CN108182452B - Aero-engine fault detection method and system based on grouping convolution self-encoding encoder - Google Patents
Aero-engine fault detection method and system based on grouping convolution self-encoding encoder Download PDFInfo
- Publication number
- CN108182452B CN108182452B CN201711472261.3A CN201711472261A CN108182452B CN 108182452 B CN108182452 B CN 108182452B CN 201711472261 A CN201711472261 A CN 201711472261A CN 108182452 B CN108182452 B CN 108182452B
- Authority
- CN
- China
- Prior art keywords
- variables
- variable
- convolution
- vector
- feature
- 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
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
The present invention relates to it is a kind of based on grouping convolution self-encoding encoder aero-engine fault detection method and system, wherein method include:Wing Newsletter addressing and the variable of reporting system data are divided into multiple set of variables based on the correlation between variable by variable grouping step;Characteristic extraction step, the feature that each set of variables is extracted using convolution denoising autocoder Model Independent;The Fusion Features of all set of variables are got up to form feature vector by fault identification step, identify fault sample using support vector machines based on this feature vector.The present invention does not need a large amount of expertise experience, avoid cumbersome data prediction work, it is a large amount of it is good have exemplar in the case where still there is preferable resultant fault detection performance, and robustness is good, it is suitable for engineering practice, calculates lower with time cost.
Description
Technical field
The present invention relates to aero-engine technology field more particularly to a kind of aviation hairs based on grouping convolution self-encoding encoder
Motivation fault detection method and system.
Background technique
Along with the development of civil aviaton's cause, safety, reliability and the economy of aircraft engine are by more and more
Ground concern.Engine failure detection is an important channel for improving above-mentioned performance, it can help manager more reasonably
Distribution monitoring resource, improves the safety and reliability to fly every time, formulates the maintenance project of science, farthest reduce fortune
Battalion and maintenance cost.
Current many engine failure detection methods all rely on aberrations in property data.But aberrations in property data source in
Manufacturers of engines (OEM), airline need to pay high expense to obtain performance deviation data.If airline with
Cooperation between OEM is interrupted because of certain unpredictable reasons, and airline will be difficult to obtain performance deviation data, can not
Engine failure detection is effectively performed, easily causes unsafe fly event and huge economic losses.
To improve autonomous engine failure detectability, a kind of condition monitoring for capableing of substitution performance deviation data is needed
Data, and aircraft communication addressing and reporting system (ACARS) data are proper selections.However, ACARS data and performance
Deviation data has differences in some aspects, and one section of smoothed out aberrations in property data is corresponding in the flight cruise stage
ACARS data difference is as illustrated in figs. 1A and ib.Firstly, the dimension of ACARS data is usually higher than aberrations in property data.Every
The cruising phase of secondary flight, ACARS data can achieve tens parameters and aberrations in property data usually only 3 parameters, such as
Shown in figure.Wherein aberrations in property data include DEGT (engine exhaust temperature deviation), GPCN25 (core engine revolving speed deviation
Value) and GWFM (fuel flow deviation).ACARS data then include at least ZALT (height), ZPCN12 (fan instruction revolving speed),
ZPCN25 (core engine instruction revolving speed), ZT1A (total Air Temperature), ZT49 (delivery temperature), ZWF36 (fuel flow) and ZXM (horse
Conspicuous number) etc. parameters.Secondly, aberrations in property data, which can eliminate operating condition and external environment variation bring, to be influenced, it is obvious
Ground reflects engine condition.And each parameter is affected by operating condition and environmental change in ACARS data, and related between each parameter
Relationship is complicated, and fault mode is typically hidden in the complicated variation of multi-parameter.It largely makes an uproar finally, being generally comprised in ACARS data
Sound is easy to interfere analysis result.
In conclusion ACARS data typically include engine condition letter more abundant compared with aberrations in property data
Breath.If handled using effective method it, fault detection effect may be than the method based on aberrations in property data
More preferably.But ACARS data dimension is high, relationship is complicated between parameter and the features such as including a large amount of noises also gives fault detection side
Method brings challenge.The method based on aberrations in property data is proposed in the prior art, they are constructed using machine learning model
The mapping relations of failure and aberrations in property data.Although they achieve good fault detection knot in aberrations in property data
Fruit, but they are difficult to find the failure being hidden among the variation of ACARS multi-parameter complexity.Another kind of method is imitative using engine
The data that true mode generates are verified.And true easy changed by operating condition and external environment of ACARS data is influenced and is sent out
Raw complicated variation.Therefore, compared with emulating data, the true ACARS quality of data is lower and complexity is higher.In addition, also
Another kind of method carries out engine failure detection using a few performance parameter, and such method is difficult to apply to higher-dimension
ACARS data.Therefore, current most of engine failure detection methods are difficult to effectively handle ACARS data.
Due to being achieved in the pattern recognition problem of many higher-dimension complexity well as a result, convolution recent years is compiled automatically
Code device (CAE) has obtained the concern of more and more researchers.It exists in the prior art a kind of using the sparse self-encoding encoder pair of convolution
The method that power transmission lines carry out fault detection and classification, when convolution denoising autocoder model in this method is to multidimensional
Between all variables of sequence synchronize and handled, however its test data set dimension is only 6.When using it come processing parameter relationship
It is complicated, dimension is high and when including the ACARS data of much noise, local feature may be faced and be submerged, character representation ability
The problems such as decline, model complexity is high, and number of parameters is huge, and computing resource and time loss are more.
Summary of the invention
The technical problem to be solved in the present invention is that for existing fault detection method be unsuitable for dealing with relationship it is complicated,
Dimension is high and the problem of include the ACARS data of much noise, by division operation, convolutional neural networks, self-encoding encoder and branch
Vector machine is held to combine, propose it is a kind of based on grouping convolution denoising self-encoding encoder aero-engine fault detection method and be
System, for handling ACARS data.
In order to solve the above-mentioned technical problem, first aspect present invention provides a kind of based on grouping convolution self-encoding encoder
Aero-engine fault detection method, includes the following steps:
Variable is grouped step, is divided Wing Newsletter addressing and the variable of reporting system data based on the correlation between variable
At multiple set of variables;
Characteristic extraction step, the feature that each set of variables is extracted using convolution denoising autocoder Model Independent;
The Fusion Features of all set of variables are got up to form feature vector by fault identification step, are adopted based on this feature vector
Fault sample is identified with support vector machines.
In the aero-engine fault detection method according to the present invention based on grouping convolution self-encoding encoder, preferably
Ground, the characteristic extraction step include following sub-step:
Unsupervised learning sub-step, the convolution kernel of each set of variables is carried out using denoising autocoder model it is unsupervised
Study;
The ACARS data slot of each set of variables is carried out convolution with corresponding convolution kernel by convolution operation sub-step, is acquired
The convolution characteristic pattern of each set of variables;
Pondization operates sub-step, the convolution characteristic pattern based on each set of variables carries out pondization operation, obtains each set of variables
Pond feature.
In the aero-engine fault detection method according to the present invention based on grouping convolution self-encoding encoder, preferably
The pond feature structure of all set of variables is combined by the pond feature vector of each set of variables in the fault identification step in ground
At the feature vector of sample, pattern-recognition is carried out using support vector machines based on feature vector, identifies the failure in ACARS data
Sample.
Second aspect of the present invention provides a kind of aero-engine fault detection system based on grouping convolution self-encoding encoder
System, including:
Variable grouping module, for Wing Newsletter to be addressed to the change with reporting system data based on the correlation between variable
Amount is divided into multiple set of variables;
Characteristic extracting module, for extracting the spy of each set of variables using convolution denoising autocoder Model Independent
Sign;
Fault identification module, for the Fusion Features of all set of variables to be got up to form feature vector, based on this feature to
Amount identifies fault sample using support vector machines.
In the aero-engine fault detection system according to the present invention based on grouping convolution self-encoding encoder, preferably
Ground, the characteristic extracting module include:
Unsupervised learning unit, for being carried out using convolution kernel of the denoising autocoder model to each set of variables without prison
Educational inspector practises;
Convolution operation unit is asked for the ACARS data slot of each set of variables to be carried out convolution with corresponding convolution kernel
Obtain the convolution characteristic pattern of each set of variables;
Pond operating unit carries out pondization operation for the convolution characteristic pattern based on each set of variables, obtains each variable
The pond feature of group.
In the aero-engine fault detection system according to the present invention based on grouping convolution self-encoding encoder, preferably
The pond feature structure of all set of variables is combined by the pond feature vector of each set of variables in the fault identification module in ground
At the feature vector of sample, pattern-recognition is carried out using support vector machines based on feature vector, identifies the failure in ACARS data
Sample.
Implement it is of the invention based on grouping convolution self-encoding encoder aero-engine fault detection method and system, have with
Lower beneficial effect:It is several groups that the present invention, which is first depending on the correlation between ACARS data variable for all variable partitions, is then adopted
Extract the feature of each set of variables with convolution denoising autocoder Model Independent, finally merge all features formed features to
Amount is based on these eigenvector recognition fault samples;The present invention does not need a large amount of expertise and experience;The invention avoids
Cumbersome data prediction work;The present invention it is a large amount of it is good have exemplar in the case where still have it is preferable comprehensive
Close fault detection performance;Robustness of the present invention is preferable, is suitable for engineering practice;The present invention calculates lower with time cost.
Detailed description of the invention
Fig. 1 a and Fig. 1 b are respectively one section of cruising phase smooth aberrations in property data and corresponding original ACARS data
Figure;
Fig. 2 is the aero-engine fault detection side based on grouping convolution self-encoding encoder according to the preferred embodiment of the present invention
The flow chart of method;
Fig. 3 is the aero-engine fault detection based on grouping convolution self-encoding encoder according to the preferred embodiment of the present invention
The overall technology frame diagram of method;
Fig. 4 is the schematic diagram of convolution kernel unsupervised learning according to the present invention;
Fig. 5 is the convolutional neural networks contrast schematic diagram that most of machine learning model and the present invention use;
Fig. 6 is the contrast schematic diagram for the partially connected network that fully-connected network and the present invention use;
Fig. 7 is the parameter sharing schematic diagram of convolutional neural networks;
Fig. 8 is the schematic diagram that the present invention extracts ACARS data slot using sliding window;
Fig. 9 is ACARS convolution characteristic pattern generating process schematic diagram of the invention;
Figure 10 is the aero-engine fault detection based on grouping convolution self-encoding encoder according to the preferred embodiment of the present invention
The module frame chart of system;
Figure 11 is the aero-engine fault detection based on grouping convolution self-encoding encoder according to the preferred embodiment of the present invention
The schematic diagram of characteristic extracting module in system;
Figure 12 a and Figure 12 b be respectively variable group result dendrogram and cost with number of packet variation diagram;
Figure 13 is the box-shaped figure of three kinds of method precision in test set;
Figure 14 is the box-shaped figure of three kinds of method recall rates in test set;
Figure 15 is the box-shaped figure of three kinds of method F1 values in test set.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, technical scheme in the embodiment of the invention is clearly and completely described.Based on the embodiments of the present invention,
Those of ordinary skill in the art's every other embodiment obtained without making creative work belongs to this
Invent the range of protection.
The present invention proposes a kind of novel and effective aviation hair based on grouping convolution self-encoding encoder based on ACARS data
Motivation fault detection method.This method not will it is a large amount of it is good have exemplar in the case where still have it is preferable comprehensive
Fault detection performance, calculating is lower with time cost, and does not need a large amount of expertise and experience, avoids cumbersome number
Data preprocess work.Firstly, this method does not extract feature directly to all variables of ACARS data, but based on variable it
Between correlation they are divided into several set of variables.Then, unsupervised feature extraction is independently carried out to each set of variables, and
Convolution kernel is converted by the feature weight of extraction.Then, convolution Feature Mapping is carried out with pondization operation to each set of variables to mention
Take the feature of ACARS set of variables.Finally the Fusion Features of all set of variables are got up to form feature vector, be adopted based on feature vector
Failure is identified with support vector machines (SVM).Fault detection test is carried out in the true ACARS data of certain type Civil Aviation Engine
Verify the method for the present invention and other several control methods.Test result shows the resultant fault detection performance of the method for the present invention most
Good, number of parameters is relatively fewer, and calculating speed is very fast, it was demonstrated that the superiority of the method for the present invention.
Referring to Fig. 2, for according to the aero-engine event based on grouping convolution self-encoding encoder of the preferred embodiment of the present invention
Hinder the flow chart of detection method.As shown in Fig. 2, the aero-engine event based on grouping convolution self-encoding encoder that the embodiment provides
Detection method includes the following steps for barrier:
Firstly, in step sl, performance variable is grouped step, based on the correlation between variable by Wing Newsletter addressing with
The variable of reporting system (ACARS) data is divided into multiple set of variables.
Then, in step s 2, characteristic extraction step is executed, is extracted using convolution denoising autocoder Model Independent
The feature of each set of variables.
Finally, in step s3, execute fault identification step, by the Fusion Features of all set of variables get up to be formed feature to
Amount identifies fault sample using support vector machines based on this feature vector.
Referring to Fig. 3, for according to the aero-engine based on grouping convolution self-encoding encoder of the preferred embodiment of the present invention
The overall technology frame diagram of fault detection method.It is situated between in detail below with reference to specific implementation process of the Fig. 3 to the method for the present invention
It continues.
S1, variable are grouped step:ACARS variable is grouped;
Invention introduces variable division operations, are grouped according to the correlation between ACARS variable to them, by phase
The stronger variable of closing property is placed on same group, and uncorrelated or weak correlated variables is put into different groups.Complete variable grouping behaviour
After work, the feature that can more preferably indicate ACARS data distribution space can be obtained to each set of variables progress unsupervised learning.
In addition to this, variable is grouped later number of parameters and originally all greatly reduced compared with time complexity, and network query function efficiency is big
It is big to improve.Currently used variable group technology having time sequence k-means cluster, is coagulated at independent variable fractional analysis (IVGA)
Poly- independent variable fractional analysis (AIVGA) etc..The present invention is grouped ACARS data variable using AIVGA method, has certainly
It is dynamic to determine the advantages such as optimal group quantity, calculating speed be fast.
Preferably, the variable of ACARS data is grouped using AIVGA method in above-mentioned variable grouping step S1,
In give an ACARS data set X=[x1x2..., xm], xiRepresent i-th of variable, variable xi=(xi(1),...,xi(T))
It is a time series, T is length of time series.The purpose of AIVGA algorithm is become to the multidimensional of ACARS data set in the step
Amount is split, and m dimension variable is divided into n mutually disjunct subset G={ Gi| j=1 ..., n } so that different set of variables
Model HiThe summation of edge log-likelihood maximize, wherein GiFor the ACARS variables collection of i-th of set of variables.
Edge log-likelihood is obtained for approximation, using variable Bayes's approximation qi(θi) the different variable group models of fitting
Posterior distrbutionpCost function C is minimized in model fitting process.Cost function C is as follows:
In formula, DKL(q | | p) indicates the Kullback-Leibler divergence between q and p.
The relationship of cost function and interactive information can be represented as:
H (x) indicates the entropy of x, IG(x) interactive information in grouping G is indicated.
S2, characteristic extraction step
Preferably, step S2 includes:
S21, unsupervised learning sub-step carry out nothing using convolution kernel of the denoising autocoder model to each set of variables
Supervised learning.
The ACARS data slot of each set of variables is carried out convolution with corresponding convolution kernel by S22, convolution operation sub-step,
Acquire the convolution characteristic pattern for taking each set of variables.
S23, pondization operate sub-step, and the convolution characteristic pattern based on each set of variables carries out pondization operation, obtain each change
The pond feature of amount group.
It describes in detail below to characteristic extraction step.
Common unsupervised feature learning model has autocoder (AE), denoising autocoder (DAE), limited bohr
Hereby graceful machine (RBM) etc..Wherein, DAE model is easy construction, and dimensionality reduction effect is good, and noise removal capability is strong, and the feature robustness of extraction is good.
Dropout can optimize DAE model, further improve the feature that DAE is extracted, therefore use and have in this research
The feature of the DAE model extraction ACARS data of dropout.
It is all using all variables as the input of a DAE model in previous method, the feature set of extraction is all changes
The set of the function f of amount, the ACARS data characteristics collection extracted at this time are as shown in Equation 3.
Features={ f (ZT49, ZPCN12, ZPCN25, ZWF36 ...) } (3)
ZT49 (delivery temperature), ZPCN12 (fan instruction revolving speed), ZPCN25 (core engine instruction revolving speed), ZWF36 in formula 3
(fuel flow) is the variable of ACARS data.
Each DAE is in some set of variables G in the present inventioniIn individually training, therefore the quantity of DAE and set of variables phase
Together, the feature set of extraction is the function set with group variable.Assuming that all variables in ACARS data are divided into G1、G2、G3
Three set of variables, the feature set of extraction are as shown in Equation 4:
Features={ f1 (G1),f2(G2),f3(G3)}
ZALT (height) in formula, ZT1A (total Air Temperature) are the variable of ACARS data.
The training convolutional core at least two benefits in each set of variables rather than on all variables.Firstly, grouping carries out
The feature that unsupervised learning obtains has stronger representativeness, can preferably indicate the distribution space of ACARS data.This be because
It can be the effective means for improving model compactness with the subproblem of Independent modeling to convert large-scale problem to.In the prior art
Verified when multidimensional variable is divided into variable subset according to correlation, independent model to variable subset can obtain more
Close and effective model has better robustness and generalization.
Secondly, network parameter quantity is greatly decreased in method proposed by the present invention, time complexity is reduced.Assuming that ACARS
Data slot length is l, dimension m, then the input dimension of DAE model is ml.To make the feature extracted not lose too many original
Beginning data information, hidden node quantity usually should be more slightly lower than input layer, it is assumed that the number of hidden nodes ml-s1, s1<<ml.
If all variables of ACARS data are averagely divided into n group, the input vector dimension of each DAE model is ml/n, hidden layer section
Points are ml/n-s2, n<<ml,s2<<ml/n.Method (Chen KJ, Hu J, the He JL.Detection and of documents 1
Classification of Transmission Line Faults Based on Unsupervised Feature
Learning and Convolutional Sparse Autoencoder.IEEE T SMART GRID,2016;(99):1-
11.) as shown in table 1 with each index comparison of the method for the present invention.
Each index of method and the method for the present invention of 1 documents 1 of table compares
The ratio between parameter total amount of two methods is:
The weight parameter total amount 1/n that is reduced to documents 1 approximate with time complexity in method proposed by the present invention,
It not only can accelerate calculating speed, it is also helpful to the parallel training of realization network.It also demonstrates in the prior art parallel
Training can break through the performance boundary of single device, and the more massive network of training simultaneously obtains higher essence in classification task
Degree.
Unsupervised learning is carried out to convolution kernel by DAE in step S21.The schematic diagram of convolution kernel unsupervised learning is as schemed
Shown in 4.Unsupervised learning sub-step S21 is specifically included:
(1) each ACARS data slot vectorization is become into one-dimensional vector A.DAE model can only receive one-dimensional vector form
Input, therefore each ACARS data slot need vectorization become one-dimensional vector A=[ZT49 (t) ZPCN12 (t) ZPCN25
(t) ZWF36 (t) ...] (t=1,2 ... T).
(2) by a random function sf (), by one-dimensional vector A pollution at
In Fig. 4, A is original input data,It is the input data after pollution.Masking methods are used in the present invention to pollute
Input data, it allows the sample of λ ratio to become 0 at random.By a random function sf (), A is contaminated intoSuch as following formula institute
Show:
(3) then, willBeing mapped to hidden layer indicates h, is shown below:
In formula, σ indicates the activation primitive between input layer and hidden layer, it is preferable that the activation primitive used in the present invention for
Sigmoid function.W is the weight matrix between input layer and hidden layer, and b is bias vector.
(4) by Dropout model Regularization Technique, one is generated comprising 0-1 random number and obeys Bernoulli Jacob's distribution
The vector is indicated that h carries out inner product operation and obtains with hidden layer by vector bv
Dropout is a kind of model Regularization Technique, it has convenience of calculation and the advantages such as regularization effect is good.
(5) willIt is mapped to the reconstruct Z of A, which can be expressed as follows:
In formula, WTIt is the transposition of the weight matrix W between input layer and hidden layer, c is the biasing between hidden layer and output layer
Vector.
(6) objective function is constructed based on input data A, output data Z and regularization term, by minimizing objective function
Obtain the optimal value W of weight matrix W and bias vector boptWith bopt, wherein optimal value WoptIt is deformed into convolution kernel, boptAs inclined
Vector is set to use in convolution operation.
The training objective of DAE is to keep output data Z and original input data A as consistent as possible.The objective function J of DAE can
To be expressed as:
In formula, m is the quantity of training sample, ZiIt is output data, AiIt is input data, WiIt is input layer and hidden layer
Weight between node, D are KL divergence functions, and ρ is Sparse parameter,It is that hidden node j is averaged to all input layers
Activity, λ and β are the coefficients of respective items.
By minimizing objective function J, the optimal value W of available weight matrix W and bias vector boptWith bopt:
Then, WoptIt is deformed into convolution kernel, boptIt is used as bias vector in convolution Feature Mapping and Chi Huazhong.
Most models in machine learning field can only accepted vector form input.It therefore can be by the ACARS time
Sequence vector is then input in machine learning model and carries out pattern-recognition.However meeting after ACARS time series vectorization
Lose the correlativity between different variables.Convolutional neural networks (CNN) can directly handle network data, can incite somebody to action
Each variate-value of ACARS and the relationship between them input together.Two methods comparison is as shown in Figure 5.
The node of each hidden layer is only connect with part input node in CNN, and this property is called partially connected.Partially connected
It is as shown in Figure 6 with the comparison of full Connection Neural Network.Partially connected can efficiently extract data local feature, be very suitable for
Detect the local fault segment in ACARS time series.There is 1 malfunctioning node x1 (red) in Fig. 6, in neural network input layer
With 4 normal nodes x2, x3, x4, x5.In fully-connected network, hidden node h1 is connected with x1, x2, x3, x4, x5, the failure of x1
Mode is covered by other normal modes, and hidden node h1, h2, h3 activity is all lower.In partially connected, h1 only with x1, x2, x3
It is connected, h1 is affected by x1, and activity is higher.
Other than it can effectively capture the local feature in ACARS data, partially connected can also greatly reduce network ginseng
Number quantity, reduces model complexity.Assuming that input layer and hidden layer have r and s node respectively, (r is had in full Connection Neural Network
+ 1) × s parameter, calculating time complexity are O (r × s+s).Each node of hidden layer is only with input layer in partially connected network
t(t<R) a node connection, number of parameters are (t+1) × s, and calculating time complexity is O (t × s+s), are reduced to original (t+
1)/(r+1).For example number of parameters is 18 in fully-connected network in Fig. 6, number of parameters is 12 in partially connected network.
In addition to partially connected, CNN also has the property of parameter sharing, i.e., all nodes of same hidden layer are all having the same
Weight, as shown in Figure 7.Model parameter quantity not only can be further reduced, moreover it is possible to be replaced calculating cost with convolution algorithm
Higher matrix multiplication operation improves network query function speed.
ACARS data slot is extracted using sliding window method in the present invention.Sliding window sldwin length is lw, and u is
The dimension of ACARS data variable group.Assuming that when the starting point of sliding window reaches the column of ACARS data i-th, the starting point of sldwin
It is pi, end point is pi+lw-1.After the ACARS data slot for obtaining current window, sldwin can move forward a step.At this point, pi
Point will be excluded from sldwin, and pi+lw point can be added to sldwin, and the starting point and end point of window become pi+1 and pi
+lw.The window continuous forward slip in ACARS time series data generates ACARS data slot, as shown in Figure 8.
Convolutional calculation is also what grouping carried out, is carried out to the ACARS data slot of each set of variables with corresponding convolution kernel
Convolution acquires the convolution characteristic pattern of each set of variables.It is worth noting that convolution algorithm is previously required to return ACARS data
One change processing, this is because convolution kernel is the training in the ACARS data after normalization.I-th of ACARS data slot
seg(i)∈Ru×lwWith j-th of convolution kernel k(j)∈Ru×lkCarry out the available convolution characteristic pattern fm ∈ R of convolution algorithm1×(lw-lk+1),
Calculation formula is as shown in Equation 12, and calculating process is as shown in Figure 9.
Fm=σ (seg(i)*k(j)+bj) (12)
In formula, bjFor biasing, " * " is convolution algorithm symbol.
Pondization operation makes feature set have local translation invariance.Local translation invariance examines ACARS piece segment fault
Survey it is highly useful because being only concerned whether failure occurs when we judge whether certain section of ACARS sequence fragment is defective segment,
The position occurred without concern for failure.In addition, pondization can also select apparent ACARS feature, characteristic dimension is reduced, improves net
Network computational efficiency.More common pond method has maximum pond, average pond, spatial pyramid pond, random pool, L2Pond
Change etc..Maximum pondization is chosen in the present invention to study with average pond method.
As convolution, pondization is also independently to carry out on the convolution characteristic pattern of each set of variables of ACARS data.
S3, fault identification step is executed
After completing pondization operation, by the pond feature vector of each set of variables of ACARS data, all set of variables are combined
Pond feature constitute sample feature vector.SVM is solving small sample, has in non-linear and high dimensional data classification problem
Stronger advantage, therefore the present invention classifies to feature vector using SVM, identifies the defective segment in ACARS data.
Referring to Fig. 10, for according to the aero-engine event based on grouping convolution self-encoding encoder of the preferred embodiment of the present invention
Hinder the module frame chart of detection system.As shown in Figure 10, the aeroplane engine based on grouping convolution self-encoding encoder which provides
Machine fault detection system 100 includes:Variable grouping module 110, characteristic extracting module 120 and fault identification module 130.
Wherein, variable grouping module 110 be used for based on the correlation between variable by Wing Newsletter address and reporting system
The variable of data is divided into multiple set of variables.The realization of variable grouping step S1 in the variable grouping module 110 and preceding method
Cheng Xiangtong, details are not described herein.
Characteristic extracting module 120 is used to extract the spy of each set of variables using convolution denoising autocoder Model Independent
Sign.
Fault identification module 130 is based on this feature for the Fusion Features of all set of variables to get up to form feature vector
Vector identifies fault sample using support vector machines.
Figure 11 is please referred to, for according to the aero-engine event based on grouping convolution self-encoding encoder of the preferred embodiment of the present invention
Hinder the schematic diagram of characteristic extracting module in detection system.As shown in figure 11, it is preferable that characteristic extracting module 120 further comprises:
Unsupervised learning unit 121, convolution operation unit 122 and pond operating unit 123.
Wherein, unsupervised learning unit 121 is used for the convolution kernel using denoising autocoder model to each set of variables
Carry out unsupervised learning.The realization process phase of the unsupervised learning unit 121 and unsupervised learning sub-step S21 in preceding method
Together, details are not described herein.
Convolution operation unit 122 is used to the ACARS data slot of each set of variables carrying out convolution with corresponding convolution kernel,
Acquire the convolution characteristic pattern of each set of variables.The reality of convolution operation sub-step S22 in the convolution operation unit 122 and preceding method
Existing process is identical, and details are not described herein.
Pond operating unit 123 carries out pondization operation for the convolution characteristic pattern based on each set of variables, obtains each change
The pond feature of amount group.The pond operating unit 123 is identical as the pondization operation realization process of sub-step S23 in preceding method,
Details are not described herein.
Fault identification module 130 the pond feature vector of each set of variables combines the pond feature of all set of variables
The feature vector for constituting sample carries out pattern-recognition using support vector machines based on feature vector, identifies the event in ACARS data
Hinder sample.
Below by being tested to compare the failure of the method for the present invention Yu other several methods to true ACARS data
Detection performance.
1, the preparation of data set
The operating characteristic of different types of engine has differences.In order to reduce this otherness as far as possible, in experiment
All ACARS data both are from CFM56-7B26 engine.In order to help airline to find engine failure, OEM in time
Customer notification report (CNR) would generally be sent to airline.CNR has recorded the anomalous variation of engine performance deviation data,
And judge the fault type that engine may occur.Therefore, ACARS data label is determined according to CNR in this experiment.It is selected in test
The 19 ACARS parameters taken are as shown in table 2.
The 19 ACARS parameters chosen in the test of table 2
According to the record of CNR, primary fault all at least occurred for all engines selected in test.Specifically, it sends out
Raw fault type has:Jet tailpipe temperature indicator failure, fuel flow indicator failure, total Air Temperature sensor instruction failure,
Core engine decline failure and fan imbalance fault.Performance parameter deviation is abnormal the engine operation number during variation
According to being considered as fault sample, other times intrinsic motivation operation data is considered as normal sample.Variable grouping and convolution
18530 normal samples have been used when the unsupervised learning of core.6606 samples have been used when training and test support vector machines
(wherein 857 are fault samples).
2, experiment process
Follow the steps below test:
(1) ACARS variable is grouped
AIVGA algorithm is using the data after Regularization as input, and variable group result is as depicted in figs. 12 a and 12b.
Figure 12 a is the dendrogram of variable group result, it gives the group result of each step, indicates best at dotted line
Grouping scheme.Figure 12 b is variation of the cost with number of packet.Obviously, cost minimization when number of packet is 4.Table 3 gives
The optimal group result of ACARS parameter.
The optimal group result of 3 ACARS parameter of table
(2) DAE model is constructed
Since input variable is divided into 4 set of variables, for each set of variables one DAE of training, 4 are established altogether
DAE model.DAE model is constructed using deep learning tool box.The hyper parameter of DAE model rely primarily on its basic principle and repeatedly
Test determines.The hyper parameter value of all DAE models is as shown in table 4.
The hyper parameter value of all DAE models of table 4
(3) convolution Feature Mapping and pond
The length of each segment is 30.It is verified through test of many times, it is final to use maximum pond method.Pond dimension is 5
When, fault detection effect is best.
(4) support vector machines
Supporting vector machine model is constructed using the powerful tool box LIBSVM in test, the parameter of support vector machines is set
It sets as shown in table 5.
The parameter setting of 5 support vector machines of table
Given data set is divided into training set and test set.Due to fault sample limited amount, tested using 5 layers of intersection
Card method selects best model.
(5) selection of evaluation index
The data set that the present invention uses in test is serious unbalanced.In this case, precision, recall rate and
3 indexs of F1 value can preferably measure fault detection performance.Their calculation formula sees document (Luo H, Zhong
SS.Gas turbine engine gas path anomaly detection using deep learning with
Gaussian distribution.PHM-Harbin:2017Prognostics and System Health Management
Conference;2017July 9-12;Harbin,China.New York:IEEE,2017.p.1-6).
All algorithms are run in MATLAB R2010b, and the computer used includes Liang Ge Intel Duo
I5 (2.3GHz dominant frequency) CPU, 8GB memory and Windows10 Professional operating system.
3, test result and discussion
Value in table 6 is the average value of five cross validation test results.In each column best result indicated with runic (in addition to
Other than infinity).
The average value of 6 five cross matching results of table
In table 6 at subscript:1, precision and F1 value infinity are because all samples are all identified as normal sample by algorithm herein
This is observed that SVM fault detection performance in training set is best from table 6, but worst in test centrality energy.This is
Because ACARS sequence samples dimension is too high, and there are redundancy between attribute, over-fitting occurs when causing using SVM.In SVM
Preceding increase DAE can reduce characteristic dimension, remove the redundancy in ACARS data, partially remove the noise in ACARS data, because
Performance of this this method in test set is better than first method.
Because partially connected network can preferably extract the local feature in data, relative to first two method, rear three
Kind method has more preferably fault detection performance, is the emphasis comparison other in this research.The precision of three kinds of methods, recall rate afterwards
And the box-shaped figure of F1 value is as illustrated in figs. 13-15.
From Figure 13 to Figure 15 it can be found that although in the method for the present invention precision almost than full dimension CDAE method,
Whether median or distribution will be substantially better than other two methods in recall rate, F1 value.Compared to precision and
Recall rate, comprehensive performance of the F1 value more representative of fault detection.Importantly, in engine health control engineering practice,
Recall rate is generally more important than precision.This is because precision reduction will increase workload, but recall rate reduction can reduce flight
Safety.The two is compared, and safety is unquestionably more importantly.Therefore, proposed by the present invention compared to other two methods
Method is more suitable for the engineering practice of engine health control.
Local dimension CDAE method can preferably extract image expression.But there are structures between ACARS sequence and picture
Difference.In general, pixel distance is closer in picture, and correlation is stronger.Point for ACARS sequence, on time shaft
Distance it is closer, correlation is stronger, this point and image than more consistent.But correlation and distance between the variable in attribute axis
It is far and near not related.Therefore this method effect in ACARS data is not fine.
Variable division operation is introduced compared with the method for documents 1, in the method for the present invention, extracts all steps of feature
It is independently carried out in rapid all each set of variables.It has analyzed above, feature is extracted in grouping can preferably indicate ACARS data
Distribution space, therefore method proposed by the present invention has better fault detection comprehensive performance, and robustness is stronger.
The calculating time and number of parameters comparison of the method for the present invention and 1 method of documents are as shown in table 7.It can be with from table
It was found that the method for the present invention number of parameters is approximately the 1/4 of documents 1, the calculating time is also considerably less than the side of documents 1
Method, the results show analysis of preceding method part of the present invention.
The calculating time and number of parameters of 7 the method for the present invention of table and 1 method of documents compare
In conclusion the aero-engine fault detection method and system of the invention based on grouping convolution self-encoding encoder,
It has the characteristics that:
(1) current most of aero-engine fault detection methods are all based on the aberrations in property data that OEM is provided.For
Autonomous engine failure detectability is improved, the present invention proposes a kind of completely new aeroplane engine based on original ACARS data
Machine fault detection method.This method gets rid of the dependence to OEM, and has preferable resultant fault detection performance.
(2) present invention is not applied for that high-dimensional, parameters relationship is complicated for previous convolution denoising self-encoding encoder
ACARS data propose a kind of grouping convolution denoising self-encoding encoder model.The model can be extracted from ACARS data more added with
Representative feature.This method also can be reduced number of parameters simultaneously, reduce model time complexity, reduce calculate and the time at
This.Finally, the method for the present invention does not need a large amount of expertise experiences, cumbersome data prediction work is also avoided.
(3) true ACARS data verification the method for the present invention and control methods of CFM56-7B26 engine are collected
Validity.Experiment results proved the method for the present invention will be better than its other party in terms of fault detection comprehensive performance and robustness
Method is more suitable for the engineering practice of engine health control.In addition to this, the number of parameters in the method for the present invention and time at
This is many less compared with the Standard convolution in documents 1 denoises self-encoding encoder method.
Claims (6)
1. a kind of aero-engine fault detection method based on grouping convolution self-encoding encoder, which is characterized in that including following step
Suddenly:
Variable is grouped step, based on the correlation between variable by the variable of aircraft communication addressing and reporting system ACARS data
It is divided into multiple set of variables;
Characteristic extraction step, the feature that each set of variables is extracted using convolution denoising autocoder Model Independent;
The Fusion Features of all set of variables are got up to form feature vector by fault identification step, based on this feature vector using branch
Vector machine is held to identify fault sample;
The characteristic extraction step includes following sub-step:
Unsupervised learning sub-step carries out unsupervised using convolution kernel of the denoising autocoder model to each set of variables
It practises;
The ACARS data slot of each set of variables is carried out convolution with corresponding convolution kernel by convolution operation sub-step, is acquired each
The convolution characteristic pattern of set of variables;
Pondization operates sub-step, the convolution characteristic pattern based on each set of variables carries out pondization operation, obtains the pond of each set of variables
Change feature;
The unsupervised learning sub-step includes:
(1) each ACARS data slot vectorization is become into one-dimensional vector A;
(2) by a random function sf (), by one-dimensional vector A pollution at
(3) willBeing mapped to hidden layer indicates h, is shown below:
σ (a)=1/ (1+e-a)
In formula, σ indicates the activation primitive between input layer and hidden layer, and W is the weight matrix between input layer and hidden layer, and b is inclined
Set vector;
(4) by Dropout model Regularization Technique, one is generated comprising 0-1 random number and obeys the vector of Bernoulli Jacob's distribution
The vector is indicated that h carries out inner product operation and obtains with hidden layer by bv
It (5) will by following formulaIt is mapped to the reconstruct Z of A:
In formula, WTIt is the transposition of the weight matrix W between input layer and hidden layer, c is the bias vector between hidden layer and output layer;
(6) objective function is constructed based on input data A, output data Z and regularization term, is obtained by minimizing objective function
The optimal value W of weight matrix W and bias vector boptWith bopt, wherein optimal value WoptIt is deformed into convolution kernel, boptAs being biased towards
Amount uses in convolution operation.
2. the aero-engine fault detection method according to claim 1 based on grouping convolution self-encoding encoder, feature
It is, by the pond feature vector of each set of variables in the fault identification step, combines the pond feature of all set of variables
The feature vector for constituting sample carries out pattern-recognition using support vector machines based on feature vector, identifies the event in ACARS data
Hinder sample.
3. the aero-engine fault detection method according to claim 1 based on grouping convolution self-encoding encoder, feature
It is, the variable of ACARS data is grouped using cohesion independent variable fractional analysis method in the variable grouping step,
Give an ACARS data set X=[x1x2..., xm], xiRepresent i-th of variable, variable xi=(xi(1),...,xi(T)) it is
One time series, T are length of time series;The multidimensional variable of ACARS data set is split in the step, m is tieed up and is become
Amount is divided into n mutually disjunct subset G={ Gi| j=1 ..., n } so that the model H of different set of variablesiEdge logarithm seemingly
The summation of right property maximizes, wherein GiFor the ACARS variables collection in i-th of set of variables.
4. a kind of aero-engine fault detection system based on grouping convolution self-encoding encoder, which is characterized in that including:
Variable grouping module, for based on the correlation between variable by aircraft communication addressing and reporting system ACARS data
Variable is divided into multiple set of variables;
Characteristic extracting module, for extracting the feature of each set of variables using convolution denoising autocoder Model Independent;
Fault identification module is adopted for the Fusion Features of all set of variables to get up to form feature vector based on this feature vector
Fault sample is identified with support vector machines;
The characteristic extracting module includes:
Unsupervised learning unit, for carrying out unsupervised using convolution kernel of the denoising autocoder model to each set of variables
It practises;
Convolution operation unit acquires every for the ACARS data slot of each set of variables to be carried out convolution with corresponding convolution kernel
The convolution characteristic pattern of a set of variables;
Pond operating unit carries out pondization operation for the convolution characteristic pattern based on each set of variables, obtains each set of variables
Pond feature;
The unsupervised learning unit carries out unsupervised learning by following steps:
(1) each ACARS data slot vectorization is become into one-dimensional vector A and is used as input data;
(2) by a random function sf (), by one-dimensional vector A pollution at
(3) willBeing mapped to hidden layer indicates h, is shown below:
σ (a)=1/ (1+e-a)
In formula, σ indicates the activation primitive between input layer and hidden layer, and W is the weight matrix between input layer and hidden layer, and b is inclined
Set vector;
(4) by Dropout model Regularization Technique, one is generated comprising 0-1 random number and obeys the vector of Bernoulli Jacob's distribution
The vector is indicated that h carries out inner product operation and obtains with hidden layer by bv
It (5) will by following formulaThe reconstruct for being mapped to A obtains output data Z:
In formula, WTIt is the transposition of the weight matrix W between input layer and hidden layer, c is the bias vector between hidden layer and output layer;
(6) objective function is constructed based on input data A, output data Z and regularization term, is obtained by minimizing objective function
The optimal value W of weight matrix W and bias vector boptWith bopt, wherein optimal value WoptIt is deformed into convolution kernel, boptAs being biased towards
Amount uses in convolution operation.
5. the aero-engine fault detection system according to claim 4 based on grouping convolution self-encoding encoder, feature
It is, by the pond feature vector of each set of variables in the fault identification module, combines the pond feature of all set of variables
The feature vector for constituting sample carries out pattern-recognition using support vector machines based on feature vector, identifies the event in ACARS data
Hinder sample.
6. the aero-engine fault detection system according to claim 4 based on grouping convolution self-encoding encoder, feature
It is, the variable grouping module is grouped the variable of ACARS data using cohesion independent variable fractional analysis method, gives
A fixed ACARS data set X=[x1x2..., xm], xiRepresent i-th of variable, variable xi=(xi(1),...,xiIt (T)) is one
A time series, T are length of time series;The variable grouping module is split the multidimensional variable of ACARS data set, by m
Dimension variable is divided into n mutually disjunct subset G={ Gi| j=1 ..., n } so that the model H of different set of variablesiEdge pair
The summation of number likelihood maximizes, wherein GiFor the ACARS variables collection in i-th of set of variables.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711472261.3A CN108182452B (en) | 2017-12-29 | 2017-12-29 | Aero-engine fault detection method and system based on grouping convolution self-encoding encoder |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711472261.3A CN108182452B (en) | 2017-12-29 | 2017-12-29 | Aero-engine fault detection method and system based on grouping convolution self-encoding encoder |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108182452A CN108182452A (en) | 2018-06-19 |
CN108182452B true CN108182452B (en) | 2018-11-20 |
Family
ID=62548968
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711472261.3A Active CN108182452B (en) | 2017-12-29 | 2017-12-29 | Aero-engine fault detection method and system based on grouping convolution self-encoding encoder |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108182452B (en) |
Families Citing this family (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109115501A (en) * | 2018-07-12 | 2019-01-01 | 哈尔滨工业大学(威海) | A kind of Civil Aviation Engine Gas path fault diagnosis method based on CNN and SVM |
CN109102005B (en) * | 2018-07-23 | 2020-08-18 | 杭州电子科技大学 | Small sample deep learning method based on shallow model knowledge migration |
CN109087259A (en) * | 2018-08-01 | 2018-12-25 | 中国石油大学(北京) | Pre stack data denoising method and system based on convolution self-encoding encoder |
CN109272499B (en) * | 2018-09-25 | 2020-10-09 | 西安电子科技大学 | Non-reference image quality evaluation method based on convolution self-coding network |
CN111046887A (en) * | 2018-10-15 | 2020-04-21 | 华北电力大学(保定) | Method for extracting characteristics of image with noise |
CN109583499B (en) * | 2018-11-30 | 2021-04-16 | 河海大学常州校区 | Power transmission line background object classification system based on unsupervised SDAE network |
CN109866931B (en) * | 2019-03-15 | 2020-10-27 | 西北工业大学 | Airplane throttle control method based on self-encoder |
CN110059601B (en) * | 2019-04-10 | 2021-04-13 | 西安交通大学 | Intelligent fault diagnosis method for multi-feature extraction and fusion |
CN110334716B (en) * | 2019-07-04 | 2022-01-11 | 北京迈格威科技有限公司 | Feature map processing method, image processing method and device |
CN110260925B (en) * | 2019-07-12 | 2021-06-25 | 重庆赛迪奇智人工智能科技有限公司 | Method and system for detecting quality of driver parking technology, intelligent recommendation method and electronic equipment |
CN110207997B (en) * | 2019-07-24 | 2021-01-19 | 中国人民解放军国防科技大学 | Liquid rocket engine fault detection method based on convolution self-encoder |
CN110378431A (en) * | 2019-07-24 | 2019-10-25 | 中国人民解放军国防科技大学 | Convolutional neural network-based supersonic combustion chamber combustion mode detection method |
CN110609477B (en) * | 2019-09-27 | 2021-06-29 | 东北大学 | Electric power system transient stability discrimination system and method based on deep learning |
CN110929765B (en) * | 2019-11-06 | 2023-09-22 | 北京工业大学 | Batch-imaging-based convolution self-coding fault monitoring method |
CN111144505B (en) * | 2019-12-30 | 2023-09-01 | 北京三快在线科技有限公司 | Variable classification method, device, equipment and medium based on dimension slice |
CN111338878A (en) * | 2020-02-21 | 2020-06-26 | 平安科技(深圳)有限公司 | Anomaly detection method and device, terminal device and storage medium |
CN112488238B (en) * | 2020-12-14 | 2022-11-15 | 桂林电子科技大学 | Hybrid anomaly detection method based on countermeasure self-encoder |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104268627A (en) * | 2014-09-10 | 2015-01-07 | 天津大学 | Short-term wind speed forecasting method based on deep neural network transfer model |
CN106529605A (en) * | 2016-11-28 | 2017-03-22 | 东华大学 | Image identification method of convolutional neural network model based on immunity theory |
US9733141B1 (en) * | 2012-06-27 | 2017-08-15 | The United States Of America As Represented By The Administrator Of National Aeronautics And Space Administration | Method for making measurements of the post-combustion residence time in a gas turbine engine |
CN107103658A (en) * | 2017-04-25 | 2017-08-29 | 哈尔滨工业大学 | Aerial engine air passage method for detecting abnormality based on deep learning and Gaussian Profile |
CN107220469A (en) * | 2017-07-14 | 2017-09-29 | 华北电力大学(保定) | A kind of method and system of fan condition estimation |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9668699B2 (en) * | 2013-10-17 | 2017-06-06 | Siemens Healthcare Gmbh | Method and system for anatomical object detection using marginal space deep neural networks |
CN106295906A (en) * | 2016-08-22 | 2017-01-04 | 万马联合新能源投资有限公司 | A kind of urban public bus lines querying method based on degree of depth study |
CN106503746B (en) * | 2016-11-03 | 2018-07-06 | 哈尔滨工业大学 | A kind of Fault Diagnosis of Aeroengines method based on offset of performance amount |
CN107506695A (en) * | 2017-07-28 | 2017-12-22 | 武汉理工大学 | Video monitoring equipment failure automatic detection method |
-
2017
- 2017-12-29 CN CN201711472261.3A patent/CN108182452B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9733141B1 (en) * | 2012-06-27 | 2017-08-15 | The United States Of America As Represented By The Administrator Of National Aeronautics And Space Administration | Method for making measurements of the post-combustion residence time in a gas turbine engine |
CN104268627A (en) * | 2014-09-10 | 2015-01-07 | 天津大学 | Short-term wind speed forecasting method based on deep neural network transfer model |
CN106529605A (en) * | 2016-11-28 | 2017-03-22 | 东华大学 | Image identification method of convolutional neural network model based on immunity theory |
CN107103658A (en) * | 2017-04-25 | 2017-08-29 | 哈尔滨工业大学 | Aerial engine air passage method for detecting abnormality based on deep learning and Gaussian Profile |
CN107220469A (en) * | 2017-07-14 | 2017-09-29 | 华北电力大学(保定) | A kind of method and system of fan condition estimation |
Also Published As
Publication number | Publication date |
---|---|
CN108182452A (en) | 2018-06-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108182452B (en) | Aero-engine fault detection method and system based on grouping convolution self-encoding encoder | |
Xuyun et al. | Aircraft engine fault detection based on grouped convolutional denoising autoencoders | |
Xie et al. | Graph neural network approach for anomaly detection | |
Zhuang et al. | Non-negative low rank and sparse graph for semi-supervised learning | |
Li et al. | Multi-view low-rank analysis for outlier detection | |
Li et al. | A novel deep autoencoder and hyperparametric adaptive learning for imbalance intelligent fault diagnosis of rotating machinery | |
Pan et al. | A fault diagnosis approach for roller bearing based on symplectic geometry matrix machine | |
Wu et al. | A transformer-based approach for novel fault detection and fault classification/diagnosis in manufacturing: A rotary system application | |
Hurst et al. | Big data analysis techniques for cyber-threat detection in critical infrastructures | |
Wei et al. | Adoption and realization of deep learning in network traffic anomaly detection device design | |
Yu et al. | Stacked denoising autoencoder‐based feature learning for out‐of‐control source recognition in multivariate manufacturing process | |
Cheng et al. | Online power system event detection via bidirectional generative adversarial networks | |
CN113378160A (en) | Graph neural network model defense method and device based on generative confrontation network | |
CN106127260A (en) | A kind of multi-source data fuzzy clustering algorithm of novelty | |
Jia et al. | An effective imbalanced JPEG steganalysis scheme based on adaptive cost-sensitive feature learning | |
Javaheripi et al. | CuRTAIL: ChaRacterizing and thwarting AdversarIal deep learning | |
Wang et al. | Single and simultaneous fault diagnosis of gearbox via wavelet transform and improved deep residual network under imbalanced data | |
Wang et al. | A hybrid 3DSE-CNN-2DLSTM model for compound fault detection of wind turbines | |
Yang et al. | Few-shot classification via ensemble learning with multi-order statistics | |
Wei et al. | GNN-ensemble: Towards random decision graph neural networks | |
CN102496027B (en) | Semi-supervised image classification method based on constrained adaptive transmission | |
Ince | A novel approach for intrusion detection systems: V-IDS | |
Cao et al. | Weighted margin sparse embedded classifier for brake cylinder detection | |
Xu-Dong et al. | Control Chart Recognition Method Based on Transfer Learning | |
Tian et al. | A cross-domain intelligent fault diagnosis method based on feature transfer with improved Inception ResNet for rolling bearings under varying working condition |
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 |