CN106980822A - A kind of rotary machinery fault diagnosis method learnt based on selective ensemble - Google Patents
A kind of rotary machinery fault diagnosis method learnt based on selective ensemble Download PDFInfo
- Publication number
- CN106980822A CN106980822A CN201710148872.6A CN201710148872A CN106980822A CN 106980822 A CN106980822 A CN 106980822A CN 201710148872 A CN201710148872 A CN 201710148872A CN 106980822 A CN106980822 A CN 106980822A
- Authority
- CN
- China
- Prior art keywords
- pnn
- weight
- psosen
- learning machine
- 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.)
- Granted
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 method learnt based on selective ensemble, including：Known input signal is pre-processed, training sequence is obtained；The training sequence is handled using Bagging algorithms, a series of base learning machine of differentiation is generated；Learnt by selective ensemble, preferred, the PSOSEN models trained are carried out to a series of base learning machine of differentiation；Rotary machinery fault diagnosis is carried out using the PSOSEN models trained.The present invention solves a series of problem for the grader for being difficult to select better performances present in prior art from graders.
Description
Technical field
The present invention relates to rotary machinery fault diagnosis field, more particularly 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 technology
Fault diagnosis technology is the equipment diagnosis techniques risen with the development of the big production of modern industry, and it is intended to
In the case of not dismantled substantially in equipment operation or when shutting down, by methods such as signal transacting, pattern-recognitions, the fortune of equipment is grasped
Row present situation, judges position, reason, the order of severity and the state of equipment fault, and then realizes the prediction of equipment life and reliability,
Effective reference is provided for the formulation of maintenance decision.In recent years, 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 industrial quarters.
Rotating machinery is used as the important component in Mechatronic Systems, reliable and stable fortune of its working condition to whole system
Row has tremendous influence.Due to the influence of high rotating 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 can generally include spy
Levy extraction and two parts of Fault Pattern Recognition.The purpose of feature extraction is using methods such as time and frequency domain analysis, from monitoring
Obtain being capable of the parameter of the current bearing malfunction of effective expression in signal, such as time domain index root mean square, set-back and frequency domain refer to
Wavelet coefficient etc. is marked, these parameters, as the feature of bearing fault, are the bases of Fault Pattern Recognition.The mesh of Fault Pattern Recognition
Be the characteristic based on extraction, using effective grader, be accurately positioned the building block broken down in rotating machinery,
It is that judging basis are made in maintenance and repair and reliability design.
As electronic product economy, reliability and security requirement are improved constantly, exist in recent years in order to meet operator
Demand in terms of health control and maintenance, gradually develops a series of related feature extractions and method for diagnosing faults.In feature
In terms of selection, prior art is concentrated mainly on time and frequency domain analysis field, by statistical mathematics or frequency-domain analysis method, extracts
Go out the signal index for being capable of characterization failure characteristic, typical method such as empirical mode decomposition (EMD) and wavelet analysis etc..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 adapted 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
Occur in that numerous deriveding analysis methods suitable for different objects.Such as local feature Scale Decomposition (LMD), compressed sensing (CS)
Deng, and generate the characteristic optimization on the basis of feature extraction and dimension reduction method, exemplary such as principal component analysis (PCA) peace
Equal influence value is assessed (MIV) etc..In terms of Fault Pattern Recognition, numerous sides such as existing neutral net, SVMs (SVM)
Method, these methods obtain rational model parameter using the feature extracted, training, by recognizing the different events of rotary mechanical part
The otherness of barrier state, extraction obtains corresponding fault category label, realizes 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 is achieved well on small sample, high-dimensional and complex data structures object
Application effect.Different from the Manufacturing resource theory of traditional " dividing and rule ", integrated study can be trained by data reconstruction
The base learning machine of differentiation is obtained, then by certain convergence strategy, strengthens the mode identificating ability to same data set.With it is individual
Body algorithm is compared, and integrated study technology can be average by model integrated and error, improves the generalization ability of classification, reduced plan
The risk of conjunction, it achieves in fields such as image recognition, medical diagnosis, biological gene detections and is widely applied.However, in collection
Into in study, research finds that base grader number is excessive, and wherein the grader of poor-performing is possible to that classification can be caused on the contrary
Accuracy declines, and consumes unnecessary storage resource.Therefore, researcher proposes the thought of selective ensemble study, i.e.,
By means such as optimization, cluster or sequences, the grader of better performances is selected from the base learning machine of structure, instead of entirety collection
Into.One advantage of selective ensemble study is that can be by the only preferable learning machine of integrated performance, with smaller resource
The precision and efficiency of consumption reinforcing diagnosis.
The content of the invention
In order to solve prior art presence rotating machinery data complexity the problem of, the present invention proposes one kind and is based on
The rotary machinery fault diagnosis method of selective ensemble study.
The rotary machinery fault diagnosis method learnt based on selective ensemble of the present invention is included：
Known input signal is pre-processed, training sequence is obtained；
The training sequence is handled using Bagging algorithms, a series of base learning machine of differentiation is generated；
PSOSEN is learnt by selective ensemble, trained from a series of base learning machine of differentiation
PSOSEN models；
Rotary machinery fault diagnosis is carried out using the PSOSEN models trained.
Wherein, it is described that known input signal pre-process to obtain training sequence and cycle tests and include：
Obtain the characteristic vector under different faults state, it is contemplated that the actual demand of fault diagnosis, this feature vector is generally
Refer to the time and frequency domain characteristics of Monitoring Data, such as root mean square, wavelet energy coefficient and entropy, so as to obtain training sequence.
Wherein, it is described that the training sequence is handled using Bagging algorithms, generate a series of base of differentiation
Learning machine includes：
The training sequence is reconstructed using the Bagging based on Bootstrap thoughts, generation base learning machine PNN
Network sequence, i.e., the base learning machine of a series of differentiation；
Weight sequence to a series of base learning machine of differentiation is initialized.
Wherein, the training sequence is reconstructed Bagging of the described utilization based on Bootstrap thoughts, generation
Base learning machine PNN network sequences include：
By Bagging processes, a sequence partially overlapped is generated from original training data, it is different for training
Individual PNN networks, and then improve ability to express of the base grader set to learning object.
Wherein, it is described to be learnt by selective ensemble, trained from a series of base learning machine of differentiation
Good PSOSEN models include：
After the base learning machine generating process of differentiation terminates, using improved optimized algorithm APSO, adapted to according to diagnosis
The weight vectors of PNN network sequences are constantly carried out optimizing by degree demand, obtain the final weight distribution D of PNN networkst；
D is distributed according to the final weightt, the PSOSEN models trained from the PNN network sequences.
Wherein, it is described that D is distributed according to final weighttThe PSOSEN models trained.Including：
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, determined described final multiple
Optimal weight vector parameter corresponding to PNN networks, so that the PSOSEN models trained.
Wherein, it is 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 predetermined threshold value, obtains the final weight more than threshold value；
Multiple PNN networks that its final weight in the PNN network sequences is more than threshold value are elected to be described final multiple
PNN networks.
Wherein, the improved optimized algorithm APSO is the optimized 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, in the weight vectors progress searching process of the PNN network sequences, the Nonlinear Dynamic inertia is utilized
Weight of successively decreasing is adaptively adjusted the size of search space regions, it is to avoid be absorbed in the mistake of suboptimization because extent of polymerization is too high
Fit solution.
The method have the benefit that, selected from a series of graders by the base grader based on APSO preferably
Go out the more preferable grader of performance, so as to solve the problem of prior art.
Brief description of the drawings
Fig. 1 is the basic block diagram of PNN networks；
Fig. 2 is PSOSEN Model Diagnosis flow charts；
Fig. 3 is rotary actuator fundamental diagram；
Fig. 4 is RBF observer schematic diagrams；
Fig. 5 is that five kinds of fault modes calculate obtained residual error schematic diagram, wherein (a) figure represents normal, (b) figure represents servo
Valve failure, (c) figure represents hydraulic motor failure, and (d) figure represents rotary actuator failure, and (e) figure represents axle 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 diagrames；
Fig. 8 is the hydraulic pump test stand schematic diagram that the present invention is used；
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 learnt based on selective ensemble of the present invention.
Embodiment
1st, the generation of base grader
The present invention is classified using probabilistic neural network (probabilistic neural networks, PNN) as base
Device, realizes the identification of rotating machinery fault pattern.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 circumstances, PNN network structures are as shown in Figure 1.
In Fig. 1, we can be found that PNN networks 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, in input layer can calculate signal first data point and train to
The distance between amount point；Then, distance can be converted into corresponding to different pieces of information point by mode layer using a RBF
Weighting parameter；In superimposed layer, weighting parameter that can be by each point in input data for different faults pattern adds up, and exports
One probability vector；Probability vector is converted into expected label form by output layer, is used as the final output result of network.
We assume that an input vector is X=(x1, x2, x3, L xn), it includes two kinds of fault mode θAAnd θB, then PNN
The assorting process of network can using general summary as：
If hAlAfA(X) ＞ hBlBfB(X), then X ∈ θA；
If hAlAfA(X) ＜ hBlBfA(X), then X ∈ θB.
Wherein hAAnd hBFault mode θ is represented respectivelyAAnd θBPrior probability.It is assumed that NAAnd NBIt is the instruction in training data
Practice sample and be belonging respectively to θAAnd θBNumber, N is the sum of training sample, then hAAnd hBCan be according to hA=NA/N,hB=NB/ N's
Mode is calculated.lAIt is the work factor in assorting process, represents in training set, belong to fault mode θASample it is wrong
Misclassification is to fault mode θBRatio, lBIn contrast.fAAnd fBFault mode θ is represented respectivelyAAnd θBProbability density function.
Theoretical more than, we can be found that PNN assorting processes are similar to one kind and calculate changeable using RBF kernel functions
The bayes classification method of the lower different faults model probabilities Density Distribution situation of amount, with fault mode θAExemplified by, its is corresponding general
Rate density function can be defined as：
In formula (1), X is input sample vector, and s is the variable number of sample vector.XAjTo belong to j-th of training of A classes
Sample vector, as weight in probabilistic neural network, m represents the vectorial number of the training sample for belonging to A classes, and δ is smooth coefficients.
2nd, the selective ensemble theories of learning
Individual PNN networks are suitable for different decision boundaries dependent on relatively good training input feature vector to construct mostly
Mapping relations, this allows for its generalization ability and is significantly limited by manual features extraction and priori.In order to further
Its Generalization Capability is improved, present invention introduces the thought of integrated study, by building the individual PNN graders with integrated 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
Boosting and Bagging etc., then obtains final predict the outcome by modes such as Nearest Neighbor with Weighted Voting.Rotating machinery fault is examined
For disconnected, the present invention builds individual learning machine using Bagging method, by way of putting back to random sampling, builds
Go out to include the PNN models of different characteristic information.It can be summarized as suitable for the integrated study framework of diagnosis：
For a mapping relationship f from input feature vector to fault mode:Rm→ D, D represent the label of test set, there is N number of
Participate in integrated individual PNN networks.It is assumed that the learning error rate of each base learning machine is q, then the extensive mistake of the entirety of integrated study
Difference can be expressed as：
As shown in formula (2), when meeting following two conditions, integrated study can effectively improve the generalization of diagnosis
Energy：(1) learning error rate is less than 50%；(2) it is separate between each base learning machine.If however, existed in integrated study
Excessive base grader, it is possible to can make it that insignificant calculating and storage resource consumption are excessive.In order to solve this problem, grind
Study carefully personnel propose selective ensemble study concept, i.e., by certain means from the grader of generation, select performance compared with
Good individual is final integrated to build, so as in the case where ensureing precision, further lift Generalization Capability and computational efficiency.
In fault diagnosis, f is madeijAnd djReality output of i-th of PNN network for j-th of characteristic vector is represented respectively
And anticipated output.By taking three classification fault diagnosises as an example, fijAnd djForm 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 characteristic vectorsiIt can be defined as：
Wherein, m represents the quantity of feature, and Error (x) is an indicative function：
On this basis, for j-th of characteristic vector, the integrated result of N number of PNN networks can be expressed as：
ωiThe weight vectors of PNN set are represented, in most common expert's ballot method, it is believed that each weight is
1/N.Whole data set is then directed to, the extensive error of integrated study can be defined as：
Formula (6) and (7) are based further on, on the premise of we can provide consideration guarantee grader Generalization Capability, choosing
The condition of selecting property integrated study is：
In formula (8), fkRepresent 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 grader bad by rejecting part learning performance,
Diagnosis performance and efficiency can be further enhanced on the premise of integrated error is not increased.
3rd, the base grader based on APSO is preferred
Although selective ensemble study possesses bigger advantage than all combining classifiers, how to be selected from grader
Go out the grader of better performances, be still a current problem.In practical problem, the method such as hand picking and priori
Need to consume substantial amounts of computing resource.However, in rotary machinery fault diagnosis, because each PNN output is identical, by
Formula (7) and formula (8) can be obtained, 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 present invention proposes a kind of improved optimized algorithm (APSO) adaptively to assess each base learning machine
Performance, and select PNN set of the weight more than threshold value, the integrated diagnosis learner final for building.
The algorithm is a kind of improved population optimizing algorithm, has used for reference the colonies such as birds and has been determined in search of food most preferably
The process in path.These particles are persistently moved in certain area, according to fitness function, it is determined that the individual in searching process
Optimal location and colony's optimal location.In each iterative process, particle can all move to a new position, and by the position
It is used as the starting point of optimizing next time.The fitness and speed of particle can be assessed with particle properties and constantly updated.In the present invention,
Basic APSO steps are as follows：
For N number of PNN networks, it is assumed that the search space of a L dimension, W=(W1,W2,…,WN) represent 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) represent
Integrated weight vectors to be optimized.Fitness function in diagnosis may be defined as：
Wherein, p and m represent the extensive error and number of features in the diagnosis defined in step (2) respectively.F can consider
It is personal best particle Gbest in each iteration, and global optimum position Zbest is one and includes the individual constantly updated
The matrix of optimal weight vector.The speed of searching optimization for making i-th of particle is Vi=(vi1,vi2,…,viL)T, rising in next update
Beginning position and speed can be calculated according to equation below：
Wherein, θ is an inertia weight, c1And c2It is the accelerated factor for adjusting Gbest and Zbest significance levels, r1And r2
It is a random constant between 0-1.Inertia weight θ is a core parameter in optimization process, usually through linear decrease plan
Slightly calculate, but frequently can lead to its situation for being easily trapped into local optimum.Therefore, the present invention proposes a kind of based on polymerization pass
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：
WhereinRepresent the average fitness value of particle.
So, it is we have proposed a kind of Nonlinear Dynamic inertia weight of successively decreasing in APSO：
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 the case of often, 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, Searching efficiency be improved, while avoiding because of particle
Aggregation extent is too high and is absorbed in the over-fitting situation of local optimum.In addition, ωkvAnd ωkaCan dynamically it be adjusted according to optimization demand
It is whole so that kvAnd kaAn equalization point can be reached.
After APSO optimizing terminates, the individual segregation that the weight parameter in Zbest is higher than given threshold λ can be selected
Device, final for building is integrated.That is, weights are higher, corresponding base classifier performance is better.Determining base classification
After device subset, although the weight vectors in Zbest are normalized, integrated weight vectors can be obtained, but the method only reflects
A kind of possibility in searching process, the to a certain extent easy certain valuable information of loss.It is all in view of Gbest
The set of optimal weights vector in iterations, includes the global information of more horn of plenty, the present invention to Gbest by carrying out SVD
Resolution process, the weight parameter being more applicable.
It is assumed that having selected K PNN network, Gbest from N number of base graderseBe this K grader it is corresponding it is global most
Control matrix L × K, each of which row represents that PNN is integrated into corresponding optimal weights vector in this iteration.We are to GbestseEnter
Row SVD is decomposed：
Gbestse=U Λ VT (15)
Wherein, Λ represents singular value matrix.U and V are corresponding left and right orthogonal matrix respectively.Make νmaxFor in right orthogonal matrix
In, maximum singular value smaxCorresponding first row characteristic vector.By νmaxIt is normalized, can obtain final integrated weight is：
Compared to Zbest is directly utilized, the advantage of this Weight Determination is by minimizing extensive error to make
Maximum probability degree must be classified close to expected sequence label, so as to further improve precision and the Shandong of rotary machinery fault diagnosis
Rod.Result based on APSO selective ensembles is output as：
4th, learnt based on selective ensemble, the present invention proposes a kind of rotary machinery fault diagnosis method, can utilized many
The characteristic parameter of individual angle, strengthens the adaptive ability and accuracy of diagnosis.Whole 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, it is main include with
Under several steps：(1) characteristic vector under different faults state is obtained.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 produced
Learning machine.Bagging algorithms based on Bootstrap thoughts are used to initialize base grader.Bootstrap is that one kind is based on
There is the data creation method for putting back to random resampling, can be from original training data X=(X by Bagging processes1,X2,…,
Xn) in one sequence for partially overlapping of generationThe individual PNN network different for training, and then improve base
Ability to express of the grader set to learning object.The weights of each generation grader carry out initial according to approximately average principle
Change.(3) the PNN networks of better performances are selected, and determine rational weight vectors.After pre-training process terminates, the present invention
Using APSO methods, according to diagnosis fitness demand, optimizing constantly is carried out to weight vectors, final weight distribution D is obtainedt。
In APSO, introduce speed and the degree of polymerization factor to improve global optimizing ability.In each iteration, DtMeeting continuous updating, most
The grader that wherein weight is more than given threshold is selected eventually.Afterwards, SVD decomposition is carried out to global optimum's matrix, it is determined that participating in
Optimal weight vector parameter corresponding to integrated grader.(4) the PSOSEN models completed using training carry out rotating machinery event
Barrier diagnosis.
5th, 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 produce polytype failure.Therefore, the present invention verifies proposed PSOSEN methods using two class object as case
Validity.
(1) rotary actuator fault diagnosis
1) data explanation
As shown in figure 3, the main control module for being responsible for adjusting angle of displacement according to feedback signal by one of rotary actuator, one
Individual PDU modules and a planetary gear train executing agency composition.PDU modules include servo valve, water brake mechanism, one
Individual hydraulic electric motor and One function switching valve.Rotary variable differential transformer is one mutual for electric signal and angular displacement signal
The angle displacement sensor of conversion.In this experiment, five kinds of fault modes are filled with altogether：(1) normal condition；(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
It is a nonlinear feedback control system in view of rotary actuator system, it is difficult to it is determined that accurate mathematical modeling,
A present invention Failure Observer as shown in Figure 4 according to RBF network structions, truly exports and expects defeated for obtaining system
Residual error between going out, so as to preferably symbolize influence of the different faults state to rotary actuator system by residual error magnitude.
Observer model structure and the residual error obtained are as shown in Figure 5.
2) data prediction
Time-Frequency Analysis is signal processing method the most frequently used at present, therefore, and the present invention is carried for the residual signals proposed
The time and frequency domain characteristics of correlation are taken, for obtaining under data fluctuations and complex working condition, there is the feature of more preferable identification to failure
Amount.The time domain charactreristic parameter that the present invention is extracted is as follows.
The Time-domain Statistics feature of table 1
Wavelet transformation is a kind of conventional feature representation method that signal energy distribution is described in time-frequency domain field, this hair
Bright extraction wavelet energy, maximum wavelet coefficient and Wavelet Entropy are as shown in the table as fault signature index.
The time-frequency domain statistical nature of table 2
In table, j represents the number of plies of wavelet decomposition, and i represents the order that the layer signal is decomposed, 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, in order to preferably evaluate performance individual in optimization process, the present invention to training set by carrying out
The mode of Bootstrap samplings obtains a checking collection.Training set is that, for initial training PNN networks, checking collection is then used for
Later stage selects the optimization process of base grader.
The rotary actuator fault diagnosis of table 3：Five kinds of diagnosis cases
3) the rotary actuator system fault diagnosis based on PSOSEN
In PSOSEN, we initialize 20 PNN base graders, and the training subset resampling ratio of each grader 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) normal, servo valve failure, hydraulic motor failure, pressurized strut failure and drive shaft failure are corresponded to respectively with (0,0,0,0,1).
In the base grader choice phase, minimum and maximum search speed is set as+1 and -1, and particle size is 0-1, population
Scale is 100, and iterations is 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 models such as following table
It is shown.
The rotary actuator of table 4 diagnoses PSOSEN model parameters
In present case, we set selection threshold value and base grader use of the weights higher than 0.2 after optimization are selected as 0.2, that is,
It is final integrated in building.SVD decomposition is carried out to obtained global optimum's matrix, takes maximum singular value in right orthogonal matrix corresponding
Characteristic vector is integrated weight.Base grader right-value optimization process is as shown in the table.
The case 1 of table 5：PSOSEN weight renewal processes
The present invention is contrasted PSOSEN models with other methods.Control methods includes the selection using genetic algorithm
Property integrated study model GASEN, based on the multi-classification algorithm AdaBoost.M1 that Boosting is integrated, Bagging Integrated Algorithms and
Individual body Model PNN and SVM.Ten times of cross validation methods are employed in diagnosis, with the diagnostic accuracy average value and mark calculated
Quasi- difference is used as measurement index.Extensive error is calculated according to equation below：
Wherein N represents 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.
The diagnostic results of table 6Case 1
From diagnostic result as can be seen that from C1 to C5, with the lifting of training data ratio, the precision of diagnosis is also therewith
Increase.Compared with individual PNN and SVM classifier, Ensemble Learning Algorithms learn to enhance the table to sample data by multi-angle
Danone power, and possible negative effect in training process is eliminated, so as in most cases all possess higher classification essence
Degree.By taking C3 as an example, PSOSEN, PNN, Bagging, AdaBoost.M1, GASEN and SVM nicety of grading is respectively 93.56%,
86.04%th, 89.02%, 92.99%, 91.00% and 87.21%.In addition, integrated learning approach than individual process have it is smaller
Standard deviation,
Namely possesses higher robustness in diagnosis.
In integrated study class method, as training data increases, integrated study class method in C5 there is highest to examine
Disconnected precision 95.55%.The advantage of such Boosting algorithms is to adjust base grader according to concatenation rules and incline to classification
To weight, and then improve diagnosis performance, reduce over-fitting risk.However, relative to PSOSEN models, AdaBoost.M1's
Standard deviation fluctuation is larger, and diagnosis performance is not sufficiently stable.In addition, PSOSEN and GASEN models are by rejecting performance bad base point
Class device, possesses diagnosis capability more more preferable than Bagging.In the present invention, APSO methods can make up genetic algorithm in high dimension
The problem of easily precocious 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 iterations, APSO methods possess more preferable convergence rate and ability of searching optimum.
For example, in C2, training data is less, APSO can quickly converge to more preferable fitness scope than GA.In C4, although
GA has just started optimizing effect preferably, but as iterations increases, it has been absorbed in the situation of local optimum, and the parameter of fitness is not
Change again, and APSO then shows a more lasting and stable searching process.
(2) Fault Diagnosis of Hydraulic Pump
1) case is described
Hydraulic pump test stand employed in present case is as shown in Figure 8.The testing stand mainly includes a plunger pump, one
Hydraulic circuit and a motor.Motor speed is 5280r/min.Acceleration transducer is arranged on end, for collecting
Vibration signal data, sample frequency is 1024Hz.
Three kinds of fault modes, including normal, swash plate abrasion and valve plate rotor abrasion are filled with experiment altogether, is defined respectively
For fault mode S1, S2 and S3.
2) data prediction
In this experiment, we decompose the method with comentropy combination to extract data characteristics using EMD.First, it is original
Signal is broken down into a series of intrinsic modal components (IMFs), then calculates these IMFs comentropy, is diagnosed as PSOSEN
The input of model.The basic procedure of data processing is as shown in table 8.
The hydraulic pump data prediction flow of table 7
To each group of IMFs, choose the first eight IE to construct eigenmatrix, that is to say, that include 300 in feature set altogether
Sample, each sample includes 8 IE components altogether.80% characteristic quantity is selected at random as training set, and 20% as test set,
Final diagnostic result is obtained using the method for ten times of cross validations to hydraulic pressure Pump data.
3) Fault Diagnosis of Hydraulic Pump based on PSOSEN
It is similar with rotary actuator system diagnostics, set integrated PNN network numbers and reconstruct ratio as 20, bagging samples
For 80%, particle size is 0-1, and population scale is 100, and iterations is 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：
The PSOSE model parameters of table 8
In three steps of integrated calculating, the renewal process of integrated weight is as shown in the table.
The case 2 of table 9:PSOSEN right value updates
Similarly, PSOSEN models are analyzed with other models in case 1 for we.The hydraulic pump of each model
Diagnostic result is as shown in Figure 9.
In figure, red represents expected label vector, and blueness represents the label vector that actual classification is obtained.It can observe
Arrive, although by optimize train, the precision of each model more than 90%, but integrated study model than PNN have it is higher
Diagnostic accuracy, for example, PSOSEN category of model accuracy has reached 100%, it was demonstrated that it passes through efficiently integrated acquisition non-thread
The ability of property fault characteristic.Significantly, since the superiority on small sample problem, SVM methods achieve with it is integrated
Study diagnosis performance similar in hydraulic pump diagnostic data set.
In addition, it is contemplated that hydraulic pump is usually operated in the larger environment of noise, easily influenceed 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 methods
Robustness and Generalization Capability under influence of noise.We concentrate 10dB, 30dB and 50dB noise added to legacy data, use
PCA method, is mapped the data into a two-dimensional space, as shown in Figure 10.
It can be seen in the drawing that as noise increases, the data obfuscation degree of different faults pattern also gradually increases, it was demonstrated that
Noise can bring very important influence to classification task really, also require that diagnosis algorithm has higher robust to data
Property.The diagnostic result of all models is as shown in the table.
Diagnostic result under the difference of table 10 SNRs
From diagnostic result it can be seen that, although noisy influence, PSOSEN models are by reducing deviation and variance
Influence, and rely on more preferable optimization process to improve hydraulic pump health status recognition capability, achieve more more preferable than other models
Classifying quality.For example, when it is SNR10dB that signal to noise ratio is relatively low, the classification that PSOSEN models can still obtain 91.86% is correct
Rate, it was demonstrated that the model can preferably handle the negative effect of grass.
In summary, as shown in figure 11, the rotary machinery fault diagnosis method of the invention learnt based on selective ensemble
Including：
Known input signal is pre-processed, training sequence is obtained；
The training sequence is handled using Bagging algorithms, a series of base learning machine of differentiation is generated；
PSOSEN is learnt by selective ensemble, trained from a series of base learning machine of differentiation
PSOSEN models；
Rotary machinery fault diagnosis is carried out using the PSOSEN models trained.
Wherein, it is described that known input signal pre-process to obtain training sequence and cycle tests and include：
Obtain the characteristic vector under different faults state, it is contemplated that the actual demand of fault diagnosis, this feature vector is generally
Refer to the time and frequency domain characteristics of Monitoring Data, such as root mean square, wavelet energy coefficient and entropy, so as to obtain training sequence.
Wherein, it is described that the training sequence is handled using Bagging algorithms, generate a series of base of differentiation
Learning machine includes：
The training sequence is reconstructed using the Bagging based on Bootstrap thoughts, generation base learning machine PNN
Network sequence, i.e., the base learning machine of a series of differentiation；
Weight sequence to a series of base learning machine of differentiation is initialized.
Wherein, the training sequence is reconstructed Bagging of the described utilization based on Bootstrap thoughts, generation
Base learning machine PNN network sequences include：
By Bagging processes, a sequence partially overlapped is generated from original training data, it is different for training
Individual PNN networks, and then improve ability to express of the base grader set to learning object.
Wherein, it is described to be learnt by selective ensemble, trained from a series of base learning machine of differentiation
Good PSOSEN models include：
After the base learning machine generating process of differentiation terminates, using improved optimized algorithm APSO, adapted to according to diagnosis
The weight vectors of PNN network sequences are constantly carried out optimizing by degree demand, obtain the final weight distribution D of PNN networkst；
D is distributed according to the final weightt, the PSOSEN models trained from the PNN network sequences.
Wherein, it is described that D is distributed according to final weighttThe PSOSEN models trained.Including：
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, determined described final multiple
Optimal weight vector parameter corresponding to PNN networks, so that the PSOSEN models trained.
Wherein, it is 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 predetermined threshold value, obtains the final weight more than threshold value；
Multiple PNN networks that its final weight in the PNN network sequences is more than threshold value are elected to be described final multiple
PNN networks.
Wherein, the improved optimized algorithm APSO is the optimized 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, in the weight vectors progress searching process of the PNN network sequences, the Nonlinear Dynamic inertia is utilized
Weight of successively decreasing is adaptively adjusted the size of search space regions, it is to avoid be absorbed in the mistake of suboptimization because extent of polymerization is too high
Fit solution.
Although the present invention has been described above with particularity, but the invention is not restricted to this, those skilled in the art can
To be modified according to the principle of the present invention, therefore, the various modifications that all principles according to the present invention are carried out all should be understood to
Fall into protection scope of the present invention.
Claims (10)
1. a kind of rotary machinery fault diagnosis method learnt based on selective ensemble, including：
Known input signal is pre-processed, training sequence is obtained；
The training sequence is handled using Bagging algorithms, a series of base learning machine of differentiation is generated；
PSOSEN, the PSOSEN trained from a series of base learning machine of differentiation are learnt by selective ensemble
Model；
Rotary machinery fault diagnosis is carried out using the PSOSEN models trained.
2. according to the method described in claim 1, wherein, it is described that known input signal is carried out to pre-process and obtain training sequence
Include with cycle tests：
Obtain the characteristic vector under different faults state, it is contemplated that the actual demand of fault diagnosis, this feature vector is often referred to prison
Survey the time and frequency domain characteristics of data, such as root mean square, wavelet energy coefficient and entropy, so as to obtain training sequence.
3. according to the method described in claim 1, wherein, it is described using Bagging algorithms to the training sequence at
Reason, generating a series of base learning machine of differentiation includes：
The training sequence is reconstructed using the Bagging based on Bootstrap thoughts, base learning machine PNN networks are generated
Sequence, i.e., the base learning machine of a series of differentiation；
Weight sequence to a series of base learning machine of differentiation is initialized.
4. method according to claim 3, wherein, described Bagging of the utilization based on Bootstrap thoughts is to described
Training sequence is reconstructed, and generation base learning machine PNN network sequences include：
By Bagging processes, a sequence partially overlapped, the different individual for training are generated from original training data
PNN networks, and then improve ability to express of the base grader set to learning object.
5. according to the method described in claim 1, wherein, it is described to be learnt by selective ensemble, from a series of differences
The PSOSEN models trained in the base learning machine of change include：
After the base learning machine generating process of differentiation terminates, using improved optimized algorithm APSO, according to diagnosis fitness need
Ask, optimizing constantly is carried out to the weight vectors of PNN network sequences, obtain the final weight distribution D of PNN networkst；
D is distributed according to the final weightt, the PSOSEN models trained from the PNN network sequences.
6. method according to claim 5, wherein, it is described that D is distributed according to final weighttThe PSOSEN trained
Model.Including：
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, final multiple PNN nets are determined
Optimal weight vector parameter corresponding to network, so that the PSOSEN models trained.
7. method according to claim 6, wherein, it is described that D is distributed according to the final weightt, select final multiple
PNN networks include：
The final weight is distributed DtIt is compared with predetermined threshold value, obtains the final weight more than threshold value；
Multiple PNN networks that its final weight in the PNN network sequences is more than threshold value are elected to be final multiple PNN nets
Network.
8. method according to claim 5, wherein, the improved optimized algorithm APSO be introduce the rate of decay factor and
The optimization algorithms SO of the degree of polymerization factor.
9. method according to claim 8, wherein, utilize 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.
10. method according to claim 9, wherein, carry out searching process in the weight vectors of the PNN network sequences
In, it is adaptively adjusted the sizes of search space regions using Nonlinear Dynamic inertia weight of successively decreasing, it is to avoid because of polymerization journey
Spend over-fitting situation that is high and being absorbed in 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 true CN106980822A (en) | 2017-07-25 |
| CN106980822B 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) |
Cited By (27)
| 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 |
| 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 |
| CN108594788A (en) * | 2018-03-27 | 2018-09-28 | 西北工业大学 | A kind of aircraft actuator fault detection and diagnosis method based on depth random forests algorithm |
| CN108985380A (en) * | 2018-07-25 | 2018-12-11 | 西南交通大学 | A kind of goat fault recognition method based on clustering ensemble |
| CN108986047A (en) * | 2018-07-13 | 2018-12-11 | 中国科学技术大学 | Image denoising method |
| CN109376413A (en) * | 2018-10-08 | 2019-02-22 | 四川大学 | Line neural network method for diagnosing faults based on data-driven |
| CN110018678A (en) * | 2019-02-14 | 2019-07-16 | 电子科技大学 | A kind of method for diagnosing faults of net connection automotive control system |
| CN110045608A (en) * | 2019-04-02 | 2019-07-23 | 太原理工大学 | Based on the twin mechanical equipment component structural dynamic state of parameters optimization method of number |
| CN110163205A (en) * | 2019-05-06 | 2019-08-23 | 网易有道信息技术（北京）有限公司 | Image processing method, device, medium and calculating equipment |
| CN110188383A (en) * | 2019-04-23 | 2019-08-30 | 华中科技大学 | A kind of power station SCR denitration modeling method based on selective ensemble model |
| CN110197190A (en) * | 2018-02-27 | 2019-09-03 | 北京猎户星空科技有限公司 | The localization method and device of model training and object |
| CN111738309A (en) * | 2020-06-03 | 2020-10-02 | 哈尔滨工业大学 | Gas sensor fault mode identification method based on multi-scale analysis and integrated learning |
| CN111783945A (en) * | 2020-06-22 | 2020-10-16 | 东北石油大学 | Long oil transmission pipeline leakage real-time detection system based on transfer learning LSTM |
| CN111914613A (en) * | 2020-05-21 | 2020-11-10 | 淮阴工学院 | Multi-target tracking and facial feature information identification method |
| CN112257335A (en) * | 2020-10-10 | 2021-01-22 | 西南交通大学 | Oil-immersed transformer fault diagnosis method combining PNN and SVM |
| CN112307965A (en) * | 2020-10-30 | 2021-02-02 | 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 |
| CN112534447A (en) * | 2018-08-03 | 2021-03-19 | 西门子医疗保健诊断公司 | Method and apparatus for training machine learning routines for controlling engineering systems |
| CN112541524A (en) * | 2020-11-18 | 2021-03-23 | 湖南大学 | BP-Adaboost multi-source information motor fault diagnosis method based on attention mechanism improvement |
| CN113447273A (en) * | 2021-06-28 | 2021-09-28 | 哈尔滨工业大学 | Gas turbine sensor and actuating mechanism fault detection method based on cross validation |
| CN113591728A (en) * | 2021-08-03 | 2021-11-02 | 天津大学 | Electric energy quality disturbance classification method based on integrated deep learning |
| CN113642433A (en) * | 2021-07-30 | 2021-11-12 | 深圳市中金岭南有色金属股份有限公司凡口铅锌矿 | Method and device for determining working state of elevator, terminal equipment and medium |
| CN113824580A (en) * | 2020-06-19 | 2021-12-21 | 北京市天元网络技术股份有限公司 | Network index early warning method and system |
| 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 |
| CN114021285A (en) * | 2021-11-17 | 2022-02-08 | 上海大学 | Rotary machine fault diagnosis method based on mutual local countermeasure transfer learning |
| CN114722866A (en) * | 2022-03-28 | 2022-07-08 | 杭州中网智慧科技有限公司 | Management and control system is tracked to equipment based on industry internet |
| CN114893390A (en) * | 2022-07-15 | 2022-08-12 | 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 |
| CN112534447B (en) * | 2018-08-03 | 2024-05-24 | 西门子医疗保健诊断公司 | Method and apparatus for training a machine learning routine for controlling an engineering system |
Citations (2)
| 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 |
| 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 (2)
| 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 |
| CN105550711A (en) * | 2015-12-16 | 2016-05-04 | 合肥工业大学 | Firefly algorithm based selective ensemble learning method |
Non-Patent Citations (2)
| Title |
|---|
| 徐卫霄: "高光谱影像集成学习分类及后处理技术研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 蒋芸 等: "基于Bagging的概率神经网络集成分类算法", 《计算机科学》 * |
Cited By (42)
| 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 |
| CN110197190A (en) * | 2018-02-27 | 2019-09-03 | 北京猎户星空科技有限公司 | The localization method and device of model training and object |
| 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 |
| CN108594788A (en) * | 2018-03-27 | 2018-09-28 | 西北工业大学 | A kind of aircraft actuator fault detection and diagnosis method based on depth random forests 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 |
| CN108986047A (en) * | 2018-07-13 | 2018-12-11 | 中国科学技术大学 | Image denoising method |
| CN108986047B (en) * | 2018-07-13 | 2022-03-01 | 中国科学技术大学 | Image noise reduction method |
| CN108985380A (en) * | 2018-07-25 | 2018-12-11 | 西南交通大学 | A kind of goat fault recognition method based on clustering ensemble |
| CN108985380B (en) * | 2018-07-25 | 2021-08-03 | 西南交通大学 | Point switch fault identification method based on cluster integration |
| CN112534447B (en) * | 2018-08-03 | 2024-05-24 | 西门子医疗保健诊断公司 | Method and apparatus for training a machine learning routine for controlling an engineering system |
| CN112534447A (en) * | 2018-08-03 | 2021-03-19 | 西门子医疗保健诊断公司 | Method and apparatus for training machine learning routines for controlling engineering systems |
| CN109376413A (en) * | 2018-10-08 | 2019-02-22 | 四川大学 | Line neural network method for diagnosing faults based on data-driven |
| CN110018678B (en) * | 2019-02-14 | 2021-05-14 | 电子科技大学 | Fault diagnosis method for networked automobile control system |
| CN110018678A (en) * | 2019-02-14 | 2019-07-16 | 电子科技大学 | A kind of method for diagnosing faults of net connection automotive 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 |
| CN110045608A (en) * | 2019-04-02 | 2019-07-23 | 太原理工大学 | Based on the twin mechanical equipment component structural dynamic state of parameters optimization method of number |
| CN110188383A (en) * | 2019-04-23 | 2019-08-30 | 华中科技大学 | A kind of power station SCR denitration modeling method based on selective ensemble model |
| 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 |
| CN110163205A (en) * | 2019-05-06 | 2019-08-23 | 网易有道信息技术（北京）有限公司 | Image processing method, device, medium and calculating equipment |
| CN111914613B (en) * | 2020-05-21 | 2024-03-01 | 淮阴工学院 | Multi-target tracking and facial feature information recognition method |
| CN111914613A (en) * | 2020-05-21 | 2020-11-10 | 淮阴工学院 | Multi-target tracking and facial feature information identification method |
| CN111738309A (en) * | 2020-06-03 | 2020-10-02 | 哈尔滨工业大学 | Gas sensor fault mode identification method based on multi-scale analysis and integrated learning |
| CN113824580A (en) * | 2020-06-19 | 2021-12-21 | 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 |
| CN111783945A (en) * | 2020-06-22 | 2020-10-16 | 东北石油大学 | Long oil transmission pipeline leakage real-time detection system based on transfer learning LSTM |
| CN112257335A (en) * | 2020-10-10 | 2021-01-22 | 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 |
| CN112307965A (en) * | 2020-10-30 | 2021-02-02 | 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 |
| CN112541524A (en) * | 2020-11-18 | 2021-03-23 | 湖南大学 | BP-Adaboost multi-source information motor fault diagnosis method based on attention mechanism improvement |
| CN113447273A (en) * | 2021-06-28 | 2021-09-28 | 哈尔滨工业大学 | Gas turbine sensor and actuating mechanism fault detection method based on cross validation |
| 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 |
| CN113642433A (en) * | 2021-07-30 | 2021-11-12 | 深圳市中金岭南有色金属股份有限公司凡口铅锌矿 | Method and device for determining working state of elevator, terminal equipment and medium |
| 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 |
| CN114021285A (en) * | 2021-11-17 | 2022-02-08 | 上海大学 | Rotary machine fault diagnosis method based on mutual local countermeasure transfer learning |
| CN114021285B (en) * | 2021-11-17 | 2024-04-12 | 上海大学 | Rotary machine fault diagnosis method based on mutual local countermeasure migration learning |
| CN114722866A (en) * | 2022-03-28 | 2022-07-08 | 杭州中网智慧科技有限公司 | Management and control system is tracked to equipment based on industry internet |
| 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 |
| CN114893390A (en) * | 2022-07-15 | 2022-08-12 | 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 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106980822B (en) | 2019-04-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106980822B (en) | A kind of rotary machinery fault diagnosis method based on selective ensemble study | |
| 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 | |
| Chen et al. | Bearing fault diagnosis base on multi-scale CNN and LSTM model | |
| Wang et al. | Modified multiscale weighted permutation entropy and optimized support vector machine method for rolling bearing fault diagnosis with complex signals | |
| Lei et al. | Fault diagnosis of wind turbine based on Long Short-term memory networks | |
| Saufi et al. | Gearbox fault diagnosis using a deep learning model with limited data sample | |
| Wang et al. | Fault diagnosis for rotary machinery with selective ensemble neural networks | |
| Lu et al. | Deep model based domain adaptation for fault diagnosis | |
| Meng et al. | Data segmentation and augmentation methods based on raw data using deep neural networks approach for rotating machinery fault diagnosis | |
| Zhang et al. | A deep learning-based recognition method for degradation monitoring of ball screw with multi-sensor data fusion | |
| Li et al. | Unsupervised rotating machinery fault diagnosis method based on integrated SAE–DBN and a binary processor | |
| Cabrera et al. | Generative adversarial networks selection approach for extremely imbalanced fault diagnosis of reciprocating machinery | |
| Lu et al. | Dominant feature selection for the fault diagnosis of rotary machines using modified genetic algorithm and empirical mode decomposition | |
| CN109918752A (en) | Mechanical failure diagnostic method, equipment and medium based on migration convolutional neural networks | |
| Liang et al. | Rolling bearing fault diagnosis based on one-dimensional dilated convolution network with residual connection | |
| CN106874957A (en) | A kind of Fault Diagnosis of Roller Bearings | |
| Zhao et al. | Fault Diagnosis of Motor in Frequency Domain Signal by Stacked De-noising Auto-encoder. | |
| Han et al. | Bearing fault diagnosis based on multiscale convolutional neural network using data augmentation | |
| Yang et al. | Fault diagnosis of reciprocating compressor based on convolutional neural networks with multisource raw vibration signals | |
| CN108932499A (en) | A kind of rolling bearing prediction technique and system based on local binary patterns and depth confidence network | |
| Pu et al. | A one-class generative adversarial detection framework for multifunctional fault diagnoses | |
| Ma et al. | Cross-domain meta learning fault diagnosis based on multi-scale dilated convolution and adaptive relation module | |
| Su et al. | Generative adversarial networks for gearbox of wind turbine with unbalanced data sets in fault diagnosis | |
| Wang et al. | A novel semi-supervised generative adversarial network based on the actor-critic algorithm for compound fault recognition | |
| Tan et al. | Fault diagnosis method for rolling bearing based on VMD and improved SVM optimized by METLBO |
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 |