CN109186973A - A kind of mechanical failure diagnostic method of unsupervised deep learning network - Google Patents
A kind of mechanical failure diagnostic method of unsupervised deep learning network Download PDFInfo
- Publication number
- CN109186973A CN109186973A CN201810949099.8A CN201810949099A CN109186973A CN 109186973 A CN109186973 A CN 109186973A CN 201810949099 A CN201810949099 A CN 201810949099A CN 109186973 A CN109186973 A CN 109186973A
- Authority
- CN
- China
- Prior art keywords
- training sample
- training
- cluster
- udln
- fault
- 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.)
- Granted
Links
Classifications
-
- 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
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M7/00—Vibration-testing of structures; Shock-testing of structures
- G01M7/02—Vibration-testing by means of a shake table
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
Abstract
It the invention discloses a kind of mechanical failure diagnostic method of unsupervised deep learning network, comprises the following steps that (1) is installed corresponding sensor near the components such as the bearing of mechanical equipment and is acquired to vibration signal, obtains mechanical vibration signal;(2) hybrid domain fault signature data set is converted by the vibration signal of acquisition, and is divided into test and training sample character subset;(3) training sample character subset is input to study and training in constructed unsupervised deep learning network (UDLN) model, which is made of two layers of improved unsupervised feature extraction layer of sparseness filtering (L12SF) and one layer of weighted euclidean distance similar affine (WE-AP) cluster layer;(4) test sample is input to trained diagnostic model, realizes whole unsupervised feature learning and fault cluster.(5) its discrimination is calculated according to the subjection degree of test sample clustering, realizes fault identification and diagnosis.The invention is simple and feasible, and this method can carry out adaptive unsupervised fault diagnosis to all kinds of failures of mechanical equipment.
Description
Technical field
The present invention relates to the fault diagnosis technology field in industrial production, especially a kind of unsupervised deep learning network
Mechanical failure diagnostic method.
Background technique
Currently, industrial equipment just develops towards the direction of enlargement, high speed, automation.Mechanical equipment is set as industry
Most commonly seen component part, is widely used in the key areas such as Aeronautics and Astronautics, communications and transportation, intelligence manufacture in standby field.
At this stage, the components such as bearing are still mechanical most important power transmitting and support member, the according to statistics production of 30% mechanical breakdown
Life is that local damage occur by components such as bearings or defect causes, it is therefore necessary to be had to mechanical equipment and its key position
The condition monitoring and fault diagnosis of effect.
The moment such as mechanical equipment such as crane, rotating machinery generate the real time information for largely reflecting its operation conditions, this
A little information include various physical quantitys such as vibration, sound, temperature.Once there is exception in mechanical equipment, will bring corresponding
The variation of physical message.In order to describe the operation conditions of mechanical equipment from different angles, more mechanical informations need to be acquired, are led to
The analysis based on data or information is crossed, effective intellectual monitoring and diagnosis are carried out to mechanical equipment.Therefore, the purpose of fault diagnosis
It is exactly the mechanical information by obtaining to the effective diagnosis of equipment realization and analysis, to reduce or reduce mechanical equipment fault institute band
The loss or harm come.
Since mechanical equipment is monitored, and measuring point big with the equipment cluster scale of diagnosis is more, sample frequency is high, working life is gone through
Duration, thus monitoring and fault diagnosis system obtain be magnanimity data, cause mechanical health monitoring to enter with management domain " big
Data " the epoch.However, " mechanical big data " phenomenon equally brings mechanical equipment label information in actual industry spot
Obtain relatively difficult problem.This is because the runing time of mechanical equipment is far longer than the time of failure generation, mechanical data
Sparsity be inevitable, and handmarking's information is relatively difficult.For this purpose, the label information of mechanical data lacks and data
The sparsity of information brings a series of challenge to fault diagnosis.Thus, in the limited or sample label information that lacks training
In the case where, carrying out effective status monitoring with fault identification to mechanical equipment is that fault diagnosis field is all the time undecided
Critical issue.
Summary of the invention
A kind of unsupervised deep learning is provided the technical problem to be solved by the present invention is to overcome the deficiencies in the prior art
The mechanical failure diagnostic method of network, it is intended to realize that the whole process from feature extraction to the pattern-recognition stage of mechanical equipment is unsupervised
Fault diagnosis, this method can provide a kind of soluble effective scheme for the mechanical fault diagnosis under class label deletion condition.
The present invention uses following technical scheme to solve above-mentioned technical problem:
A kind of mechanical failure diagnostic method of the unsupervised deep learning network proposed according to the present invention, including following step
It is rapid:
Unit under test in step 1, preselect mechanical equipment, the vibration signal of unit under test in collection machinery equipment;
Step 2 converts the vibration signal of acquisition to hybrid domain fault signature data set, and is classified as test sample spy
Subset and training sample character subset are levied, test sample character subset is used as test sample, and training sample character subset is made
For training sample;
The parameter of step 3, the unsupervised deep learning network UDLN model of initialization, is input to UDLN model for training sample
Middle carry out pre-training, obtains the parameter of UDLN model;
Pre-training is specific as follows:
UDLN model is made of two stages study: firstly, it is special that training sample is input to L12 norm sparseness filtering L12SF
Extract layer is levied, L12 norm is to merge the norm expression of 1 norm and 2 norms, feature, which generated, using the L12SF competes special efficacy, thus
Extract characteristic value;Then, the characteristic value of the extraction is fed to the similar affine WE-AP cluster layer of weighted euclidean distance, obtains
The parameter of UDLN model after training;
Step 4, the UDLN model that test sample is input to training completion again, obtain fault cluster and recognition result;
Step 5, degree of membership situation and cluster centre according to fault cluster calculate all kinds of fault recognition rates, realize mechanical
The fault diagnosis of equipment.
As a kind of mechanical failure diagnostic method side of advanced optimizing of unsupervised deep learning network of the present invention
Case, the unit under test in step 1 includes crane, the bearing of rotating machinery, gear.
As a kind of mechanical failure diagnostic method side of advanced optimizing of unsupervised deep learning network of the present invention
Case, the feature of hybrid domain fault signature data set includes time domain, frequency domain and time and frequency domain characteristics in step 2.
As a kind of mechanical failure diagnostic method side of advanced optimizing of unsupervised deep learning network of the present invention
Case, the process that training sample is input to progress pre-training in UDLN model are as follows:
Step 3.1, the parameter for initializing UDLN model;
Step 3.2, being originally inputted using training sample character subset as two layers L12SF, successively extract hybrid domain failure
The unsupervised feature of the low-dimensional of characteristic data set;Wherein, the objective function of L12SF is as follows:
Wherein, | | * | |1Indicate 1 norm, | | * | |22 norms are indicated, for M training sampleIt indicates
First of characteristic value of i-th of training sample, xiIndicate i-th of training sample, wiIt is the weighting parameter of i-th of training sample,
It indicates to minimize wi, Wl TFor the transposition of the weighting parameter of the l characteristic value of UDLN model, | | * | |L12For 1 norm of fusion and 2 models
Several L12;Merge the L12 norm R (w) of 1 norm Yu 2 norms are as follows:
Wherein,For regularization norm adjustment factor, i-th new of training sample is extracted by optimization object function
L characteristic value fl iFor
fl i=G (Wl Txi)
Wherein, G (*) is activation primitive;
By fl i=G (Wl Txi) optimize the objective function of L12SF using L-BFGS algorithm until convergence;
The progress fault cluster study of cluster layer and failure of step 3.3, the characteristic value input WE-AP for extracting step 3.2
It divides, WE-AP cluster initialization first here has the Weighted Similarity matrix S of N number of training samplew
Wherein, x1kWith x2kIndicate k-th of feature of two different training samples, SkIndicate the variance of two training samples, n
It is characterized value number, N=M;
Meanwhile the confidence level and availability of i-th of training sample are calculated, specific calculation is
R (i, k)=S (i, k)-max { A (i, j)+S (i, j) }
St.j=1,2 ..., N and j ≠ i, k
St.j=1,2 ..., N and j ≠ i, k
R (k, k)=B (k)-max { A (k, j)+S (k, j) }
St.j=1,2 ..., N and j ≠ k
Wherein, B (k) is priori numerical value, represents the tendentiousness that each training sample is chosen as cluster centre point;S(i,k)
Indicate that the Weighted Similarity matrix of i-th of training sample and k-th of training sample, A (i, j) illustrate i-th of training sample choosing
Availability of j-th of training sample as its cluster centre is selected, S (i, j) indicates i-th of training sample and j-th of training sample
Weighted Similarity matrix, A (i, k) illustrates that i-th of training sample selection, k-th of training sample is the available of its cluster centre
Degree, A (i, j) illustrate availability of i-th of training sample selection, j-th of the training sample as its cluster centre, R (K, k) table
Confidence level of k-th of the training sample of k-th training sample selection as its cluster centre is shown, R (i, k) illustrates i-th of instruction
Practice confidence level of k-th of the training sample of samples selection as its cluster centre, R (k, k) illustrates k-th of training sample selection the
Confidence level of the k training sample as its cluster centre, A (k, j) illustrate k-th of training sample selection, j-th of training sample
As the availability of its cluster centre, S (k, j) indicates the Weighted Similarity square of k-th of training sample and j-th of training sample
Battle array.
As a kind of mechanical failure diagnostic method side of advanced optimizing of unsupervised deep learning network of the present invention
Case, in step 3.2, G (*) is the activation primitive using soft absolute value:
Wherein, σ is activation threshold, then willOptimize the mesh of L12SF using L-BFGS algorithm
Scalar functions are until convergence.
As a kind of mechanical failure diagnostic method side of advanced optimizing of unsupervised deep learning network of the present invention
Case, σ=10-8。
As a kind of mechanical failure diagnostic method side of advanced optimizing of unsupervised deep learning network of the present invention
Case calculates following formula according to UDLN model to determine condition and person in servitude of k-th of training sample as cluster centre in step 5
Category degree situation
R (k, k)+A (k, k) > 0
Wherein, R (k, k) indicates confidence level of k-th of training sample selection, k-th of the training sample as its cluster centre, A
(k, k) indicates availability of k-th of training sample selection, k-th of the training sample as its cluster centre;
According to preset maximum number of iterations tmax, update the confidence level and availability t of each training samplemaxIt is secondary, it updates
Mode is
R (i, k)=(1-lam) * R (i, k)+lam*R (i-1, k)
A (i, k)=(1-lam) * A (i, k)+lam*A (i-1, k)
Wherein, lam is damping factor, wherein R (i-1, k) indicates k-th of trained sample of (i-1) a training sample selection
This confidence level as its cluster centre, A (i-1, k) indicate k-th of training sample of (i-1) a training sample selection as it
The availability of cluster centre;
Finally, calculating all kinds of fault recognition rates according to the degree of membership situation and cluster centre of fault cluster, realize that machinery is set
Standby fault diagnosis.
As a kind of mechanical failure diagnostic method side of advanced optimizing of unsupervised deep learning network of the present invention
Case, the model parameter of the UDLN of step 3 include the weighting parameter w of i-th of training sampleiAnd regularization norm adjustment factor
The invention adopts the above technical scheme compared with prior art, has following technical effect that
(1) the improved unsupervised feature extraction network of L12 norm sparseness filtering (L12SF), can extract number of faults layer by layer
According to unsupervised special medical treatment information, the L12 norm newly defined enhances the generalization ability of the low-dimensional feature of extraction;
(2) improved weighting European similar affine (WE-AP) cluster can more accurately realize the adaptive of fault category
Unsupervised clustering is answered, traditional clustering method is overcome and needs to preset cluster numbers and cluster centre, while prominent different samples
Percentage contribution of the feature to cluster;
(3) on the basis of L12SF and WE-AP is clustered, a kind of completely new unsupervised deep learning network-is constructed
The whole unsupervised learning from feature extraction to pattern-recognition may be implemented in UDLN, the neural network;
(4) a kind of new mechanical failure diagnostic method based on UDLN model is invented, is realized under no class label information
Unsupervised intelligent machine fault diagnosis, and the feasible of mentioned inventive method is demonstrated by the vibration signal of mechanical equipment
Property.In addition, the invention is simple and feasible, the mechanical equipment in the case where lacking suitable for class label is online or field failure is examined
It is disconnected.
Detailed description of the invention
Fig. 1 is the flow chart of the technology of the present invention.
Fig. 2 is the structural schematic diagram of multilayer L12SF network.
Fig. 3 is UDLN model structure schematic diagram.
Fig. 4 is the time-domain and frequency-domain waveform envelope of all kinds of fault-signals of bearing.
Fig. 5 is the confusion matrix of the mechanical breakdown identification based on UDLN fault diagnosis.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with the accompanying drawings and the specific embodiments
The present invention will be described in detail.
In the diagnosis of actual industry spot, multilayer unsupervised learning can lack feelings for training sample and its class label
One of soluble scheme is provided under condition, the present invention is quasi- thus passes through proposed L12 norm sparseness filtering (L12SF) and weighting
A kind of new unsupervised deep learning network (UDLN) model of similar affine (WE-AP) joint mapping of Euclidean distance, and in this base
The mechanical failure diagnostic method based on UDLN model has been invented on plinth.
A kind of mechanical failure diagnostic method process of unsupervised deep learning network is as shown in Figure 1, step may be summarized as follows:
(1) is firstly, the key position in mechanical equipment (such as crane) arranges corresponding sensor to mechanical equipment
Vibration signal is picked up;
(2) vibration signal is converted multiple domain mixed fault characteristic data set by, and is classified as test sample character subset
With training sample character subset;
Currently, most common feature includes time domain, frequency domain, time and frequency domain characteristics etc., these features feature machinery well
The health status of fault data collection.Mixing multi-domain characteristics collection as shown in table 1 is constructed by bearing vibration signal sample, is denoted as feature
Data set H.
1 time domain of table, frequency spectrum, time-frequency domain statistical nature parameter
(3) training sample character subset is input in UDLN model and is trained study by initialization network parameter;
Specific method is described as follows:
(3.1) to the explanation of L12SF method
Sparseness filtering is considered as the unsupervised double-deck neural network, and compare other feature learning methods, sparse mistake
Filter needs not try to construct a model to data distribution, it only needs to optimize a simple 2 norm standard of loss function
Change the sparsity of feature.The given sample set with M training sample is as follows
fl i=Wl Txi
Wherein, fl iIndicate first of characteristic value of i-th of sample (column), Wl TIndicate turning for the weight matrix of l characteristic value
It sets.
Sparse filtering uses the sparsity of 2 norm normalization datas.These features form an eigenmatrix.We are first
2 norms of every a line of eigenmatrix in all samples are standardized.
Wherein,Indicate the l characteristic value of training sample after 2 norms standardize, flIndicate training sample l characteristic value.
Then each column are standardized according to its 2 norms
Wherein,Indicate normalized i-th of the sample of 2 norms,Indicate i-th of sample of 2 norm normalization characteristic values.
Finally, weight matrix WT lIt can be solved by optimizing cost function 1 norm of constraint of each sample, as follows.Standardization
It is characterized in the optimization by using 1 punishment as sparsity afterwards.So for the data for having M sample for one,
The objective function of given sparse filtering is:
Wherein, M indicates training sample number, | | * | |1Indicate 1 norm, | | * | |2Indicate 2 norms.Meanwhile it is understood that L1
Norm can make the more sparse of data, and 2 norms can prevent the over-fitting of data, the generalization ability of lift scheme.Therefore,
If the advantages of merging 1 norm and 2 norm, can sufficiently adjust the sparse characteristic of data and the Generalization Capability of neural network.
So the present invention redefines a kind of norm R (w):
Wherein,For regularization norm adjustment factor, wiFor the weighting parameter of i-th of training sample of model.On finally,
The change of formula objective function are as follows:
Wherein, | | * | |L12Indicate L12 norm, by the objective function in optimized-type, learning characteristic can be from input sample
The middle more nonlinear transformations of discovery, have preferable generalization ability.
fl i=G (Wl Txi)
Wherein, G (*) indicates the activation primitive of soft absolute value.So L12SF structural schematic diagram is as shown in Figure 2.
(3.2) to the explanation of WE-AP clustering method
That steps are as follows is shown for the specific calculating of WE-AP cluster
(1): calculating the similarity matrix S of weightingw, define greatest iteration tmax=1000.
(2): the confidence level R and availability A, specific calculation for calculating each sample point are
R (i, k)=S (i, k)-max { A (i, j)+S (i, j) }
St.j=1,2 ..., N and j ≠ i, k
St.j=1,2 ..., N and j ≠ i, k
R (k, k)=B (k)-max { A (k, j)+S (k, j) }
St.j=1,2 ..., N and j ≠ k
Wherein, B (k) is priori numerical value, represents the tendentiousness that each training sample is chosen as cluster centre point;S(k,j)
Indicate that the Weighted Similarity matrix of training sample, R (i, k) indicate that training sample i is suitable as the cluster centre of training sample k
Degree, A (k, j) illustrate the appropriateness that training sample k selects training sample j as its cluster centre.
(3): determining that can k-th of training sample as the condition of cluster centre point, for instruction according to confidence level is calculated
Practice sample itself, similarity numerical value is set as B (k):
R (k, k)+A (k, k) > 0
Wherein, R (k, k) indicates that training sample k is suitable as the degree of the cluster centre of training sample k, and A (k, k) is indicated
Training sample k selects the training sample k as the appropriateness of its cluster centre.
(4): updating confidence level R and availability A.Update mode is
R (i, k)=(1-lam) * R (i, k)+lam*R (i-1, k)
A (i, k)=(1-lam) * A (i, k)+lam*A (i-1, k)
Wherein, lam is damping factor, and R (i-1, k) indicates that training sample i-1 is suitable as the cluster centre of training sample k
Degree, A (i-1, k) illustrates the appropriateness that training sample i-1 selects training sample k as its cluster centre.It is acted on
It is the balance front and back confidence level R and availability A in iteration twice in order to avoid shaking.
(5): checking whether that meeting termination condition i.e. the number of iterations reaches greatest iteration tmax, 3.2 are jumped to if being unsatisfactory for
In step (2).
(3.3) to the explanation of UDLN method
UDLN is made of two stages study: firstly, modified L12 sparseness filtering (L12SF) is passed through in sample input, utilizing this
Algorithm extracts further feature expression;Then, which is admitted to based on WE-AP neural network, further by non-linear
Function Mapping establishes the inside distributed model of input data, the network weight of the entire UDLN model after being learnt.Knowing
When other equally first to test sample carry out sparseness filtering, then by the feature of extraction be input to it is trained based on WE-AP it
UDLN directly with classifying quality can be obtained afterwards.Relative to deep learning and its other method for diagnosing faults, entire model
Without class label, it can be achieved that whole unsupervised feature extraction and Fault Pattern Recognition.The training process of UDLN model
It is as follows: the feature extraction layer and one layer of unsupervised WE-AP cluster that the algorithm pre-training stage is made of two layers of L12SF structure
Composition.UDLN model structure is as shown in Figure 3.
Specific embodiment:
Case study on implementation 1: verifying the performance of the Mechanical Fault Diagnosis Model based on UDLN proposed in this section, needs
Simulate all kinds of rolling bearing faults.The experiment is that (ABLT-1A is studied by Hangzhou bearing test accelerating bearing life tester
Center provides), the chief component of experimental bench are as follows: computer control system, test headstock, experiment head, lubricating system, transmission
System, loading system, test and data collection system.It is mounted on the tester of design there are four bearing and is driven by alternating current generator
Axis on, transmission system is by rubber strip support for using two belt pulleys connection alternating current generators and axis.At the same time, it tests
Simulate all kinds of failures of 6205 single deep groove ball bearings.The failure that experiment simulates 6205 rolling bearings with wire cutting respectively has
Inner ring failure, Internal and external cycle combined failure, the compound Weak fault of Internal and external cycle, rolling element outer ring combined failure, rolling element outer ring are compound
The five class failure such as Weak fault.One group was acquired every 5 minutes by NI9234 data collecting card and PCB acceleration transducer
20480 points of vibration data.Sample frequency is 10240Hz, revolving speed 1050r/min.Each data collection system includes four
A acceleration transducer and data collecting card.The test condition of 6205 bearings is specific as shown in table 2 with test data explanation.
2 6205 bearing test experiments situation of table
In bearing fault data, having Weak fault in all kinds of bearing states also has catastrophe failure while also having combined failure
Etc. all kinds of fault conditions under different situations, the time domain of vibration signal and unilateral spectrum frequency-domain waveform figure are as shown in Figure 4.It can see
Out, it is influenced greatly relative to catastrophe failure, Weak fault vibration signal relative weak and the degree by noise jamming, impact is special
Property it is unobvious, resonance bands are also unobvious in frequency spectrum, and early-stage weak fault, which is difficult to observe, to be come.Traditional Time-frequency method amount of being difficult to
Change fault degree and classification, needs to rely on a large amount of expertise and field experience, lead to physical fault difficult diagnosis.Therefore,
Need a kind of intelligent method for diagnosing faults, to quantify fault diagnosis result, the intelligent failure diagnosis method based on machine learning
It is widely used.In order to show more diagnostic messages, Fig. 5 gives mentioned fault diagnosis model according to fault diagnosis flow scheme
For Fig. 1 to the processing result of bearing fault data characteristics collection H, the test sample confusion matrix of fault diagnosis result is as shown in Figure 5.From
Fig. 5 can be seen that proposed method and the test sample mistake of several second class samples be classified as the 4th class, and reason may be
Discrimination is unobvious between two class combined failures, and is all easy to produce and obscures comprising inner ring failure, other kinds all to cluster more
Success.
Feature extraction and mode identificating ability based on UDLN are mentioned for verifying.The present invention is by L12SF+AP, L12SF+
Kmeans;Comparative test of the L12SF+FCM as UDLN diagnostic model remembers { UDLN=FD1 respectively;L12SF+AP=FD2;
L12SF+Kmeans=FD3;L12SF+FCM=FD4 }.According to shown in Troubleshooting Flowchart 1 as above, using it is above-mentioned other
Diagnostic model is tested as a comparison, time domain, frequency domain, mixing characteristic of field is inputted above-mentioned 4 class fault diagnosis model respectively, according to person in servitude
Category degree sorts out that obtain corresponding recognition result as shown in table 3 with dividing:
All kinds of clustering recognition results of the table 3 based on different characteristic data set
To sum up, in order to enable knowledge and diagnostician experience of the intelligent trouble diagnosis independent of priori.The present invention proposes
A kind of mechanical failure diagnostic method being based on unsupervised learning neural network (ULDN).The ULDN model that this method is mentioned be by
L12SF and one layer of improvement AP cluster composition of two layers of improvement.In the method in ULDN, improving sparseness filtering can be adaptively with nothing
The mode of supervision has the feature of authentication information from vibration signal capture;Then these features improved AP is input to cluster,
And classified in unsupervised mode to health status.Show that this method can not only by the case study of bearing data set
Enough study can effectively realize unsupervised whole-course automation fault diagnosis to the high-level characteristic with identification.Unfavorable
In the case where label information, the advantages of this method can make full use of unsupervised learning, the accurate of mechanical fault diagnosis is improved
Property with automatic identification of defective health status.
The above description is merely a specific embodiment, but scope of protection of the present invention is not limited thereto, any
In the technical scope disclosed by the present invention, any changes or substitutions that can be easily thought of by those familiar with the art, all answers
It is included within the scope of protection of the present invention.
Claims (8)
1. a kind of mechanical failure diagnostic method of unsupervised deep learning network, which comprises the following steps:
Unit under test in step 1, preselect mechanical equipment, the vibration signal of unit under test in collection machinery equipment;
Step 2 converts the vibration signal of acquisition to hybrid domain fault signature data set, and is classified as test sample feature
Collection and training sample character subset, test sample character subset are used as test sample, and training sample character subset is i.e. as instruction
Practice sample;
The parameter of step 3, the unsupervised deep learning network UDLN model of initialization, by training sample be input in UDLN model into
Row pre-training obtains the parameter of UDLN model;
Pre-training is specific as follows:
UDLN model is made of two stages study: being mentioned firstly, training sample is input to L12 norm sparseness filtering L12SF feature
Layer is taken, L12 norm is to merge the norm expression of 1 norm and 2 norms, generates feature using the L12SF and competes special efficacy, to extract
To characteristic value;Then, the characteristic value of the extraction is fed to the similar affine WE-AP cluster layer of weighted euclidean distance, is trained
The parameter of UDLN model afterwards;
Step 4, the UDLN model that test sample is input to training completion again, obtain fault cluster and recognition result;
Step 5, degree of membership situation and cluster centre according to fault cluster calculate all kinds of fault recognition rates, realize mechanical equipment
Fault diagnosis.
2. a kind of mechanical failure diagnostic method of unsupervised deep learning network according to claim 1, which is characterized in that
Unit under test in step 1 includes crane, the bearing of rotating machinery, gear.
3. a kind of mechanical failure diagnostic method of unsupervised deep learning network according to claim 1, which is characterized in that
The feature of hybrid domain fault signature data set includes time domain, frequency domain and time and frequency domain characteristics in step 2.
4. a kind of mechanical failure diagnostic method of unsupervised deep learning network according to claim 1, which is characterized in that
The process that training sample is input to progress pre-training in UDLN model is as follows:
Step 3.1, the parameter for initializing UDLN model;
Step 3.2, being originally inputted using training sample character subset as two layers L12SF, successively extract hybrid domain fault signature
The unsupervised feature of the low-dimensional of data set;Wherein, the objective function of L12SF is as follows:
Wherein, | | * | |1Indicate 1 norm, | | * | |22 norms are indicated, for M training sample It indicates i-th
First of characteristic value of training sample, xiIndicate i-th of training sample, wiIt is the weighting parameter of i-th of training sample,It indicates
Minimize wi, Wl TFor the transposition of the weighting parameter of the l characteristic value of UDLN model, | | * | |L12For 1 norm of fusion and 2 norms
L12;Merge the L12 norm R (w) of 1 norm Yu 2 norms are as follows:
Wherein,For regularization norm adjustment factor, l of i-th new of training sample are extracted by optimization object function
Characteristic value fl iFor
fl i=G (Wl Txi)
Wherein, G (*) is activation primitive;
By fl i=G (Wl Txi) optimize the objective function of L12SF using L-BFGS algorithm until convergence;
The cluster layer progress fault cluster study of step 3.3, the characteristic value input WE-AP for extracting step 3.2 is divided with failure,
Here WE-AP cluster initialization first has the Weighted Similarity matrix S of N number of training samplew
Wherein, x1kWith x2kIndicate k-th of feature of two different training samples, SkIndicate the variance of two training samples, n is spy
Value indicative number, N=M;
Meanwhile the confidence level and availability of i-th of training sample are calculated, specific calculation is
R (i, k)=S (i, k)-max { A (i, j)+S (i, j) }
St.j=1,2 ..., Nandj ≠ i, k
St.j=1,2 ..., Nandj ≠ i, k
R (k, k)=B (k)-max { A (k, j)+S (k, j) }
St.j=1,2 ..., Nandj ≠ k
Wherein, B (k) is priori numerical value, represents the tendentiousness that each training sample is chosen as cluster centre point;S (i, k) is indicated
The Weighted Similarity matrix of i-th of training sample and k-th of training sample, A (i, j) illustrate i-th of training sample selection jth
Availability of a training sample as its cluster centre, S (i, j) indicate the weighting of i-th of training sample and j-th of training sample
Similarity matrix, A (i, k) illustrate that i-th of training sample selection, k-th of training sample is the availability of its cluster centre, A
(i, j) illustrates availability of i-th of training sample selection, j-th of the training sample as its cluster centre, and R (K, k) is illustrated
Confidence level of k-th of the training sample of k-th training sample selection as its cluster centre, R (i, k) illustrate i-th of trained sample
Confidence level of k-th of the training sample of this selection as its cluster centre, R (k, k) illustrate k-th of training sample selection k-th
Confidence level of the training sample as its cluster centre, A (k, j) illustrate k-th of training sample selection, j-th of training sample conduct
The availability of its cluster centre, S (k, j) indicate the Weighted Similarity matrix of k-th of training sample and j-th of training sample.
5. a kind of mechanical failure diagnostic method of unsupervised deep learning network according to claim 4, which is characterized in that
In step 3.2, G (*) is the activation primitive using soft absolute value:
Wherein, σ is activation threshold, then willOptimize the target letter of L12SF using L-BFGS algorithm
Number is until convergence.
6. a kind of mechanical failure diagnostic method of unsupervised deep learning network according to claim 5, which is characterized in that
σ=10-8。
7. a kind of mechanical failure diagnostic method of unsupervised deep learning network according to claim 1, which is characterized in that
Following formula is calculated to determine k-th of training sample as the condition of cluster centre and be subordinate to according to UDLN model in step 5
Spend situation
R (k, k)+A (k, k) > 0
Wherein, confidence level of R (k, k) expression k-th of training sample selection, k-th of the training sample as its cluster centre, A (k,
K) availability of k-th of training sample selection, k-th of the training sample as its cluster centre is indicated;
According to preset maximum number of iterations tmax, update the confidence level and availability t of each training samplemaxIt is secondary, update mode
For
R (i, k)=(1-lam) * R (i, k)+lam*R (i-1, k)
A (i, k)=(1-lam) * A (i, k)+lam*A (i-1, k)
Wherein, lam is damping factor, wherein R (i-1, k) indicates that k-th of training sample of (i-1) a training sample selection is made
For the confidence level of its cluster centre, A (i-1, k) indicates k-th of training sample of (i-1) a training sample selection as its cluster
The availability at center;
Finally, calculating all kinds of fault recognition rates according to the degree of membership situation and cluster centre of fault cluster, realizing mechanical equipment
Fault diagnosis.
8. a kind of mechanical failure diagnostic method of unsupervised deep learning network according to claim 1, which is characterized in that
The model parameter of the UDLN of step 3 includes the weighting parameter w of i-th of training sampleiAnd regularization norm adjustment factor
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810949099.8A CN109186973B (en) | 2018-08-20 | 2018-08-20 | Mechanical fault diagnosis method of unsupervised deep learning network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810949099.8A CN109186973B (en) | 2018-08-20 | 2018-08-20 | Mechanical fault diagnosis method of unsupervised deep learning network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109186973A true CN109186973A (en) | 2019-01-11 |
CN109186973B CN109186973B (en) | 2020-03-31 |
Family
ID=64918447
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810949099.8A Active CN109186973B (en) | 2018-08-20 | 2018-08-20 | Mechanical fault diagnosis method of unsupervised deep learning network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109186973B (en) |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110082106A (en) * | 2019-04-17 | 2019-08-02 | 武汉科技大学 | A kind of Method for Bearing Fault Diagnosis of the depth measure study based on Yu norm |
CN110132554A (en) * | 2019-04-17 | 2019-08-16 | 东南大学 | A kind of rotary machinery fault diagnosis method that depth Laplce encodes certainly |
CN110705114A (en) * | 2019-10-10 | 2020-01-17 | 辽宁工程技术大学 | Ventilation fault diagnosis method without training sample |
CN110849627A (en) * | 2019-11-27 | 2020-02-28 | 哈尔滨理工大学 | Width migration learning network and rolling bearing fault diagnosis method based on same |
CN110864887A (en) * | 2019-11-19 | 2020-03-06 | 北京瑞莱智慧科技有限公司 | Method, device, medium and computing equipment for determining operating condition of mechanical equipment |
CN111089720A (en) * | 2020-01-16 | 2020-05-01 | 山东科技大学 | Regularization sparse filtering method suitable for gear fault diagnosis under variable rotating speed |
CN111734669A (en) * | 2020-07-02 | 2020-10-02 | 重庆大学 | Multi-source information layered fusion centrifugal blower fault diagnosis method |
CN111811818A (en) * | 2020-06-02 | 2020-10-23 | 桂林电子科技大学 | Rolling bearing fault diagnosis method based on AP clustering algorithm of specified clustering number |
CN112131787A (en) * | 2020-09-18 | 2020-12-25 | 江西兰叶科技有限公司 | Unsupervised self-evolving motor design method and system |
CN112488983A (en) * | 2019-09-11 | 2021-03-12 | 中信戴卡股份有限公司 | Defect identification network obtaining method, defect identification method and grade determining method |
CN112631250A (en) * | 2020-12-15 | 2021-04-09 | 中国计量大学 | Fault isolation and identification method in nonlinear process based on denoising autoencoder |
CN113935443A (en) * | 2021-12-20 | 2022-01-14 | 广东电网有限责任公司广州供电局 | Heterogeneous equipment multi-domain joint fault prediction method, device, terminal and storage medium |
CN114414227A (en) * | 2021-12-29 | 2022-04-29 | 华电电力科学研究院有限公司 | Equipment collision sensing method, device, equipment and storage medium |
CN115096627A (en) * | 2022-06-16 | 2022-09-23 | 中南大学 | Method and system for fault diagnosis and operation and maintenance in manufacturing process of hydraulic forming intelligent equipment |
CN115563575A (en) * | 2022-09-28 | 2023-01-03 | 江苏省特种设备安全监督检验研究院 | Crane fault prediction method, system and storage medium based on neural network |
WO2023044978A1 (en) * | 2021-09-27 | 2023-03-30 | 苏州大学 | Adversarial-flow-model-based unsupervised fault diagnosis method for mechanical device |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102902961A (en) * | 2012-09-21 | 2013-01-30 | 武汉大学 | Face super-resolution processing method based on K neighbor sparse coding average value constraint |
CN104751191A (en) * | 2015-04-23 | 2015-07-01 | 重庆大学 | Sparse self-adaptive semi-supervised manifold learning hyperspectral image classification method |
CN105930629A (en) * | 2016-07-13 | 2016-09-07 | 大唐南京环保科技有限责任公司 | On-line fault diagnosis method based on massive amounts of operating data |
CN105975443A (en) * | 2016-05-04 | 2016-09-28 | 西南大学 | Lasso-based anomaly detection method and system |
CN106017876A (en) * | 2016-05-11 | 2016-10-12 | 西安交通大学 | Wheel set bearing fault diagnosis method based on equally-weighted local feature sparse filter network |
CN106295694A (en) * | 2016-08-05 | 2017-01-04 | 浙江工业大学 | A kind of face identification method of iteration weight set of constraints rarefaction representation classification |
CN106555788A (en) * | 2016-11-11 | 2017-04-05 | 河北工业大学 | Application of the deep learning based on Fuzzy Processing in hydraulic equipment fault diagnosis |
CN107944454A (en) * | 2017-11-08 | 2018-04-20 | 国网电力科学研究院武汉南瑞有限责任公司 | A kind of semanteme marking method based on machine learning for substation |
-
2018
- 2018-08-20 CN CN201810949099.8A patent/CN109186973B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102902961A (en) * | 2012-09-21 | 2013-01-30 | 武汉大学 | Face super-resolution processing method based on K neighbor sparse coding average value constraint |
CN104751191A (en) * | 2015-04-23 | 2015-07-01 | 重庆大学 | Sparse self-adaptive semi-supervised manifold learning hyperspectral image classification method |
CN105975443A (en) * | 2016-05-04 | 2016-09-28 | 西南大学 | Lasso-based anomaly detection method and system |
CN106017876A (en) * | 2016-05-11 | 2016-10-12 | 西安交通大学 | Wheel set bearing fault diagnosis method based on equally-weighted local feature sparse filter network |
CN105930629A (en) * | 2016-07-13 | 2016-09-07 | 大唐南京环保科技有限责任公司 | On-line fault diagnosis method based on massive amounts of operating data |
CN106295694A (en) * | 2016-08-05 | 2017-01-04 | 浙江工业大学 | A kind of face identification method of iteration weight set of constraints rarefaction representation classification |
CN106555788A (en) * | 2016-11-11 | 2017-04-05 | 河北工业大学 | Application of the deep learning based on Fuzzy Processing in hydraulic equipment fault diagnosis |
CN107944454A (en) * | 2017-11-08 | 2018-04-20 | 国网电力科学研究院武汉南瑞有限责任公司 | A kind of semanteme marking method based on machine learning for substation |
Cited By (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110082106A (en) * | 2019-04-17 | 2019-08-02 | 武汉科技大学 | A kind of Method for Bearing Fault Diagnosis of the depth measure study based on Yu norm |
CN110132554A (en) * | 2019-04-17 | 2019-08-16 | 东南大学 | A kind of rotary machinery fault diagnosis method that depth Laplce encodes certainly |
CN110082106B (en) * | 2019-04-17 | 2021-08-31 | 武汉科技大学 | Bearing fault diagnosis method based on Yu norm deep measurement learning |
CN110132554B (en) * | 2019-04-17 | 2020-10-09 | 东南大学 | Rotary machine fault diagnosis method based on deep Laplace self-coding |
CN112488983A (en) * | 2019-09-11 | 2021-03-12 | 中信戴卡股份有限公司 | Defect identification network obtaining method, defect identification method and grade determining method |
CN110705114B (en) * | 2019-10-10 | 2023-04-07 | 辽宁工程技术大学 | Ventilation fault diagnosis method without training sample |
CN110705114A (en) * | 2019-10-10 | 2020-01-17 | 辽宁工程技术大学 | Ventilation fault diagnosis method without training sample |
CN110864887A (en) * | 2019-11-19 | 2020-03-06 | 北京瑞莱智慧科技有限公司 | Method, device, medium and computing equipment for determining operating condition of mechanical equipment |
CN110849627A (en) * | 2019-11-27 | 2020-02-28 | 哈尔滨理工大学 | Width migration learning network and rolling bearing fault diagnosis method based on same |
CN111089720A (en) * | 2020-01-16 | 2020-05-01 | 山东科技大学 | Regularization sparse filtering method suitable for gear fault diagnosis under variable rotating speed |
CN111089720B (en) * | 2020-01-16 | 2021-10-08 | 山东科技大学 | Regularization sparse filtering method suitable for gear fault diagnosis under variable rotating speed |
CN111811818A (en) * | 2020-06-02 | 2020-10-23 | 桂林电子科技大学 | Rolling bearing fault diagnosis method based on AP clustering algorithm of specified clustering number |
CN111811818B (en) * | 2020-06-02 | 2022-02-01 | 桂林电子科技大学 | Rolling bearing fault diagnosis method based on AP clustering algorithm of specified clustering number |
CN111734669A (en) * | 2020-07-02 | 2020-10-02 | 重庆大学 | Multi-source information layered fusion centrifugal blower fault diagnosis method |
CN112131787B (en) * | 2020-09-18 | 2022-05-27 | 江西兰叶科技有限公司 | Unsupervised self-evolving motor design method and system |
CN112131787A (en) * | 2020-09-18 | 2020-12-25 | 江西兰叶科技有限公司 | Unsupervised self-evolving motor design method and system |
CN112631250A (en) * | 2020-12-15 | 2021-04-09 | 中国计量大学 | Fault isolation and identification method in nonlinear process based on denoising autoencoder |
WO2023044978A1 (en) * | 2021-09-27 | 2023-03-30 | 苏州大学 | Adversarial-flow-model-based unsupervised fault diagnosis method for mechanical device |
CN113935443A (en) * | 2021-12-20 | 2022-01-14 | 广东电网有限责任公司广州供电局 | Heterogeneous equipment multi-domain joint fault prediction method, device, terminal and storage medium |
CN114414227A (en) * | 2021-12-29 | 2022-04-29 | 华电电力科学研究院有限公司 | Equipment collision sensing method, device, equipment and storage medium |
CN114414227B (en) * | 2021-12-29 | 2023-05-26 | 华电电力科学研究院有限公司 | Equipment collision sensing method, device, equipment and storage medium |
CN115096627A (en) * | 2022-06-16 | 2022-09-23 | 中南大学 | Method and system for fault diagnosis and operation and maintenance in manufacturing process of hydraulic forming intelligent equipment |
CN115096627B (en) * | 2022-06-16 | 2023-04-07 | 中南大学 | Method and system for fault diagnosis and operation and maintenance in manufacturing process of hydraulic forming intelligent equipment |
CN115563575A (en) * | 2022-09-28 | 2023-01-03 | 江苏省特种设备安全监督检验研究院 | Crane fault prediction method, system and storage medium based on neural network |
CN115563575B (en) * | 2022-09-28 | 2023-04-21 | 江苏省特种设备安全监督检验研究院 | Crane fault prediction method, system and storage medium based on neural network |
Also Published As
Publication number | Publication date |
---|---|
CN109186973B (en) | 2020-03-31 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109186973A (en) | A kind of mechanical failure diagnostic method of unsupervised deep learning network | |
Han et al. | An adaptive spatiotemporal feature learning approach for fault diagnosis in complex systems | |
CN110110768B (en) | Rolling bearing fault diagnosis method based on parallel feature learning and multiple classifiers | |
CN109781411B (en) | Bearing fault diagnosis method combining improved sparse filter and KELM | |
CN106980822B (en) | A kind of rotary machinery fault diagnosis method based on selective ensemble study | |
CN110334764B (en) | Rotary machine intelligent fault diagnosis method based on integrated depth self-encoder | |
CN107631867B (en) | A kind of rotating machinery fault intelligent method for classifying based on deep learning | |
CN112257530B (en) | Rolling bearing fault diagnosis method based on blind signal separation and support vector machine | |
CN107505133A (en) | The probability intelligent diagnosing method of rolling bearing fault based on adaptive M RVM | |
CN112763214B (en) | Rolling bearing fault diagnosis method based on multi-label zero-sample learning | |
CN110674892A (en) | Fault feature screening method based on weighted multi-feature fusion and SVM classification | |
CN112304614B (en) | Intelligent fault diagnosis method for end-to-end rolling bearing by adopting multi-attention mechanism | |
CN105738109A (en) | Bearing fault classification diagnosis method based on sparse representation and ensemble learning | |
CN109858352A (en) | A kind of method for diagnosing faults based on compressed sensing and the multiple dimensioned network of improvement | |
CN110132554A (en) | A kind of rotary machinery fault diagnosis method that depth Laplce encodes certainly | |
CN110044623A (en) | The rolling bearing fault intelligent identification Method of empirical mode decomposition residual signal feature | |
CN106326915B (en) | A kind of Fault Diagnosis for Chemical Process method based on improvement core Fisher | |
CN113837000A (en) | Small sample fault diagnosis method based on task sequencing meta-learning | |
CN108509701A (en) | A kind of direct intelligent diagnosing method of rotating machinery fault based on vibration signal | |
CN111753891A (en) | Rolling bearing fault diagnosis method based on unsupervised feature learning | |
CN112819059A (en) | Rolling bearing fault diagnosis method based on popular retention transfer learning | |
CN114997211A (en) | Cross-working-condition fault diagnosis method based on improved countermeasure network and attention mechanism | |
CN114429152A (en) | Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption | |
CN113916535B (en) | Bearing diagnosis method, system, equipment and medium based on time frequency and CNN | |
CN113221946B (en) | Method for diagnosing fault types of mechanical equipment |
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 |