CN105426889A - PCA mixed feature fusion based gas-liquid two-phase flow type identification method - Google Patents

Abstract

The invention relates to the technical field of gas-liquid two-phase flow measurement, in particular to a PCA mixed feature fusion based gas-liquid two-phase flow type identification method. The PCA mixed feature fusion based gas-liquid two-phase flow type identification method comprises: acquiring a flow type image by utilizing a high-speed camera; performing feature analysis on a preprocessed image; extracting three features including an invariant moment feature, a gray-level cooccurrence matrix feature and an LBP feature; performing fusion and dimension reduction on the three features, and performing classification identification by adopting a BP neutral network; and finally, applying a classification identification result to online automatic detection of two-phase flow. The three-feature fusion based method can better reflect image information. The mixed feature dimension number is relatively high, the calculation complexity is relatively high, and the consumed time is long, so that online detection is not facilitated. With the adoption of a PCA technology, the data dimensions can be effectively reduced, the calculation amount can be reduced, original information can be kept from being lost to a great extent, namely, little information is used for representing the original information as far as possible, and data are compressed.

Description

The Method for Discriminating Gas-liquid Two Phase Flow that Based PC A composite character merges

Technical field

The present invention relates to biphase gas and liquid flow field of measuring technique, especially relate to the flow pattern of gas-liquid two-phase flow knowledge method for distinguishing that Based PC A composite character merges.

Background technology

Extensively there is two-phase flow in nature and engineering field, all have a wide range of applications in fields such as oil, power, refrigeration, nuclear energy, metallurgy, water conservancy, environmental protection, building and space flight.And biphase gas and liquid flow is one of form the most common in diphasic flow.The heat transfer of two-phase fluid, mass transfer characteristic can be subject to the impact of flow pattern, and the Measurement accuracy of other parameters of two-phase flow also often depends on the understanding of convection.It is the basis determining heattransfer and fluid flow, and especially whether flow pattern accuracy of judgement decides the precision that pressure drop calculates, and selects corresponding pressure drop computing formula, can improve the accuracy of pressure drop result of calculation according to flow pattern.In runner the change of flow pattern often also can cause flow resistance, flowing stability change and there is bad heat transfer crisis.Therefore the differentiation studying flow pattern of gas-liquid two-phase flow just seems particularly important.

The research of two phase flow pattern and conversion characteristic thereof, be two-phase flow research in be substantially the most also one of sixty-four dollar question.At present, flow type identification method can be divided into two classes: a class directly determines flow pattern according to the form of two phase flow image, and as ocular estimate, human eye catches and identifies flow pattern, and in high-velocity duct, accuracy rate is not high; Attenuation sensors, utilizes ray by the principle determination flow pattern of medium generation attenuation by absorption, but be difficult to obtain stable radiographic source, and the radiativity of ray itself also should be noted that protection; Process tomographic imaging method, mainly utilize sensor to carry out image reconstruction and realize on-line checkingi, its speed is lower, can not meet required requirement of real-time.Another kind of is indirect method, fluctuation signal analytic approach, by carrying out Treatment Analysis to the fluctuation signal of reflection two phase flow characteristic, extracting flow pattern features, and then identifying flow pattern.No matter utilize the extraction of Differential Pressure Fluctuation signal of Gas or the extraction of conductance fluctuation signal all existed to the problem of interference flowing field; Digital image processing method is a kind of comparatively emerging method, and obtain analyze qualitatively by carrying out process to picture signal, by force visual, informative, objectively responds flow phenomenon, and not interference flowing field.But in digital image processing method, during for feature extraction, characteristic information is comparatively single, only includes a kind of information, can not comprehensive representation image information.And composite character dimension is higher, complexity is comparatively large, and not easily calculate, the application thus in online automatic detection is still limited, requires further study.

Summary of the invention

The object of the invention is to overcome the deficiencies in the prior art, extract three kinds of features, image information is comparatively comprehensive, and effectively reduce composite character dimension, reduce calculated amount, recognition speed is high, identifies the recognition methods of the flow pattern of gas-liquid two-phase flow that accuracy is high, reliability is high, the Based PC A composite character of highly versatile, support online automatic detection merges.

For achieving the above object, be below the technical scheme of this invention:

The Method for Discriminating Gas-liquid Two Phase Flow that Based PC A composite character merges, it comprises the following steps:

1) flow pattern image of the biphase gas and liquid flow in rectification column under different conditions is obtained by high-speed camera;

2) extraction interesting image regions (ROI region), carries out medium filtering to image and contrast strengthen convection image carries out the pre-service of removal noise;

3) analytical procedure 2) shape of pretreated two-phase flow image and textural characteristics, extract Image Moment Invariants, gray level co-occurrence matrixes and LBP feature;

4) by step 3) after three kinds of Fusion Features, utilize PCA technology, dimensionality reduction is carried out to the composite character after merging, obtain a new proper vector, this new proper vector comprises textural characteristics and geometric properties information;

5) adopt step 4) in new proper vector carry out contrast experiment with three kinds of independent features respectively, utilize support vector machine, BP neural network and probabilistic neural network to carry out training and identifying respectively.

Preferably, described step 1) in flow pattern image be the flow pattern image of slug flow under different gas-liquid flow, wave flow, mist annular flow 3 kinds of typical flow patterns.

Described step 1) the typical flow pattern image that size is 600 × 600 is collected to slug flow, wave flow, mist annular flow 3 kinds of typical flow patterns.

Described step 2) image is carried out to the process of region of interesting extraction, size is 300 × 60.

Described step 2) described medium filtering process, the pixel of image local area is sorted by gray shade scale, gets the gray-scale value of intermediate value as current pixel of gray scale in this neighborhood; Make the actual value that the pixel value of neighborhood is close, thus eliminate isolated noise spot, its expression formula is as follows:

g(x,y)＝med{f(x-k,y-l),(k,l∈W)}

In formula, f (x, y) and g (x, y) is respectively original image and the rear image of process, and W is two dimension pattern plate, selects 3 × 3 regions;

Described step 2) described contrast strengthen pre-service is the histogram equalization method adopted in contrast strengthen: establish original image at (x, y) gray scale at place is g, and the image after changing is h, then can be expressed as the method for image enhaucament and the gray scale g at (x, y) place is mapped as h.In gray-level histogram equalization process, the mapping function of image be may be defined as:

h＝EQ(g)。

Preferably, described step 3) the described invariant moment features main shape facility of image-region, it has the invariant features of the characteristics such as rotation, translation, yardstick, gray level co-occurrence matrixes is that the spatial correlation characteristic by studying gray scale carrys out Description Image texture, the texture information of image local can be measured and extract to LBP feature, has unchangeability to illumination.

Preferably, described step 3) extract the not bending moment of image, gray level co-occurrence matrixes and LBP feature composition flow pattern features vector respectively, utilize the principle of not bending moment, gray level co-occurrence matrixes and LBP feature, extract image 7 respectively and tie up invariant moment features vector, be designated as image 8 ties up gray level co-occurrence matrixes proper vector, is designated as f ₁, f ₂..., f ₈, image 59 ties up LBP proper vector, is designated as p ₁, p ₂..., p ₅₉.

Wherein, described not bending moment, gray level co-occurrence matrixes and LBP feature are respectively:

A) geometric properties of invariant moment features main image-region, because have the invariant features such as rotation, translation, yardstick, there is global property, strong interference immunity, in image procossing, object can be represented as an important feature, thus feature can carry out the operations such as classification to image accordingly;

For the image that intensity profile is h (x, y), its common square in (p+q) rank and centre distance are defined as:

$m_{pq}=\underset{x=1}{\overset{M}{\Σ}}\underset{y=1}{\overset{N}{\Σ}}{x}^p{y}^{q}h(x,y)$

${\mathrm{\μ}}_{pq}=\underset{x=1}{\overset{M}{\Σ}}\underset{y=1}{\overset{N}{\Σ}}{(x-x_0)}^p{(y-y_0)}^{q}h(x,y)$

In formula, p, q=0,1,2 ..., the centre of moment (x ₀, y ₀) be

For two dimensional image, x ₀represent gradation of image grey scale centre of gravity in the horizontal direction, y ₀represent gradation of image grey scale centre of gravity in vertical direction;

(p+q) normalization center square is defined as:

$y_{pq}=\frac{{\mathrm{\μ}}_{pq}}{{\mathrm{\μ}}_{00}^r}$

In formula, $r=\frac{p+q+2}{2},p+q=2,3,\mathrm{....}$

Utilize second order and three normalization center, rank squares can derive 7 not bending moment groups as proper vector;

B) gray level co-occurrence matrixes

Gray level co-occurrence matrixes describes texture by the spatial correlation characteristic studying gray scale, can reflect the integrated information of gradation of image about direction, adjacent spaces, amplitude of variation.Characterize the feature of gray level co-occurrence matrixes with some scalars, the feature that gray level co-occurrence matrixes is conventional has to make G represent:

Energy $ASM={\Σ}_{i=1}^k{\Σ}_{j=1}^k{\left(G\right(i,j\left)\right)}^2$

Entropy $ENT=-{\mathrm{\Σ}}_{i=1}^k{\mathrm{\Σ}}_{j=1}^{k}G(i,j)logG(i,j)$

Auto-correlation $COR={\mathrm{\Σ}}_{i=1}^k{\mathrm{\Σ}}_{j=1}^k\frac{(i,j)G(i,j)-{\mathrm{\μ}}_i{\mathrm{\μ}}_j}{s_i{s}_j}$

Unfavourable balance square $IDM={\mathrm{\Σ}}_{i=1}^k{\mathrm{\Σ}}_{j=1}^k\frac{G(i,j)}{1+{(i-j)}^2}$

Calculate selecting 0 °, 45 °, 90 °, 135 ° four directions respectively to the same characteristic parameter of same piece image, the textural characteristics parameter of invariable rotary can be obtained, so just, inhibit durection component on the impact of result, the integrated information of gradation of image about direction, adjacent spaces, amplitude of variation can be reflected; By energy, entropy, unfavourable balance square, autocorrelative average and standard deviation, be respectively f ₁, f ₂, f ₃, f ₄, f ₅, f ₆, f ₇, f ₈, as final 8 dimension textural characteristics;

C) LBP feature

LBP is used for Description Image Local textural feature, has the significant advantage such as rotational invariance and gray scale unchangeability, for texture feature extraction; The window of definition 3 × 3, with window center pixel for threshold value, the gray-scale value of adjacent 8 pixels is compared with it, if surrounding pixel values is greater than the value of central point, then this location of pixels is labeled 1, otherwise is 0, finally the binary number around central pixel point is turned to decimal number, obtain LBP value, its computing formula is as follows:

${\mathrm{LBP}}_{P,R}=\underset{p=0}{\overset{P-1}{\Σ}}s(g_p-g_c)2^p$

$s\left(x\right)=\left\{\begin{array}{c}1,x\&GreaterEqual;0\\ 0,x<0\end{array}\right.$

Adopt LBP equivalent formulations, namely on the basis of traditional LBP algorithm, select from 0 to 1 or be no more than the eight-digit binary number sequence of 2 times from the saltus step of 1 to 0, be converted into the LBP value as this window after the decimal system, the feature of extraction is the textural characteristics p of the local of image ₁, p ₂..., p ₅₉; Such histogram becomes 59 dimensions from original 256 dimensions, serves the effect of dimensionality reduction;

$U\left({\mathrm{LBP}}_{P,R}\right)=|s(g_{P-1}-g_c)-s(g_0-g_c)|+\underset{p=1}{\overset{P-1}{\&Sigma;}}|s(g_p-g_c)-s(g_{P-1}-g_c)|$

LBP equivalent formulations greatly reduces the kind of binary mode, and does not lose any information, reduces the dimension of proper vector, decreases the impact that high frequency noise produces.

Preferably, described step 4) Based PC A technology, three kinds of principal eigenvector in the luv space data of said extracted are merged, obtains 74 dimensional feature vectors, linear transformation is carried out to this vector, i.e. eigencenter, ask for covariance matrix and eigenwert thereof and proper vector, the descending arrangement of eigenwert, due to front 7 eigenwerts and exceeded 95% of all eigenwert sums, choose front 7 eigenwert characteristics of correspondence vector, extract major component and obtain 7 dimensional feature vector t ₁, t ₂..., t ₈, then this vector is normalized, form the data set in new lower dimensional feature space.

Preferred, step 4) key step is as follows:

A) feature extraction is carried out to training sample, the totally 74 dimensional feature vector { m of three kinds of features after extraction ₁, m ₂..., m ₇₄, as this matrix M of PCA former state,

B) average to matrix M, namely every one-dimensional data all deducts this dimension average, obtains matrix B,

$\stackrel{\&OverBar;}{m}=\frac{1}{M}\underset{i=1}{\overset{M}{\&Sigma;}}{m}_i$

C) the covariance matrix C of B is calculated,

$C=\frac{1}{M}\underset{i=1}{\overset{M}{\&Sigma;}}(x_i-\stackrel{\&OverBar;}{m}){(x_i-\stackrel{\&OverBar;}{m})}^T$

D) eigenvalue λ and the proper vector y of covariance matrix C is calculated,

(λE _M-C)y＝0

E) eigenwert is sorted according to order from big to small, selects wherein maximum 7, then 7 of its correspondence proper vectors are formed eigenvectors matrix as column vector, obtain principal component transform matrix D,

F) projected to by sample point in the proper vector chosen, this matrix M of former state and principal component transform matrix D convert, and former state notebook data dimension is reduced to 7 dimensions.

Preferably, two-phase flow image recognition and classification, based on support vector cassification identification, using the support vector of the proper vector of training sample as support vector machine, select Radial basis kernel function, training is carried out to whole training sample and obtains supporting vector machine model, utilize the model obtained to carry out testing and predict; Based on the Classification and Identification of BP neural network and probabilistic neural network, all using the training sample of proper vector as neural network, be configured to the neural network structure of Flow Regime Ecognition, relevant training parameter is set, and carry out learning training, use the test sample book of this neural network to different flow pattern to identify.

Using the proper vector of 7 dimensional feature vectors after 7 dimension invariant moment features obtained above, 8 dimension gray scale symbiosis moment characteristics and 59 dimension LBP Fusion Features dimensionality reductions as piece image, be training sample, training pattern is constructed by the input and output relation of support vector machine, BP neural network and probabilistic neural network self, wherein support vector machine selects Radial basis kernel function, and the gamma function coefficients in kernel function is set to 1.When setting up BP neural network, 10 hidden neurons, 3 output neurons, hidden layer and output layer transition function are logarithm S shape transfer function and linear transfer function respectively, and training method adopts learning rate changing momentum gradient descent algorithm.When creating probabilistic neural network, select radial basis function, its propagation coefficient is set to 0.6.Respectively test sample book is detected by above-mentioned training pattern.

Preferably, also comprise convection image recognition evaluation: according to recognition result, by calculating is compared with the relevant position in classified image and picture of classifying in the position of each actual measurement pixel and classification, the precision of classification results is presented at inside a confusion matrix, and match stop result and actual measured value are evaluated.

In the precision of images is evaluated, confusion matrix is used for match stop result and actual measured value, better visual, better in this, as evaluation criterion, composite character is merged the eigenvector recognition result after dimensionality reduction and single eigenvector recognition result compares, result shows that the recognition methods precision of composite character is the highest, and the time is shorter.So, the method is applied to the detection of two-phase flow, and accuracy rate is higher, reliability is stronger, speed is faster, support on-line automatic identification.

Train as training sample with 90 samples (3 kinds flow pattern each 30), then by the training pattern obtained, 60 samples (3 kinds flow pattern each 20) are predicted as test sample book.To obtain a result and to represent with confusion matrix, each row of confusion matrix represent prediction classification, the sum of each row represents the number being predicted as such other data, every a line represents the true belonging kinds of data, the data count of every a line represents the number of such other data instance, and the numeric representation True Data in each row is predicted to be such number.As shown in table 1,2,3, each numerical value represents discrimination, for correctly to identify number in bracket.

In order to verify the reliability of this method, adopt composite character respectively with other three kinds of independent feature (not bending moments, gray level co-occurrence matrixes, LBP feature) carry out contrast experiment, meanwhile, when feature is identical, we have employed three kinds of sorter (BP neural networks, support vector machines, probabilistic neural network PNN) carry out contrasting with the best sorter of Selection effect.Recognition result as proper vector after three kinds of composite characters fusion dimensionality reductions is best, be best suited for the characteristic of division as two phase flow pattern identification, and the method accuracy of identification combined with BP neural network is the highest, BP neural network is in identifying, good than other two kinds of sorter performances, accuracy of identification is high.Approximately, but dimension significantly reduces the recognition result of composite character and LBP feature, and therefore, the method that composite character and BP neural network combine is better than other several combinations, and the training time is also shorter, is applicable to flow pattern ONLINE RECOGNITION.

The present invention will be further described below:

The Method for Discriminating Gas-liquid Two Phase Flow that 1 Based PC A composite character merges, it comprises the following steps:

1) biphase gas and liquid flow image acquisition, utilize high-speed camera, the flow proportional of biphase gas and liquid flow is changed in horizontal pipe, obtain the image of gas-liquid two-phase flow pattern under different operating mode, obtain the slug flow under different gas-liquid flows, wave flow, mist annular flow 3 kinds of typical flow pattern images, 3 kinds of typical flow pattern images are gathered;

2) biphase gas and liquid flow Image semantic classification, first extracts area-of-interest (ROI region) to the flow pattern image collected, then carries out the pre-service of removal noise, utilizes medium filtering and contrast strengthen convection image to carry out the pre-service of removal noise;

3) biphase gas and liquid flow image characteristic analysis and extraction, analyze shape and the textural characteristics of pretreated two-phase flow image, the invariant moment features main shape facility of flow pattern regions, it has the invariant features of the characteristics such as rotation, translation, yardstick, gray level co-occurrence matrixes is that the spatial correlation characteristic by studying gray scale describes two phase flow pattern image texture, the texture information of image local can be measured and extract to LBP feature, has unchangeability to illumination.According to these characteristics, extract the not bending moment of flow pattern image, gray level co-occurrence matrixes and LBP feature composition flow pattern features vector respectively, utilize its principle to extract flow pattern image 7 respectively and tie up invariant moment features vector, be designated as image 8 ties up gray level co-occurrence matrixes proper vector, is designated as f ₁, f ₂..., f ₈, flow pattern image 59 ties up LBP proper vector, is designated as p ₁, p ₂..., p ₅₉;

4) biphase gas and liquid flow multi-features, Based PC A technology, three kinds of principal eigenvector in the luv space data of said extracted are merged, obtain 74 dimensional feature vectors, linear transformation is carried out to this vector, i.e. eigencenter, ask for covariance matrix and eigenwert thereof and proper vector, the descending arrangement of eigenwert, due to front 7 eigenwerts and exceeded 95% of all eigenwert sums, choose front 7 eigenwert characteristics of correspondence vector, extract major component and obtain 7 dimensional feature vector t ₁, t ₂..., t ₇, then this vector is normalized, form the data set in new lower dimensional feature space, decrease the redundancy of data, retain the characteristic information of the overwhelming majority in original feature space simultaneously, thus the problem that the intrinsic dimensionality solving extraction is too high;

5) biphase gas and liquid flow image recognition and classification, based on support vector cassification identification, using the support vector of the proper vector of training sample as support vector machine, select Radial basis kernel function, training is carried out to whole training sample and obtains supporting vector machine model, utilize the model obtained to carry out testing and predict; Based on the Classification and Identification of BP neural network and probabilistic neural network, all using the training sample of proper vector as neural network, be configured to the neural network structure of Flow Regime Ecognition, relevant training parameter is set, and carry out learning training, use the test sample book of this neural network to different flow pattern to identify;

6) flow pattern image identification and evaluation, according to recognition result, by calculating is compared with the relevant position in classified image and picture of classifying in the position of each actual measurement pixel and classification, is presented at the precision of classification results inside a confusion matrix.In the precision of images is evaluated, confusion matrix is used for match stop result and actual measured value, better visual, better in this, as evaluation criterion, composite character is merged the eigenvector recognition result after dimensionality reduction and single eigenvector recognition result compares, result shows that the recognition methods precision of composite character is the highest, and the time is shorter.So, the method is applied to the detection of two-phase flow, and accuracy rate is higher, reliability is stronger, speed is faster, support on-line automatic identification.

Technical conceive of the present invention is: utilize high-speed camera to gather flow pattern image, signature analysis is carried out to pretreated image, extract three kinds of features, not bending moment, gray level co-occurrence matrixes and LBP feature, after three kinds of Fusion Features dimensionality reductions, adopt BP neural network to carry out Classification and Identification, be finally applied to the online automatic detection of two-phase flow.

Beneficial effect of the present invention is mainly manifested in:

1) based on the method for digital image processing techniques, during for feature extraction, characteristic information is comparatively single, only includes a kind of information, can not comprehensive representation image information.And the present invention adopts the method for three kinds of Fusion Features, image information can be reflected better.And composite character dimension is higher, computation complexity is comparatively large, and length consuming time, is unfavorable for on-line checkingi.Adopt PCA technology, significantly reduce data dimension, simplify calculated amount, original information can also be retained largely and be lost, namely characterize original information by the least possible information, data are compressed.

2) composite character selected of the method, merges not bending moment, gray level co-occurrence matrixes and LBP feature, can describe the essential information of flow pattern from texture and shape preferably, reflects the feature between flow pattern more all sidedly.The geometric properties in invariant moment features token image region, has the invariant features such as rotation, translation, yardstick.Gray level co-occurrence matrixes reflects the integrated information of gray scale about direction, adjacent spaces, amplitude of variation of flow pattern image.LBP Description Image Local textural feature, has the significant advantage such as rotational invariance and gray scale unchangeability.

Accompanying drawing explanation

Fig. 1 is the Method for Discriminating Gas-liquid Two Phase Flow process flow diagram that Based PC A composite character merges.

Embodiment

Drawings and Examples are utilized to be described further the Method for Discriminating Gas-liquid Two Phase Flow that the present invention is based on the fusion of PCA composite character below.

As shown in Figure 1, the Method for Discriminating Gas-liquid Two Phase Flow that the present invention is based on the fusion of PCA composite character comprises the steps:

1) biphase gas and liquid flow image acquisition, utilize high-speed camera, the flow proportional of biphase gas and liquid flow is changed in horizontal pipe, obtain the image of gas-liquid two-phase flow pattern under different operating mode, obtain the slug flow under different gas-liquid flows, wave flow, mist annular flow 3 kinds of typical flow pattern images, 3 kinds of typical flow pattern images are gathered, collects the typical flow pattern image that size is 600 × 600.

2) biphase gas and liquid flow Image semantic classification, first extracts area-of-interest (ROI region) to the flow pattern image collected, then carries out the pre-service of removal noise, utilizes medium filtering and contrast strengthen convection image to carry out the pre-service of removal noise.

A) the image non-critical areas owing to collecting is more, so carry out the process of region of interesting extraction to image, size is 300 × 60.

B) medium filtering process, sorts the pixel of image local area by gray shade scale, gets the gray-scale value of intermediate value as current pixel of gray scale in this neighborhood.Make the actual value that the pixel value of neighborhood is close, thus eliminate isolated noise spot, its expression formula is as follows:

g(x,y)＝med{f(x-k,y-l),(k,l∈W)}

In formula, f (x, y) and g (x, y) is respectively original image and the rear image of process, and W is two dimension pattern plate, selects 3 × 3 regions.

After medium filtering process, the impulsive noise in image is by filtering well, and fog-level obviously reduces, and protects the marginal information of signal, makes it not fuzzy.Its method comparison is simple, consuming time shorter.

C) histogram equalization method in contrast strengthen is adopted, after treatment, add the dynamic range of grey scale pixel value, make gradation of image distribution uniformity, eliminate in image and cross bright or excessively dark noise, thus reach the effect strengthening integral image contrast.

If the gray scale of original image at (x, y) place is g, and the image after changing is h, then can be expressed as the method for image enhaucament and the gray scale g at (x, y) place is mapped as h.In gray-level histogram equalization process, the mapping function of image be may be defined as:

h＝EQ(g)。

3) biphase gas and liquid flow image characteristic analysis and extraction, analyze shape and the textural characteristics of pretreated two-phase flow image, the invariant moment features main shape facility of image-region, it has the invariant features of the characteristics such as rotation, translation, yardstick, gray level co-occurrence matrixes is that the spatial correlation characteristic by studying gray scale carrys out Description Image texture, the texture information of image local can be measured and extract to LBP feature, has unchangeability to illumination.According to these characteristics, extract the not bending moment of image, gray level co-occurrence matrixes and LBP feature composition flow pattern features vector respectively, utilize its principle to extract image 7 respectively and tie up invariant moment features vector, be designated as image 8 ties up gray level co-occurrence matrixes proper vector, is designated as f ₁, f ₂..., f ₈, image 59 ties up LBP proper vector, is designated as p ₁, p ₂..., p ₅₉.

A) geometric properties of invariant moment features main image-region, because have the invariant features such as rotation, translation, yardstick, there is global property, strong interference immunity, in image procossing, object can be represented as an important feature, thus feature can carry out the operations such as classification to image accordingly.

For the image that intensity profile is h (x, y), its common square in (p+q) rank and centre distance are defined as:

$m_{pq}=\underset{x=1}{\overset{M}{\&Sigma;}}\underset{y=1}{\overset{N}{\&Sigma;}}{x}^p{y}^{q}h(x,y)$

${\mathrm{\&mu;}}_{pq}=\underset{x=1}{\overset{M}{\&Sigma;}}\underset{y=1}{\overset{N}{\&Sigma;}}{(x-x_0)}^p{(y-y_0)}^{q}h(x,y)$

In formula, p, q=0,1,2 ..., the centre of moment (x ₀, y ₀) be

For two dimensional image, x ₀represent gradation of image grey scale centre of gravity in the horizontal direction, y ₀represent gradation of image grey scale centre of gravity in vertical direction;

(p+q) normalization center square is defined as:

$y_{pq}=\frac{{\mathrm{\&mu;}}_{pq}}{{\mathrm{\&mu;}}_{00}^r}$

In formula, $r=\frac{p+q+2}{2},p+q=2,3,\mathrm{....}$

Utilize second order and three normalization center, rank squares can derive 7 not bending moment groups as proper vector.

B) gray level co-occurrence matrixes

Gray level co-occurrence matrixes describes texture by the spatial correlation characteristic studying gray scale, can reflect the integrated information of gradation of image about direction, adjacent spaces, amplitude of variation.Characterize the feature of gray level co-occurrence matrixes with some scalars, the feature that gray level co-occurrence matrixes is conventional has to make G represent:

Energy $ASM={\&Sigma;}_{i=1}^k{\&Sigma;}_{j=1}^k{\left(G\right(i,j\left)\right)}^2$

Entropy $ENT=-{\mathrm{\&Sigma;}}_{i=1}^k{\mathrm{\&Sigma;}}_{j=1}^{k}G(i,j)logG(i,j)$

Auto-correlation $COR={\mathrm{\&Sigma;}}_{i=1}^k{\mathrm{\&Sigma;}}_{j=1}^k\frac{(i,j)G(i,j)-{\mathrm{\&mu;}}_i{\mathrm{\&mu;}}_j}{s_i{s}_j}$

Unfavourable balance square $IDM={\mathrm{\&Sigma;}}_{i=1}^k{\mathrm{\&Sigma;}}_{j=1}^k\frac{G(i,j)}{1+{(i-j)}^2}$

0 °, 45 °, 90 °, 135 ° four directions are selected to calculate respectively to the same characteristic parameter of same piece image, the textural characteristics parameter of invariable rotary can be obtained, so just, inhibit durection component on the impact of result, the integrated information of gradation of image about direction, adjacent spaces, amplitude of variation can be reflected; By energy, entropy, unfavourable balance square, autocorrelative average and standard deviation, be respectively f ₁, f ₂, f ₃, f ₄, f ₅, f ₆, f ₇, f ₈, as final 8 dimension textural characteristics.

C) LBP feature

LBP is used for Description Image Local textural feature, has the significant advantage such as rotational invariance and gray scale unchangeability, for texture feature extraction.The window of definition 3 × 3, with window center pixel for threshold value, the gray-scale value of adjacent 8 pixels is compared with it, if surrounding pixel values is greater than the value of central point, then this location of pixels is labeled 1, otherwise is 0, finally the binary number around central pixel point is turned to decimal number, obtain LBP value, its computing formula is as follows.

${\mathrm{LBP}}_{P,R}=\underset{p=0}{\overset{P-1}{\&Sigma;}}s(g_p-g_c)2^p$

$s\left(x\right)=\left\{\begin{array}{c}1,x\&GreaterEqual;0\\ 0,x<0\end{array}\right.$

Adopt LBP equivalent formulations, namely on the basis of traditional LBP algorithm, select from 0 to 1 or be no more than the eight-digit binary number sequence of 2 times from the saltus step of 1 to 0, be converted into the LBP value as this window after the decimal system, the feature of extraction is the textural characteristics p of the local of image ₁, p ₂..., p ₅₉.Such histogram becomes 59 dimensions from original 256 dimensions, serves the effect of dimensionality reduction.

$U\left({\mathrm{LBP}}_{P,R}\right)=|s(g_{P-1}-g_c)-s(g_0-g_c)|+\underset{p=1}{\overset{P-1}{\&Sigma;}}|s(g_p-g_c)-s(g_{P-1}-g_c)|$

LBP equivalent formulations greatly reduces the kind of binary mode, and does not lose any information, reduces the dimension of proper vector, decreases the impact that high frequency noise produces.

4) biphase gas and liquid flow multi-features, Based PC A technology, three kinds of principal eigenvector in the luv space data of said extracted are merged, obtain 74 dimensional feature vectors, linear transformation is carried out to this vector, i.e. eigencenter, ask for covariance matrix and eigenwert thereof and proper vector, the descending arrangement of eigenwert, due to front 7 eigenwerts and exceeded 95% of all eigenwert sums, choose front 7 eigenwert characteristics of correspondence vector, extract major component and obtain 7 dimensional feature vector t ₁, t ₂..., t ₇, then this vector is normalized, form the data set in new lower dimensional feature space, decrease the redundancy of data, retain the characteristic information of the overwhelming majority in original feature space simultaneously, thus the problem that the intrinsic dimensionality solving extraction is too high.

High dimensional data is projected to comparatively lower dimensional space by principal component analysis (PCA), the data dimension that effectively compression is original, and remains original information largely and be not lost, thus reaches the object of dimensionality reduction.Its key step is as follows:

A) feature extraction is carried out to training sample, the totally 74 dimensional feature vector { m of three kinds of features after extraction ₁, m ₂..., m ₇₄, as this matrix M of PCA former state,

B) average to matrix M, namely every one-dimensional data all deducts this dimension average, obtains matrix B,

$\stackrel{\&OverBar;}{m}=\frac{1}{M}\underset{i=1}{\overset{M}{\&Sigma;}}{m}_i$

C) the covariance matrix C of B is calculated,

$C=\frac{1}{M}\underset{i=1}{\overset{M}{\&Sigma;}}(x_i-\stackrel{\&OverBar;}{m}){(x_i-\stackrel{\&OverBar;}{m})}^T$

D) eigenvalue λ and the proper vector y of covariance matrix C is calculated,

(λE _M-C)y＝0

E) eigenwert is sorted according to order from big to small, selects wherein maximum 7, then 7 of its correspondence proper vectors are formed eigenvectors matrix as column vector, obtain principal component transform matrix D,

F) projected to by sample point in the proper vector chosen, this matrix M of former state and principal component transform matrix D convert, and former state notebook data dimension is reduced to 7 dimensions.

5) biphase gas and liquid flow image recognition and classification, based on support vector cassification identification, using the support vector of the proper vector of training sample as support vector machine, select Radial basis kernel function, training is carried out to whole training sample and obtains supporting vector machine model, utilize the model obtained to carry out testing and predict; Based on the Classification and Identification of BP neural network and probabilistic neural network, using the training sample of proper vector as neural network, be configured to the neural network structure of Flow Regime Ecognition, relevant training parameter is set, and carry out learning training, use the test sample book of this neural network to different flow pattern to identify.

Using the proper vector of 7 dimensional feature vectors after 7 dimension invariant moment features obtained above, 8 dimension gray scale symbiosis moment characteristics and 59 dimension LBP Fusion Features dimensionality reductions as piece image, be training sample, training pattern is constructed by the input and output relation of support vector machine, BP neural network and probabilistic neural network self, wherein support vector machine selects Radial basis kernel function, and the gamma function coefficients in kernel function is set to 1.When setting up BP neural network, 10 hidden neurons, 3 output neurons, hidden layer and output layer transition function are logarithm S shape transfer function and linear transfer function respectively, and training method adopts learning rate changing momentum gradient descent algorithm.When creating probabilistic neural network, select radial basis function, its propagation coefficient is set to 0.6.Respectively test sample book is detected by above-mentioned training pattern.

6) flow pattern image identification and evaluation, according to recognition result, by calculating is compared with the relevant position in classified image and picture of classifying in the position of each actual measurement pixel and classification, is presented at the precision of classification results inside a confusion matrix.In the precision of images is evaluated, confusion matrix is used for match stop result and actual measured value, better visual, better in this, as evaluation criterion, composite character is merged the eigenvector recognition result after dimensionality reduction and single eigenvector recognition result compares, result shows that the recognition methods precision of composite character is the highest, and the time is shorter.So, the method is applied to the detection of two-phase flow, and accuracy rate is higher, reliability is stronger, speed is faster, support on-line automatic identification.

Train as training sample with 90 samples (3 kinds flow pattern each 30), then by the training pattern obtained, 60 samples (3 kinds flow pattern each 20) are predicted as test sample book.To obtain a result and to represent with confusion matrix, each row of confusion matrix represent prediction classification, the sum of each row represents the number being predicted as such other data, every a line represents the true belonging kinds of data, the data count of every a line represents the number of such other data instance, and the numeric representation True Data in each row is predicted to be such number.As shown in table 1,2,3, each numerical value represents discrimination, for correctly to identify number in bracket.

In order to verify the reliability of this method, adopt composite character respectively with other three kinds of independent feature (not bending moments, gray level co-occurrence matrixes, LBP feature) carry out contrast experiment, meanwhile, when feature is identical, we have employed three kinds of sorter (BP neural networks, support vector machines, probabilistic neural network PNN) carry out contrasting with the best sorter of Selection effect.Recognition result as proper vector after three kinds of composite characters fusion dimensionality reductions is best, be best suited for the characteristic of division as two phase flow pattern identification, and the method accuracy of identification combined with BP neural network is the highest, BP neural network is in identifying, good than other two kinds of sorter performances, accuracy of identification is high.Approximately, but dimension significantly reduces the recognition result of composite character and LBP feature, and therefore, the method that composite character and BP neural network combine is better than other several combinations, and the training time is also shorter, is applicable to flow pattern ONLINE RECOGNITION.

The recognition result of table 1BP neural network

The recognition result of table 2 support vector machines

The recognition result of table 3 probabilistic neural network PNN

Claims (10)

1. the Method for Discriminating Gas-liquid Two Phase Flow of Based PC A composite character fusion, it comprises the following steps:

1) flow pattern image of the biphase gas and liquid flow in rectification column under different conditions is obtained by high-speed camera;

2) extraction interesting image regions (ROI region), carries out medium filtering to image and contrast strengthen convection image carries out the pre-service of removal noise;

3) analytical procedure 2) shape of pretreated two-phase flow image and textural characteristics, extract Image Moment Invariants, gray level co-occurrence matrixes and LBP feature;

4) by step 3) after three kinds of Fusion Features, utilize PCA technology, dimensionality reduction is carried out to the composite character after merging, obtain a new proper vector, this new proper vector comprises textural characteristics and geometric properties information;

5) adopt step 4) in new proper vector carry out contrast experiment with three kinds of independent features respectively, utilize support vector machine, BP neural network and probabilistic neural network to carry out training and identifying respectively.

2. the Method for Discriminating Gas-liquid Two Phase Flow that merges of Based PC A composite character according to claim 1, is characterized in that: described step 1) in flow pattern image be the flow pattern image of slug flow under different gas-liquid flow, wave flow, mist annular flow 3 kinds of typical flow patterns.

3. the Method for Discriminating Gas-liquid Two Phase Flow of Based PC A composite character fusion according to claim 2, is characterized in that:

Described step 1) the typical flow pattern image that size is 600 × 600 is collected to slug flow, wave flow, mist annular flow 3 kinds of typical flow patterns.

Described step 2) image is carried out to the process of region of interesting extraction, size is 300 × 60.

4. the Method for Discriminating Gas-liquid Two Phase Flow of Based PC A composite character fusion according to claim 1, is characterized in that:

Described step 2) described medium filtering process, the pixel of image local area is sorted by gray shade scale, gets the gray-scale value of intermediate value as current pixel of gray scale in this neighborhood; Make the actual value that the pixel value of neighborhood is close, thus eliminate isolated noise spot, its expression formula is as follows:

g(x,y)＝med{f(x-k,y-l),(k,l∈W)}

In formula, f (x, y) and g (x, y) is respectively original image and the rear image of process, and W is two dimension pattern plate, selects 3 × 3 regions;

Described step 2) described contrast strengthen pre-service is the histogram equalization method adopted in contrast strengthen: establish original image at (x, y) gray scale at place is g, and the image after changing is h, then can be expressed as the method for image enhaucament and the gray scale g at (x, y) place is mapped as h.In gray-level histogram equalization process, the mapping function of image be may be defined as:

h＝EQ(g)。

5. the Method for Discriminating Gas-liquid Two Phase Flow of Based PC A composite character fusion according to claim 1, it is characterized in that: described step 3) the described invariant moment features main shape facility of image-region, it has the invariant features of the characteristics such as rotation, translation, yardstick, gray level co-occurrence matrixes is that the spatial correlation characteristic by studying gray scale carrys out Description Image texture, the texture information of image local can be measured and extract to LBP feature, has unchangeability to illumination.

6. the Method for Discriminating Gas-liquid Two Phase Flow of Based PC A composite character fusion according to claim 1, it is characterized in that: described step 3) extract the not bending moment of image, gray level co-occurrence matrixes and LBP feature composition flow pattern features vector respectively, utilize the principle of not bending moment, gray level co-occurrence matrixes and LBP feature, extract image 7 respectively and tie up invariant moment features vector, be designated as image 8 ties up gray level co-occurrence matrixes proper vector, is designated as f ₁, f ₂..., f ₈, image 59 ties up LBP proper vector, is designated as p ₁, p ₂..., p ₅₉.

Wherein, described not bending moment, gray level co-occurrence matrixes and LBP feature are respectively:

A) geometric properties of invariant moment features main image-region, because have the invariant features such as rotation, translation, yardstick, there is global property, strong interference immunity, in image procossing, object can be represented as an important feature, thus feature can carry out the operations such as classification to image accordingly;

For the image that intensity profile is h (x, y), its common square in (p+q) rank and centre distance are defined as:

$m_{pq}=\underset{x=1}{\overset{M}{\&Sigma;}}\underset{y=1}{\overset{N}{\&Sigma;}}{x}^p{y}^{q}h(x,y)$

${\mathrm{\&mu;}}_{pq}=\underset{x=1}{\overset{M}{\&Sigma;}}\underset{y=1}{\overset{N}{\&Sigma;}}{(x-x_0)}^p{(y-y_0)}^{q}h(x,y)$

In formula, p, q=0,1,2 ..., the centre of moment (x ₀, y ₀) be

For two dimensional image, x ₀represent gradation of image grey scale centre of gravity in the horizontal direction, y ₀represent gradation of image grey scale centre of gravity in vertical direction;

(p+q) normalization center square is defined as:

$y_{pq}=\frac{{\mathrm{\&mu;}}_{pq}}{{\mathrm{\&mu;}}_{00}^r}$

In formula, $r=\frac{p+q+2}{2},p+q=2,3,....$

Utilize second order and three normalization center, rank squares can derive 7 not bending moment groups as proper vector;

B) gray level co-occurrence matrixes

Gray level co-occurrence matrixes describes texture by the spatial correlation characteristic studying gray scale, can reflect the integrated information of gradation of image about direction, adjacent spaces, amplitude of variation.Characterize the feature of gray level co-occurrence matrixes with some scalars, the feature that gray level co-occurrence matrixes is conventional has to make G represent:

Energy $ASM={\&Sigma;}_{i=1}^k{\&Sigma;}_{j=1}^k{\left(G\left(i,j\right)\right)}^2$

Entropy $ENT=-{\&Sigma;}_{i=1}^k{\&Sigma;}_{j=1}^{k}G\left(i,j\right)\log G\left(i,j\right)$

Auto-correlation $COR={\&Sigma;}_{i=1}^k{\&Sigma;}_{j=1}^k\frac{(i,j)G(i,j)-{\mathrm{\&mu;}}_i{\mathrm{\&mu;}}_j}{s_i{s}_j}$

Unfavourable balance square $IDM={\&Sigma;}_{i=1}^k{\&Sigma;}_{j=1}^k\frac{G(i,j)}{1+{(i-j)}^2}$

0 °, 45 °, 90 °, 135 ° four directions are selected to calculate respectively to the same characteristic parameter of same piece image, the textural characteristics parameter of invariable rotary can be obtained, so just, inhibit durection component on the impact of result, the integrated information of gradation of image about direction, adjacent spaces, amplitude of variation can be reflected; By energy, entropy, unfavourable balance square, autocorrelative average and standard deviation, be respectively f ₁, f ₂, f ₃, f ₄, f ₅, f ₆, f ₇, f ₈, as final 8 dimension textural characteristics;

C) LBP feature

LBP is used for Description Image Local textural feature, has the significant advantage such as rotational invariance and gray scale unchangeability, for texture feature extraction; The window of definition 3 × 3, with window center pixel for threshold value, the gray-scale value of adjacent 8 pixels is compared with it, if surrounding pixel values is greater than the value of central point, then this location of pixels is labeled 1, otherwise is 0, finally the binary number around central pixel point is turned to decimal number, obtain LBP value, its computing formula is as follows:

${\mathrm{LBP}}_{P,R}=\underset{p=0}{\overset{P-1}{\&Sigma;}}s(g_p-g_c)2^p$

$s\left(x\right)=\left\{\begin{array}{c}1,x\&GreaterEqual;0\\ 0,x<0\end{array}\right.$

Adopt LBP equivalent formulations, namely on the basis of traditional LBP algorithm, select from 0 to 1 or be no more than the eight-digit binary number sequence of 2 times from the saltus step of 1 to 0, be converted into the LBP value as this window after the decimal system, the feature of extraction is the textural characteristics p of the local of image ₁, p ₂..., p ₅₉; Such histogram becomes 59 dimensions from original 256 dimensions, serves the effect of dimensionality reduction;

$U\left({\mathrm{LBP}}_{P,R}\right)=|s(g_{P-1}-g_c)-s(g_0-g_c)|+\underset{p=1}{\overset{P-1}{\&Sigma;}}|s(g_p-g_c)-s(g_{P-1}-g_c)|$

LBP equivalent formulations greatly reduces the kind of binary mode, and does not lose any information, reduces the dimension of proper vector, decreases the impact that high frequency noise produces.

7. the Method for Discriminating Gas-liquid Two Phase Flow of Based PC A composite character fusion according to claim 6, it is characterized in that: described step 4) Based PC A technology, three kinds of principal eigenvector in the luv space data of said extracted are merged, obtain 74 dimensional feature vectors, linear transformation is carried out to this vector, i.e. eigencenter, ask for covariance matrix and eigenwert thereof and proper vector, the descending arrangement of eigenwert, due to front 7 eigenwerts and exceeded 95% of all eigenwert sums, choose front 7 eigenwert characteristics of correspondence vector, extract major component and obtain 7 dimensional feature vector t ₁, t ₂..., t ₇, then this vector is normalized, form the data set in new lower dimensional feature space.

8. the Method for Discriminating Gas-liquid Two Phase Flow that merges of Based PC A composite character according to claim 7, is characterized in that: step 4) key step is as follows:

A) feature extraction is carried out to training sample, the totally 74 dimensional feature vector { m of three kinds of features after extraction ₁, m ₂..., m ₇₄, as this matrix M of PCA former state,

B) average to matrix M, namely every one-dimensional data all deducts this dimension average, obtains matrix B,

$\stackrel{\&OverBar;}{m}=\frac{1}{M}\underset{i=1}{\overset{M}{\&Sigma;}}{m}_i$

C) the covariance matrix C of B is calculated,

$C=\frac{1}{M}\underset{i=1}{\overset{M}{\&Sigma;}}(x_i-\stackrel{\&OverBar;}{m}){(x_i-\stackrel{\&OverBar;}{m})}^T$

D) eigenvalue λ and the proper vector y of covariance matrix C is calculated,

(λE _M-C)y＝0

E) eigenwert is sorted according to order from big to small, selects wherein maximum 7, then 7 of its correspondence proper vectors are formed eigenvectors matrix as column vector, obtain principal component transform matrix D,

F) projected to by sample point in the proper vector chosen, this matrix M of former state and principal component transform matrix D convert, and former state notebook data dimension is reduced to 7 dimensions.

9. the Method for Discriminating Gas-liquid Two Phase Flow of Based PC A composite character fusion according to claim 1, it is characterized in that: two-phase flow image recognition and classification, based on support vector cassification identification, using the support vector of the proper vector of training sample as support vector machine, select Radial basis kernel function, training is carried out to whole training sample and obtains supporting vector machine model, utilize the model obtained to carry out testing and predict; Based on the Classification and Identification of BP neural network and probabilistic neural network, all using the training sample of proper vector as neural network, be configured to the neural network structure of Flow Regime Ecognition, relevant training parameter is set, and carry out learning training, use the test sample book of this neural network to different flow pattern to identify.

10. the Method for Discriminating Gas-liquid Two Phase Flow of Based PC A composite character fusion according to claim 1, it is characterized in that: also comprise convection image recognition evaluation: according to recognition result, by calculating is compared with the relevant position in classified image and picture of classifying in the position of each actual measurement pixel and classification, the precision of classification results is presented at inside a confusion matrix, and match stop result and actual measured value are evaluated.