Deep learning model based on recurrence plot and the application in the measurement of profit phase content
Technical field
The present invention relates to a kind of deep learning model.More particularly to one kind is obtained for four sector distributing triggers reorganizations
To multi-channel data build the deep learning model based on recurrence plot and profit phase content measurement in application.
Background technology
Oil-water two-phase flow is widely present in oil exploitation and transportation industry.In water-oil phase streaming system, the distribution of each phase
It is being continually changing with space over time, is defining different nowed formings, referred to as flow pattern.The flow pattern of two phase flow is complicated more
Become, local flow information is difficult to accurately seizure so that the measurement of the two-phase flow parameter such as phase content has many difficult points.This is right
Many impacts are caused in oil exploitation and technological transformation.At present, for the research of flow pattern mainly adopts observation method, little baud
Levy analysis and fuzzy C-means clustering, fuzzy logic and genetic algorithm, Digital Image Processing algorithm etc..For phase content measurement more adopt
With conductance method, capacitance method, optical method and ray method etc..The annular conductivity sensor and double helix electric capacity that traditional measurement is adopted is passed
Sensor etc., is single channel sensor, is easily lost the local flow information of microcosmic.And distributed conductivity sensor and excitation are followed
Ring stimulus sensor etc. then can simultaneously gather multi channel signals, capture more rich micro flow information, be flow pattern and phase
Research containing rate provides important technology support.
Neutral net has obtained in the past few decades tremendous development as a kind of grader.It is based primarily upon logarithm
According to feature learnt with realize classification.But past training pattern great majority are shallow-layers, and classifying quality is unable to reach
Very accurate stage.As the problem of gradient disappearance in neutral net is by effectively solving, the neutral net of profound level is built as deeply
Degree confidence network is possibly realized.Deep learning model is obtained significantly compared to shallow-layer network in feature extraction and classificatory performance
Lifted.
Recurrence plot is an important and effective instrument in Nonlinear Time Series Analysis, and is obtained extensively multi-field
Using especially for unstable, short time series, its analytical effect is very notable.It can be realized to unitary signal
The visualization of the dynamic characteristic of the trajectory of phase space that reconstruct is obtained.Different time sequence is disclosed using picture from intuitively angle
Between intrinsic characteristic.
The content of the invention
The technical problem to be solved is to provide a kind of deep learning model based on recurrence plot and in profit phase
Containing the application in rate measurement.
The technical solution adopted in the present invention is:A kind of foundation of the deep learning model based on recurrence plot, including it is as follows
Step:
1) recurrence plot is built based on four sector distributing triggers reorganization data, including:
(1) for each channel data of four sector distributing triggers reorganizations is regarded as unitary time series uk, to described
Unitary time series ukPhase space reconfiguration is carried out, a sequence vector is obtainedAny one vector in the sequence vector
It is expressed as:
Wherein, m is Embedded dimensions, is determined using wrong nearest neighbor method;τ is time delay, is determined using mutual information method;
(2) for unitary time series ukObtain a sequence vectorFor sequence vectorDefine recurrence plot:
Wherein, Ri,jAny point in the recursion matrix corresponding to recurrence plot is represented, ε is threshold value, using 15% unitary
Seasonal effect in time series standard deviation is determined,Represent the distance between two vectors;Represent ifThen its value is 1, ifThen its value is 0;So, if Ri,jValue be 1, then in recurrence plot
In be black, if Ri,jValue be 0, then be in recurrence plot white;
2) deep learning model training, including:
(1) by step 1) in the recurrence plot that obtains, after carrying out one-dimensional, as the depth confidence net in deep learning model
The input of network;
(2) using greedy successively training algorithm, generation model parameter is obtained to depth confidence network training, described is greedy
Greedy successively training is, using the training algorithm of unsupervised learning, in training from the bottom to top, a limited Bohr to be constituted hereby per two-layer
Graceful machine, with limited Boltzmann machine parameter is obtained to sdpecific dispersion Algorithm for Training, then fixes the layer parameter, continues to last layer structure
The limited Boltzmann machine made is trained, until top terminates;
(3) the limited Boltzmann machine parameter for being obtained with (2nd) step as initial value, for reset deep neural network just
Beginning weight, global training is carried out using the method for having supervision to whole network parameter, is finally given and be can be used for Accurate classification
Neural network parameter.
Step 2) described in depth confidence network be the generative probabilistic model being made up of the implicit stochastic variable of multilayer, by one
The limited Boltzmann machine stacking of fixed number purpose is formed, and then Down-Up carries out successively pre-training;Stacking process is as follows:Training one
After the limited Boltzmann machine of individual Bernoulli Jacob-Bernoulli Jacob, it is limited the activation probability of hidden unit as next layer of Bernoulli Jacob-Bernoulli Jacob
The input data of Boltzmann machine;The activation primitive of the limited Boltzmann machine of second layer Bernoulli Jacob-Bernoulli Jacob is used as third layer primary
The visible input data of the limited Boltzmann machine of Nu Li-Bernoulli Jacob, later each layer is by that analogy;Described limited Boltzmann machine
It is to be built based on the theory of successively greedy learning strategy.
Step 2) described in greed successively to train be training algorithm using unsupervised learning, i.e. training process need not
Know the classification belonging to training sample, be so very easy to obtain substantial amounts of training sample;In training from the bottom to top, per two-layer structure
Into a limited Boltzmann machine, with limited Boltzmann machine parameter is obtained to sdpecific dispersion Algorithm for Training, this layer of ginseng is then fixed
Number, continues to be trained the limited Boltzmann machine of last layer construction, until top terminates.
Step 2) (3rd) step described in global training be, using the training algorithm of supervised learning, to use band phase content
The training sample of label is trained, at the beginning of being the limited Boltzmann machine parameter as deep neural network obtained using (2nd) step
Beginning parameter, and it is finely adjusted training with back-propagation algorithm.
Application of a kind of deep learning model based on recurrence plot in the measurement of water-oil phase flow containing rate, using four sectors
Distributing triggers reorganization carries out the proportioning of vertical oil-water two-phase flow experiment, fixed oil phase and water phase, change oil phase and water phase
Flow is tested;Experiment comprises the steps:
1) proportioning of fixed water phase and oil phase, is passed through a certain amount of water, then gradually to pipeline toward vertical ascent pipeline
In be passed through oil phase, when water-oil phase fully merge and it is gradually stable after, led to using four sector distributing triggers reorganizations collections more
Road measurement signal, and record flow pattern with high-speed camera instrument;After once collection terminates, change the flow of oil phase and water phase, by above-mentioned
Process continues to gather, until designed operating mode under fixed proportioning is all completed, then changes the proportioning of oil phase and water phase, repeats
Said process completes different water phases and the measurement under oil phase proportioning, until all operating modes of design are all measured;
2) based on step 1) multi-channel measurement signal, for each channel signal, to be based on four sectors distributed according to above-mentioned
Conductivity sensor data build the method for recurrence plot and build recurrence plot;
3) after obtaining the recurrence plot of all operating modes, using the recurrence plot of a portion as depth confidence network training
Collection, the recurrence plot of another part as depth confidence network test set, using recurrence plot as depth confidence network input, institute
State depth confidence network to be made up of the limited Boltzmann machine of multilayer, the input of the limited Boltzmann machine of each layer is the defeated of last layer
Go out, be trained by the unsupervised weight to each layer of Boltzmann machine of the mechanism successively trained;
4) weight of the weights initialisation deep neural network of each layer after training of limited Boltzmann machine is used;
5) deep neural network will be input into as the recurrence plot of training set, the weight employing to deep neural network has supervision
Training be finely adjusted;
6) by the unsupervised learning and supervised learning to sample, a depth confidence network based on recurrence plot is obtained
Model;By the model realization in oilfield exploitation under unknown operating mode to effective measurement of phase content.
The deep learning model based on recurrence plot of the present invention and the application in the measurement of profit phase content, by fanning to four
The multichannel sensor data that area's distributing triggers reorganization is obtained are visualized using the method for recurrence plot, and recurrence plot is made
For the input of deep learning model, allow model to be trained by the unsupervised and supervised learning to great amount of samples, obtain one
The individual deep learning model based on recurrence plot, and use it for the measurement to oil-water two-phase flow flow parameter such as phase content.
Description of the drawings
Fig. 1 is deep learning model schematic of the present invention based on recurrence plot.
Specific embodiment
With reference to embodiment and accompanying drawing to the deep learning model based on recurrence plot of the present invention and in profit phase content
Application in measurement is described in detail.
The deep learning model based on recurrence plot of the present invention and the application in the measurement of profit phase content, there is provided Yi Zhongji
In recurrence plot deep learning model and use it for water-oil phase flow containing rate measurement in, by four sector distributed electricals
The multi-channel data that derivative sensor is obtained is visualized using the method for recurrence plot, using recurrence plot as deep learning model
Input, allows the model to be trained by the unsupervised and supervised learning to great amount of samples, obtains one based on recurrence plot
Deep learning model, and use it for the measurement of oil-water two-phase flow flow parameter such as phase content.
As shown in figure 1, the foundation of the deep learning model based on recurrence plot of the present invention, comprises the steps:
1) recurrence plot is built based on four sector distributing triggers reorganization data, including:
(1) for each channel data of four sector distributing triggers reorganizations is regarded as unitary time series uk, to described
Unitary time series ukPhase space reconfiguration is carried out, a sequence vector is obtainedAny one vector in the sequence vector
It is represented by:
Wherein, m is Embedded dimensions, is determined using wrong nearest neighbor method;τ is time delay, is determined using mutual information method;
(2) for unitary time series ukObtain a sequence vectorFor sequence vectorDefine recurrence plot:
Wherein, Ri,jAny point in the recursion matrix corresponding to recurrence plot is represented, ε is threshold value, using 15% unitary
Seasonal effect in time series standard deviation is determined,Represent the distance between two vectors;Represent ifThen its value is 1, ifThen its value is 0;So, if Ri,jValue be 1, then in recurrence plot
In be black, if Ri,jValue be 0, then be in recurrence plot white;
2) deep learning model training, including:
(1) by step 1) in the recurrence plot that obtains, after carrying out one-dimensional, as the depth confidence net in deep learning model
The input of network;
Described depth confidence network is the generative probabilistic model being made up of the implicit stochastic variable of multilayer, by certain amount
Limited Boltzmann machine (RBM) stacking is formed, and then Down-Up carries out successively pre-training;Stacking process is as follows:Training one
After Bernoulli Jacob-Bernoulli Jacob RBM, hidden unit is activated into probability as the input data of next layer of Bernoulli Jacob-Bernoulli Jacob RBM;The
The activation primitive of two layers of Bernoulli Jacob-Bernoulli Jacob RBM as third layer Bernoulli Jacob-Bernoulli Jacob RBM visible input data, later respectively
Layer is by that analogy;Described limited Boltzmann machine is built based on the theory of successively greedy learning strategy.
(2) using greedy successively training algorithm, generation model parameter is obtained to depth confidence network training;
It is that using the training algorithm of unsupervised learning, i.e. training process requires no knowledge about training that described greediness is successively trained
Classification belonging to sample, is so very easy to obtain substantial amounts of training sample;In training from the bottom to top, one is constituted per two-layer
RBM, with to sdpecific dispersion Algorithm for Training RBM parameters are obtained, and then fix the layer parameter, continue to carry out the RBM of last layer construction
Training, until top terminates.
(3) the RBM parameters for being obtained with (2nd) step, for resetting the initial weight of deep neural network, are adopted as initial value
The method for having supervision carries out global training to whole network parameter, finally gives the neutral net ginseng that can be used for Accurate classification
Number;
Described global training is, using the training algorithm of supervised learning, to be carried out with the training sample with phase content label
Training, be the RBM parameters obtained using (2nd) step as deep neural network initial parameter, and carried out with back-propagation algorithm micro-
Adjust training.Due to have passed through the greed of the first step successively pre-training, therefore when global training is carried out, can solve well to pass
The problem of local best points is easily converged on system to deep layer network training mode.
Application of the deep learning model based on recurrence plot of the present invention in the measurement of water-oil phase flow containing rate, adopts four
Sector distributing triggers reorganization carries out the proportioning of vertical oil-water two-phase flow experiment, fixed oil phase and water phase, changes oil phase and water
The flow of phase is tested;Experiment comprises the steps:
1) proportioning of fixed water phase and oil phase, is passed through a certain amount of water, then gradually to pipeline toward vertical ascent pipeline
In be passed through oil phase, when water-oil phase fully merge and it is gradually stable after, led to using four sector distributing triggers reorganizations collections more
Road measurement signal, and record flow pattern with high-speed camera instrument;Sample frequency is 4000Hz, and the sampling time is 30s.Once collection terminates
Afterwards, change the flow of oil phase and water phase, continue to gather by said process, until designed operating mode under fixed proportioning is all complete
Into, then change the proportioning of oil phase and water phase, repeat said process and complete different water phases and the measurement under oil phase proportioning, until design
All operating modes be all measured;
2) based on step 1) multi-channel measurement signal, for each channel signal, to be based on four sectors distributed according to above-mentioned
Conductivity sensor data build the method for recurrence plot and build recurrence plot;
3) after obtaining the recurrence plot of all operating modes, using the recurrence plot of a portion as depth confidence network training
Collection, the recurrence plot of another part as depth confidence network test set, using recurrence plot as depth confidence network input, institute
State depth confidence network to be made up of the limited Boltzmann machine of multilayer, the input of the limited Boltzmann machine of each layer is the defeated of last layer
Go out, be trained by the unsupervised weight to each layer of Boltzmann machine of the mechanism successively trained;
4) weight of the weights initialisation deep neural network of each layer after training of limited Boltzmann machine is used;
5) deep neural network will be input into as the recurrence plot of training set, the weight employing to deep neural network has supervision
Training be finely adjusted;
6) by the unsupervised learning and supervised learning to sample, a depth confidence network based on recurrence plot is obtained
Model;By the model realization in oilfield exploitation under unknown operating mode to effective measurement of phase content.
Using said method, can be with the inherence of the sensor electric signal of the inherent texture structure of recurrence plot reflection non-stationary
Nonlinear dynamic characteristic.Deep learning model then can pass through to extract the feature of recurrence plot and be trained, and realize that classification is distinguished
Know.Using recurrence plot as bridge, effective identification of the deep learning model to non-stationary electric signal is realized.
The multi channel signals obtained for four sector distributing triggers reorganization measurements build recurrence plot, using recurrence plot as
Deep learning mode input, by training and study to great amount of samples, obtains the depth of an achievable phase content measurement
Model is practised, realization is in oilfield exploitation to the accurate measurement of phase content under unknown operating mode.
Above to the description of the present invention and embodiment, it is not limited to which this, the description in embodiment is only the reality of the present invention
One of mode is applied, it is any to design and the technical scheme without creative in the case of without departing from the invention objective
Similar structure or embodiment, belong to protection scope of the present invention.