CN108897947A - A kind of equipment degeneration multi-source data fusion method based on improvement variation autocoding - Google Patents
A kind of equipment degeneration multi-source data fusion method based on improvement variation autocoding Download PDFInfo
- Publication number
- CN108897947A CN108897947A CN201810681053.2A CN201810681053A CN108897947A CN 108897947 A CN108897947 A CN 108897947A CN 201810681053 A CN201810681053 A CN 201810681053A CN 108897947 A CN108897947 A CN 108897947A
- Authority
- CN
- China
- Prior art keywords
- variation
- source data
- degeneration
- model
- data fusion
- 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
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/04—Ageing analysis or optimisation against ageing
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Computational Mathematics (AREA)
- Mathematical Optimization (AREA)
- General Engineering & Computer Science (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Analysis (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Operations Research (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Computation (AREA)
- Geometry (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Hardware Design (AREA)
- Algebra (AREA)
- Probability & Statistics with Applications (AREA)
- Databases & Information Systems (AREA)
- Software Systems (AREA)
- Compression, Expansion, Code Conversion, And Decoders (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of based on the equipment degeneration multi-source data fusion method for improving variation autocoding, using logarithm normal distribution as the prior distribution of the hidden variable of variation autocoder, construct corresponding cost function regularization expression, combined data normalization and model batch processing training, obtain the fusion results of degeneration multi-source data.The prior distribution for the hidden variable that the present invention is encoded using logarithm normal distribution as variation certainly, and corresponding regularization expression is constructed, to retain the Singular variance prior information of degenerate state.
Description
Technical field
The present invention relates to modeling and the predicting residual useful life fields of degenerating, and in particular to one kind is based on improving variation autocoding
Equipment degeneration multi-source data fusion method.
Background technique
With the continuous development of science and technology and industrial level, mechanical equipment structure is increasingly complicated, operating environment requirements day
Beneficial harsh, user is also higher and higher to the reliability of mechanical equipment and security requirement.Therefore, to the degenerate state of mechanical equipment
It is effectively extracted, the abnormal state information of timely discovering device operation has weight to guarantee mechanical equipment safe and reliable operation
Want meaning.
The degradation information of mechanical equipment often resides in the device status information that the more physical property sensor monitorings of multi-source obtain,
How information is obtained to multisensor and carry out effective integration, obtain comprehensive degradation information, is to carry out the extraction of equipment degenerate state
With one step of key of predicting residual useful life.Common multi-sources Information Fusion Method is based primarily upon dimension reduction method, as principal component analysis,
Core principle component analysis and autocoder etc..These methods focus primarily on the reservation multi-source data for making fusion results more as far as possible
All information.However, being used for the degenerative character of predicting residual useful life, often have uncertainty early period bigger, latter end is not true
Surely smaller Singular variance data structure is spent.The above method is often lost the prior information of fusion results itself Singular variance.
Summary of the invention
It is an object of the invention to it is a kind of based on improve variation autocoding equipment degeneration multi-source data fusion method, with
Overcome the problems of the above-mentioned prior art, the priori for the hidden variable that the present invention is encoded using logarithm normal distribution as variation certainly
Distribution, and corresponding regularization expression is constructed, to retain the Singular variance prior information of degenerate state.
In order to achieve the above objectives, the present invention adopts the following technical scheme that:
A kind of equipment degeneration multi-source data fusion method based on improvement variation autocoding, is made using logarithm normal distribution
For the prior distribution of the hidden variable of variation autocoder, corresponding cost function regularization expression, combined data normalizing are constructed
Change and model batch processing is trained, obtains the fusion results of degeneration multi-source data.
Further, include the following steps:
Step 1:Construct the logarithm normal distribution variation autocoder network model of Three Tiered Network Architecture, the three-layer network
Network structure includes coding layer, resampling layer and decoding layer;
Step 2:Degeneration multi-source data is normalized, foundation is used for the logarithm normal distribution variation of step 1 building certainly
The training of encoder network model, test sample collection;
Step 3:Using variation Bayes principle, the corresponding KL of logarithm normal distribution variation self-encoding encoder network model is derived
Divergence regularization expression;
Step 4:The training parameter for the logarithm normal distribution variation self-encoding encoder network model that setting steps 1 construct, and just
Beginningization Model Weight;
Step 5:Model is trained using training sample set data, iteration updates model training parameter, and passes through survey
Examination collection verifies training result, obtains multisource data fusion result.
Further, coding layer described in step 2 includes mean value network and variance network, the weight and biasing of mean value network
It is expressed as:Wmean、bmean, the weight of variance network and biasing are expressed as:WvarAnd bvar, the corresponding parameter μ of logarithm normal distribution and
σ is expressed as respectively:
μ=Wmean+bmean
σ=Wvar+bvar
The resampling layer, which can obtain implicit layer state by unitary sampling and be sent to decoding layer, carries out original sample solution
Code.
Further, the corresponding KL divergence regularization of logarithm normal distribution variation autocoder is expressed as in step 3:
Wherein, pθIt (z) is the prior distribution of degeneration multisource data fusion result, θ={ 0, σ1It is the corresponding parameter being distributed
Collection, qφ(z) point of the data fusion result calculated for the logarithm normal distribution variation self-encoding encoder network model constructed by step 1
Cloth, φ={ μ, σ } are the parameter set that corresponding coding layer calculates.
Further, model training parameter includes model learning rate, input and output dimension, batch processing scale in step 4
And training the number of iterations.
Further, weights initialisation is in step 4Being uniformly distributed in range,
Middle ninAnd noutFor the number of nodes of input layer and output layer.
Further, the model learning rate is gradually reduced with model iterative process.
Further, multisource data fusion results expression is in step 5:
Wherein z fusion results, { Wmean,Wvarbmean,bvarIt is the obtained corresponding weight of coding layer of training and biasing, x are
Input sample.
Compared with prior art, the invention has the following beneficial technical effects:
The method of the present invention remains the prior information of fusion results, and compared to existing method, this method being capable of effective acquisition
More reasonable fusion results, by analyzing result, the fluctuation that the method for the present invention obtains result is smaller, can guarantee rear
The uncertainty of calculated result is smaller when phase life prediction, and in addition the method for the present invention can guarantee that certain right avertence is presented in fusion results
Property, it is more in line with the prior information of fusion results itself, is more in line with engineering demand.
Further, the fusion results of the method for the present invention are simultaneously containing variation from the letter of encoded mean value network and variance network
Breath, can make full use of all weights and offset information of cataloged procedure, ensure that the robustness of algorithm.
Further, model learning rate is gradually reduced with model iterative process, is guaranteed in model training early period, calculated result
When apart from optimal solution farther out, single step obtains larger update amplitude, in the model training later period, comparison of computational results close to optimal solution,
Single step, which updates, carries out lesser improvement.
Detailed description of the invention
Fig. 1 is original multi-source data sample instance described in present example;
Fig. 2 is that degenerative character described in present example extracts flow chart;
Fig. 3 is data characteristics fusion results described in present example;
Fig. 4 is fusion results overall distribution histogram described in present example;
Fig. 5 is conventional method fusion results distribution histogram.
Specific embodiment
Present invention is further described in detail below:
A kind of equipment degeneration multi-source data fusion method based on improvement variation autocoding is more in equipment non-linear degradation
The Prior Fusion result of source data replaces original standard normal point using logarithm normal distribution there are in the case of obvious right avertence
Cloth, constructs corresponding cost function regularization expression, and the processes such as combined data normalization and model batch processing training are obtained and degenerated
The fusion results of multi-source data, this method make the dextrality of data be guaranteed.
Specifically include following steps:
Step 1:Building includes " coding-resampling-decoding " Three Tiered Network Architecture logarithm normal distribution variation autocoding
Device network model, " coding-resampling-decoding " Three Tiered Network Architecture, coding layer include " mean value " network and " variance " variance
Network, weight and biasing are represented sequentially as:Wmean、bmean、WvarAnd bvar, the corresponding parameter μ of logarithm normal distribution and σ can divide
μ=W is not expressed as itmean+bmean, σ=Wvar+bvar, resampling layer can be obtained by unitary sampling to imply layer state and transmits
Original sample decoding is carried out to decoding layer;
Step 2:Degeneration multi-source data is normalized, the sample set for model training, test is established;
Step 3:Using variation Bayes principle, the corresponding KL divergence of logarithm normal distribution variation autocoder is being derived just
Then change expression, formula is as follows:
Wherein, pθIt (z) is the prior distribution of degeneration multisource data fusion result, θ={ 0, σ1It is the corresponding parameter being distributed
Collection, qφIt (z) is the distribution of the data fusion result calculated by model, φ={ μ, σ } is the parameter set that corresponding coding layer calculates;
Step 4:The model trainings such as model learning rate, input and output dimension, batch processing scale, training the number of iterations are set
Parameter, and initialization model weight, weights initialisation areBeing uniformly distributed in range, wherein
ninAnd noutFor the number of nodes of front and back two-tier network, to guarantee that information can effectively be transmitted in a network with appropriate amplitude, learning rate
Parameter is gradually reduced with model iterative process, is guaranteed in model training early period, calculated result apart from optimal solution farther out when, single step
Larger update amplitude is obtained, in the model training later period, for comparison of computational results close to optimal solution, single step, which updates, carries out lesser change
It is kind;
Step 5:Model is trained using training sample set data, iteration updates model parameter, and passes through test set
Training result is verified, and exports accordingly result, export results expression is:
Wherein z fusion results, { Wmean,Wvarbmean,bvar, for the corresponding weight of coding layer and biasing that training obtains, x
For input sample.
The embodiment of the present invention is described in detail with reference to the accompanying drawing:
Certain aero-engine degeneration multi-source data is analyzed, which is made of 21 dimension degeneration indexs, such as Fig. 1 institute
Show, can effectively be fused to single index using the present invention, this example corresponding operating flow chart is as shown in Fig. 2, specific steps
It is as follows:
(1) variation autoencoder network model is constructed, input layer and output layer are 21 number of nodes (initial data dimension), are implied
Layer is 1 node, and it is 0.01 that initial learning rate, which is arranged,.
(2) max min method for normalizing is used, training set and test set data are normalized, make to count
According to being distributed in [0,1] section;
(3) network model is trained using training set data, and using stochastic gradient descent method to model parameter
It is iterated update.
(4) model training result is tested using test set data, testing model training result, verifying model is effective
Property, Fusion Features result as shown in figure 3, the distribution histogram of fusion results as shown in figure 4, in order to illustrate the superior of innovatory algorithm
Property, fusion results are compared with traditional Feature fusion (Fig. 5) based on autocoder, as seen from Figure 3
Improved method, the fluctuation for obtaining result is smaller, can guarantee the uncertainty of the calculated result when remanent life is predicted
It is smaller, by Fig. 4 comparison it can be seen that improved method, can guarantee that certain dextrality is presented in fusion results, be more in line with and melt
The prior information for closing result itself, is more in line with engineering demand.
Claims (8)
1. a kind of based on the equipment degeneration multi-source data fusion method for improving variation autocoding, which is characterized in that utilize logarithm
Prior distribution of the normal distribution as the hidden variable of variation autocoder constructs corresponding cost function regularization expression, knot
Data normalization and model batch processing training are closed, the fusion results of degeneration multi-source data are obtained.
2. a kind of equipment degeneration multi-source data fusion method based on improvement variation autocoding according to claim 1,
It is characterized by comprising the following steps:
Step 1:Construct the logarithm normal distribution variation autocoder network model of Three Tiered Network Architecture, the three-layer network knot
Structure includes coding layer, resampling layer and decoding layer;
Step 2:Degeneration multi-source data is normalized, the logarithm normal distribution variation constructed for step 1 is established and encodes certainly
The training of device network model, test sample collection;
Step 3:Using variation Bayes principle, the corresponding KL divergence of logarithm normal distribution variation self-encoding encoder network model is derived
Regularization expression;
Step 4:The training parameter for the logarithm normal distribution variation self-encoding encoder network model that setting steps 1 construct, and initialize
Model Weight;
Step 5:Model is trained using training, test sample collection data, iteration updates model training parameter, and passes through survey
Examination collection verifies training result, obtains multisource data fusion result.
3. a kind of equipment degeneration multi-source data fusion method based on improvement variation autocoding according to claim 2,
It is characterized in that, coding layer described in step 2 includes mean value network and variance network, the weight of mean value network and biasing are indicated
For:Wmean、bmean, the weight of variance network and biasing are expressed as:WvarAnd bvar, the corresponding parameter μ of logarithm normal distribution and σ divide
It is not expressed as:
μ=Wmean+bmean
σ=Wvar+bvar
The resampling layer, which can obtain implicit layer state by unitary sampling and be sent to decoding layer, carries out original sample decoding.
4. a kind of equipment degeneration multi-source data fusion method based on improvement variation autocoding according to claim 3,
It is characterized in that, the corresponding KL divergence regularization of logarithm normal distribution variation autocoder is expressed as in step 3:
Wherein, pθIt (z) is the prior distribution of degeneration multisource data fusion result, θ={ 0, σ1It is the corresponding parameter set being distributed, qφ
(z) distribution of the data fusion result calculated for the logarithm normal distribution variation self-encoding encoder network model constructed by step 1, φ
={ μ, σ } is the parameter set that corresponding coding layer calculates.
5. a kind of equipment degeneration multi-source data fusion method based on improvement variation autocoding according to claim 4,
It is characterized in that, model training parameter includes model learning rate, input and output dimension, batch processing scale and instruction in step 4
Practice the number of iterations.
6. a kind of equipment degeneration multi-source data fusion method based on improvement variation autocoding according to claim 5,
It is characterized in that, weights initialisation is in step 4Being uniformly distributed in range, wherein ninWith
noutFor the number of nodes of input layer and output layer.
7. a kind of equipment degeneration multi-source data fusion method based on improvement variation autocoding according to claim 5,
It is characterized in that, the model learning rate is gradually reduced with model iterative process.
8. a kind of equipment degeneration multi-source data fusion method based on improvement variation autocoding according to claim 3,
It is characterized in that, multisource data fusion results expression is in step 5:
Wherein z fusion results, { Wmean,Wvarbmean,bvarIt is the corresponding weight of coding layer and biasing that training obtains, x is input
Sample.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810681053.2A CN108897947B (en) | 2018-06-27 | 2018-06-27 | Equipment degradation multi-source data fusion method based on improved variational automatic coding |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810681053.2A CN108897947B (en) | 2018-06-27 | 2018-06-27 | Equipment degradation multi-source data fusion method based on improved variational automatic coding |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108897947A true CN108897947A (en) | 2018-11-27 |
CN108897947B CN108897947B (en) | 2021-02-19 |
Family
ID=64346534
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810681053.2A Active CN108897947B (en) | 2018-06-27 | 2018-06-27 | Equipment degradation multi-source data fusion method based on improved variational automatic coding |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108897947B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112599246A (en) * | 2021-03-03 | 2021-04-02 | 四川华迪信息技术有限公司 | Vital sign data processing method, system, device and computer readable medium |
CN112926505A (en) * | 2021-03-24 | 2021-06-08 | 重庆大学 | Rotating machine health index construction method based on DTC-VAE neural network |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080215298A1 (en) * | 2006-10-10 | 2008-09-04 | Haney Philip J | Parameterization of non-linear/non-gaussian data distributions for efficient information sharing in distributed sensor networks |
CN101539241A (en) * | 2009-05-07 | 2009-09-23 | 北京航空航天大学 | Hierarchical multi-source data fusion method for pipeline linkage monitoring network |
CN107341220A (en) * | 2017-06-28 | 2017-11-10 | 阿里巴巴集团控股有限公司 | A kind of multi-source data fusion method and device |
CN108110790A (en) * | 2017-12-15 | 2018-06-01 | 国网江苏省电力有限公司南京供电分公司 | Active distribution network operating status appraisal procedure based on multi-source data staged dynamic fuzzy system |
-
2018
- 2018-06-27 CN CN201810681053.2A patent/CN108897947B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080215298A1 (en) * | 2006-10-10 | 2008-09-04 | Haney Philip J | Parameterization of non-linear/non-gaussian data distributions for efficient information sharing in distributed sensor networks |
CN101539241A (en) * | 2009-05-07 | 2009-09-23 | 北京航空航天大学 | Hierarchical multi-source data fusion method for pipeline linkage monitoring network |
CN107341220A (en) * | 2017-06-28 | 2017-11-10 | 阿里巴巴集团控股有限公司 | A kind of multi-source data fusion method and device |
CN108110790A (en) * | 2017-12-15 | 2018-06-01 | 国网江苏省电力有限公司南京供电分公司 | Active distribution network operating status appraisal procedure based on multi-source data staged dynamic fuzzy system |
Non-Patent Citations (3)
Title |
---|
ZHILIANG FAN 等: "Degradation data-driven approach for remaining useful life estimation", 《JOURNAL OF SYSTEMS ENGINEERING AND ELECTRONICS》 * |
何正嘉 等: "机械设备运行可靠性评估的发展与思考", 《机械工程学报》 * |
王帅 等: "含裂纹离心压缩机叶轮结构的振动局部化", 《振动与冲击》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112599246A (en) * | 2021-03-03 | 2021-04-02 | 四川华迪信息技术有限公司 | Vital sign data processing method, system, device and computer readable medium |
CN112599246B (en) * | 2021-03-03 | 2021-06-29 | 四川华迪信息技术有限公司 | Vital sign data processing method, system, device and computer readable medium |
CN112926505A (en) * | 2021-03-24 | 2021-06-08 | 重庆大学 | Rotating machine health index construction method based on DTC-VAE neural network |
CN112926505B (en) * | 2021-03-24 | 2022-11-11 | 重庆大学 | Rotating machine health index construction method based on DTC-VAE neural network |
Also Published As
Publication number | Publication date |
---|---|
CN108897947B (en) | 2021-02-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108985929B (en) | Training method, business data classification processing method and device, and electronic equipment | |
CN109784015B (en) | Identity authentication method and device | |
CN111368920A (en) | Quantum twin neural network-based binary classification method and face recognition method thereof | |
CN112761628B (en) | Shale gas yield determination method and device based on long-term and short-term memory neural network | |
CN112702329B (en) | Traffic data anomaly detection method and device and storage medium | |
CN113221104B (en) | Detection method of abnormal behavior of user and training method of user behavior reconstruction model | |
CN110647916B (en) | Pornographic picture identification method and device based on convolutional neural network | |
CN110084609B (en) | Transaction fraud behavior deep detection method based on characterization learning | |
CN112291098B (en) | Network security risk prediction method and related device thereof | |
CN113569243A (en) | Deep semi-supervised learning network intrusion detection method based on self-supervised variation LSTM | |
CN113395276A (en) | Network intrusion detection method based on self-encoder energy detection | |
CN108897947A (en) | A kind of equipment degeneration multi-source data fusion method based on improvement variation autocoding | |
CN107609589A (en) | A kind of feature learning method of complex behavior sequence data | |
CN112699998A (en) | Time series prediction method and device, electronic equipment and readable storage medium | |
CN107240100B (en) | Image segmentation method and system based on genetic algorithm | |
CN115659254A (en) | Power quality disturbance analysis method for power distribution network with bimodal feature fusion | |
CN108491719A (en) | A kind of Android malware detection methods improving NB Algorithm | |
CN117041017A (en) | Intelligent operation and maintenance management method and system for data center | |
CN115293235A (en) | Method for establishing risk identification model and corresponding device | |
CN114679310A (en) | Network information security detection method | |
CN112464281B (en) | Network information analysis method based on privacy grouping and emotion recognition | |
CN115438753B (en) | Method for measuring security of federal learning protocol data based on generation | |
Wang et al. | Research on feature selection method of intrusion detection based on deep belief network | |
CN111209567B (en) | Method and device for judging perceptibility of improving robustness of detection model | |
CN115878391A (en) | Method and device for detecting disk abnormality |
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 |