault diagnosis is greatly improved.
2. According to the construction method of the cross-size motor bearing fault diagnosis model, the originally acquired bearing signals are influenced by complex working conditions, contain a large amount of noise and belong to non-stable process signals.
3. Aiming at the problem that the existing bearing fault diagnosis method based on deep learning usually needs to collect a large amount of data of a fault bearing, the invention introduces a time sequence generation countermeasure network, expands samples of the fault bearing, provides more data for the training process of subsequent fault diagnosis, solves the problem of poor model training effect caused by the shortage of abnormal samples, and saves the labor cost and the time cost for collecting the samples.
4. In the method for constructing the cross-size motor bearing fault diagnosis model, acceleration data used for model training is provided with rotating speed labels (working condition labels), and a plurality of cross-size motor bearing fault diagnosis models can be trained by respectively adopting training samples at different rotating speeds, so that the distribution of normal and abnormal samples at different rotating speeds under the cross-size condition is obtained. By highly refining a plurality of small model tasks, the calculation complexity is effectively reduced, the model precision is improved, and the practicability of generated data is enhanced.
5. According to the construction method of the cross-size motor bearing fault diagnosis model provided by the invention, through the proposed multi-scale (loss function can be applied through single or multiple distances) constraint that the distance between normal samples and the distance between fault samples of the cross-size bearing are minimum, the generation quality of a time countermeasure generation network can be effectively improved, and the generated samples can better meet the prior physical constraint.
Drawings
Fig. 1 is a flowchart of a method for constructing a cross-size motor bearing fault diagnosis model according to embodiment 1 of the present invention;
FIG. 2 is a schematic structural diagram of a cross-dimension motor bearing fault diagnosis model provided in embodiment 1 of the present invention;
fig. 3 is a schematic structural diagram of a TimeGAN network according to embodiment 1 of the present invention;
fig. 4 is a schematic structural diagram of a DAAN network according to embodiment 1 of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
Examples 1,
A method for constructing a cross-dimension motor bearing fault diagnosis model is shown in FIG. 1, and comprises the following steps:
s1, building a cross-size motor bearing fault diagnosis model; specifically, as shown in fig. 2, the cross-size motor bearing fault diagnosis model includes: the system comprises a cascaded cross-size data generation module, a feature extraction module and a DAAN network; the cross-size data generation module comprises a first TimeGAN network, a second TimeGAN network, a third TimeGAN network and a fourth TimeGAN network which are connected in parallel; the feature extraction module comprises a first feature extraction unit and a second feature extraction unit which are connected in parallel; the output ends of the first TimeGAN network and the third TimeGAN network are respectively connected with the input end of the first characteristic extraction unit; the output ends of the second TimeGAN network and the fourth TimeGAN network are respectively connected with the input end of the second characteristic extraction unit; the DAAN network is used for carrying out transfer learning on the input bearing data characteristics so as to predict the health state of the bearing data characteristics.
S2, obtaining a training sample set; wherein the training samples in the training sample set include bearing data of two different sizes, denoted as (N)A,FA,NB)；NAThe vibration acceleration of the bearing A in a normal state and a working condition corresponding to the vibration acceleration are obtained; fAThe vibration acceleration and the corresponding working condition of the bearing A in the fault state are obtained; n is a radical ofBThe vibration acceleration and the corresponding working condition of the bearing B in the normal state are obtained; the working conditions of all the training samples are the same;
specifically, in this embodiment, data are collected for a single bearing fault each time through an acceleration sensor on the motor, and the data collected each time include collection time, vibration acceleration of each channel of the bearing, bearing rotation speed (working condition), and a health state label of the bearing, and are uploaded through hardware equipment. The training sample set can be constructed by carrying out angular domain resampling after carrying out data interception on the collected original vibration acceleration signals. Furthermore, the collected original vibration acceleration signals can be subjected to data interception, and then noise reduction processing and angular domain resampling are sequentially performed. The method comprises the following specific steps:
1) intercepting the existing data based on a sliding time window: the known time sequence is framed through a window with unit length of 2048, a vibration signal section is intercepted, statistical indexes such as an average value in the frame are calculated, data in a sliding block of a unit time window are obtained, and a data range is reduced.
2) And (3) performing denoising processing by using wavelet transform: in order to remove noise in original data, the wavelet transform method adopted in this embodiment performs denoising processing, including the following steps: a. and selecting a group of orthogonal wavelet bases and the number N of decomposition layers, performing N-layer wavelet decomposition on the original vibration signal, and dividing the original vibration signal into a plurality of small frequency bands. b. And processing the high-frequency coefficients of each layer from the 1 st layer to the Nth layer by using a soft threshold quantization method. c. And performing signal reconstruction according to the low-frequency coefficient of the Nth layer of the wavelet decomposition and the processed high-frequency coefficients of the 1 st layer to the Nth layer, thereby obtaining an estimated value of the original vibration signal.
3) Angular domain resampling of the resulting signal: in order to solve the problem of data asynchrony on a time scale caused by the fact that collected data have no calibration signals, the method carries out angular domain resampling on a fault sample. Specifically, the relationship between the vibration signal and the rotating speed signal is established by using the assumption of linear change of the signal between adjacent sampling points, so that the vibration signal is resampled according to the angle signal, and the angle signal is aligned. It should be noted that the angular domain resampled signal obtained by the above method has a stable characteristic, and also has an actual physical meaning of time domain sampling, so that it is easier to extract time domain and frequency domain features in subsequent steps.
In order to provide more abnormal sample data for training in subsequent transfer learning, the invention provides an intelligent and automatic data generation method for abnormal sample data based on a TimeGAN network, wherein the static and driving-end acceleration of the abnormal sample is respectively extracted.
S3, training sample (N) in training sample setA,FA,NB) Inputting the fault diagnosis model into a cross-size motor bearing for training:
respectively adding NAVia a first TimeGAN network, NBInputting the parameters into a cross-size motor bearing fault diagnosis model through a second TimeGAN network to train the first TimeGAN network, the second TimeGAN network and the DAAN network (in the process, parameters in the first TimeGAN network, the second TimeGAN network, the first feature extraction unit, the second feature extraction unit and the DAAN network are updated), and calculating the distance dis between the features output by the first feature extraction unit and the second feature extraction unit after the training is finished12；
F is to beAInputting the signals into a cross-size motor bearing fault diagnosis model through a third TimeGAN network and a fourth TimeGAN network respectively, and calculating the distance dis between the features output by the first feature extraction unit and the second feature extraction unit at the moment34(ii) a In minimizing dis34、dis12And dis34And dis12On the premise of the distance difference, training the third TimeGAN network, the fourth TimeGAN network and the DAAN network (in the process, updating parameters in the third TimeGAN network, the fourth TimeGAN network, the first feature extraction unit, the second feature extraction unit and the DAAN network);
and S4, repeating the step S3 until all training samples in the training sample set are input into the cross-size motor bearing fault diagnosis model for training.
It should be noted that the first TimeGAN network, the second TimeGAN network, and the third TimeGAN network are all used to generate bearing data similar to the input bearing data, in consideration of the timing correlation;
the fourth TimeGAN network is used for generating bearing data in the fault state of the bearing B based on the difference between the bearing data in the normal state of the bearing a and the bearing data in the fault state of the bearing B and considering the time sequence correlation.
It should be noted that the TimeGAN network considers the operating condition factor rotation speed as a static characteristic and considers the driving end acceleration as a temporal characteristic, so that a series of values of the driving end acceleration of the fault sample at a certain rotation speed can be obtained. As shown in fig. 3, the TimeGAN network includes four components, an embedding function, a recovery function, a sequence generator, and a sequence arbiter. Mapping the bearing rotating speed and the driving end acceleration to potential spaces by using an embedding function respectively to realize the dimension reduction of the characteristics; the generator firstly reduces the dimension of the rotating speed characteristics sampled from the distribution, and then generates a new low-dimensional acceleration characteristic vector according to the rotating speed characteristics after the dimension reduction, the low-dimensional acceleration characteristics at the previous moment and the current driving end acceleration; respectively checking the generated rotation speed characteristic and acceleration characteristic by using a sequence discriminator, and passing fault sample data capable of representing a real situation; after learning is completed, the recovery function performs inverse mapping on the rotation speed and the driving end acceleration characteristics in the potential space, and restores and reconstructs the characteristics to the original characteristics.
Wherein the function (e) is embeddedS(. and e)X(. -) and a recovery function (r)S(. and r)X(. -) provides a mapping between the features and the potential space; specifically, the embedding function realizes the dimension reduction of the characteristic, and the bearing rotating speed s is mapped to h with lower dimensionsThen the driving end acceleration x is settMapping to features h of lower dimensionstThe process is as follows:
hs＝eS(s)
ht＝eX(hS,ht-1,xt)
further, after the network learns the features with lower dimensionality, the recovery function performs inverse mapping on the features in the potential space again to restore the features to the original features, and specifically, the inverse mapping process for the bearing rotating speed and the driving end acceleration is as follows:
Figure BDA0003495647660000101
Figure BDA0003495647660000102
in order to ensure the accuracy of original data reconstruction, the loss function constructed by the invention is specifically as follows:
Figure BDA0003495647660000103
wherein L isRRepresents the distance reconstruction loss, s, xtRespectively representing the original rotating speed and the acceleration at the time t,
Figure BDA0003495647660000104
the reconstructed rotation speed and the acceleration at time t are shown. The present invention ensures the accuracy of the reconstruction by minimizing this loss function.
Further, the generator in TimeGAN is divided into gS(. and g)XTwo parts, wherein the generator gS(. random vector z of bearing rotational speed obtained by random sampling of the distributionsConversion to static variables of low dimensionality
Figure BDA0003495647660000105
Generator gXAccording to the generated characteristic vector of the rotating speed of the bearing
Figure BDA0003495647660000106
Generated last
Figure BDA0003495647660000107
And the current drive-end acceleration z obtained by the Gaussian distribution and the wiener processtGenerating low-dimensional drive-end acceleration eigenvectors
Figure BDA0003495647660000108
The process is as follows:
Figure BDA0003495647660000109
Figure BDA00034956476600001010
the invention selects CNN as the bearing rotating speed
Figure BDA00034956476600001011
Using a bidirectional RNN as the driving-end acceleration
Figure BDA0003495647660000111
The discriminator of (1) outputs 0 for the false data and 1 for the real data, and the process is as follows:
Figure BDA0003495647660000112
Figure BDA0003495647660000113
further, the RNN used in the above formula is obtained
Figure BDA0003495647660000114
The process of (2) is as follows:
Figure BDA0003495647660000115
Figure BDA0003495647660000116
the TimeGAN network adopts unsupervised learning for the static characteristic bearing rotating speed and introduces the output generated by two discriminators
Figure BDA0003495647660000117
The acceleration of the temporal characteristic driving end adopts a supervision learning mode; to nothingThe loss functions constructed by supervised learning and supervised learning are respectively as follows:
Figure BDA0003495647660000118
Figure BDA0003495647660000119
wherein L isU、LSRespectively representing loss functions adopted by unsupervised learning and supervised learning; y iss、ytThe label is the original label of the rotating speed and the acceleration of the sample;
Figure BDA00034956476600001110
the output generated by judging the generated rotating speed and acceleration by the discriminator is shown; h issRepresenting the rotation speed characteristics after dimensionality reduction, htAnd representing the acceleration characteristic at the time t after dimensionality reduction.
The embedding function, the recovery function, the generator and the arbiter are optimized according to the calculated loss function as follows:
Figure BDA00034956476600001111
Figure BDA00034956476600001112
wherein, thetae、θr、θg、θdParameters of the embedding network, the recovery network, the generator and the discriminator are respectively expressed, wherein lambda and eta are hyper-parameters introduced for balancing losses.
The TimeGAN network can capture potential complex dynamics of acceleration over time, generate new time series data of acceleration under different rotating speed conditions simultaneously, and generate additional fault samples for subsequent data generation over size.
Further, in order to obtain cross-size bearing fault data, for example, the data of the bearing a with the known fault sample label generated by training is applied to the bearing B with another size, the present invention uses the DAAN network to migrate the original fault data, and the purpose of the present invention is to migrate and generate the fault sample data distribution of the bearings with different sizes at a certain rotation speed according to the bearings with the known normal and fault sample data and combining the normal sample data of the bearings with different sizes. Specifically, as shown in fig. 4, the DAAN network includes: the system comprises a feature extractor, a label classifier, a global area discriminator and a local area discriminator; the label classifier, the whole local area discriminator and the local area discriminator are connected in parallel and are respectively connected with the output end of the feature extractor; the local domain discriminator comprises C local domain discriminators which are mutually connected in parallel; c is the health state category number of the bearing;
in the training process of the model, bearing data characteristics of a bearing A input by a characteristic extraction module are used as source domain data, and bearing data characteristics of a bearing B are used as target domain data, and the source domain data and the target domain data are alternately input into a DAAN network; the characteristic extractor is used for respectively extracting the domain invariant characteristics of the source domain data and the target domain data, which are related to the bearing health state category; the label classifier is used for mapping the domain invariant features of the source domain data into corresponding health state classes and measuring the difference between the health state classes and the real health state classes; the global area discriminator is used for aligning the edge distribution between the domain invariant features of the source domain data and the target domain data; the local domain discriminator is used for respectively aligning the condition distribution between the domain invariant features of the source domain data and the target domain data under each type of health state;
during the application of the model: the characteristic extractor is used for extracting the domain invariant characteristics of the bearing data to be diagnosed, which are related to the bearing health state category; the label classifier is used for mapping the domain invariant features of the bearing data to be diagnosed into corresponding health status categories.
In this embodiment, the health status of the bearing includes: a normal state and a fault state; wherein the fault states include an outer ring fault, an inner ring fault, a rolling element fault and a retainer fault. The DAAN network mainly comprises a deep feature extractor (in the embodiment, a ResNet network), a label classifier, a global domain discriminator and 5 local domain discriminators, wherein the five local domain discriminators respectively process data in a normal state, an outer ring fault, an inner ring fault, a rolling body fault and a retainer fault and in 5 types of states.
In this embodiment, the training process of the DAAN network is as follows:
1. extracting domain invariant features by using a one-dimensional ResNet network, and passing acceleration data of a bearing with known fault data size through a feature extractor to obtain a depth feature vector; specifically, the invention performs feature extraction on the time-frequency domain features of the vibration acceleration through a one-dimensional ResNet network to obtain the driving end acceleration features of the fault bearing which is most effective in classification and identification.
2. Distinguishing labels of input samples from a source domain by using a label classifier, constructing cross entropy loss, and judging normal data and fault data of an original bearing;
3. aligning the edge distribution between the source domain and the target domain by using a global local area discriminator, and integrally aligning the normal abnormal data distribution of the cross-size bearing;
4. and aligning the multi-mode structure of the cross-size bearing positive abnormal data distribution by using a local domain discriminator to align the conditional distribution between the source domain and the target domain, thereby performing domain adaptation with finer granularity.
Specifically, the tag classifier distinguishes normal and abnormal sample data of an original bearing, classifies source domain data (sample data of an A bearing) by adopting a softmax activation function, and is realized by constructing the following cross entropy loss:
Figure BDA0003495647660000131
wherein n issFor the number of source domain data in the training dataset, DsC represents the number of bearing health status categories for the set of source domain data in the training dataset,
Figure BDA0003495647660000132
representing the ith input sample xiProbability of belonging to the c-th health State class, GyRepresenting a tag classifier, GfA feature extractor is represented.
It should be noted that the selected ResNet network can solve the problems of gradient disappearance, explosion and network degradation caused by the increase of the number of network layers as much as possible. The softmax activation function is used, and bearing data can be divided into four types including a normal state, an outer ring fault, an inner ring fault and a rolling body fault. The cross entropy loss function can effectively improve the convergence speed of the model.
Further, in order to confuse A, B the health status data fields of two bearings with different sizes and maximize the domain classification error, the present invention aligns the edge distribution between the source domain and the target domain (A, B two bearing health status data fields) using a global domain arbiter and constructs the loss function as follows:
Figure BDA0003495647660000141
wherein n istFor the number of target domain data in the training dataset, DtSet of target domain data in training dataset, LdRepresents the cross-entropy loss of the global-area arbiter, GfIs a feature extractor, diIs a domain label of the input data that is used to predict whether the input data is characterized as belonging to a source domain or a target domain.
Further, using the local domain arbiter to align the condition distribution between the source domain and the target domain, i.e. using the local domain arbiter to align A, B the condition distribution corresponding to each type of fault for the two bearings, a loss function is constructed as follows:
Figure BDA0003495647660000142
wherein the content of the first and second substances,
Figure BDA0003495647660000143
and
Figure BDA0003495647660000144
is the local domain arbiter associated with the c-th health state class and the corresponding cross-entropy penalty,
Figure BDA0003495647660000145
is input data xiA conditional probability distribution over the c-th health state class, diIs input data xiThe domain tag of (1). In this embodiment, c is 1,2,3,4, and 5, respectively corresponding to a normal operation of the bearing, a failure of the outer ring, a failure of the inner ring, a failure of the rolling element, and a failure of the cage.
In order to reduce the domain classification precision, a gradient reversal layer GRL is added in both the global domain discriminator and the local domain discriminator so as to reduce the fault difference of A, B bearing data of two sizes and achieve the aim of aligning edge distribution and condition distribution.
Combining the above steps, the total loss values generated for the cross-dimensional bearing fault samples are as follows:
Figure BDA0003495647660000146
wherein, thetaf、θy、θdAnd
Figure BDA0003495647660000151
parameters in a local domain discriminator corresponding to the feature extractor, the label classifier, the global domain discriminator and the c-th health state category respectively; λ is a trade-off parameter. w is a dynamic countermeasure factor, reflects the relative importance degree of edge distribution and condition distribution, and can be obtained by self-calculation of the network. Preferably, during the training process, the present embodiment optimizes the above parameters by using a random gradient descent method.
It should be noted that the above dynamic countermeasure factor w can be estimated as:
Figure BDA0003495647660000152
the inter-domain distance between the global domain discriminator and the local domain discriminator is defined as follows:
dA,g(Ds,Dt)＝2(1-2(Lg))
Figure BDA0003495647660000153
wherein the content of the first and second substances,
Figure BDA0003495647660000154
and
Figure BDA0003495647660000155
representing source domain data samples and target domain data samples from category c,
Figure BDA0003495647660000156
is the cross entropy loss of the local domain arbiter corresponding to the c-th health state class.
The present embodiment preferably learns and updates w in a deep confrontation manner without using shallow features, making DAAN more robust and accurate. In addition, the DAAN neural network selected by the invention can directly utilize the loss of the domain discriminator to carry out automatic fine adjustment on w, and is simpler and more effective.
It should be noted that the above process is a process of selecting a training sample at a certain specific rotation speed to train the cross-size motor bearing fault diagnosis model at the rotation speed. In order to obtain bearing state sample distribution of bearing B at different rotating speeds, different rotating speeds [ n ] of bearing A can be adopted1,n2,n3,...,nn]Corresponding acceleration data trains a plurality of cross-size motor bearing fault diagnosis models [ M1,M2,M3,...,Mn]The training process is the same as above, and is not described herein.
Examples 2,
A cross-dimension motor bearing fault diagnostic method comprising: inputting bearing data to be diagnosed into a DAAN network in a cross-size motor bearing fault diagnosis model established by the method for establishing the cross-size motor bearing fault diagnosis model provided by the embodiment 1 of the invention, so as to obtain the health state of the bearing data to be diagnosed.
In the embodiment, firstly, the acceleration signal is detected by using monitoring equipment, the signal is intercepted, the angular domain resampling is carried out, the angular domain data of the fault bearing data is obtained, secondly, the time-frequency domain feature extraction is carried out on the obtained data, the most effective acceleration feature is obtained, then, a cross-size motor bearing fault diagnosis model is trained, and finally, the fault data migration is realized through the model.
The following examples illustrate the invention in detail:
1) data acquisition and preprocessing:
the method uses monitoring equipment to obtain data such as the rotating speed of the bearing, one-dimensional acceleration signals and the like, and the obtained signals contain a large amount of noise due to the influence of complex working conditions, belong to non-stable process signals, and carry out preprocessing on the collected data in the modes of sliding time window interception, wavelet noise reduction, angular domain resampling and the like.
Wherein, the sliding time window intercepting process is as follows: framing a known time series by a window with unit length of 2048, cutting a vibration signal segment, calculating statistical indexes such as an average value in the frame, and acquiring data in a sliding block of a unit time window to narrow a data range, for example, for an input signal x ═ x1,x2,...,xNIn terms of xiFor the value of the ith input signal, N is the length of the input signal, framing is performed by sliding a "window" of unit length 2048, each 2048 windows are classified into one class, the mean value is calculated, and the signal representation y ═ y at different scales is obtained1,y2,...,ynAnd each y represents the mean value in each window, and high-frequency disturbance and random noise are filtered to a certain extent.
The wavelet denoising process comprises the following steps: according to the fact that when wavelet decomposition is conducted on signals and noise under different scales, the mode of the noise and the maximum value of the mode of the signals are opposite to each other along with the change of the wavelet scales, N-layer wavelet decomposition is conducted on original vibration signals, and the original vibration signals are divided into a plurality of small frequency bands { x }1,x2,...,xNAnd then, processing each layer of high-frequency coefficients from the 1 st layer to the Nth layer by using a soft threshold quantization method, not processing the low-frequency coefficients, and finally, performing signal reconstruction according to the low-frequency coefficients of the Nth layer of wavelet decomposition and the processed high-frequency coefficients from the 1 st layer to the Nth layer to obtain an estimated value of an original vibration signal, thereby achieving the purpose of signal noise reduction.
The angular domain resampling process is as follows: introducing vibration signals into two adjacent acquisition points (t)n,gn) And (t)n+1,gn+1) And a rotational speed signal acquisition point (t)M,gM) The relationship is obtained by considering the signal between the vibration signal sample points n and n +1 to be linear
Figure BDA0003495647660000171
And then obtaining a vibration acceleration signal sequence t0,t1,t2...tm-2,tm-1,tmAnd the sequence obtained at the moment has a stable characteristic, and also has the actual physical significance of time domain sampling, so that the time domain and frequency domain characteristics can be extracted more easily in the subsequent steps.
2) The bearing fault diagnosis model trained based on the cross-size motor in the embodiment 1 of the invention processes the preprocessed bearing fault data to obtain the health state of the bearing fault data.
The related technical scheme is the same as embodiment 1, and is not described herein.
Examples 3,
A cross-dimension motor bearing fault diagnostic method comprising: the cross-size motor bearing fault diagnosis device comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to execute the cross-size motor bearing fault diagnosis method provided by the embodiment 2 of the invention.
The related technical scheme is the same as embodiment 2, and is not described herein.
Examples 4,
A machine readable storage medium storing machine executable instructions which, when invoked and executed by a processor, cause the processor to carry out a method of building a cross-dimension motor bearing fault diagnosis model as provided by embodiment 1 of the invention and/or a cross-dimension motor bearing fault diagnosis method as provided by embodiment 2 of the invention.
The related technical solutions are the same as those in embodiments 1 and 2, and are not described herein.
It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (9)

1. A construction method of a cross-size motor bearing fault diagnosis model is characterized by comprising the following steps:
s1, building a cross-size motor bearing fault diagnosis model; the cross-dimension motor bearing fault diagnosis model comprises: the system comprises a cascaded cross-size data generation module, a feature extraction module and a DAAN network; the cross-size data generation module comprises a first TimeGAN network, a second TimeGAN network, a third TimeGAN network and a fourth TimeGAN network which are connected in parallel; the feature extraction module comprises a first feature extraction unit and a second feature extraction unit which are connected in parallel; the output ends of the first TimeGAN network and the third TimeGAN network are respectively connected with the input end of the first feature extraction unit; the output ends of the second TimeGAN network and the fourth TimeGAN network are respectively connected with the input end of the second feature extraction unit; the DAAN network is used for carrying out transfer learning on input bearing data characteristics so as to predict the health state of the DAAN network;
s2, obtaining a training sample set; the training samples in the set of training samples include bearing data of two different sizes, denoted as (N)A,FA,NB)；NAThe vibration acceleration of the bearing A in a normal state and a working condition corresponding to the vibration acceleration are obtained; fAThe vibration acceleration and the corresponding working condition of the bearing A in the fault state are obtained; n is a radical ofBThe vibration acceleration and the corresponding working condition of the bearing B in the normal state are obtained; each training sampleThe working conditions are the same;
s3, training samples (N) in the training sample setA,FA,NB) Inputting the fault diagnosis model into the cross-size motor bearing for training:
respectively adding NAVia the first TimeGAN network, NBInputting the calculated distance dis into the cross-dimension motor bearing fault diagnosis model via the second TimeGAN network to train the first TimeGAN network, the second TimeGAN network and the DAAN network, and calculating the distance dis between the features output by the first feature extraction unit and the second feature extraction unit at that time12；
F is to beAInputting the calculated distance dis between the features output by the first feature extraction unit and the second feature extraction unit into the cross-size motor bearing fault diagnosis model through the third TimeGAN network and the fourth TimeGAN network respectively, and calculating the distance dis between the features output by the first feature extraction unit and the second feature extraction unit at the moment34(ii) a In minimizing said dis34Dis described12And the dis34And said dis12Training the third TimeGAN network, the fourth TimeGAN network, and the DAAN network on the premise of the distance difference;
and S4, repeating the step S3 until all the training samples in the training sample set are input into the cross-size motor bearing fault diagnosis model for training.
2. The method of constructing a cross-size motor bearing fault diagnosis model of claim 1, wherein the first TimeGAN network, the second TimeGAN network, and the third TimeGAN network are all configured to generate bearing data similar to the input bearing data in consideration of timing correlations;
the fourth TimeGAN network is used for generating bearing data in the fault state of the bearing B based on the difference between the bearing data in the normal state of the bearing a and the bearing data in the fault state of the bearing B and considering the time sequence correlation.
3. The method of constructing a cross-scale motor bearing fault diagnosis model of claim 2, wherein the DAAN network comprises: the system comprises a feature extractor, a label classifier, a global area discriminator and a local area discriminator; the label classifier, the whole local area discriminator and the local area discriminator are connected in parallel and are respectively connected with the output end of the feature extractor; the local domain discriminator comprises C local domain discriminators which are connected in parallel; c is the health state category number of the bearing;
during the training of the model:
the bearing data characteristics of the bearing A input by the characteristic extraction module are used as source domain data, and the bearing data characteristics of the bearing B are used as target domain data;
the feature extractor is used for respectively extracting domain-invariant features of the source domain data and the target domain data, wherein the domain-invariant features are related to the bearing health state category;
the label classifier is used for mapping the domain invariant features of the source domain data into corresponding health state classes and measuring the difference between the health state classes and the real health state classes;
the global area discriminator is used for aligning the edge distribution between the domain invariant features of the source domain data and the target domain data;
the local domain discriminator is used for respectively aligning the condition distribution between the domain invariant features of the source domain data and the target domain data under each type of health state;
during the application of the model:
the characteristic extractor is used for extracting the domain invariant characteristics of the bearing data to be diagnosed, which are related to the bearing health state category;
the label classifier is used for mapping the domain invariant features of the bearing data to be diagnosed into corresponding health state classes.
4. The method for constructing a cross-dimensional motor bearing fault diagnosis model according to any one of claims 1-3, wherein the health status of the bearing comprises: a normal state and a fault state; wherein the fault states include an outer ring fault, an inner ring fault, a rolling element fault and a cage fault.
5. The method for constructing the cross-size motor bearing fault diagnosis model according to any one of claims 1 to 3, wherein the method for acquiring the vibration acceleration in the training data set comprises the following steps: and after data interception is carried out on the collected original vibration acceleration signals, angular domain resampling is carried out.
6. The method for constructing the cross-size motor bearing fault diagnosis model according to any one of claims 1 to 3, wherein the method for acquiring the vibration acceleration in the training data set comprises the following steps: and after data interception is carried out on the acquired original vibration acceleration signals, noise reduction processing and angular domain resampling are carried out in sequence.
7. A cross-dimension motor bearing fault diagnosis method is characterized by comprising the following steps: inputting bearing data to be diagnosed into a DAAN network in a cross-size motor bearing fault diagnosis model established by the method for establishing the cross-size motor bearing fault diagnosis model according to any one of claims 1 to 6, so as to obtain the health state of the bearing data to be diagnosed.
8. A cross-dimension motor bearing fault diagnosis method is characterized by comprising the following steps: a memory storing a computer program and a processor executing the computer program to perform the cross-dimensional motor bearing fault diagnosis method of claim 7.
9. A machine readable storage medium having stored thereon machine executable instructions which, when invoked and executed by a processor, cause the processor to carry out the method of building a cross-dimensional motor bearing fault diagnosis model according to any one of claims 1 to 6 and/or the method of building a cross-dimensional motor bearing fault diagnosis according to claim 7.
CN202210113864.9A 2022-01-30 2022-01-30 Construction method and application of cross-size motor bearing fault diagnosis model Pending CN114492534A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210113864.9A CN114492534A (en) 2022-01-30 2022-01-30 Construction method and application of cross-size motor bearing fault diagnosis model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210113864.9A CN114492534A (en) 2022-01-30 2022-01-30 Construction method and application of cross-size motor bearing fault diagnosis model

Publications (1)

Publication Number Publication Date
CN114492534A true CN114492534A (en) 2022-05-13

Family

ID=81477826

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210113864.9A Pending CN114492534A (en) 2022-01-30 2022-01-30 Construction method and application of cross-size motor bearing fault diagnosis model

Country Status (1)

Country Link
CN (1) CN114492534A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115095953A (en) * 2022-06-16 2022-09-23 青岛海信日立空调系统有限公司 Training method and device of fault diagnosis model
CN116383757A (en) * 2023-03-09 2023-07-04 哈尔滨理工大学 Bearing fault diagnosis method based on multi-scale feature fusion and migration learning

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115095953A (en) * 2022-06-16 2022-09-23 青岛海信日立空调系统有限公司 Training method and device of fault diagnosis model
CN115095953B (en) * 2022-06-16 2024-04-09 青岛海信日立空调系统有限公司 Training method and device for fault diagnosis model
CN116383757A (en) * 2023-03-09 2023-07-04 哈尔滨理工大学 Bearing fault diagnosis method based on multi-scale feature fusion and migration learning
CN116383757B (en) * 2023-03-09 2023-09-05 哈尔滨理工大学 Bearing fault diagnosis method based on multi-scale feature fusion and migration learning

Similar Documents

Publication Publication Date Title
Chen et al. A deep learning method for bearing fault diagnosis based on cyclic spectral coherence and convolutional neural networks
CN111504676B (en) Equipment fault diagnosis method, device and system based on multi-source monitoring data fusion
CN114492534A (en) Construction method and application of cross-size motor bearing fault diagnosis model
CN112766342A (en) Abnormity detection method for electrical equipment
CN109946080B (en) Mechanical equipment health state identification method based on embedded circulation network
CN111562108A (en) Rolling bearing intelligent fault diagnosis method based on CNN and FCMC
CN111947928A (en) Multi-source information fusion bearing fault prediction system and method
CN114358124B (en) New fault diagnosis method for rotary machinery based on deep countermeasure convolutional neural network
CN114167838B (en) Multi-scale health assessment and fault prediction method for servo system
Son et al. Deep learning-based anomaly detection to classify inaccurate data and damaged condition of a cable-stayed bridge
CN115791174A (en) Rolling bearing abnormity diagnosis method and system, electronic equipment and storage medium
Prosvirin et al. Intelligent rubbing fault identification using multivariate signals and a multivariate one-dimensional convolutional neural network
CN115526515A (en) Safety monitoring system of gate for water conservancy and hydropower
CN113610188A (en) Bow net contact force non-section abnormity identification method and device
Chou et al. SHM data anomaly classification using machine learning strategies: A comparative study
CN116977708B (en) Bearing intelligent diagnosis method and system based on self-adaptive aggregation visual view
CN117591950A (en) Rolling bearing variable working condition fault diagnosis method, device, terminal and storage medium
Jian et al. LAFD-net: learning with noisy pseudo-labels for semisupervised bearing fault diagnosis
Castro-Cabrera et al. Adaptive classification using incremental learning for seismic-volcanic signals with concept drift
Lova et al. Intrusion detection toward feature reconstruction using Huber conditional variational AutoEncoder
Wu et al. Semi-supervised fault diagnosis of wheelset bearings in high-speed trains using autocorrelation and improved flow Gaussian mixture model
US20220187789A1 (en) Equipment failure diagnosis apparatus, equipment failure diagnosis method, smart factory system and application agent
CN112269778B (en) Equipment fault diagnosis method
CN114841196A (en) Mechanical equipment intelligent fault detection method and system based on supervised learning
CN114004144A (en) Method for predicting remaining service life of aircraft engine

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