CN108334907A - A kind of complex equipment point method for detecting abnormality and system based on deep learning - Google Patents
A kind of complex equipment point method for detecting abnormality and system based on deep learning Download PDFInfo
- Publication number
- CN108334907A CN108334907A CN201810131253.0A CN201810131253A CN108334907A CN 108334907 A CN108334907 A CN 108334907A CN 201810131253 A CN201810131253 A CN 201810131253A CN 108334907 A CN108334907 A CN 108334907A
- Authority
- CN
- China
- Prior art keywords
- feature
- sdae
- abnormality detection
- complex equipment
- training sample
- 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
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
Abstract
The complex equipment point method for detecting abnormality and system, wherein method that the present invention relates to a kind of based on deep learning include：Training sample processing step, the monitoring performance parameter for choosing complex equipment, and obtain the normal sample and exceptional sample of monitoring performance parameter, composing training sample set；Characteristic extraction step, Feature Selection Model of the structure based on SDAE, input the training sample set and carry out model training；Detection model training step, according to abnormality detection model of the feature construction trained through Feature Selection Model based on GSM, characteristic mean and the characteristic standard for obtaining abnormality detection model are poor；Sample to be tested is input to the Feature Selection Model based on SDAE by anomalies detecting step, and obtained feature is input in the abnormality detection model based on GSM, carries out an abnormality detection.The point abnormality detection that complex equipment especially aero-engine occurs in the present invention is with obvious effects.
Description
Technical field
The present invention relates to aero-engine technology field more particularly to a kind of complex equipment point based on deep learning are abnormal
Detection method and system.
Background technology
All the time, complex equipment state is difficult to quantitative evaluation, cannot achieve the problems such as failure predication be industrial quarters and
One of the direction that art circle emphasis solves.The more efficient complex equipment method for detecting abnormality of design is the health of Large Complex Equipment
An important ring in management system.Aero-engine is the most important core of aircraft as a kind of typical Large Complex Equipment
Center portion part, operating status it is normal whether it is most important.
By taking aero-engine as an example, traditional abnormality detection and method for diagnosing faults based on data-driven mainly pass through
To a small number of parameters in OEM data, for example, take off the parameters such as delivery temperature nargin (EGTM), fuel flow deviation (Δ FF) into
Row analysis obtains.Apparent disadvantage there are two ways to this tradition, one is these parameters in the OEM data used are not
It is the initial data most started, is by the processed data of producer's system, therefore obtained data cannot be guaranteed real-time；Its
Second is that judging that the normal abnormal result of aero-engine is limited only according to two to three parameters.
In addition, abnormality detection is based primarily upon Hawkins to abnormal definition at present：Exception be far from other observation data and
The observation data generated suspected of different mechanisms.Whether independent between each other according to data point, anomaly pattern can be roughly divided into two
Class：Point form is abnormal and time series form is abnormal.If an individual data point is considered relative to its remainder
According to exception, then the example is referred to as an exception.This is simplest Exception Type and most of abnormality detection researchs
Emphasis.If a data point or one section of sequence number strong point are abnormal relative to other segment datas in sequence context, this
The exception of sample is referred to as time series exception.Similarly, the anomaly pattern of aero-engine can be equally divided into a little different in this way
Often and time series is abnormal.According to the different form of aero-engine abnormal data, need to separately design for various forms of
Method for detecting abnormality.
Invention content
The technical problem to be solved in the present invention is, for the detection result of existing complex equipment point method for detecting abnormality
Bad defect, it is proposed that a kind of complex equipment point method for detecting abnormality and system based on deep learning.
In order to solve the above-mentioned technical problem, first aspect present invention provides a kind of complex equipment based on deep learning
Point method for detecting abnormality, includes the following steps：
Training sample processing step, the monitoring performance parameter for choosing complex equipment, and obtain the normal of monitoring performance parameter
Sample and exceptional sample, composing training sample set；
Characteristic extraction step, Feature Selection Model of the structure based on SDAE, input the training sample set and carry out model instruction
Practice；
Detection model training step is examined according to exception of the feature construction trained through Feature Selection Model based on GSM
Model is surveyed, characteristic mean and the characteristic standard for obtaining abnormality detection model are poor；
Sample to be tested is input to the Feature Selection Model based on SDAE by anomalies detecting step, and obtained feature is input to
In abnormality detection model based on GSM, an abnormality detection is carried out.
In the complex equipment point method for detecting abnormality according to the present invention based on deep learning, it is preferable that described
The data that training sample processing step also concentrates training sample are normalized.
In the complex equipment point method for detecting abnormality according to the present invention based on deep learning, it is preferable that described
The stacking number of SDAE is two layers in characteristic extraction step.
In the complex equipment point method for detecting abnormality according to the present invention based on deep learning, it is preferable that described
The structure of the Feature Selection Model of SDAE used in characteristic extraction step is determined by following steps：
1) SDAE of 200 different structures is built, the structure of input layer-hidden layer-output layer of wherein DAE is a-X-Y,
In 1≤X≤20,1≤Y≤10, X, Y is respectively hidden layer and the number of output layer, and is integer, and a is the number of input layer, with
The number of the monitoring performance parameter for the complex equipment chosen in the training sample processing step is consistent；
2) initial parameter of training sample set is input in the SDAE of 200 different structures, obtains corresponding 200 groups of spies
Collection；
3) it is analyzed for the obtained features of the SDAE of each structure, passes through the correlation analyzed between feature, screening
Invalid feature；
4) after screening invalid feature, according to the feature set degree of correlation of each SDAE, the degree of correlation is small between choosing each feature
SDAE structures, as the Feature Selection Model in abnormality detection model.
5, the complex equipment point method for detecting abnormality described in any one of claim 1 to 3 based on deep learning,
It is characterized in that, the abnormality detection model based on GSM established in the detection model training step is multivariate Gaussian models,
Conditional probability is：
Wherein μ is the mean value of training sample, and ∑ is the covariance matrix of training sample；For sample to be tested x, grader is fixed
Justice is as follows：
Wherein θ is the threshold value for judging whether sample to be tested abnormal, and h (x) is discriminant function, and 1 represents normal value, 0 represent it is different
Constant value.
In the complex equipment point method for detecting abnormality according to the present invention based on deep learning, it is preferable that described
Complex equipment is aero-engine, and the monitoring performance parameter for the complex equipment that the training sample processing step is chosen includes：Sea
Face temperature, flying height, fan instruction rotating speed, core engine instruction rotating speed, oil liquid pressure, import total temperature, EGT, oil liquid temperature,
Fuel flow and Mach number.
Second aspect of the present invention provides a kind of complex equipment point abnormality detection system based on deep learning, including：
Training sample processing module, the monitoring performance parameter for choosing complex equipment, and obtain monitoring performance parameter
Normal sample and exceptional sample, composing training sample set；
Characteristic extracting module inputs the training sample set and carries out mould for building the Feature Selection Model based on SDAE
Type training；
Detection model training module, for according to the feature construction trained through Feature Selection Model based on the different of GSM
Normal detection model, characteristic mean and the characteristic standard for obtaining abnormality detection model are poor；
Abnormality detection module, for sample to be tested to be input to the Feature Selection Model based on SDAE, obtained feature is defeated
Enter into the abnormality detection model based on GSM, carries out an abnormality detection.
In the complex equipment point abnormality detection system according to the present invention based on deep learning, it is preferable that described
The structure of the Feature Selection Model of SDAE used in characteristic extracting module is determined by following steps：
1) SDAE of 200 different structures is built, the structure of input layer-hidden layer-output layer of wherein DAE is a-X-Y,
In 1≤X≤20,1≤Y≤10, X, Y is respectively hidden layer and the number of output layer, and is integer, and a is the number of input layer, with
The number of the monitoring performance parameter for the complex equipment chosen in the training sample processing step is consistent；
2) initial parameter of training sample set is input in the SDAE of 200 different structures, obtains corresponding 200 groups of spies
Collection；
3) it is analyzed for the obtained features of the SDAE of each structure, passes through the correlation analyzed between feature, screening
Invalid feature；
4) after screening invalid feature, according to the feature set degree of correlation of each SDAE, the degree of correlation is small between choosing each feature
SDAE structures, as the Feature Selection Model in abnormality detection model.
In the complex equipment point abnormality detection system according to the present invention based on deep learning, it is preferable that described
The abnormality detection model based on GSM established in detection model training module is multivariate Gaussian models, and conditional probability is：
Wherein μ is the mean value of training sample, and ∑ is the covariance matrix of training sample；For sample to be tested x, grader is fixed
Justice is as follows：
Wherein θ is the threshold value for judging whether sample to be tested abnormal, and h (x) is discriminant function, and 1 represents normal value, 0 represent it is different
Constant value.
In the complex equipment point abnormality detection system according to the present invention based on deep learning, it is preferable that described
Complex equipment is aero-engine, and the monitoring performance parameter for the complex equipment that the training sample processing module is chosen includes：Sea
Face temperature, flying height, fan instruction rotating speed, core engine instruction rotating speed, oil liquid pressure, import total temperature, EGT, oil liquid temperature,
Fuel flow and Mach number.
Implement the complex equipment point method for detecting abnormality and system based on deep learning of the present invention, has below beneficial to effect
Fruit：The present invention builds the Feature Selection Model based on SDAE first, and the feature then extracted according to SDAE passes through single Gauss model
It carries out abnormality detection, the point abnormality detection occurred for complex equipment especially aero-engine is with obvious effects.
Description of the drawings
Fig. 1 is the flow according to the complex equipment point method for detecting abnormality based on deep learning of the preferred embodiment of the present invention
Figure；
Fig. 2 is the schematic diagram of autocoder structure；
Fig. 3 is the structural schematic diagram of DAE；
Fig. 4 is the block mold structure according to the complex equipment point method for detecting abnormality based on deep learning of the present invention.
Specific implementation mode
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,
The every other embodiment that those of ordinary skill in the art are obtained without making creative work belongs to this
Invent the range of protection.
Referring to Fig. 1, for according to the complex equipment point abnormality detection side based on deep learning of the preferred embodiment of the present invention
The flow chart of method.As shown in Figure 1, the embodiment provide the complex equipment point method for detecting abnormality based on deep learning include with
Lower step：
First, in step sl, training sample processing step is executed, chooses the monitoring performance parameter of complex equipment, and obtain
Take the normal sample and exceptional sample of monitoring performance parameter, composing training sample set.The complex device of the present invention includes but unlimited
In aero-engine, numerically-controlled machine tool, gas turbine and nuclear power generating equipment etc..Preferably, when for complex equipment be aeroplane engine
When machine, the monitoring performance parameter for the complex equipment that training sample processing step S1 chooses includes：Sea Level Temperature, flying height,
Fan indicates rotating speed, core engine instruction rotating speed, oil liquid pressure, import total temperature, EGT, oil liquid temperature, fuel flow and Mach number.
Then, in step s 2, characteristic extraction step, spy of the structure based on SDAE (stacking denoising autocoder) are executed
Extraction model is levied, the training sample set that input step S1 is obtained carries out model training.Feature extraction mould based on SDAE herein
Type may be simply referred to as SDAE models again.
Then, in step s3, detection model training step is executed, according to the feature trained through Feature Selection Model
Abnormality detection model of the structure based on GSM (single Gauss model), characteristic mean and the characteristic standard for obtaining abnormality detection model are poor.
The abnormality detection model based on GSM may be simply referred to as GSM models again herein.
Finally, in step s 4, it executes anomalies detecting step, sample to be tested is input to the feature extraction mould based on SDAE
Type, obtained feature are input in the abnormality detection model based on GSM, carry out an abnormality detection.
The deep learning of the present invention plays good effect to the hiding information in depth mining data, therefore can carry out
The abnormality detection of complex appts especially aero-engine.
Preferably, the data that training sample processing step S1 also concentrates training sample are normalized.Correspondingly,
Sample to be tested in anomalies detecting step is also required to input again in SDAE models after being normalized.
Preferably, the structure of the Feature Selection Model of the SDAE used in characteristic extraction step S2 is true by following steps
It is fixed：
1) SDAE of 200 different structures is built, the structure of input layer-hidden layer-output layer of wherein DAE is a-X-Y,
In 1≤X≤20,1≤Y≤10, X, Y is respectively hidden layer and the number of output layer, and is integer, and a is the number of input layer, i.e.,
The training sample set that will be inputted has the monitoring for the complex equipment chosen in a kind parameters, that is, training sample processing step S1
The number of performance parameter；
2) initial parameter of training sample set is input in the SDAE of 200 different structures, obtains corresponding 200 groups of spies
Collection；
3) it is analyzed for the obtained features of the SDAE of each structure, passes through the correlation analyzed between feature, screening
Invalid feature；
4) after screening invalid feature, according to the feature set degree of correlation of each SDAE, the degree of correlation is small between choosing each feature
SDAE structures, as the Feature Selection Model in abnormality detection model.
The complex equipment point method for detecting abnormality based on deep learning of the present invention is described in detail below.
There are two major parts for method for detecting abnormality proposed by the present invention：Feature Selection Model and abnormality detection model.It is first
The Feature Selection Model based on SDAE is first built, the feature then extracted according to SDAE, training list Gauss model carries out abnormal inspection
It surveys.SDAE and single Gauss model are introduced separately below.
1, model foundation
1.1 Feature Selection Models based on SDAE
Referring to Fig. 2, for the schematic diagram of autocoder structure.It is exactly autocoder (AE) mould in dotted line frame
Type, it is made of encoder (Encoder) and decoder (Decoder) two parts, is inherently to do certain to input signal
Transformation.Input signal x is transformed into encoded signal y by encoder, and decoder is converted into output signal by y is encoded
The purpose of self-encoding encoder is to allow outputReproduction input x as far as possible, compared to output, the coding of middle layer, i.e. y are more
It is important.If f and g are identical mappings, have with regard to perseveranceBut such transformation is without in all senses.M signal y is done
Certain constraint generally chooses nonlinear sigmoid or tanh and is activation primitive f.Forcing coding y different with input x
In the case of, system can also go to restore original signal x, then illustrate to encode all information that y has carried initial data, but
By it is a kind of it is different in the form of.Here it is realize automatically feature extraction.
On the basis of autocoder, noise is added in training data, so autocoder must be learned by removing this
It plants noise and obtains the input that do not crossed by noise pollution really.Here it is DAE, as shown in Figure 3.And DAE is forced to go to learn
The more robust expression of input signal, this is also its generalization ability reason better than general autocoder effect.Assuming that
Initial data x " is deliberately destroyed ", for example Gauss white noise is added, or certain dimension datas are erased, and is becomeThen again
It is rightCoding, decoding, be restored signalThe recovery signal approaches unpolluted data x as far as possible.At this point,
The error of supervised training becomes from L (x, g (f (x)))
DAE is stacked, is exactly SDAE, as shown in Fig. 4 left-halfs.I.e. using the hidden layer of first DAE as second
A input layer.Shown in Fig. 4 is the block mold structure of proposition method of the present invention.
1.2 single Gauss models
It is analyzed by the characteristic parameter to extraction, Unimodal Distribution is presented in characteristic parameter.Thus, it is supposed that extraction obtained
Each characteristic parameter Gaussian distributed, is carried out abnormality detection using single Gaussian distribution model.
If only extracting one group of feature, using unitary Gaussian distribution model,
Wherein it is the mean value of μ training samples, h is the inverse of standard deviation sigma.According to 2 σ and 3 σ criterion, if test sample away from
Within 2 times from mean value or 3 times of standard deviations, then the correct probability of test sample is respectively：
P (| x- μ | 2 σ of ＜)=95% (2)
P (| x- μ | 3 σ of ＜)=99.73% (3)
In this way from the angle of probability theory, it is believed that be normal.Conversely, test sample is apart from 2 times or 3 times standard deviations of mean value
Other than difference, test sample is considered abnormal.
If extracting multigroup feature, using multivariate Gaussian models, conditional probability：
Wherein μ is the mean value of training sample, and ∑ is the covariance matrix of training sample.New sample x, grader define such as
Under：
Wherein θ is the threshold value for judging whether sample to be tested is abnormal, and h (x) is discriminant function, and 1 to represent be target class, i.e., normally
Value, 0 represents exceptional value.It is main so that the minimum criterion of error rate on target class sample to set θ.Single Gaussian distribution model is two
Dimension space should be similar to ellipse, and ellipsoid is similar on three dimensions.Wherein the setting of threshold value will at least ensure test sample
In exceptional value be all measured and just have application value.
2, the complex equipment method for detecting abnormality based on deep learning
2.1 method flow
In order to give full play to the effect of proposition method, it is necessary to explicitly for the specific knot of the abnormality detection model of complex equipment
Structure.Below by taking aero-engine as an example, the method for the present invention is described in detail.Wherein, to the choosing of Engine Parameter
It selects, the SDAE models concrete structure i.e. determination of the number of plies and number of nodes and the selection of single Gauss model threshold value are to abnormality detection model
Effect it is most important.
Step 1：The structure of sample set.The meaning of abnormality detection is rapidly, in time to report in producer CNR
Before notice, airline finds that the exception of aircraft engine.In aircraft operational process, most fast obtain is initial parameter.Cause
This, selects the original gas path parameter of aircraft engine that can ensure the promptness of abnormality detection.
The monitoring performance parameter of sample set chooses nine groups of aero-engine initial parameters during aircraft cruise as ginseng
It counts, is respectively：Sea Level Temperature, flying height, fan instruction rotating speed, core engine instruction rotating speed, oil liquid pressure, import total temperature,
EGT, oil liquid temperature, fuel flow and Mach number.
CNR (the Customer Notification Report) reports provided according to producer determine different in OEM data
Regular data.CNR public lectures provide the time that model, number, failure cause, the failure of sick engine occur.Normal data,
Fleet data are provided according to airline to determine.Fleet public lecture records the time residing for every engine condition of fleet, state
And situations such as the reason of Status Change.State include new hair it is initial, in the wing, to be repaired and throw a lease.It is considered that new hair is initial
It is normal condition to arrive afterwards and send the stage repaiied for the first time, to ensure the accuracy of normal data, when initial latter section is newly sent out in selection
Between data as normal sample.Normal sample according to a certain percentage and exceptional sample composing training sample set.In order to verify
The abnormality detection effect of the present invention, while the sample for having chosen some known states waits for test sample as in anomalies detecting step
This, hereon referred to as test sample collection.
Step 2：The pretreatment of data.Data in each sample set of structure are normalized.Due to each ginseng
Numberical range difference between number is larger, is directly placed into if being trained in model, the smaller parameter of numerical value can by numerical value compared with
It is approximately zero that big parameter, which is diluted to, and the information for being diluted to zero parameter also will loss.Normalized purpose is to reduce
Abnormal influence between this each parameter.
Step 3：The concrete structure of SDAE models.The number of plies that SDAE is stacked is excessive, not only cumbersome, but also is easily lost letter
Breath.The number of plies is very few, will lose the meaning of deep learning.When building three layers or SDAE more than three layers for the data of selection, mould
The reconstructed error of type is excessive.Again because SDAE at least needs two layers, consider, stacks two layers of DAE.
It is critical issue that how the number of nodes of every layer of SDAE, which determines,.The invention also provides the SDAE based on Controlling UEP
Model structure determines method.The specific steps of method and analysis will be illustrated in subsequent step.
Step 4：Build the Feature Selection Model based on SDAE.After the concrete structure of Feature Selection Model is determined,
Use MATLAB construction feature extraction models.Then training sample set, training pattern are inputted, trained model retains corresponding
Weight matrix it is constant.
Step 5：Build the abnormality detection model based on GSM.According to the feature trained through Feature Selection Model, structure
GSM models are built, characteristic mean is obtained and characteristic standard is poor.The key of structure GSM models is the determination of threshold value, and automatic threshold is arranged
Threshold value of the highest threshold value of accuracy of detection as model is chosen in selection.
Step 6：Test sample collection is carried out abnormality detection.Test sample collection is first inputted to the feature based on SDAE
Extraction model, obtained feature are input in the abnormality detection model based on GSM, carry out an abnormality detection.And analyze its test
Accuracy rate..
2.2 based on the SDAE model structures of Controlling UEP determination method
So far, there are no the model structures of the SDAE of generally acknowledged strong applicability to determine method in the prior art.It consults
The journal article for using SDAE both at home and abroad, is the model structure for directly giving SDAE mostly, does not provide and how to determine that model has
The specific method or thinking of body structure.In the research process to SDAE models, the sample set built for the present invention is found
Stronger linear dependence is presented by SDAE feature extractions in data mostly between obtained feature.According to this phenomenon, originally
Invention proposes the SDAE model structures based on Controlling UEP and determines method.
It is critical issue that how the number of nodes of every layer of SDAE, which determines,.The present invention uses the SDAE models based on Controlling UEP
Structure determination methodology builds the SDAE of multigroup different structure, inputs identical initial parameter.It analyzes obtaining feature, selects
Take the SDAE structures that feature behaves oneself best.The SDAE of the structure can be as the Feature Selection Model of experiment.
It is as follows：
Step 1：SDAE, the i.e. 9-X-Y for building 200 different structures, wherein 1≤X≤20,1≤Y≤10, X, Y are equal
For integer.It is excessively high and model will be increased if first hidden layer number X and the second too low information that will lose data of hidden layer number Y
Reconstructed error.Therefore, the SDAE of this 200 structures is had chosen, is analyzed to obtain the best model structure of effect.
Step 2：The initial parameter of 9 groups of training sample sets is input in the SDAE of this 200 different structure, obtains 200
Group feature.Every group include 1 to 10 unequal numbers feature.
Step 3：It is analyzed for the obtained features of the SDAE of each concrete structure, by analyzing the phase between feature
Guan Xing, with the quality of judging characteristic.For example, the SDAE of some structure obtains 4 features.The 1st feature and the 2nd, 3,4 is sought respectively
Then the degree of correlation of a feature seeks the degree of correlation of the 2nd feature and the 3rd, 4 feature, finally ask the 3rd feature and the 4th spy
The degree of correlation of sign.The degree of correlation is similar higher than the information that two features of threshold value retain, and only need to retain one between them, be same
One validity feature.Similarly, the degree of correlation all retains less than two features of threshold value.
Step 4：After having screened invalid feature, the selection of concrete structure is carried out.It, can be with according to the purpose of feature extraction
Think, retain the information of initial data, and the lower feature set that influences each other between each feature is preferably feature set.Institute
It is better with the lower characteristic effect of the degree of correlation between characteristic.Therefore, it according to the feature set degree of correlation situation of each SDAE, chooses each
A smaller SDAE structure of the degree of correlation between feature, as the Feature Selection Model in abnormality detection model.
Obtained feature is analyzed, following result is obtained：
(1) linear correlation (only exports a characteristic parameter to remove between the characteristic parameter that the SDAE of some structures is obtained
Outside), therefore the characteristic parameter of output can be regarded as only there are one effective characteristic parameters.For example, the SDAE that model structure is 9-1-6 is defeated
Go out 6 groups of characteristic parameters, the degree of correlation is as shown in table 1 between characteristic parameter.
1 model structure of table is the degree of correlation between the characteristic parameter that the SDAE of 9-1-6 is obtained
Correlation is too strong between the characteristic parameter obtained for the SDAE of above-mentioned 9-1-6 structures, the degree of correlation between each parameter
More than 0.9, therefore it is considered as only 1 effective characteristic parameters.Because the first hidden layer number of 9-1-6 is too small (only 1), damage
Bulk information is lost, therefore 6 groups of obtained features are linearly related each other.Therefore, five features are filtered out, only need to retain one
Feature is as validity feature.
(2) characteristic parameter of the SDAE outputs of some structure, by Feature Selection, the phase between remaining feature
Pass degree absolute value is still higher.Here it is considered that such SDAE is not preferable Feature Selection Model.As shown in Table 2 is 9-5-
The degree of correlation between the parameter that the SDAE of 6 structures is obtained.As can be seen from the table, the 3rd feature and the 4th feature needs screen out
One of them.Although other remaining features are not above threshold value, but the degree of correlation is still larger, the feature which goes out
Effect is bad.
2 structure of table is the degree of correlation between the characteristic parameter that 9-4-4SDAE is obtained
(3) there are also the characteristic parameters of the SDAE of structure outputs remains by Feature Selection (or there is no screen)
Degree of correlation absolute value between remaining feature is very low.The Feature Selection Model that such SDAE models have been considered as.Table 3 is 9-
The degree of correlation between the parameter that the SDAE of 13-5 structures is obtained.Degree of correlation absolute value between each feature is smaller, it is believed that special
It is preferable to levy extraction effect.
3 structure of table is the degree of correlation between the characteristic parameter that 9-7-8SDAE is obtained
In view of the purpose of feature extraction is on the basis of retaining initial parameter information, it is easier to realize different classes of
Classification.The SDAE for choosing 9-13-5 structures exports characteristic parameter as the feature carried out abnormality detection.
3, experimental analysis
It chooses test sample and concentrates 100 sample points, and affirm as 900 samples in normal time into test sample
Collection.In order to preferably be compared proposition method, setting adds the control group of single Gaussian distribution model with manual extraction feature.It surveys
It in this input model of sample, chooses under different threshold conditions, experimental group and the highest three groups of experiment knots of control group accuracy rate
Fruit such as following table：
Aero-engine point form method for detecting abnormality result of the table 4 based on SDAE
It can be seen that aero-engine point form method for detecting abnormality effect proposed by the present invention is preferable, main cause exists
It is effective in the characteristic extraction procedure based on deep learning.It is found during experiment, this method occurs for aero-engine
Vibrations failure (AVM) effect it is best, for other class failure effect unobvious.By learning CNR reports, conjecture is because of shake
Dynamic failure belongs to an exception, and other class failures belong to the exception of time series form mostly.And each sample point of parameter inputted
Independently of each other, therefore it is not easy the exception of detection time sequence form.For such case, when needing design for aero-engine
Between sequence form abnormal detection method.
Based on same inventive concept, the complex equipment point abnormality detection based on deep learning that the present invention also provides a kind of
System, including：Training sample processing module, characteristic extracting module, detection model training module and abnormality detection module.
Wherein, training sample processing module is used to choose the monitoring performance parameter of complex equipment, and obtains monitoring performance ginseng
Several normal samples and exceptional sample, composing training sample set.The training sample processing module and training sample in preceding method
The realization process of processing step S1 is identical, and details are not described herein.
Characteristic extracting module inputs the training sample set and carries out mould for building the Feature Selection Model based on SDAE
Type training.This feature extraction module is identical as the realization process of characteristic extraction step S2 in preceding method, and details are not described herein.
Detection model training module is used for according to exception of the feature construction trained through Feature Selection Model based on GSM
Detection model, characteristic mean and the characteristic standard for obtaining abnormality detection model are poor.The detection model training module and preceding method
The realization process of middle detection model training step S3 is identical, and details are not described herein.
Abnormality detection module is used to sample to be tested being input to the Feature Selection Model based on SDAE, obtained feature input
Into the abnormality detection model based on GSM, an abnormality detection is carried out.The abnormality detection module is walked with abnormality detection in preceding method
The realization process of rapid S4 is identical, and details are not described herein.
In conclusion the present invention is directed to the point abnormal conditions of complex equipment especially aero-engine, first according to aviation
The physical characteristic of the operation of engine and existing data set situation carry out parameter selection to the data that needs are analyzed, that is, select
Which input of the kind parameter as detection method selected；Then the pretreatment of the input data of selection is normalized；So
Processed data are input in Feature Selection Model afterwards, Feature Selection Model is based on stacking denoising autocoder
It is built-up, and propose a kind of characteristic model optimization method according to the feature degree of correlation；Finally the feature of extraction is inputted
To being carried out abnormality detection in grader, the abnormality detection model based on single Gauss model is established.Pass through experimental verification, card
Real complex equipment point form method for detecting abnormality effect proposed by the present invention is effective.
Claims (10)
1. a kind of complex equipment point method for detecting abnormality based on deep learning, which is characterized in that include the following steps：
Training sample processing step, the monitoring performance parameter for choosing complex equipment, and obtain the normal sample of monitoring performance parameter
And exceptional sample, composing training sample set；
Characteristic extraction step, Feature Selection Model of the structure based on SDAE, input the training sample set and carry out model training；
Detection model training step, according to abnormality detection mould of the feature construction trained through Feature Selection Model based on GSM
Type, characteristic mean and the characteristic standard for obtaining abnormality detection model are poor；
Sample to be tested is input to the Feature Selection Model based on SDAE by anomalies detecting step, and obtained feature, which is input to, to be based on
In the abnormality detection model of GSM, an abnormality detection is carried out.
2. the complex equipment point method for detecting abnormality according to claim 1 based on deep learning, which is characterized in that described
The data that training sample processing step also concentrates training sample are normalized.
3. the complex equipment point method for detecting abnormality according to claim 1 based on deep learning, which is characterized in that described
The stacking number of SDAE is two layers in characteristic extraction step.
4. the complex equipment point method for detecting abnormality described in any one of claim 1 to 3 based on deep learning, special
Sign is that the structure of the Feature Selection Model of the SDAE used in the characteristic extraction step is determined by following steps：
1) SDAE of 200 different structures is built, the structure of input layer-hidden layer-output layer of wherein DAE is a-X-Y, wherein 1
≤ X≤20,1≤Y≤10, X, Y are respectively hidden layer and the number of output layer, and are integer, and a is the number of input layer, with institute
The number for stating the monitoring performance parameter for the complex equipment chosen in training sample processing step is consistent；
2) initial parameter of training sample set is input in the SDAE of 200 different structures, obtains corresponding 200 groups of features
Collection；
3) it is analyzed for the obtained features of the SDAE of each structure, by analyzing the correlation between feature, screening is invalid
Feature；
4) after screening invalid feature, according to the feature set degree of correlation of each SDAE, the small SDAE of the degree of correlation between each feature is chosen
Structure, as the Feature Selection Model in abnormality detection model.
5. the complex equipment point method for detecting abnormality described in any one of claim 1 to 3 based on deep learning, special
Sign is that the abnormality detection model based on GSM established in the detection model training step is multivariate Gaussian models, condition
Probability is：
Wherein μ is the mean value of training sample, and ∑ is the covariance matrix of training sample；For sample to be tested x, grader defines such as
Under：
Wherein θ is the threshold value for judging whether sample to be tested is abnormal, and h (x) is discriminant function, and 1 represents normal value, and 0 represents exceptional value.
6. the complex equipment point method for detecting abnormality described in any one of claim 1 to 3 based on deep learning, special
Sign is that the complex equipment is aero-engine, the monitoring performance for the complex equipment that the training sample processing step is chosen
Parameter includes：Sea Level Temperature, flying height, fan instruction rotating speed, core engine instruction rotating speed, oil liquid pressure, import total temperature,
EGT, oil liquid temperature, fuel flow and Mach number.
7. a kind of complex equipment point abnormality detection system based on deep learning, which is characterized in that including：
Training sample processing module, the monitoring performance parameter for choosing complex equipment, and obtain the normal of monitoring performance parameter
Sample and exceptional sample, composing training sample set；
Characteristic extracting module inputs the training sample set and carries out model instruction for building the Feature Selection Model based on SDAE
Practice；
Detection model training module, for being examined according to exception of the feature construction trained through Feature Selection Model based on GSM
Model is surveyed, characteristic mean and the characteristic standard for obtaining abnormality detection model are poor；
Abnormality detection module, for sample to be tested to be input to the Feature Selection Model based on SDAE, obtained feature is input to
In abnormality detection model based on GSM, an abnormality detection is carried out.
8. the complex equipment point abnormality detection system according to claim 7 based on deep learning, which is characterized in that described
The structure of the Feature Selection Model of SDAE used in characteristic extracting module is determined by following steps：
1) SDAE of 200 different structures is built, the structure of input layer-hidden layer-output layer of wherein DAE is a-X-Y, wherein 1
≤ X≤20,1≤Y≤10, X, Y are respectively hidden layer and the number of output layer, and are integer, and a is the number of input layer, with institute
The number for stating the monitoring performance parameter for the complex equipment chosen in training sample processing step is consistent；
2) initial parameter of training sample set is input in the SDAE of 200 different structures, obtains corresponding 200 groups of features
Collection；
3) it is analyzed for the obtained features of the SDAE of each structure, by analyzing the correlation between feature, screening is invalid
Feature；
4) after screening invalid feature, according to the feature set degree of correlation of each SDAE, the small SDAE of the degree of correlation between each feature is chosen
Structure, as the Feature Selection Model in abnormality detection model.
9. the complex equipment point abnormality detection system according to claim 6 based on deep learning, which is characterized in that described
The abnormality detection model based on GSM established in detection model training module is multivariate Gaussian models, and conditional probability is：
Wherein μ is the mean value of training sample, and ∑ is the covariance matrix of training sample；For sample to be tested x, grader defines such as
Under：
Wherein θ is the threshold value for judging whether sample to be tested is abnormal, and h (x) is discriminant function, and 1 represents normal value, and 0 represents exceptional value.
10. the complex equipment point abnormality detection system based on deep learning according to any one of claim 6~9,
It is characterized in that, the complex equipment is aero-engine, the monitoring for the complex equipment that the training sample processing module is chosen
Can parameter include：Sea Level Temperature, flying height, fan instruction rotating speed, core engine instruction rotating speed, oil liquid pressure, import total temperature,
EGT, oil liquid temperature, fuel flow and Mach number.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810131253.0A CN108334907B (en) | 2018-02-09 | 2018-02-09 | A kind of complex equipment point method for detecting abnormality and system based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810131253.0A CN108334907B (en) | 2018-02-09 | 2018-02-09 | A kind of complex equipment point method for detecting abnormality and system based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108334907A true CN108334907A (en) | 2018-07-27 |
| CN108334907B CN108334907B (en) | 2019-05-17 |
Family
ID=62927314
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810131253.0A Active CN108334907B (en) | 2018-02-09 | 2018-02-09 | A kind of complex equipment point method for detecting abnormality and system based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108334907B (en) |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109035488A (en) * | 2018-08-07 | 2018-12-18 | 哈尔滨工业大学（威海） | Aero-engine time series method for detecting abnormality based on CNN feature extraction |
| CN109188211A (en) * | 2018-07-30 | 2019-01-11 | 上海交通大学 | A kind of high voltage equipment insulation method for diagnosing faults and system |
| CN109523171A (en) * | 2018-11-20 | 2019-03-26 | 未必然数据科技(北京)有限公司 | A kind of gas turbine inlet air system health degree appraisal procedure based on SVDD |
| CN109829903A (en) * | 2019-01-28 | 2019-05-31 | 合肥工业大学 | A kind of chip surface defect inspection method based on convolution denoising self-encoding encoder |
| CN110727669A (en) * | 2019-10-10 | 2020-01-24 | 国网辽宁省电力有限公司电力科学研究院 | Device and method for cleaning sensor data of power system |
| CN110991779A (en) * | 2018-09-30 | 2020-04-10 | 北京国双科技有限公司 | Anomaly detection method and device for oil pumping well |
| CN111064721A (en) * | 2019-12-11 | 2020-04-24 | 中国科学院深圳先进技术研究院 | Training method and detection method of network flow abnormity detection model |
| CN111080651A (en) * | 2019-12-12 | 2020-04-28 | 西南科技大学 | Automatic monitoring method for petroleum drilling polluted gas based on water flow segmentation |
| CN111160454A (en) * | 2019-12-27 | 2020-05-15 | 北京旋极信息技术股份有限公司 | Method and device for detecting speed change signal |
| CN111159961A (en) * | 2019-12-13 | 2020-05-15 | 杭州广立微电子有限公司 | Abnormity detection method and system based on curve data |
| CN111175593A (en) * | 2018-11-12 | 2020-05-19 | 株式会社东芝 | Method of detecting anomalies in a waveform and system therefor |
| CN111274543A (en) * | 2020-01-17 | 2020-06-12 | 北京空间飞行器总体设计部 | Spacecraft system anomaly detection method based on high-dimensional space mapping |
| CN112446002A (en) * | 2020-11-13 | 2021-03-05 | 天津大学 | Abnormity detection method for time sequence KPI data |
| CN112735094A (en) * | 2020-12-17 | 2021-04-30 | 中国地质环境监测院 | Geological disaster prediction method and device based on machine learning and electronic equipment |
| CN114037012A (en) * | 2021-11-09 | 2022-02-11 | 四川大学 | Flight data anomaly detection method based on correlation analysis and deep learning |
| CN114790991A (en) * | 2022-04-14 | 2022-07-26 | 国家电投集团电站运营技术（北京）有限公司 | Cavitation detection system and method for water feed pump |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104166787A (en) * | 2014-07-17 | 2014-11-26 | 南京航空航天大学 | Aero-engine remaining life prediction method based on multi-stage information fusion |
| CN106202668A (en) * | 2016-07-01 | 2016-12-07 | 中国航天标准化研究所 | Complex equipment quality risk appraisal procedure based on quality problems data and reverse conduction neutral net |
| CN106199174A (en) * | 2016-07-01 | 2016-12-07 | 广东技术师范学院 | Extruder energy consumption predicting abnormality method based on transfer learning |
| CN106707099A (en) * | 2016-11-30 | 2017-05-24 | 国网上海市电力公司 | Monitoring and locating method based on abnormal electricity consumption detection module |
| CN106778841A (en) * | 2016-11-30 | 2017-05-31 | 国网上海市电力公司 | The method for building up of abnormal electricity consumption detection model |
| 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 |
-
2018
- 2018-02-09 CN CN201810131253.0A patent/CN108334907B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104166787A (en) * | 2014-07-17 | 2014-11-26 | 南京航空航天大学 | Aero-engine remaining life prediction method based on multi-stage information fusion |
| CN106202668A (en) * | 2016-07-01 | 2016-12-07 | 中国航天标准化研究所 | Complex equipment quality risk appraisal procedure based on quality problems data and reverse conduction neutral net |
| CN106199174A (en) * | 2016-07-01 | 2016-12-07 | 广东技术师范学院 | Extruder energy consumption predicting abnormality method based on transfer learning |
| CN106707099A (en) * | 2016-11-30 | 2017-05-24 | 国网上海市电力公司 | Monitoring and locating method based on abnormal electricity consumption detection module |
| CN106778841A (en) * | 2016-11-30 | 2017-05-31 | 国网上海市电力公司 | The method for building up of abnormal electricity consumption detection model |
| 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 |
Non-Patent Citations (2)
| Title |
|---|
| SHASHAKANG: ""异常事件监测研究之路（一）"", 《CSDN—HTTPS://BLOG.CSDN.NET/SHASHAKANG/ARTICLE/DETAILS/68556539》 * |
| 小灰兔呼噜噜: ""Stanford机器学习---第十一讲.异常检测"", 《CSDN博客—HTTPS://BLOG.CSDN.NET/1281865263/ARTICLE/DETAILS/46654353》 * |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109188211A (en) * | 2018-07-30 | 2019-01-11 | 上海交通大学 | A kind of high voltage equipment insulation method for diagnosing faults and system |
| CN109035488A (en) * | 2018-08-07 | 2018-12-18 | 哈尔滨工业大学（威海） | Aero-engine time series method for detecting abnormality based on CNN feature extraction |
| CN110991779A (en) * | 2018-09-30 | 2020-04-10 | 北京国双科技有限公司 | Anomaly detection method and device for oil pumping well |
| CN111175593A (en) * | 2018-11-12 | 2020-05-19 | 株式会社东芝 | Method of detecting anomalies in a waveform and system therefor |
| CN111175593B (en) * | 2018-11-12 | 2022-06-28 | 株式会社东芝 | Method of detecting anomalies in waveforms and system therefor |
| CN109523171B (en) * | 2018-11-20 | 2022-11-25 | 未必然数据科技(北京)有限公司 | SVDD-based gas turbine air inlet system health degree evaluation method |
| CN109523171A (en) * | 2018-11-20 | 2019-03-26 | 未必然数据科技(北京)有限公司 | A kind of gas turbine inlet air system health degree appraisal procedure based on SVDD |
| CN109829903A (en) * | 2019-01-28 | 2019-05-31 | 合肥工业大学 | A kind of chip surface defect inspection method based on convolution denoising self-encoding encoder |
| CN110727669A (en) * | 2019-10-10 | 2020-01-24 | 国网辽宁省电力有限公司电力科学研究院 | Device and method for cleaning sensor data of power system |
| CN110727669B (en) * | 2019-10-10 | 2023-10-13 | 国网辽宁省电力有限公司电力科学研究院 | Electric power system sensor data cleaning device and cleaning method |
| CN111064721A (en) * | 2019-12-11 | 2020-04-24 | 中国科学院深圳先进技术研究院 | Training method and detection method of network flow abnormity detection model |
| CN111080651A (en) * | 2019-12-12 | 2020-04-28 | 西南科技大学 | Automatic monitoring method for petroleum drilling polluted gas based on water flow segmentation |
| CN111159961B (en) * | 2019-12-13 | 2024-02-02 | 杭州广立微电子股份有限公司 | Anomaly detection method and system based on curve data |
| CN111159961A (en) * | 2019-12-13 | 2020-05-15 | 杭州广立微电子有限公司 | Abnormity detection method and system based on curve data |
| CN111160454B (en) * | 2019-12-27 | 2023-12-22 | 北京旋极信息技术股份有限公司 | Quick change signal detection method and device |
| CN111160454A (en) * | 2019-12-27 | 2020-05-15 | 北京旋极信息技术股份有限公司 | Method and device for detecting speed change signal |
| CN111274543A (en) * | 2020-01-17 | 2020-06-12 | 北京空间飞行器总体设计部 | Spacecraft system anomaly detection method based on high-dimensional space mapping |
| CN112446002A (en) * | 2020-11-13 | 2021-03-05 | 天津大学 | Abnormity detection method for time sequence KPI data |
| CN112446002B (en) * | 2020-11-13 | 2022-11-15 | 天津大学 | Abnormity detection method for time sequence KPI data |
| CN112735094A (en) * | 2020-12-17 | 2021-04-30 | 中国地质环境监测院 | Geological disaster prediction method and device based on machine learning and electronic equipment |
| CN114037012B (en) * | 2021-11-09 | 2023-04-07 | 四川大学 | Flight data anomaly detection method based on correlation analysis and deep learning |
| CN114037012A (en) * | 2021-11-09 | 2022-02-11 | 四川大学 | Flight data anomaly detection method based on correlation analysis and deep learning |
| CN114790991A (en) * | 2022-04-14 | 2022-07-26 | 国家电投集团电站运营技术（北京）有限公司 | Cavitation detection system and method for water feed pump |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108334907B (en) | 2019-05-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108334907B (en) | A kind of complex equipment point method for detecting abnormality and system based on deep learning | |
| CN106895975B (en) | Method for Bearing Fault Diagnosis based on Stacked SAE deep neural network | |
| CN110779746B (en) | Diagnosis method for improving composite fault of deep sparse self-encoder network rotating machinery | |
| CN107657250B (en) | Bearing fault detection and positioning method and detection and positioning model implementation system and method | |
| CN110132598A (en) | Slewing rolling bearing fault noise diagnostics algorithm | |
| CN108363382A (en) | A kind of complex equipment fault diagnosis method and system | |
| DE102016107064A1 (en) | System level fault diagnosis for the air control system of an aircraft | |
| CN108537259A (en) | Train control on board equipment failure modes and recognition methods based on Rough Sets Neural Networks model | |
| Fu et al. | A re-optimized deep auto-encoder for gas turbine unsupervised anomaly detection | |
| CN110110768A (en) | Fault Diagnosis of Roller Bearings based on Concurrent Feature study and multi-categorizer | |
| CN109813542A (en) | The method for diagnosing faults of air-treatment unit based on production confrontation network | |
| CN113642754A (en) | Complex industrial process fault prediction method based on RF noise reduction self-coding information reconstruction and time convolution network | |
| CN110647830A (en) | Bearing fault diagnosis method based on convolutional neural network and Gaussian mixture model | |
| Sakthivel et al. | Use of histogram features for decision tree-based fault diagnosis of monoblock centrifugal pump | |
| CN115587290A (en) | Aero-engine fault diagnosis method based on variational self-coding generation countermeasure network | |
| Åsnes et al. | Predictive maintenance and life cycle estimation for hydro power plants with real-time analytics | |
| CN110554682A (en) | fault detection inference machine based on fault correlation analysis | |
| CN106909136A (en) | A kind of method for diagnosing faults based on index regularization kernel linear discriminant analysis | |
| Xuyun et al. | A new point anomaly detection method about aero engine based on deep learning | |
| CN115356599B (en) | Multi-mode urban power grid fault diagnosis method and system | |
| CN116521406A (en) | Method for detecting anomaly of non-overrun flight parameter data of aero-engine based on residual gate GRU-VAE model | |
| JiaJun et al. | Self-supervised representation learning for first-shot unsupervised anomalous sound detection | |
| Baggeröhr et al. | Novel bearing fault detection using generative adversarial networks | |
| CN115392109A (en) | LSTM multivariable time series anomaly detection method based on generative model | |
| Klein et al. | Decision and fusion for diagnostics of mechanical components |
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 |