CN113191245B - Migration intelligent diagnosis method for multi-source rolling bearing health state fusion - Google Patents
Migration intelligent diagnosis method for multi-source rolling bearing health state fusion Download PDFInfo
- Publication number
- CN113191245B CN113191245B CN202110449135.6A CN202110449135A CN113191245B CN 113191245 B CN113191245 B CN 113191245B CN 202110449135 A CN202110449135 A CN 202110449135A CN 113191245 B CN113191245 B CN 113191245B
- Authority
- CN
- China
- Prior art keywords
- rolling bearing
- vibration signal
- signal sample
- sample set
- target
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000005096 rolling process Methods 0.000 title claims abstract description 259
- 230000036541 health Effects 0.000 title claims abstract description 77
- 238000003745 diagnosis Methods 0.000 title claims abstract description 58
- 238000013508 migration Methods 0.000 title claims abstract description 52
- 230000005012 migration Effects 0.000 title claims abstract description 52
- 238000000034 method Methods 0.000 title claims abstract description 34
- 230000004927 fusion Effects 0.000 title claims abstract description 13
- 238000009826 distribution Methods 0.000 claims abstract description 50
- 230000006978 adaptation Effects 0.000 claims abstract description 16
- 238000012549 training Methods 0.000 claims abstract description 14
- 230000006870 function Effects 0.000 claims description 19
- 230000004913 activation Effects 0.000 claims description 6
- 210000002569 neuron Anatomy 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 3
- 238000012360 testing method Methods 0.000 description 7
- 238000005070 sampling Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 3
- 238000013526 transfer learning Methods 0.000 description 3
- 230000015556 catabolic process Effects 0.000 description 2
- 238000006731 degradation reaction Methods 0.000 description 2
- 238000012631 diagnostic technique Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000003137 locomotive effect Effects 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 238000012546 transfer Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000004744 fabric Substances 0.000 description 1
- 238000009776 industrial production Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Cheminformatics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
Firstly, simultaneously training a plurality of local distribution adaptation submodels constructed by a single source rolling bearing-target rolling bearing vibration signal sample set pair to obtain a health state diagnosis result of the target rolling bearing based on each source rolling bearing, wherein for each local distribution adaptation submodel, the method comprises a domain sharing depth residual error network for extracting depth migration characteristics and a domain confusion network for extracting parameter sharing of the domain confusion characteristics; finally, through training of a multi-rolling bearing fusion migration diagnosis model, a final diagnosis result of the health state of a target rolling bearing vibration signal sample set is obtained by fusing diagnosis results of different source rolling bearings to the target rolling bearing; the method overcomes the influence that the diagnosis knowledge of the source rolling bearing cannot cover the fault type of the target rolling bearing and the unbalance of the vibration signal sample of the target rolling bearing, and obviously improves the diagnosis precision of the migration diagnosis model.
Description
Technical Field
The invention belongs to the technical field of rolling bearing fault diagnosis, and particularly relates to a migration intelligent diagnosis method for multi-source rolling bearing health state fusion.
Background
The rolling bearing is one of the key parts with the highest use frequency in various rotary mechanical equipment, plays an important role in the mechanical industrial production, is used as a mechanical 'joint' and bears the important work of maintaining the normal rotating function, so the health condition of the rolling bearing is directly related to the performance of the mechanical equipment. The working environment of mechanical equipment is complex and variable, so that the health condition of the rolling bearing is easy to generate faults along with the extension of the running time, and the research on the fault diagnosis technology of the rolling bearing is necessary for improving the safety and the reliability of the mechanical equipment.
In recent years, attention has been paid to intelligent fault diagnosis of rolling bearings, which can provide intuitive diagnosis results without relying too much on expert diagnosis knowledge, and particularly, with the rapid development of machine learning, deep learning is introduced into intelligent fault diagnosis of rolling bearings, and some excellent results are obtained. However, the basic assumption of the conventional rolling bearing intelligent diagnosis method is that the training data and the test data are required to obey the same probability distribution, and in many practical problems, the assumption cannot be met, so that the performance of the methods is reduced remarkably. The transfer learning method can alleviate this problem by realizing knowledge transfer between different fields, and therefore the rolling bearing fault diagnosis method based on the transfer learning has attracted much attention in recent years.
However, the existing rolling bearing migration diagnostic techniques have significant limitations: 1) The health state set of the source rolling bearing should overlap with the health state set of the target rolling bearing; 2) The number of target rolling bearing samples is balanced between the healthy states. However, in practical applications, such assumptions are difficult to hold: first, the target rolling bearing inevitably suffers from the type of failure that the source rolling bearing has never experienced, and therefore diagnostic knowledge from a single source rolling bearing is insufficient to identify target rolling bearing samples taken from all health states; secondly, the target rolling bearing is in a normal state for a long time in the life cycle, so the collected target rolling bearing data set consists of a large number of healthy samples and a small number of fault samples, and the two factors reduce the diagnosis precision of the traditional migration diagnosis technology on the faults of the rolling bearing.
The performance of existing migration intelligent diagnostic techniques is significantly degraded by the effect that diagnostic knowledge from the source rolling bearing may not cover all fault types of the target rolling bearing and target rolling bearing sample imbalances.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention aims to provide a migration intelligent diagnosis method for multi-source rolling bearing health state fusion, which improves the precision of the migration diagnosis of the rolling bearing under the condition that the diagnosis knowledge of the source rolling bearing possibly cannot cover all fault types of the target rolling bearing and the target rolling bearing sample is unbalanced, and promotes the practical application of the intelligent diagnosis technology.
In order to achieve the purpose, the invention adopts the technical scheme that:
a migration intelligent diagnosis method for multi-source rolling bearing health state fusion comprises the following steps:
1) Acquiring a plurality of source rolling bearing vibration signal sample sets and a target rolling bearing vibration signal sample set;
2) Constructing a local distribution adaptation sub-model by using a single-source rolling bearing-target rolling bearing vibration signal sample set pair;
3) Training a plurality of local distribution adapter models simultaneously;
4) Repeating the step 3 to the network parameter [ theta ] sk |s k E is converged by S;
5) And fusing the diagnosis results of the target rolling bearing based on different source rolling bearings to obtain the final diagnosis result of the health state of the target rolling bearing vibration signal sample set:
wherein,
representing the health state diagnosis result of the ith sample in the target rolling bearing vibration signal sample set,
denotes the s th k Status labels of the individual source rolling bearing vibration signal sample sets,
the step 1) is specifically as follows: obtaining N source rolling bearing vibration signal sample sets
Wherein it is present>
Denotes the s th k A sample set of vibration signals of a source rolling bearing, the sample set comprising samples of the amount
Denotes the s th k Status label of the individual source rolling bearing vibration signal sample set, based on the status label>
Wherein->
Denotes the th s k The total number of state types of the vibration signal sample set of the source rolling bearing; acquisition target rolling bearing vibration signal sample set>
Wherein->
A sample set of vibration signals representing a target rolling bearing, containing a sample number n t 。
The step 2) is specifically as follows: pairing N source rolling bearing vibration signal sample sets with a single target rolling bearing vibration signal sample set to form N pairs
And constructing N local distribution adaptation submodels which correspond to the N source rolling bearing-target rolling bearing vibration signal sample sets one to one.
The step 3) is specifically as follows: the following steps are executed to train the N local distribution adaptation submodels simultaneously to the s th k Vibration signal sample set pair of source rolling bearing and target rolling bearing
Each partial section is illustratedThe training process of the cloth adaptation submodel comprises the following steps:
3.1 Calculate the s-th k Deep migration fault characteristics of individual source rolling bearing-target rolling bearing vibration signal sample set pairs:
at the same time from the s k Method for extracting deep migration fault features from vibration signal samples of individual source rolling bearing and target rolling bearing in concentrated mode
Wherein it is present>
Is the s k Depth migration fault feature of ith sample of vibration signal sample set of individual source rolling bearing>
For the depth migration fault characteristics of the ith sample of the target rolling bearing vibration signal sample set, the upper/lower index F 2 F representing a domain-shared deep residual network 2 A layer;
3.2 Predict the probability distribution of the healthy state:
simultaneously inputting the extracted features into F 3 Layer, wherein F corresponds to the target rolling bearing 3 The number of layer neurons is
Denotes the s th k The total number of status categories of the individual rolling bearing vibration signal sample sets is then ^ H>
Each neuron represents the unknown health state of the target rolling bearing; prediction of domain-shared depth residual error network F by using Softmax activation function 3 The layer characteristics belong to the probability distribution of the healthy state of the vibration signal sample set of the source rolling bearing>
Probability distribution pertaining to an unknown health state of a target rolling bearing>
Wherein it is present>
Is the s k The ith sample in the vibration signal sample set of the rolling bearing belongs to the probability distribution of the healthy state of the vibration signal sample set of the source rolling bearing, and is matched with the healthy state of the vibration signal sample set of the source rolling bearing>
The probability distribution that the ith sample in the target rolling bearing vibration signal sample set belongs to the healthy state of the source rolling bearing signal set is determined, and the ith sample in the target rolling bearing vibration signal sample set belongs to the healthy state of the source rolling bearing signal set>
The ith sample in the target rolling bearing vibration signal sample set belongs to the probability distribution of unknown health state, and the upper/lower index F 3 Output layer F representing a domain-shared deep residual network 3 A layer; outputting a diagnosis that the target sample belongs to an unknown health state according to the following formula:
wherein m represents the sample size input by the network in each training;
the diagnosis result indicating that the ith sample in the target rolling bearing vibration signal sample set belongs to the unknown health state, 1 indicates that the ith sample belongs to the unknown health state, and 0 indicates that the ith sample does not belong to the unknown health state;
3.3 Compute domain shared depth residual network prediction loss:
3.3.1 Respectively calculating the cross entropy and the information entropy loss of the domain-shared depth residual error network prediction source rolling bearing vibration signal sample set and the target rolling bearing vibration signal sample set belonging to the health state of the source rolling bearing vibration signal sample set:
wherein,
represents the predicted s k The vibration signal sample set of the source rolling bearing belongs to the cross entropy loss of the health state of the vibration signal sample set of the source rolling bearing, and is subjected to judgment>
Representing the information entropy loss of the health state of the vibration signal sample set of the source rolling bearing belonging to the vibration signal sample set of the prediction target rolling bearing, j representing the j-th health state in the vibration signal sample set of the source rolling bearing, I (·) representing an indication function, and/or>
Denotes the s th k The ith sample in the vibration signal sample set of the source rolling bearing belongs to the diagnosis result of the health state of the vibration signal sample set of the source rolling bearing;
3.3.2 Computing cross entropy loss of a domain sharing depth residual error network prediction target rolling bearing vibration signal sample set belonging to an unknown health state:
wherein,
representing the cross entropy loss of predicting that the target rolling bearing vibration signal sample set belongs to an unknown health state;
3.3.3 Minimizing cross-entropy loss of the above equation, updating domain-shared depth residual network parameters
Namely: />
3.4 ) constructing a field confusion network with shared parameters and updating network parameters by iteration for a plurality of times:
constructing a field confusion network with shared parameters, wherein the parameter to be trained of the field confusion network is theta adv The input of the network is a deep migration fault feature
Output as a field obfuscating feature>
Wherein it is present>
Is the s k Field confusion characteristics of the ith sample in the vibration signal sample set of the individual source rolling bearing>
For the field confusion characteristics of the ith sample of the target rolling bearing vibration signal sample set, the upper/lower index adv represents a field confusion network; maximizing the parameter θ of the objective function update domain confusion network adv Namely:
iteratively updating n adv Sub-domain confusion network parameter θ adv After each iteration update, the parameter theta to be trained of the domain confusion network adv Truncation is in the range { - ξ, ξ };
3.5 Calculating local distribution difference of the depth migration fault characteristics of the source rolling bearing vibration signal sample set and the target rolling bearing vibration signal sample set:
3.5.1 Calculating a weighting coefficient of the target rolling bearing vibration signal sample set balance guide:
wherein,
a weighting factor representing the ith sample of the set of samples of vibration signals for the target rolling bearing>
For the field confusion feature of the ith sample of the target rolling bearing vibration signal sample set, based on the comparison result>
The ith sample in the target rolling bearing vibration signal sample set belongs to the probability distribution of unknown health state, and the upper/lower index F 3 Output layer (F) representing a domain-shared deep residual network 3 Layer), σ s () represents an activation function;
3.5.2 Weighting the deep migration fault characteristics of the target rolling bearing vibration signal sample set, and calculating the maximum mean difference of local distribution:
wherein,
representing a polynomial kernel maximum mean difference function;
3.6 Multi-rolling bearing fusion migration diagnosis model training:
3.6.1 Carrying out feature distribution adaptation on the depth migration fault features of the target rolling bearing learned based on the health state knowledge of the vibration signal sample sets of different source rolling bearings, and calculating the maximum mean difference as follows:
wherein S = { S = 1 ,s 2 ,...,s N Denotes N sample sets of source rolling bearing vibration signals,
are respectively expressed as being based on i 、s j Depth migration fault characteristics, upper/lower index F, of target rolling bearing vibration signal sample set extracted from individual source rolling bearing vibration signal sample set knowledge 2 F representing a domain-shared deep residual network 2 A layer;
3.6.2 Minimizing an objective function to simultaneously update parameter sets of an N-domain shared depth residual network
Namely: />
Where λ, β, γ are trade-off parameters.
The beneficial effects of the invention are as follows: the invention provides a migration intelligent diagnosis method for multi-source rolling bearing health state fusion, on one hand, the health state of a target rolling bearing is diagnosed by fusing the health state knowledge of a plurality of source rolling bearings, and the problem that the diagnosis knowledge of the source rolling bearing possibly cannot cover all fault types of the target rolling bearing is solved; on the other hand, the problem of unbalanced target rolling bearing samples is solved by calculating the weighting coefficient of the balance guide of the target rolling bearing vibration signal sample set and weighting the deep migration fault characteristics of the target rolling bearing vibration signal sample set in the characteristic distribution adaptation, and the diagnosis precision of the existing migration intelligent diagnosis technology is further improved.
Drawings
FIG. 1 is an overall flow chart of the present invention.
FIG. 2 is based on the s k And (3) constructing a local distribution adapter model frame diagram by the source rolling bearing-target rolling bearing vibration signal sample set.
FIG. 3 is a frame diagram of a multi-rolling bearing fusion migration diagnostic model.
Detailed Description
The invention is described in further detail below with reference to the figures and examples.
As shown in fig. 1, a migration intelligent diagnosis method for multi-source rolling bearing health state fusion comprises the following steps:
1) Obtaining a plurality of source rolling bearing vibration signal sample sets and a target rolling bearing vibration signal sample set:
obtaining N source rolling bearing vibration signal sample sets
Wherein +>
Denotes the s th k Individual source rolling bearing sample sets which contain a number of samples in->
Denotes the th s k A status label for each sample set of source rolling bearings,
wherein->
Denotes the th s k The total number of state types of the sample set of the source rolling bearings; acquiring vibration signal sample set of target rolling bearing>
Wherein->
Sample set, bag, representing target rolling bearingContaining a sample of n t ；
2) Constructing a local distribution adaptation submodel by using a single-source rolling bearing-target rolling bearing vibration signal sample set pair, as shown in FIG. 2;
pairing N source rolling bearing vibration signal sample sets with a single target rolling bearing vibration signal sample set to form N pairs
Constructing N local distribution adaptation submodels as shown in FIG. 2, and corresponding to the N source rolling bearing-target rolling bearing vibration signal sample sets in a one-to-one manner;
3) Training a plurality of local distribution adapter models simultaneously:
the following steps are executed to train the N local distribution adaptation submodels simultaneously to the s th k Vibration signal sample set pair of source rolling bearing and target rolling bearing
The training process of each local distribution adapter submodel is explained in detail, and comprises the following steps:
3.1 Calculate the s th k Deep migration fault characteristics of individual source rolling bearing-target rolling bearing vibration signal sample set pairs:
at the same time from the s k Method for extracting deep migration fault features from vibration signal samples of individual source rolling bearing and target rolling bearing in concentrated mode
Wherein it is present>
Is the th s k Depth migration fault feature of ith sample of vibration signal sample set of individual source rolling bearing>
For the depth migration fault characteristics of the ith sample of the target rolling bearing vibration signal sample set, the upper/lower index F 2 Representing domain shared depth residualsF of a bad network 2 A layer;
3.2 Predict the probability distribution of healthy states:
inputting the extracted features into F at the same time 3 Layer, wherein F corresponds to the target rolling bearing 3 The number of layer neurons is
Denotes the s th k The total number of status categories of the individual rolling bearing vibration signal sample sets is then ^ H>
Each neuron represents the unknown health state of the target rolling bearing; predicting domain-shared depth residual network F using Softmax activation function 3 The layer characteristics belong to the probability distribution of the healthy state of the vibration signal sample set of the source rolling bearing>
Probability distribution belonging to an unknown healthy state of a target rolling bearing>
Wherein it is present>
Is the th s k The ith sample in the vibration signal sample set of the rolling bearing belongs to the probability distribution of the healthy state of the vibration signal sample set of the source rolling bearing, and is matched with the healthy state of the vibration signal sample set of the source rolling bearing>
The probability distribution that the ith sample in the target rolling bearing vibration signal sample set belongs to the healthy state of the source rolling bearing signal set is determined, and the result is combined>
The ith sample in the target rolling bearing vibration signal sample set belongs to the probability distribution of unknown health state, and the upper/lower index F 3 Output layer F representing a domain-shared deep residual network 3 A layer; according to the following formula, output the objectiveThe standard sample belongs to the diagnosis result of unknown health state:
wherein m represents the sample size input by each training of the network;
the diagnosis result indicating that the ith sample in the target rolling bearing vibration signal sample set belongs to the unknown health state, 1 indicates that the ith sample belongs to the unknown health state, and 0 indicates that the ith sample does not belong to the unknown health state;
3.3 Computing domain shared depth residual network prediction loss:
3.3.1 Respectively calculating the cross entropy and the information entropy loss of the domain-shared depth residual error network prediction source rolling bearing vibration signal sample set and the target rolling bearing vibration signal sample set belonging to the health state of the source rolling bearing vibration signal sample set:
wherein,
represents the predicted s k The vibration signal sample set of the source rolling bearing belongs to the cross entropy loss of the health state of the vibration signal sample set of the source rolling bearing, and is subjected to judgment>
Information entropy loss representing the health state of a predicted target rolling bearing vibration signal sample set belonging to a source rolling bearing vibration signal sample set, j represents the jth health state in the source rolling bearing vibration signal sample set, I (·) represents an indication function, and/or>
Denotes the s th k The ith sample in the vibration signal sample set of the source rolling bearing belongs to the diagnosis result of the health state of the vibration signal sample set of the source rolling bearing;
3.3.2 Computing cross entropy loss of a domain-shared depth residual error network prediction target rolling bearing vibration signal sample set belonging to an unknown health state:
wherein,
representing the cross entropy loss of the prediction target rolling bearing vibration signal sample set belonging to the unknown health state;
3.3.3 Minimizing cross-entropy loss of the above equation, updating domain-shared depth residual network parameters
Namely:
3.4 ) construct a domain confusion network for parameter sharing and update network parameters for multiple iterations:
constructing a field confusion network with shared parameters, wherein the parameter to be trained of the field confusion network is theta adv The input of the network is a deep migration fault feature
Output as a field obfuscating feature>
Wherein +>
Is the s k Field confusion characteristics of the ith sample in the vibration signal sample set of the individual source rolling bearing>
For the field confusion characteristics of the ith sample of the target rolling bearing vibration signal sample set, the upper/lower index adv represents a field confusion network; maximizing the parameter θ of the objective function update domain confusion network adv Namely:
iteratively updating n adv Sub-domain confusion network parameter θ adv After each iteration update, the parameter theta to be trained of the domain confusion network adv Truncation is in the range { - ξ, ξ };
3.5 Calculating the local distribution difference of the depth migration fault characteristics of the source rolling bearing vibration signal sample set and the target rolling bearing vibration signal sample set:
3.5.1 Calculating a weighting coefficient of the target rolling bearing vibration signal sample set balance guide:
wherein,
a weighting factor representing the ith sample of the set of samples of the vibration signal of the target rolling bearing, < >>
For the field confusion feature of the ith sample of the target rolling bearing vibration signal sample set, based on the comparison result>
Belongs to unknown health state for ith sample in target rolling bearing vibration signal sample setProbability distribution, upper/lower index F 3 Output layer (F) representing a domain-shared deep residual network 3 Layer), σ s (. -) represents an activation function;
3.5.2 Weighting the deep migration fault characteristics of the target rolling bearing vibration signal sample set, and calculating the maximum mean difference of local distribution:
wherein,
representing a polynomial kernel maximum mean difference function;
3.6 As shown in fig. 3), the multi-rolling bearing fusion migration diagnostic model trains:
3.6.1 Carrying out feature distribution adaptation on the depth migration fault features of the target rolling bearing learned based on the health state knowledge of the vibration signal sample sets of different source rolling bearings, and calculating the maximum mean difference as follows:
wherein S = { S = 1 ,s 2 ,...,s N Denotes N sample sets of source rolling bearing vibration signals,
are respectively expressed as being based on i 、s j Depth migration fault characteristics, upper/lower index F, of target rolling bearing vibration signal sample set extracted from individual source rolling bearing vibration signal sample set knowledge 2 F representing a domain-shared deep residual network 2 A layer;
3.6.2 Minimizing an objective function to simultaneously update parameter sets of an N-domain shared depth residual network
Namely:
where λ, β, γ are trade-off parameters.
4) Repeating the step 3) to the network parameters
Converging;
5) And fusing the diagnosis results of the target rolling bearing based on different source rolling bearings to obtain the final diagnosis result of the health state of the target rolling bearing vibration signal sample set:
wherein,
representing the health state diagnosis result of the ith sample in the target rolling bearing vibration signal sample set,
denotes the th s k Status labels of the individual source rolling bearing vibration signal sample sets,
the embodiment is as follows: and the feasibility of the invention is verified by applying three groups of rolling bearing vibration signal sample sets.
Data set D comes from an accelerated life test stand OF a rolling bearing, an accelerometer is used for monitoring the degradation condition OF a tested bearing (LDK UER 204), during the test, the speed OF an alternating current motor for driving the tested bearing is set to 2400r/min, a hydraulic cylinder provides 11kN radial load on an outer ring OF the bearing, a sampling frequency is set to 25.6kHz, an experimenter stops and inspects the tested bearing after the test stand fails, cracks OF the outer ring OF the bearing are confirmed, therefore, collected data have degradation information from a normal (N) stage to an outer ring failure (OF) stage, data OF the whole early life stage OF the bearing in a healthy state and data OF a final stage OF the bearing in a failure state are marked, 724 samples exist in the data set D, the 724 samples are balanced in the healthy state and the failure state, and each sample comprises 1200 sampling points.
Data set E was provided by the mechanical failure prevention technical association, which was equipped with RBC NICE bearings that were healthy or contained inner ring failure (IF), engineers simulated rolling bearing inner ring wear in the laboratory, input shaft speed was set to 1500r/min, normal samples were taken under a load of 270lbs at a sampling frequency of 97.656kHz, IF samples were taken under loads of 200lbs, 250lbs and 300lbs at a sampling frequency of 48.828kHz. Data set E contains 732 samples, with a balanced number of samples.
Data set F comes from a locomotive bearing automatic test bench (197726) which can directly test a locomotive wheel pair provided with a bearing, the tested bearing is driven by a hydraulic motor and loaded by a hydraulic cylinder, and three wheel pairs are provided: the test rotating speed of a common wheel set and the wheel set with a polished surface on the inner ring or the outer ring of a bearing are 450r/min, the radial load is 680kg, a vibration sensor collects monitoring data at the sampling frequency of 76.8kHz, as shown in Table 1, a data set F is composed of three samples in health states, wherein 832 samples in health states and 1664 x m% in fault states, attention is paid to that m belongs to% and (0,1) enables the data set F to be in different imbalance levels, and detailed information is shown in Table 1:
TABLE 1 detailed information of three sets of bearing data sets
Note that: the m% ∈ (0,1) error sample is randomly selected to unbalance data set F.
A multi-source transfer learning task (D, E) → F is created to validate the proposed method. It is used to simulate the transfer of diagnostic knowledge from the laboratory to the real world situation, because there are not enough labeled samples at all times in the engineering scenario, the source bearing dataset D or dataset E cannot provide sufficient diagnostic knowledge required for the target bearing dataset F, and furthermore, the samples in dataset F are unbalanced between health states, taking m =20, i.e. dataset F has a sample imbalance of 20%. F-scan, mean average precision (mAP) and AUC classification indexes are selected to quantify the effect of the method on the migration diagnosis task, the experiment is repeated for 10 times, and the statistical value of the diagnosis result is calculated, as shown in Table 2, the three indexes of the method are 0.936, 0.983 and 0.968 respectively, the indexes are close to 1, which shows that the method has high diagnosis accuracy, and the feasibility of the method in the migration diagnosis under the condition that the source domain health state type cannot cover the target domain health state type and the target domain sample is unbalanced in engineering practice is verified.
TABLE 2 comparison of the diagnostic results of the different methods
And comparing the effect of the method of the invention with that of a common ResNet method by selecting P-ResNet, wherein the average precision mean value of P-ResNet is 0.516, which is close to a random diagnosis model, F-score and AUC are both obviously lower than that of the method of the invention, and the average precision mean value, F-score and AUC of the common ResNet method are the lowest among the three methods.
By comparing the method with a common migration diagnosis method (P-ResNet) and a common deep intelligent diagnosis method (ResNet), the method provided by the invention is shown to effectively overcome the influence that the source domain health state type cannot cover the target domain health state type and the target domain sample is unbalanced, and improve the performance of a migration diagnosis model.
Claims (1)
1. A multi-source rolling bearing health state fusion migration intelligent diagnosis method is characterized by comprising the following steps:
1) Acquiring a plurality of source rolling bearing vibration signal sample sets and a target bearing vibration signal sample set;
2) Constructing a local distribution adaptation sub-model by using a single-source rolling bearing-target rolling bearing vibration signal sample set pair;
3) Training a plurality of local distribution adapter models simultaneously;
4) Repeating the step 3 to the network parameters
Converging;
5) And fusing the diagnosis results of the target rolling bearing based on different source rolling bearings to obtain the final diagnosis result of the health state of the target rolling bearing vibration signal sample set:
wherein,
representing the health state diagnosis result of the ith sample in the target rolling bearing vibration signal sample set,
denotes the th s k A status signature of the individual source rolling bearing vibration signal sample set,
the step 1) is specifically as follows: obtaining N source rolling bearing vibration signal sample sets
Wherein +>
Denotes the th s k A sample set of vibration signals of a source rolling bearing, which sample set comprises samples in a quantity of->
Denotes the th s k Status label of the individual source rolling bearing vibration signal sample set, based on the status label>
Wherein +>
Denotes the th s k The state category total number of the vibration signal sample set of the source rolling bearing; acquiring a vibration signal sample set of a target rolling bearing>
Wherein +>
A sample set of vibration signals representing a target rolling bearing, containing a sample number n t ；
The step 2) is specifically as follows: pairing N source rolling bearing vibration signal sample sets with a single target rolling bearing vibration signal sample set to form N pairs
Constructing N local distribution adapter submodels which are in one-to-one correspondence with the N source rolling bearing-target rolling bearing vibration signal sample sets;
the step 3) is specifically as follows: the following steps are executed to train the N local distribution adaptation submodels simultaneously to the s th k Individual source rolling bearing-target rolling bearing vibration signal sample set pair
The training process of each local distribution adaptation submodel is explained, and comprises the following steps:
3.1 Calculate the s th k Deep migration fault characteristics of individual source rolling bearing-target rolling bearing vibration signal sample set pairs:
at the same time from the s k Source rolling bearing and target rolling shaftMethod for extracting deep migration fault features from vibration bearing signal samples in concentrated mode
Wherein +>
Is the s k Depth migration fault feature of ith sample of vibration signal sample set of individual source rolling bearing>
For the depth migration fault characteristics of the ith sample of the target rolling bearing vibration signal sample set, the upper/lower index F 2 F representing a domain-shared deep residual network 2 A layer;
3.2 Predict the probability distribution of healthy states:
inputting the extracted features into F at the same time 3 Layer, wherein F corresponds to the target rolling bearing 3 The number of layer neurons is
Denotes the s th k The total number of status categories of individual rolling bearing vibration signal sample sets>
Each neuron represents the unknown health state of the target rolling bearing; predicting domain-shared depth residual network F using Softmax activation function 3 Layer characteristics belong to a probability distribution in the healthy state of a sample set of vibration signals of a source rolling bearing>
Probability distribution pertaining to an unknown health state of a target rolling bearing>
Wherein +>
Is the th s k The ith sample in the vibration signal sample set of the rolling bearing belongs to the probability distribution of the healthy state of the vibration signal sample set of the source rolling bearing, and the probability distribution is combined>
The probability distribution that the ith sample in the target rolling bearing vibration signal sample set belongs to the healthy state of the source rolling bearing signal set is determined, and the ith sample in the target rolling bearing vibration signal sample set belongs to the healthy state of the source rolling bearing signal set>
The ith sample in the target rolling bearing vibration signal sample set belongs to the probability distribution of unknown health state, and the upper/lower index F 3 Output layer F representing a domain-shared deep residual network 3 A layer; outputting a diagnosis that the target sample belongs to an unknown health state according to the following formula:
wherein m represents the sample size input by the network in each training;
the diagnosis result that the ith sample in the target rolling bearing vibration signal sample set belongs to the unknown health state is shown, 1 represents that the ith sample belongs to the unknown health state, and 0 represents that the ith sample does not belong to the unknown health state;
3.3 Compute domain shared depth residual network prediction loss:
3.3.1 Respectively calculating cross entropy and information entropy loss of a domain sharing depth residual error network prediction source rolling bearing vibration signal sample set and a target rolling bearing vibration signal sample set belonging to the health state of the source rolling bearing vibration signal sample set:
wherein,
represents the predicted s k The vibration signal sample set of the source rolling bearing belongs to the cross entropy loss of the health state of the vibration signal sample set of the source rolling bearing, and is subjected to judgment>
Information entropy loss representing the health state of a predicted target rolling bearing vibration signal sample set belonging to a source rolling bearing vibration signal sample set, j represents the jth health state in the source rolling bearing vibration signal sample set, I (·) represents an indication function, and/or>
Denotes the s th k The ith sample in the vibration signal sample set of the source rolling bearing belongs to the diagnosis result of the health state of the vibration signal sample set of the source rolling bearing;
3.3.2 Computing cross entropy loss of a domain sharing depth residual error network prediction target rolling bearing vibration signal sample set belonging to an unknown health state:
wherein,
representing the cross entropy loss of predicting that the target rolling bearing vibration signal sample set belongs to an unknown health state;
3.3.3)minimizing cross entropy loss in the above manner, updating domain shared depth residual network parameters
Namely: />
3.4 ) construct a domain confusion network for parameter sharing and update network parameters for multiple iterations:
constructing a field confusion network with shared parameters, wherein the parameter to be trained of the field confusion network is theta adv The input of the network is a deep migration fault feature
Output as a field obfuscation feature>
Wherein it is present>
Is the th s k Field confusion characteristics of the ith sample in the vibration signal sample set of the individual source rolling bearing>
For the field confusion characteristics of the ith sample of the target rolling bearing vibration signal sample set, the upper/lower mark adv represents a field confusion network; maximizing the parameter θ of the objective function update domain confusion network adv Namely:
iteratively updating n adv Sub-domain confusion network parameter θ adv After each iteration update, the parameter theta to be trained of the domain confusion network adv Truncation is in the range { - ξ, ξ };
3.5 Calculating the local distribution difference of the depth migration fault characteristics of the source rolling bearing vibration signal sample set and the target rolling bearing vibration signal sample set:
3.5.1 Calculating a weighting coefficient of the target rolling bearing vibration signal sample set balance guide:
wherein,
a weighting factor representing the ith sample of the set of samples of the vibration signal of the target rolling bearing, < >>
For the field confusion feature of the ith sample of the target rolling bearing vibration signal sample set, based on the comparison result>
The ith sample in the target rolling bearing vibration signal sample set belongs to the probability distribution of unknown health state, and the upper/lower index F 3 Output layer F representing a domain-shared deep residual network 3 Layer, σ s (. -) represents an activation function;
3.5.2 Weighting the deep migration fault characteristics of the target rolling bearing vibration signal sample set, and calculating the maximum mean difference of local distribution:
wherein,
representing a polynomial kernel maximum mean difference function;
3.6 Multiple rolling bearing fusion migration diagnostic model training:
3.6.1 Carrying out feature distribution adaptation on the depth migration fault features of the target rolling bearing learned based on the health state knowledge of the vibration signal sample sets of different source rolling bearings, and calculating the maximum mean difference as follows:
wherein S = { S = 1 ,s 2 ,...,s N Denotes N sample sets of source rolling bearing vibration signals,
are respectively expressed as being based on i 、s j Depth migration fault characteristics, upper/lower index F, of target rolling bearing vibration signal sample set extracted from individual source rolling bearing vibration signal sample set knowledge 2 F representing a domain-shared deep residual network 2 A layer;
3.6.2 Minimizing an objective function to simultaneously update parameter sets of an N-domain shared depth residual network
Namely: />
Where λ, β, γ are trade-off parameters.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110449135.6A CN113191245B (en) | 2021-04-25 | 2021-04-25 | Migration intelligent diagnosis method for multi-source rolling bearing health state fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110449135.6A CN113191245B (en) | 2021-04-25 | 2021-04-25 | Migration intelligent diagnosis method for multi-source rolling bearing health state fusion |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113191245A CN113191245A (en) | 2021-07-30 |
| CN113191245B true CN113191245B (en) | 2023-04-11 |
Family
ID=76978831
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110449135.6A Active CN113191245B (en) | 2021-04-25 | 2021-04-25 | Migration intelligent diagnosis method for multi-source rolling bearing health state fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113191245B (en) |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105758644A (en) * | 2016-05-16 | 2016-07-13 | 上海电力学院 | Rolling bearing fault diagnosis method based on variation mode decomposition and permutation entropy |
| CN106596107B (en) * | 2016-12-27 | 2019-03-26 | 华南理工大学 | A kind of flexibility precise thin-wall bearing failure diagnosis life test machine |
| CN110285969B (en) * | 2019-07-10 | 2020-05-26 | 西安交通大学 | Rolling bearing fault migration diagnosis method with polynomial nuclear implantation characteristic distribution adaptation |
| CN111337256B (en) * | 2020-03-27 | 2020-12-29 | 西安交通大学 | Method for diagnosing fault depth local migration of rolling bearing weighted by domain asymmetry factor |
| CN112308147B (en) * | 2020-11-02 | 2024-02-09 | 西安电子科技大学 | Rotary machinery fault diagnosis method based on multi-source domain anchor adapter integrated migration |
| AU2020103681A4 (en) * | 2020-11-26 | 2021-02-04 | Anhui University Of Technology | Rolling Bearing Fault Diagnosis Method Based on Fourier Decomposition and Multi-scale Arrangement Entropy Partial Mean Value |
-
2021
- 2021-04-25 CN CN202110449135.6A patent/CN113191245B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN113191245A (en) | 2021-07-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110285969B (en) | Rolling bearing fault migration diagnosis method with polynomial nuclear implantation characteristic distribution adaptation | |
| CN110428004B (en) | Mechanical part fault diagnosis method based on deep learning under data imbalance | |
| CN110361176B (en) | Intelligent fault diagnosis method based on multitask feature sharing neural network | |
| Nasiri et al. | Fracture mechanics and mechanical fault detection by artificial intelligence methods: A review | |
| CN109376620A (en) | A kind of migration diagnostic method of gearbox of wind turbine failure | |
| CN110108456A (en) | A kind of rotating machinery health evaluating method of depth convolutional neural networks | |
| CN113128338B (en) | Intelligent diagnosis method for printer roller faults under small sample | |
| CN113065581B (en) | Vibration fault migration diagnosis method for reactance domain self-adaptive network based on parameter sharing | |
| CN111046945A (en) | Fault type and damage degree diagnosis method based on combined convolutional neural network | |
| CN112163640A (en) | Planetary gearbox fault diagnosis method and system based on deep convolution migration learning | |
| CN111337256A (en) | Method for diagnosing fault depth local migration of rolling bearing weighted by domain asymmetry factor | |
| CN113188794B (en) | Gearbox fault diagnosis method and device based on improved PSO-BP neural network | |
| CN113076834B (en) | Rotating machine fault information processing method, processing system, processing terminal, and medium | |
| CN114429152A (en) | Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption | |
| CN114755017B (en) | Variable-speed bearing fault diagnosis method of cross-domain data driving unsupervised field shared network | |
| CN115859077A (en) | Multi-feature fusion motor small sample fault diagnosis method under variable working conditions | |
| CN109635879B (en) | Coal mining machine fault diagnosis system with optimal parameters | |
| CN111860839A (en) | Shore bridge fault monitoring method based on multi-signal fusion and Adam optimization algorithm | |
| CN104677629A (en) | Fault detection method for vehicle transmission | |
| CN111814728A (en) | Method for recognizing wear state of cutting tool of numerical control machine tool and storage medium | |
| CN115587290A (en) | Aero-engine fault diagnosis method based on variational self-coding generation countermeasure network | |
| Wang et al. | A graph neural network-based data cleaning method to prevent intelligent fault diagnosis from data contamination | |
| CN115876467A (en) | Pseudo label transfer type two-stage field self-adaptive rolling bearing fault diagnosis method | |
| CN101539137A (en) | Residual life gray prediction method of aerospace hydraulic pump based on delta filter | |
| CN113191245B (en) | Migration intelligent diagnosis method for multi-source rolling bearing health state fusion |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |