CN106980822B - A kind of rotary machinery fault diagnosis method based on selective ensemble study - Google Patents
A kind of rotary machinery fault diagnosis method based on selective ensemble study Download PDFInfo
- Publication number
- CN106980822B CN106980822B CN201710148872.6A CN201710148872A CN106980822B CN 106980822 B CN106980822 B CN 106980822B CN 201710148872 A CN201710148872 A CN 201710148872A CN 106980822 B CN106980822 B CN 106980822B
- Authority
- CN
- China
- Prior art keywords
- pnn
- learning machine
- sequence
- weight
- final
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
- G06F17/148—Wavelet transforms
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Mathematical Optimization (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Databases & Information Systems (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Algebra (AREA)
- Evolutionary Biology (AREA)
- Signal Processing (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses a kind of rotary machinery fault diagnosis methods based on selective ensemble study, comprising: pre-processes to known input signal, obtains training sequence；The training sequence is handled using Bagging algorithm, generates a series of base learning machine of differentiation；Learnt by selective ensemble, a series of base learning machine of differentiation is carried out preferably, to obtain trained PSOSEN model；Rotary machinery fault diagnosis is carried out using trained PSOSEN model.The present invention solves the problems, such as a series of classifier existing in the prior art for being difficult to select better performances from classifiers.
Description
Technical field
The present invention relates to rotary machinery fault diagnosis field, in particular to a kind of selection based on adaptive particle swarm optimization
Property integrated study (particle swarm optimization based selective ensemble learning,
PSOSEN rotating machinery fault mode identification method and device).
Background technique
Fault diagnosis technology is the equipment diagnosis techniques risen with the development of modern industry mass production, is intended to
In the case where not dismantling substantially in equipment operation or when shutting down, by the methods of signal processing, pattern-recognition, the fortune of equipment is grasped
Row status determines position, reason, severity and the state of equipment fault, and then the prediction of realization equipment life and reliability,
Effective reference is provided for the formulation of maintenance decision.In recent years, it is examined using bearing, gear-box, pump class as the rotating machinery fault of representative
Disconnected technical research has obtained more extensive concern in academia and industry.
Rotating machinery is as the important component in Mechatronic Systems, reliable and stable fortune of the working condition to whole system
Row has tremendous influence.Due to the influence of high revolving speed and complex working condition, the troubleshooting issue of rotating machinery is always electromechanics
One important research direction of product rotating machinery health maintenance.Wherein, the fault diagnosis of rotating machinery usually may include spy
Sign is extracted and two parts of Fault Pattern Recognition.The purpose of feature extraction is using the methods of time and frequency domain analysis, from monitoring
The parameter for capableing of the current bearing malfunction of effective expression is obtained in signal, such as time domain index root mean square, set-back and frequency domain refer to
Wavelet coefficient etc. is marked, feature of these parameters as bearing fault is the basis of Fault Pattern Recognition.The mesh of Fault Pattern Recognition
Be the characteristic based on extraction, using effective classifier, the building block to break down in rotating machinery is accurately positioned,
Judging basis is made for maintenance and repair and reliability design.
As electronic product economy, reliability and security requirement are continuously improved, exist in recent years in order to meet operator
The demand of health control and maintenance aspect, gradually develops a series of relevant feature extractions and method for diagnosing faults.In feature
Selection aspect, the prior art are concentrated mainly on time and frequency domain analysis field, by statistical mathematics or frequency-domain analysis method, extract
It is capable of the signal index of characterization failure characteristic, typical method such as empirical mode decomposition (EMD) and wavelet analysis etc. out.EMD can
Any sophisticated signal is decomposed into the superposition of several basic friction angle components (IMF), and then wink is obtained by Hilbert-Huang transform
When amplitude and instantaneous frequency, be relatively suitble to time varying signal.Wavelet analysis is the expansion of Fourier transformation, can be one by signal integration
The combination of wavelets, the useful feature from frequency domain angle extraction to signal.In addition, with the development of information sensing technology, also
There are numerous deriveding analysis methods suitable for different objects.Such as local feature Scale Decomposition (LMD), compressed sensing (CS)
Deng, and characteristic optimization and dimension reduction method on the basis of feature extraction are produced, typical example such as principal component analysis (PCA) peace
Equal influence value assessment (MIV) etc..In terms of Fault Pattern Recognition, have numerous sides such as neural network, support vector machine (SVM)
Method, these methods obtain reasonable model parameter using the feature extracted, training, pass through the different events of identification rotary mechanical part
The otherness of barrier state, extraction obtain corresponding fault category label, realize fault distinguishing.
In view of the demand of fault diagnosis, integrated study is due to its significant advantage on generalization ability, in recent years
Area of pattern recognition has obtained rapid popularization, and achieves well on small sample, high-dimensional and complex data structures object
Application effect.Different from traditional Manufacturing resource theory of " dividing and rule ", integrated study can be by data reconstruction, training
The base learning machine of differentiation is obtained, then by certain convergence strategy, strengthens the mode identificating ability to the same data set.With it is a
Body algorithm is compared, and integrated study technology can be average by model integrated and error, improves the generalization ability of classification, is reduced quasi-
The risk of conjunction is achieved in fields such as image recognition, medical diagnosis, biological gene detections and is widely applied.However, collecting
At in study, research finds that base classifier number is excessive, and wherein the poor classifier of performance is possible to will lead to classification instead
Accuracy decline, and consume unnecessary storage resource.Therefore, researcher proposes the thought of selective ensemble study, i.e.,
By means such as optimization, cluster or sequences, the classifier of better performances is selected from the base learning machine of building, is collected instead of entirety
At.One advantage of selective ensemble study is that can be by only integrating the preferable learning machine of performance, with smaller resource
The precision and efficiency of diagnosis are strengthened in consumption.
Summary of the invention
In order to solve the problems, such as the complexity of rotating machinery data of the existing technology, the invention proposes one kind to be based on
The rotary machinery fault diagnosis method of selective ensemble study.
The rotary machinery fault diagnosis method based on selective ensemble study of the invention includes:
Known input signal is pre-processed, training sequence is obtained；
The training sequence is handled using Bagging algorithm, generates a series of base learning machine of differentiation；
Learn PSOSEN by selective ensemble, is obtained from a series of base learning machine of differentiation trained
PSOSEN model；
Rotary machinery fault diagnosis is carried out using trained PSOSEN model.
Wherein, described known input signal to be pre-processed to obtain training sequence and cycle tests includes:
Obtain the feature vector under different faults state, it is contemplated that the actual demand of fault diagnosis, this feature vector are usual
Refer to the time and frequency domain characteristics of monitoring data, such as root mean square, wavelet energy coefficient and entropy, to obtain training sequence.
Wherein, described that the training sequence is handled using Bagging algorithm, generate a series of base of differentiation
Learning machine includes:
The training sequence is reconstructed using the Bagging based on Bootstrap thought, generates base learning machine PNN
Network sequence, i.e., the base learning machine of a series of differentiation；
A series of weight sequence of the base learning machine of differentiation is initialized.
Wherein, the training sequence is reconstructed in the Bagging of the utilization based on Bootstrap thought, generates
Base learning machine PNN network sequence includes:
By Bagging process, a sequence to partially overlap is generated from original training data, it is different for training
Individual PNN network, and then base classifier set is improved to the ability to express of learning object.
Wherein, described to be learnt by selective ensemble, it is trained from a series of base learning machine of differentiation
Good PSOSEN model includes:
After the base learning machine generating process of differentiation, using improved optimization algorithm APSO, adapted to according to diagnosis
Degree demand constantly carries out optimizing to the weight vectors of PNN network sequence, obtains the final weight distribution D of PNN networkt；
D is distributed according to the final weightt, trained PSOSEN model is obtained from the PNN network sequence.
Wherein, described that D is distributed according to final weighttObtain trained PSOSEN model.Include:
D is distributed according to the final weightt, select final multiple PNN networks；
SVD decomposition is carried out to global optimum's matrix of final multiple PNN networks, is determined described final multiple
Optimal weight vector parameter corresponding to PNN network, to obtain trained PSOSEN model.
Wherein, described that D is distributed according to the final weightt, selecting final multiple PNN networks includes:
The final weight is distributed DtIt is compared with scheduled threshold value, obtains the final weight greater than threshold value；
Multiple PNN networks that its final weight in the PNN network sequence is greater than threshold value are elected to be described final multiple
PNN network.
Wherein, the improved optimization algorithm APSO is the optimization algorithm for introducing the rate of decay factor and the degree of polymerization factor
PSO。
Wherein, it is adjusted using the initial inertia weight of PSO described in the rate of decay factor and degree of polymerization factor pair,
Nonlinear Dynamic inertia is obtained to successively decrease weight.
Wherein, it carries out utilizing the Nonlinear Dynamic inertia in searching process in the weight vectors of the PNN network sequence
Weight of successively decreasing is adaptively adjusted the size of search space regions, avoids falling into the mistake of suboptimization because extent of polymerization is excessively high
Fit solution.
The method have the benefit that it is preferred by the base classifier based on APSO, it is selected from a series of classifiers
The better classifier of performance out, to solve the problems, such as the prior art.
Detailed description of the invention
Fig. 1 is the basic block diagram of PNN network；
Fig. 2 is PSOSEN Model Diagnosis flow chart；
Fig. 3 is rotary actuator working principle diagram；
Fig. 4 is RBF observer schematic diagram；
Fig. 5 is the residual error schematic diagram that five kinds of fault modes are calculated, wherein (a) figure indicates normal, (b) figure indicates servo
Valve failure, (c) figure indicates hydraulic motor failure, and (d) figure indicates rotary actuator failure, and (e) figure indicates axis failure；
Fig. 6 is the schematic diagram for showing the fault diagnosis result under five kinds of cases；
Fig. 7 is PSOSEN and GASEN optimization process schematic diagram；
Fig. 8 is the hydraulic pump test stand schematic diagram that the present invention uses；
Fig. 9 is the schematic diagram for showing different diagnosis algorithm Fault Diagnosis of Hydraulic Pump results；
Figure 10 is the schematic diagram for showing the signal-to-noise ratio distribution under different SNRs；
Figure 11 is the schematic diagram of the rotary machinery fault diagnosis method of the invention based on selective ensemble study.
Specific embodiment
1, the generation of base classifier
The present invention is classified using probabilistic neural network (probabilistic neural networks, PNN) as base
Device realizes the identification of rotating machinery fault mode.PNN is a kind of based on Bayesian Smallest Risk criterion and the differentiation of core fischer
Feedforward neural network is that D.F.Specht is invented in nineteen ninety.Under normal conditions, PNN network structure is as shown in Figure 1.
In Fig. 1, we can be found that PNN network can be generally divided into four layers: i.e. input layer, mode layer, cumulative layer and
Output layer.When giving one Test input signal of network, input layer can calculate in signal first data point and training to
The distance between amount point；Then, mode layer can utilize a radial basis function, convert distance to corresponding to different data point
Weighting parameter；In superimposed layer, each point in input data can be added up for the weighting parameter of different faults mode, and exported
One probability vector；Probability vector is converted expected label form by output layer, the final output as network.
We assume that an input vector is X=(x1, x2, x3, L xn) comprising two kinds of fault mode θAAnd θB, then PNN
The assorting process of network can be with general summary are as follows:
If hAlAfA(X) > hBlBfB(X), then X ∈ θA；
If hAlAfA(X) < hBlBfA(X), then X ∈ θB.
Wherein hAAnd hBRespectively indicate fault mode θAAnd θBPrior probability.It is assumed that NAAnd NBIt is the instruction in training data
Practice sample and is belonging respectively to θAAnd θBNumber, N is the sum of training sample, then hAAnd hBIt can be according to hA=NA/N,hB=NB/ N's
Mode is calculated.lAIt is the work factor in assorting process, indicates in training set, belong to fault mode θASample it is wrong
Misclassification is to fault mode θBRatio, lBIn contrast.fAAnd fBRespectively indicate fault mode θAAnd θBProbability density function.
According to the above theory, we can be found that PNN assorting process be similar to it is a kind of calculated using RBF kernel function it is changeable
The bayes classification method of lower different faults model probabilities Density Distribution situation is measured, with fault mode θAFor, it is corresponding general
Rate density function can be with is defined as:
In formula (1), X is input sample vector, and s is the variable number of sample vector.XAjFor j-th of training for belonging to A class
Sample vector is used as weight in probabilistic neural network, and m indicates the training sample vector number for belonging to A class, and δ is smooth coefficients.
2, the selective ensemble theories of learning
Individual PNN network depends on relatively good training input feature vector to be suitable for different decision boundaries to construct mostly
Mapping relations, this allows for its generalization ability and is significantly limited by manual features extraction and priori knowledge.In order to further
Its Generalization Capability is improved, present invention introduces the thoughts of integrated study, by constructing and integrating the individual PNN classifier of differentiation,
Improve the trouble diagnosibility of rotating machinery.In integrated study, it is important that a principle be exactly otherness, i.e., different
Should be relatively independent between body learning machine, and then the characteristic information implied in data is excavated from different perspectives, typical method is such as
Then Boosting and Bagging etc. obtains final prediction result by modes such as Nearest Neighbor with Weighted Voting.Rotating machinery fault is examined
For disconnected, the present invention constructs individual learning machine using the method for Bagging, by way of putting back to random sampling, building
It out include the PNN model of different characteristic information.Integrated study frame suitable for diagnosis can conclude are as follows:
For a mapping relationship f from input feature vector to fault mode: Rm→ D, D indicate the label of test set, have N number of
Participate in integrated individual PNN network.It is assumed that the learning error rate of each base learning machine is q, the then extensive mistake of the entirety of integrated study
Difference can indicate are as follows:
As shown in formula (2), when meeting two following conditions, integrated study can effectively improve the generalization of diagnosis
Can: (1) learning error rate less than 50%；(2) mutually indepedent between each base learning machine.However, if existing in integrated study
Excessive base classifier, it is possible to meaningless calculating and storage resource consumption can be made excessive.In order to solve this problem, it grinds
Study carefully personnel propose selective ensemble study concept, i.e., by certain means from the classifier of generation, select performance compared with
Good individual is final integrated to construct, to further promote Generalization Capability and computational efficiency in the case where guaranteeing precision.
In fault diagnosis, f is enabledijAnd djRespectively indicate the reality output that i-th of PNN network is directed to j-th of feature vector
And anticipated output.By taking three classification fault diagnosises as an example, fijAnd djFormat can be characterized as { 1,0,0 }, wherein the location of 1
Represent the corresponding fault mode type of rotating machinery.
So, i-th of PNN network is directed to the extensive error p of all feature vectorsiIt can be with is defined as:
Wherein, m indicates the quantity of feature, and Error (x) is an indicative function:
On this basis, for j-th of feature vector, the integrated result of N number of PNN network can be indicated are as follows:
ωiThe weight vectors for indicating PNN set, in most common expert's ballot method, it is believed that each weight is
1/N.Then it is directed to entire data set, the extensive error of integrated study can be with is defined as:
Formula (6) and (7) are based further on, under the premise of we can provide consideration guarantee classifier Generalization Capability, choosing
The condition of selecting property integrated study are as follows:
In formula (8), fkIndicate the bad base learning machine of the performance being removed, ωkIt is the corresponding power of these learning machines
Weight, dj-kIt is the corresponding label vector of remaining learning machine.The formula shows the classifier bad by rejecting part learning performance,
Diagnosis performance and efficiency can be further enhanced under the premise of not increasing integrated error.
3, the base classifier based on APSO is preferred
Although selective ensemble study has bigger advantage than all combining classifiers, how to be selected from classifier
The classifier of better performances out is still a current problem.In practical problem, the methods of hand picking and priori knowledge
Need to consume a large amount of computing resource.However, in rotary machinery fault diagnosis, due to the output of each PNN be it is identical, by
Formula (7) and formula (8) can obtain, and the preferred process of base learning machine may be considered one using integrated weight sequence as the excellent of object
Change process.Therefore, the invention proposes a kind of improved optimization algorithms (APSO) adaptively to assess each base learning machine
Performance, and select weight and gather greater than the PNN of threshold value, for constructing final integrated diagnosis learner.
The algorithm is a kind of improved population optimizing algorithm, has used for reference the groups such as birds and has determined most preferably in search of food
The process in path.These particles persistently move within a certain area, according to fitness function, determine the individual in searching process
Optimal location and group's optimal location.In each iterative process, particle can all move to a new position, and by the position
Starting point as optimizing next time.The fitness and speed of particle can be assessed with particle properties and be constantly updated.In the present invention,
Steps are as follows by basic APSO:
For N number of PNN network, it is assumed that the search space of L dimension, W=(W1,W2,…,WN) indicate population scale, WiIt is
I-th of particle in search space, may be defined as Wi=(ωi1,ωi2,…,ωiL)T, { ωij(i=1:L, j=1:N) expression
Integrated weight vectors to be optimized.Fitness function in diagnosis may be defined as:
Wherein, p and m respectively indicates the extensive error and number of features in diagnosis defined in step (2).F can consider
It is personal best particle Gbest in each iteration, and global optimum position Zbest is the individual comprising constantly updating
The matrix of optimal weight vector.The speed of searching optimization for enabling i-th of particle is Vi=(vi1,vi2,…,viL)T, rising in next update
Beginning position and speed can be calculated according to following formula:
Wherein, θ is an inertia weight, c1And c2It is the accelerated factor for adjusting Gbest and Zbest significance level, r1And r2
It is a random constant between 0-1.Inertia weight θ is a core parameter in optimization process, often passes through linear decrease plan
It slightly calculates, but frequently can lead to the case where it is easily trapped into local optimum.Therefore, the invention proposes one kind is closed based on polymerization
The adaptive optimization method of system.
Define F (Zbestt) be the t times iteration when optimization aim, a rate of decay factor kvIt may be defined as:
One polymerizing factor kaIt may be defined as:
WhereinIndicate the average fitness magnitude of particle.
In this way, successively decreasing weight in APSO we have proposed a kind of Nonlinear Dynamic inertia are as follows:
θ(t)nonlinear=θinitial-ωkv/kv+ωka/ka (14)
Wherein, θinitialIt is initial inertia weight, ωkvAnd ωkaIt is for kvAdjust kaThe incidence coefficient of significance level.It is logical
In normal situation, kvBigger, weight decline rate is bigger, and kaBigger, the extent of polymerization of particle is lower.On this basis, optimizing
Journey can be adaptively adjusted the size of search space regions according to convergence rate, improve Searching efficiency, while avoiding because of particle
Aggregation extent is excessively high and falls into the over-fitting situation of local optimum.In addition, ωkvAnd ωkaIt can be adjusted according to optimization demand dynamic
It is whole, so that kvAnd kaAn equalization point can be reached.
After APSO optimizing, the individual segregation that the weight parameter in Zbest is higher than given threshold λ can choose out
Device, for constructing final integrate.That is, weight is higher, corresponding base classifier performance is better.Determining base classification
After device subset, although normalizing to the weight vectors in Zbest, available integrated weight vectors, the method only reflects
A kind of possibility in searching process is easy to lose certain valuable information to a certain extent.In view of Gbest is all
The set of optimal weights vector in the number of iterations includes more abundant global information, and the present invention is by carrying out SVD to Gbest
Resolution process, the weight parameter being more applicable in.
It is assumed that having selected K PNN network, Gbest from N number of base classifierseBe this K classifier it is corresponding it is global most
Control matrix L × K, wherein every a line indicates that PNN is integrated into corresponding optimal weights vector in the secondary iteration.We are to GbestseInto
Row SVD is decomposed:
Gbestse=U Λ VT (15)
Wherein, Λ indicates singular value matrix.U and V is corresponding left and right orthogonal matrix respectively.Enable νmaxFor in right orthogonal matrix
In, maximum singular value smaxCorresponding first row feature vector.By νmaxIt is normalized, final integrated weight can be obtained are as follows:
Compared to Zbest is directly utilized, the advantage of this Weight Determination is to make by minimizing extensive error
Maximum probability degree must be classified close to expected sequence label, to further increase precision and the Shandong of rotary machinery fault diagnosis
Stick.Result output based on APSO selective ensemble are as follows:
4, learnt based on selective ensemble, the invention proposes a kind of rotary machinery fault diagnosis methods, can utilize more
The characteristic parameter of a angle strengthens the adaptive ability and accuracy of diagnosis.Entire diagnostic process is as shown in Figure 2.
Referring to fig. 2, disaggregated model may be considered from input space X to one of space Y mapping of output, mainly include with
Under several steps: (1) obtain different faults state under feature vector.In view of the actual demand of fault diagnosis, this feature to
Amount is often referred to the time and frequency domain characteristics of monitoring data, such as root mean square, wavelet energy coefficient and entropy；(2) individual of differentiation is generated
Learning machine.Bagging algorithm based on Bootstrap thought is used to initialization base classifier.Bootstrap is that one kind is based on
There is the data creation method for putting back to random resampling, it, can be from original training data X=(X by Bagging process1,X2,…,
Xn) one sequence to partially overlap of middle generationFor training different individual PNN networks, and then improve base
Ability to express of the classifier set to learning object.Each weight for generating classifier carries out initial according to approximate average principle
Change.(3) the PNN network of better performances is selected, and determines reasonable weight vectors.After pre-training process, the present invention
Optimizing constantly is carried out to weight vectors, obtains final weight distribution D according to diagnosis fitness demand using APSO methodt。
In APSO, speed and the degree of polymerization factor are introduced to improve global optimizing ability.In each iteration, DtMeeting continuous updating, most
The classifier that wherein weight is greater than given threshold is selected eventually.Later, SVD decomposition is carried out to global optimum's matrix, determines and participates in
Optimal weight vector parameter corresponding to integrated classifier.(4) rotating machinery event is carried out using the PSOSEN model that training is completed
Barrier diagnosis.
5, case study
Rotary actuator and plunger hydraulic pump are the critical components in rotating machinery, due to the influence of complex work environment,
It is easy to generate a plurality of types of failures.Therefore, the present invention verifies proposed PSOSEN method using two class object as case
Validity.
(1) rotary actuator fault diagnosis
1) data explanation
As shown in figure 3, the control module that rotary actuator mainly is responsible for adjusting angle of displacement based on the feedback signal by one, one
A PDU module and a planetary gear train executing agency composition.PDU module includes a servo valve, water brake mechanism, one
A hydraulic electric motor and a power and energy valve.Rotary variable differential transformer is one mutual for electric signal and angular displacement signal
The angle displacement sensor of conversion.In this experiment, it is filled with five kinds of fault modes: (1) normal condition altogether；(2) drive shaft event
Barrier；(3) rotary actuator failure；(4) servo valve failure；(5) hydraulic electric motor failure.Detailed trouble information is as follows.
Failure failure mode information
In view of rotary actuator system is a nonlinear feedback control system, it is difficult to determine accurate mathematical model,
A present invention Failure Observer as shown in Figure 4 according to RBF network struction really exports and it is expected defeated for obtaining system
Residual error between out, to preferably symbolize influence of the different faults state to rotary actuator system by residual error magnitude.
Observer model structure and the residual error of acquisition are as shown in Figure 5.
2) data prediction
Time-Frequency Analysis is most common signal processing method, and therefore, the present invention is mentioned for the residual signals proposed
Relevant time and frequency domain characteristics are taken, for obtaining under data fluctuations and complex working condition, there is the feature of more preferable identification to failure
Amount.The extracted time domain charactreristic parameter of the present invention is as follows.
1 Time-domain Statistics feature of table
Wavelet transformation is a kind of feature representation method that signal energy distribution is commonly described in time-frequency domain field, this hair
Bright extraction wavelet energy, maximum wavelet coefficient and Wavelet Entropy are as fault signature index, as shown in the table.
2 time-frequency domain statistical nature of table
In table, j indicates the number of plies of wavelet decomposition, and i indicates the sequence that the layer signal decomposes, and k is location tags, and s is SVD
Decompose the dimension of obtained diagnostic matrix.
In order to verify the generalization ability of proposed method, the training of different proportion and test data are separated from feature set
Out, as shown in table 4.In addition, the present invention is by carrying out training set in order to preferably evaluate performance individual in optimization process
The mode of Bootstrap sampling obtains a verifying collection.Training set is for initial training PNN network, and verifying collection is then used for
The optimization process of later period selection base classifier.
3 rotary actuator fault diagnosis of table: five kinds of diagnosis cases
3) the rotary actuator system fault diagnosis based on PSOSEN
In PSOSEN, we initialize 20 PNN base classifiers, and the training subset resampling ratio of each classifier is
80%.Tag along sort is set as binary form, i.e., (1,0,0,0,0), (0,1,0,0,0), (0,0,1,0,0), (0,0,0,1,
0) and (0,0,0,0,1) respectively corresponds normal, servo valve failure, hydraulic motor failure, pressurized strut failure and drive shaft failure.
In the base classifier choice phase, minimum and maximum search speed is set as+1 and -1, particle size 0-1, population
Scale is 100, the number of iterations 50, and the relative weighting of decline rate and the degree of polymerization factor is respectively 0.7 and 0.3.Each iteration
In all global optimum position Gbest and personal best particle Zbest are updated.The detail parameters of PSOSEN model such as following table
It is shown.
4 rotary actuator of table diagnoses PSOSEN model parameter
In present case, we, which set, selects threshold value as 0.2, that is, base classifier of the weight higher than 0.2 is used after selecting optimization
In building integrating finally.SVD decomposition is carried out to obtained global optimum's matrix, takes maximum singular value in right orthogonal matrix corresponding
Feature vector is integrated weight.Base classifier right-value optimization process is as shown in the table.
5 case 1:PSOSEN weight renewal process of table
The present invention is compared PSOSEN model and other methods.Control methods includes the selection using genetic algorithm
Property integrated study model GASEN, based on Boosting integrate multi-classification algorithm AdaBoost.M1, Bagging Integrated Algorithm and
A body Model PNN and SVM.Ten times of cross validation methods are used, in diagnosis with calculated diagnostic accuracy average value and mark
Quasi- difference is used as measurement index.Extensive error is calculated according to following formula:
Wherein N indicates the number of input sample.Error(xi) it is indicative function defined in formula (5).Difference training and
Under test ratio, the training of each algorithm and measuring accuracy are as shown in Figure 6.
1 diagnostic result of table 6Case
From diagnostic result as can be seen that from C1 to C5, with the promotion of training data ratio, the precision of diagnosis is also therewith
Increase.Compared with individual PNN and SVM classifier, Ensemble Learning Algorithms enhance the table to sample data by multi-angle study
Danone power, and possible negative effect in training process is eliminated, in most cases all have higher classification essence
Degree.By taking C3 as an example, the nicety of grading of PSOSEN, PNN, Bagging, AdaBoost.M1, GASEN and SVM is respectively 93.56%,
86.04%, 89.02%, 92.99%, 91.00% and 87.21%.In addition, integrated learning approach is than individual process with smaller
Standard deviation,
Namely has higher robustness in diagnosis.
In integrated study class method, as training data increases, integrated study class method has highest examine in C5
Disconnected precision 95.55%.The advantage of such Boosting algorithm is to adjust base classifier according to concatenation rules and incline to classification
To weight, and then improve diagnosis performance, reduce over-fitting risk.However, relative to PSOSEN model, AdaBoost.M1's
Standard deviation fluctuation is larger, and diagnosis performance is not sufficiently stable.In addition, PSOSEN and GASEN model is by rejecting performance bad base point
Class device has diagnosis capability more better than Bagging.In the present invention, APSO method can make up genetic algorithm in high dimension
Precocious problem is easy in, by taking C2 and C4 as an example, the optimization process of algorithm is as shown in Figure 7.
Referring to Fig. 7, in identical the number of iterations, APSO method has better convergence rate and ability of searching optimum.
For example, training data is less in C2, APSO can quickly converge to better fitness range than GA.In C4, although
It is preferable that GA has just started optimizing effect, but the case where increase with the number of iterations, fallen into local optimum, the parameter of fitness is not
It changes again, and APSO then shows a more lasting and stable searching process.
(2) Fault Diagnosis of Hydraulic Pump
1) case describes
Hydraulic pump test stand employed in present case is as shown in Figure 8.The testing stand mainly includes a plunger pump, and one
Hydraulic circuit and a driving motor.Motor speed is 5280r/min.Acceleration transducer is mounted on endface, for collecting
Vibration signal data, sample frequency 1024Hz.
Three kinds of fault modes, including normal, swash plate abrasion and valve plate rotor abrasion are filled in experiment altogether, is defined respectively
For fault mode S1, S2 and S3.
2) data prediction
In this experiment, we decomposed using EMD and comentropy in conjunction with method extract data characteristics.Firstly, original
Signal is broken down into a series of intrinsic modal components (IMFs), then calculates the comentropy of these IMFs, diagnoses as PSOSEN
The input of model.The basic procedure of data processing is as shown in table 8.
7 hydraulic pump data prediction process of table
It to each group of IMFs, chooses the first eight IE and carrys out construction feature matrix, that is to say, that altogether include 300 in feature set
Sample, each sample include 8 IE components altogether.80% characteristic quantity is selected at random as training set, and 20% is used as test set,
Final diagnostic result is obtained using the method for ten times of cross validations to hydraulic Pump data.
3) based on the Fault Diagnosis of Hydraulic Pump of PSOSEN
It is similar with rotary actuator system diagnostics, integrated PNN network number, which is set, as 20, bagging sample reconstructs ratio
It is 80%, particle size 0-1, population scale 100, the number of iterations 100, initial decline rate and degree of polymerization Factor minute
Not Wei 0.5 and 0.5, the faulty tag of normal, swash plate abrasion and valve plate rotor abrasion is respectively (1,0,0), (0,1,0) and
(0,0,1).Model detail parameters information is as shown in the table:
8 PSOSE model parameter of table
In integrated three steps calculated, the renewal process for integrating weight is as shown in the table.
9 case 2:PSOSEN right value update of table
Similarly, we compare and analyze PSOSEN model and other models in case 1.The hydraulic pump of each model
Diagnostic result is as shown in Figure 9.
It is red in the figure to indicate that expected label vector, blue indicate the label vector that actual classification obtains.It can observe
It arrives, although 90% or more, integrated study model ratio PNN has higher the precision of each model by optimizing training
Diagnostic accuracy, for example, PSOSEN category of model accuracy has reached 100%, it was demonstrated that its by it is efficiently integrated obtain it is non-thread
The ability of property fault characteristic.Significantly, since the superiority on small sample problem, SVM method is achieved and is integrated
Study similar diagnosis performance in hydraulic pump diagnostic data set.
In addition, it is contemplated that hydraulic pump is usually operated in the biggish environment of noise, it is easy to be influenced by X factor, from
And reduce diagnostic accuracy.The present invention has carried out noise pressure test in basic hydraulic pump data, to verify PSOSEN method
Robustness and Generalization Capability under influence of noise.The noise of 10dB, 30dB and 50dB are added to legacy data and concentrated by us, are used
The method of PCA maps the data into a two-dimensional space, as shown in Figure 10.
It can be seen in the drawing that the data obfuscation degree of different faults mode is also gradually increased as noise increases, it was demonstrated that
Noise can bring very important influence to classification task really, and diagnosis algorithm is also required to have higher robust to data
Property.The diagnostic result of all models is as shown in the table.
Diagnostic result under 10 difference SNRs of table
Although PSOSEN model is by reducing deviation and variance as can be seen that noisy influence from diagnostic result
It influences, and relies on better optimization process to improve hydraulic pump health status recognition capability, achieve more better than other models
Classifying quality.For example, noise it is relatively low be SNR10dB when, PSOSEN model still available 91.86% classification it is correct
Rate, it was demonstrated that the model can preferably handle the negative effect of grass.
In conclusion as shown in figure 11, the rotary machinery fault diagnosis method of the invention based on selective ensemble study
Include:
Known input signal is pre-processed, training sequence is obtained；
The training sequence is handled using Bagging algorithm, generates a series of base learning machine of differentiation；
Learn PSOSEN by selective ensemble, is obtained from a series of base learning machine of differentiation trained
PSOSEN model；
Rotary machinery fault diagnosis is carried out using trained PSOSEN model.
Wherein, described known input signal to be pre-processed to obtain training sequence and cycle tests includes:
Obtain the feature vector under different faults state, it is contemplated that the actual demand of fault diagnosis, this feature vector are usual
Refer to the time and frequency domain characteristics of monitoring data, such as root mean square, wavelet energy coefficient and entropy, to obtain training sequence.
Wherein, described that the training sequence is handled using Bagging algorithm, generate a series of base of differentiation
Learning machine includes:
The training sequence is reconstructed using the Bagging based on Bootstrap thought, generates base learning machine PNN
Network sequence, i.e., the base learning machine of a series of differentiation；
A series of weight sequence of the base learning machine of differentiation is initialized.
Wherein, the training sequence is reconstructed in the Bagging of the utilization based on Bootstrap thought, generates
Base learning machine PNN network sequence includes:
By Bagging process, a sequence to partially overlap is generated from original training data, it is different for training
Individual PNN network, and then base classifier set is improved to the ability to express of learning object.
Wherein, described to be learnt by selective ensemble, it is trained from a series of base learning machine of differentiation
Good PSOSEN model includes:
After the base learning machine generating process of differentiation, using improved optimization algorithm APSO, adapted to according to diagnosis
Degree demand constantly carries out optimizing to the weight vectors of PNN network sequence, obtains the final weight distribution D of PNN networkt；
D is distributed according to the final weightt, trained PSOSEN model is obtained from the PNN network sequence.
Wherein, described that D is distributed according to final weighttObtain trained PSOSEN model.Include:
D is distributed according to the final weightt, select final multiple PNN networks；
SVD decomposition is carried out to global optimum's matrix of final multiple PNN networks, is determined described final multiple
Optimal weight vector parameter corresponding to PNN network, to obtain trained PSOSEN model.
Wherein, described that D is distributed according to the final weightt, selecting final multiple PNN networks includes:
The final weight is distributed DtIt is compared with scheduled threshold value, obtains the final weight greater than threshold value；
Multiple PNN networks that its final weight in the PNN network sequence is greater than threshold value are elected to be described final multiple
PNN network.
Wherein, the improved optimization algorithm APSO is the optimization algorithm for introducing the rate of decay factor and the degree of polymerization factor
PSO。
Wherein, it is adjusted using the initial inertia weight of PSO described in the rate of decay factor and degree of polymerization factor pair,
Nonlinear Dynamic inertia is obtained to successively decrease weight.
Wherein, it carries out utilizing the Nonlinear Dynamic inertia in searching process in the weight vectors of the PNN network sequence
Weight of successively decreasing is adaptively adjusted the size of search space regions, avoids falling into the mistake of suboptimization because extent of polymerization is excessively high
Fit solution.
Although the present invention has been described above with particularity, however, the present invention is not limited thereto, those skilled in the art can
It is modified with principle according to the present invention, therefore, the various modifications that all principles according to the invention carry out all should be understood as
Fall into protection scope of the present invention.
Claims (8)
1. a kind of rotary machinery fault diagnosis method based on selective ensemble study, comprising:
Known rotating machinery input signal is pre-processed, training sequence is obtained comprising is obtained under different faults state
Feature vector, it is contemplated that the actual demand of fault diagnosis, this feature vector refer to the time and frequency domain characteristics of monitoring data, to be instructed
Practice sequence；
The training sequence is handled using Bagging algorithm, generates a series of base of the base learning machine as differentiation
Learning machine PNN network sequence；
Using improved optimization algorithm APSO, according to rotary machinery fault diagnosis fitness demand, constantly to the base learning machine
The weight vectors of PNN network sequence carry out optimizing, obtain the final weight distribution Dt of base learning machine PNN network sequence；
It is distributed Dt according to the final weight,
Trained PSOSEN model is obtained from the base learning machine PNN network sequence；And
By the way that rotating machinery input signal is sent into trained PSOSEN model, fault diagnosis is carried out to rotating machinery；
Wherein, the PNN network is probabilistic neural network；
Wherein, the improved optimization algorithm APSO is the optimization algorithms SO for introducing the rate of decay factor and the degree of polymerization factor.
2. according to the method described in claim 1, wherein, the time and frequency domain characteristics include root mean square, wavelet energy coefficient and entropy.
3. according to the method described in claim 1, wherein, it is described using Bagging algorithm to the training sequence at
Reason, generation include: as a series of base learning machine PNN network sequence of differentiation base learning machines
The training sequence is reconstructed using the Bagging based on Bootstrap thought, generates and is used as a series of differentiation
The base learning machine PNN network sequence of base learning machine；
A series of weight sequence of the base learning machine of differentiation is initialized.
4. according to the method described in claim 3, wherein, the Bagging of the utilization based on Bootstrap thought is to described
Training sequence is reconstructed, and generates base learning machine PNN network sequence and includes:
By Bagging process, a sequence to partially overlap is generated from original training data, for training different individuals
PNN network, and then base classifier set is improved to the ability to express of learning object.
It is described Dt is distributed according to final weight to obtain trained PSOSEN 5. according to the method described in claim 1, wherein
Model includes:
It is distributed Dt according to the final weight, selects final multiple PNN networks；
SVD decomposition is carried out to global optimum's matrix of final multiple PNN networks, determines final multiple PNN nets
Optimal weight vector parameter corresponding to network, to obtain trained PSOSEN model.
6. described to be distributed Dt according to the final weight according to the method described in claim 5, wherein, select final more
A base learning machine PNN network includes:
Final weight distribution Dt is compared with scheduled threshold value, obtains the final weight greater than threshold value；
Multiple base learning machine PNN networks that its final weight in the base learning machine PNN network sequence is greater than threshold value are elected to be institute
State final multiple base learning machine PNN networks.
7. according to the method described in claim 1, wherein, utilizing PSO described in the rate of decay factor and degree of polymerization factor pair
Initial inertia weight be adjusted, obtain Nonlinear Dynamic inertia and successively decrease weight.
8. it is carried out in searching process in the weight vectors of the PNN network sequence according to the method described in claim 7, wherein,
It is adaptively adjusted the sizes of search space regions using Nonlinear Dynamic inertia weight of successively decreasing, is avoided because of extent of polymerization mistake
Over-fitting situation that is high and falling into suboptimization.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710148872.6A CN106980822B (en) | 2017-03-14 | 2017-03-14 | A kind of rotary machinery fault diagnosis method based on selective ensemble study |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710148872.6A CN106980822B (en) | 2017-03-14 | 2017-03-14 | A kind of rotary machinery fault diagnosis method based on selective ensemble study |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106980822A CN106980822A (en) | 2017-07-25 |
| CN106980822B true CN106980822B (en) | 2019-04-05 |
Family
ID=59339453
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710148872.6A Active CN106980822B (en) | 2017-03-14 | 2017-03-14 | A kind of rotary machinery fault diagnosis method based on selective ensemble study |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106980822B (en) |
Families Citing this family (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107728602A (en) * | 2017-09-28 | 2018-02-23 | 合肥工业大学 | A kind of Personalized service method of hydroforming equipment failure |
| CN110197190B (en) * | 2018-02-27 | 2022-11-01 | 北京猎户星空科技有限公司 | Model training and object positioning method and device |
| CN108594788B (en) * | 2018-03-27 | 2020-09-22 | 西北工业大学 | Airplane actuator fault detection and diagnosis method based on deep random forest algorithm |
| CN108444715A (en) * | 2018-05-29 | 2018-08-24 | 内蒙古工业大学 | Bearing state diagnostic method, device, storage medium and electronic equipment |
| CN108986047B (en) * | 2018-07-13 | 2022-03-01 | 中国科学技术大学 | Image noise reduction method |
| CN108985380B (en) * | 2018-07-25 | 2021-08-03 | 西南交通大学 | Point switch fault identification method based on cluster integration |
| EP3605404B1 (en) * | 2018-08-03 | 2023-10-04 | Siemens Healthcare Diagnostics, Inc. | Method and device for the training of a machining learning routine for controlling a technical system |
| CN109376413B (en) * | 2018-10-08 | 2022-08-02 | 四川大学 | Online neural network fault diagnosis method based on data driving |
| CN110018678B (en) * | 2019-02-14 | 2021-05-14 | 电子科技大学 | Fault diagnosis method for networked automobile control system |
| CN110045608B (en) * | 2019-04-02 | 2022-04-05 | 太原理工大学 | Mechanical equipment part structure parameter dynamic optimization method based on digital twinning |
| CN110188383B (en) * | 2019-04-23 | 2021-01-05 | 华中科技大学 | Selective integration model-based power station SCR denitration modeling method |
| CN110163205B (en) * | 2019-05-06 | 2021-05-28 | 网易有道信息技术（北京）有限公司 | Image processing method, device, medium and computing equipment |
| CN111914613B (en) * | 2020-05-21 | 2024-03-01 | 淮阴工学院 | Multi-target tracking and facial feature information recognition method |
| CN111738309B (en) * | 2020-06-03 | 2024-07-12 | 哈尔滨工业大学 | Multi-scale analysis and integrated learning gas sensor fault mode identification method |
| CN113824580B (en) * | 2020-06-19 | 2024-08-16 | 北京市天元网络技术股份有限公司 | Network index early warning method and system |
| CN111783945B (en) * | 2020-06-22 | 2022-08-16 | 东北石油大学 | Long oil pipeline leakage real-time detection system based on transfer learning LSTM |
| CN112329804B (en) * | 2020-06-30 | 2024-10-15 | 中国石油大学(北京) | Feature-random-based naive Bayesian lithology classification integrated learning method and device |
| CN112257335B (en) * | 2020-10-10 | 2023-04-07 | 西南交通大学 | Oil-immersed transformer fault diagnosis method combining PNN and SVM |
| CN112307965B (en) * | 2020-10-30 | 2021-05-25 | 哈尔滨市科佳通用机电股份有限公司 | Rolling bearing sealing lock loss and fracture fault detection method |
| CN112541524B (en) * | 2020-11-18 | 2024-04-02 | 湖南大学 | BP-Adaboost multisource information motor fault diagnosis method based on attention mechanism improvement |
| CN113447273B (en) * | 2021-06-28 | 2022-08-05 | 哈尔滨工业大学 | Gas turbine sensor and actuating mechanism fault detection method based on cross validation |
| CN113642433B (en) * | 2021-07-30 | 2024-04-02 | 深圳市中金岭南有色金属股份有限公司凡口铅锌矿 | Method, device, terminal equipment and medium for determining working state of elevator |
| CN113591728A (en) * | 2021-08-03 | 2021-11-02 | 天津大学 | Electric energy quality disturbance classification method based on integrated deep learning |
| CN113820123A (en) * | 2021-08-18 | 2021-12-21 | 北京航空航天大学 | Gearbox fault diagnosis method based on improved CNN and selective integration |
| CN114021285B (en) * | 2021-11-17 | 2024-04-12 | 上海大学 | Rotary machine fault diagnosis method based on mutual local countermeasure migration learning |
| CN114118169A (en) * | 2021-12-08 | 2022-03-01 | 国网北京市电力公司 | Characteristic quantity extraction method, system, device and medium based on empirical mode decomposition |
| CN114722866B (en) * | 2022-03-28 | 2022-11-15 | 杭州中网智慧科技有限公司 | Management and control system is tracked to equipment based on industry internet |
| CN114893390B (en) * | 2022-07-15 | 2023-08-04 | 安徽云磬科技产业发展有限公司 | Pump equipment fault detection method based on attention and integrated learning mechanism |
| CN118413009B (en) * | 2024-07-03 | 2024-09-17 | 浙江优电云物联网技术有限公司 | Intelligent monitoring method and system for high-voltage fuse |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104484571A (en) * | 2014-12-22 | 2015-04-01 | 深圳先进技术研究院 | Integrated learning machine clip method and system based on edge distance sorting |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105550711A (en) * | 2015-12-16 | 2016-05-04 | 合肥工业大学 | Firefly algorithm based selective ensemble learning method |
-
2017
- 2017-03-14 CN CN201710148872.6A patent/CN106980822B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104484571A (en) * | 2014-12-22 | 2015-04-01 | 深圳先进技术研究院 | Integrated learning machine clip method and system based on edge distance sorting |
Non-Patent Citations (2)
| Title |
|---|
| 基于Bagging的概率神经网络集成分类算法;蒋芸 等;《计算机科学》;20130531;第40卷(第5期);第224页 |
| 高光谱影像集成学习分类及后处理技术研究;徐卫霄;《中国优秀硕士学位论文全文数据库 信息科技辑》;20120731(第2012年第07期);正文标题1.2.1、3.4.2、5.3.2、5.4.2内容 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106980822A (en) | 2017-07-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106980822B (en) | A kind of rotary machinery fault diagnosis method based on selective ensemble study | |
| Wu et al. | A hybrid classification autoencoder for semi-supervised fault diagnosis in rotating machinery | |
| Qin et al. | The optimized deep belief networks with improved logistic sigmoid units and their application in fault diagnosis for planetary gearboxes of wind turbines | |
| Han et al. | An adaptive spatiotemporal feature learning approach for fault diagnosis in complex systems | |
| Wang et al. | Modified multiscale weighted permutation entropy and optimized support vector machine method for rolling bearing fault diagnosis with complex signals | |
| Zhao et al. | Intelligent fault diagnosis of multichannel motor–rotor system based on multimanifold deep extreme learning machine | |
| Lei et al. | Fault diagnosis of wind turbine based on Long Short-term memory networks | |
| Li et al. | Fault diagnosis of rotating machinery based on combination of deep belief network and one-dimensional convolutional neural network | |
| Meng et al. | Data segmentation and augmentation methods based on raw data using deep neural networks approach for rotating machinery fault diagnosis | |
| CN109582003B (en) | Bearing fault diagnosis method based on pseudo label semi-supervised kernel local Fisher discriminant analysis | |
| Lu et al. | Deep model based domain adaptation for fault diagnosis | |
| Xiao et al. | Fault diagnosis of asynchronous motors based on LSTM neural network | |
| Liang et al. | Rolling bearing fault diagnosis based on one-dimensional dilated convolution network with residual connection | |
| Lu et al. | Dominant feature selection for the fault diagnosis of rotary machines using modified genetic algorithm and empirical mode decomposition | |
| Li et al. | Unsupervised rotating machinery fault diagnosis method based on integrated SAE–DBN and a binary processor | |
| CN109902399A (en) | Rolling bearing fault recognition methods under a kind of variable working condition based on ATT-CNN | |
| Ma et al. | Cross-domain meta learning fault diagnosis based on multi-scale dilated convolution and adaptive relation module | |
| Su et al. | Hierarchical diagnosis of bearing faults using branch convolutional neural network considering noise interference and variable working conditions | |
| CN109765333A (en) | A kind of Diagnosis Method of Transformer Faults based on GoogleNet model | |
| Wu et al. | A transformer-based approach for novel fault detection and fault classification/diagnosis in manufacturing: A rotary system application | |
| Li et al. | Multiscale dynamic fusion prototypical cluster network for fault diagnosis of planetary gearbox under few labeled samples | |
| Zhao et al. | Fault Diagnosis of Motor in Frequency Domain Signal by Stacked De-noising Auto-encoder. | |
| CN112883994A (en) | Rotating machinery variable working condition fault diagnosis method with balanced distribution adaptation | |
| CN114358124B (en) | New fault diagnosis method for rotary machinery based on deep countermeasure convolutional neural network | |
| Su et al. | A multi-level adaptation scheme for hierarchical bearing fault diagnosis under variable working conditions |
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 |