CN105740866A - Rotary kiln sintering state recognition method with artificial feedback regulation mechanism - Google Patents

Abstract

The invention discloses a rotary kiln sintering state recognition method with an artificial feedback regulation mechanism. The method is characterized by comprising the following steps: the first stage of extracting a cognition feature of a flame image to establish a complete recognition feature space of a sample; the second stage of establishing a reduction cognition feature space based on information granularity; the third stage of designing a cognition system through the adoption of an integrated RVFL classifier; the fourth stage of reliably evaluating a cognition result and freely regulating the information granularity according to an evaluation result, re-activating the second to the fourth stage, updating the reduction cognition feature space and feeding back the cognition repeatedly. Through the adoption of the method disclosed by the invention, the multi-level feedback cognition can be performed through the artificial free regulation of the complete cognition feature level, the cognition precision and stability are obviously improved, thereby really realizing the operation of replacing the manual fire watching by machine fire watching; the method lays the foundation for realizing the closed ring control of a rotary kiln clinker quality index.

Description

A kind of rotary kiln with imitative feed-back regulatory mechanism burns till state identification method

Technical field

Originally the invention belongs to rotary kiln flame image cognitive techniques field, be specifically related to a kind of rotary kiln with imitative feed-back regulatory mechanism and burn till state identification method.

Background technology

Rotary kiln is a kind of wide variety of large scale sintering equipment in metallurgy, environmental protection, chemical industry and cement etc..It is difficult to on-line measurement and the closely-related key process parameter state of burning till of clinker quality index is difficult to difficult problems such as accurately identifying owing to rotary kiln sintered process also exists clinker quality index, causes that existing rotary kiln sintered process is still in " manually seeing fire " the operated open-loop stage.But, artificial cognition is burnt till state outcome and is operated the restriction of the subjective factorss such as personnel's experience, responsibility and attention rate, easily causes the problems such as clinker quality index is unstable, kiln running rate is low, production capacity is low, energy consumption is high, hand labor intensity is big.

Clinkering zone flame image contains clinkersintering condition information, is that operator identify that the state of burning till regulable control variable then guarantees the important evidence of clinker quality.The color table of flame image understands combustion zone and thermo parameters method, and form means heat supply, good ventilation condition and satisfied clinker quality in suitable kiln.Therefore, by flame image, rotary kiln being burnt till state and identify continuously, so that it is determined that the running status of rotary kiln, tool is of great significance.At present, this has been done deep theory analysis by the existing scientific paper of studying that rotary kiln flame image burns till state identification method, also there is the engineering method of practical application, such as application for a patent for invention " rotary kiln based on flame image structural similarity burns till state identification method " (CN103578111A) and application for a patent for invention " a kind of rotary kiln flame combustion state ONLINE RECOGNITION method-application " (CN103914709A).

Wherein Chinese invention patent application prospectus CN103578111A is in " rotary kiln based on flame image structural similarity burns till state identification method " disclosed in 12 days February in 2014, a kind of method burning till state recognition for flame image is provided, by flame image to be measured and a width standard flame source image block, then the structural similarity coefficient calculated between corresponding blocks is averaged, and obtains flame image to be measured and the average structure likeness coefficient of a width standard flame source image.The relatively average structure likeness coefficient of flame image to be measured and all standard flame source images, standard flame source image corresponding to maximum burns till status categories and is flame image to be measured and burns till status categories, and this invention has the advantage that computation complexity is low, processing procedure is short.But this invention also exists following deficiency:

This cognitive system belongs to traditional feedback-less open loop recognition methods, adopts the simple feature extracting method under specified particle size level to set up the feature space of cog-nitive target in different samples, and feature space is once set up and no longer updating.Sample variation due to flame image, simple feature is not enough to the complete feature space representing flame image, and different feature relevance grades corresponding to sample differ widely, the feature that sample separating capacity near cluster centre is strong is not often suitable for the similar sample near classifying face, causes that accuracy of identification is not high.Additionally, only set up the cognitive criterion of classification by the mode of average structure likeness coefficient numeric ratio pair, grader robustness is poor.

What Chinese invention patent application prospectus CN103914709A adopted in " a kind of rotary kiln flame combustion state ONLINE RECOGNITION method " disclosed in 9 days July in 2014 is calculate the Y-PSNR between flame image to be measured and each width standard flame source image, chooses the fired state fired state as flame image to be measured of standard flame source image corresponding to Y-PSNR maximum.This invention has the advantage that computation complexity is low, interference factor is few and processing procedure is short.But this invention also exists following deficiency:

Different flame images are extracted simple feature by open loop cognitive system under specified particle size level, and feature space is once set up and no longer updating, it is easy to cause that simple feature insufficient space represents flame image with complete.Additionally, only set up the cognitive criterion of classification by the mode of Y-PSNR numeric ratio pair, grader robustness is not strong.Institute's employing method does not meet when the mankind differentiate and burn till state and adopts multi-stage characteristics to deliberate repeatedly the characteristic of comparison, and recognition correct rate is unsatisfactory.

Summary of the invention

The present invention is the weak point overcoming prior art to exist, propose a kind of rotary kiln with imitative feed-back regulatory mechanism and burn till state identification method, to can freely regulating cognitive characteristics grade and carry out multi-level feedback cognition by manual imitation, to significantly improve cognitive precision and reliability, thus really realizing " fire seen by machine " replacement " manually seeing fire ", the closed loop control for realizing rotary kiln clinker quality index lays the foundation.

For solving above-mentioned technical problem, the technical solution adopted in the present invention is:

It is being carry out as follows that a kind of rotary kiln with imitative feed-back regulatory mechanism of the present invention burns till the feature of state identification method:

Step 1, the rotary kiln under running status is shot, obtain rotary kiln and burn till state video, burn till from described rotary kiln and state video is selected v width normal condition flame image, be designated as NI, v width low temperature state flame image, it is designated as LI and v panel height temperature state flame image, is designated as HI；Formed 3v width training flame image set TI by described v width normal condition flame image NI, v width low temperature state flame image LI and v panel height temperature state flame image HI, and have TI={I₁,I₂,…,I_τ,…,I_3v}；I_τRepresent τ width training flame image；1≤τ≤3v；

Step 2, cognitive characteristics extract

Described τ width is trained flame image I by step 2.1, employing multivariate image analysis method_τProcess, it is thus achieved that τ width training flame image I_τD₁Dimension color coefficients f_1,τ；

Described τ width is trained flame image I by step 2.2, employing Principal Component Analysis Method_τProcess, it is thus achieved that τ width training flame image I_τD₂Dimension overall situation shape factor f_2,τ；

Described τ width is originally trained flame image I in conjunction with latent semantic analysis method by step 2.3, employing vision word_τProcess, it is thus achieved that τ width training flame image I_τD₃Dimension overall situation shape factor f_3,τ；

Step 2.4, to described τ width train flame image I_τD₁Dimension color coefficients f_1,τ、d₂Dimension overall situation shape factor f_2,τAnd d₃Dimension overall situation shape factor f_3,τMerge, it is thus achieved that d=d₁+d₂+d₃Maintain several C '_τ, and to described coefficient C '_τIt is normalized, it is thus achieved that described τ width training flame image I_τTraining cognitive characteristics vector C_τ；Thus obtaining the training cognitive characteristics vector set of 3v width training flame image, it is designated as C={C₁,C₂,…,C_τ,…,C_3v}；

Step 3, cognitive system design

Step 3.1, defined variable w；And initialize w=1, initialization information granularity δ_w=1；

Step 3.2, in the w time feedback identifying, based on Information Granularity δ_w, adopt rough set theory that described training cognitive characteristics vector set C is carried out yojan process, it is thus achieved that the training cognitive characteristics vector set B of the w time yojan_w；

Step 3.3, the w time yojan training cognitive characteristics vector set B_wAs the input of integrated RVFL grader and be trained, obtain the integrated RVFL grader of optimum and the cognitive result of training of the w time cognitive process；

Step 4, burn till from the rotary kiln of online real time collecting and state video obtains flame image to be identified as test image P；

Step 5, cognitive result assessment

Step 5.1, according to step 2, described flame image P to be identified is carried out cognitive characteristics extraction, it is thus achieved that the d dimension test cognitive characteristics vector FB of described flame image P to be identified；

Step 5.2, in the w time feedback identifying, based on Information Granularity δ_w, adopt rough set theory that described d dimension test cognitive characteristics vector FB is carried out yojan process, it is thus achieved that the test cognitive characteristics vector FB of the w time yojan_w；

Step 5.3, utilize the optimum integrated RVFL grader of described the w time cognitive process described flame image P to be identified is carried out classification cognition, obtain the test cognition result DL of the w time cognitive process_w；

Step 5.4, test cognition result DL according to described the w time cognitive process_w, set up the w time control information system SR_w, and the test cognition result of the w time cognitive process is carried out reliability assessment, if meeting reliability thresholds, then by the test cognition result DL of the w time cognitive process_wFinal cognitive result as described flame image P to be identified；Otherwise, w+1 is assigned to w；And return step 3.2 and perform；

Step 6, according to described final cognitive result judges described in treat that the state of burning till of cognition flame image P is whether as normal condition flame image, if so, then maintains and currently burns till state；And return step 4, obtain another flame image to be identified as test image；Otherwise, after the currently state of burning till is carried out alarm, return again to step 4 and perform.

The feature that the rotary kiln with imitative feed-back regulatory mechanism of the present invention burns till state identification method lies also in,

The w time control information system SR in described step 5.4_wIt is utilize formula (1) to set up:

${\mathrm{SR}}_w=(I_l,\stackrel{\‾}{M}(w,l\left)\right)---\left(1\right)$

In formula (1), I_lRepresent l class training flame image number；{ 1,2,3} represents the 1st class normal condition flame image, the 2nd class low temperature state flame image and the 3rd type high temp state flame image to l ∈ respectively；And have:

$\stackrel{\‾}{M}(w,l)={\∪}_{q=1}^{I_l}|{\mathrm{FB}}_w-B_{q,w}|---\left(2\right)$

In formula (2),Represent the w time test cognitive characteristics vector FB_wWith the poor matrix of feature that l class trains flame image, B_q,wRepresent the training cognitive characteristics vector of the w time yojan of q-th training flame image in l class training flame image；1≤q≤I_l。

Reliability assessment is carry out as follows by the test cognition result of described the w time cognitive process:

Step a, formula (3) is utilized to obtain the w time cognitive result DL of test_wError alpha entropyThus obtaining all previous α Entropy sequence $H_w=\{H_1^{\mathrm{\α}},H_2^{\mathrm{\α}},...,H_w^{\mathrm{\α}}\}:$

$H_w^{\mathrm{\α}}=\frac{{\mathrm{\Σ}}_{s=1}^h\frac{|E_s{|}^{\mathrm{\α}}}{|I_l{|}^{\mathrm{\α}}}-1}{1-2^{(\mathrm{\α}-1)}}---\left(3\right)$

In formula (3), E_sRepresent, based on equivalence relation, l class is trained flame image I_lAbout feature difference matrixThe equivalence class divided after carrying out granulation dividing processing, all equivalence classes composition quotient set { E₁,E₂,…,E_h, h represents the number of equivalence class in quotient set；1≤s≤h；

Step b, formula (4) is utilized to obtain the w time cognitive result DL of test_wγ tie up error alpha Entropy sequence

$H_w^{\mathrm{\γ}}=\{(H_{w-\mathrm{\γ}+1}^{\mathrm{\α}}-{\stackrel{\‾}{H}}_w^{\mathrm{\α}}),...,(H_w^{\mathrm{\α}}-{\stackrel{\‾}{H}}_w^{\mathrm{\α}})\}---\left(4\right)$

In formula (4),Represent the cognitive result DL of the w time test_wγ tie up error alpha entropy vector average；And have:

$H_w^{\mathrm{\γ}}=\frac{1}{\mathrm{\γ}}{\mathrm{\Σ}}_{k=0}^{\mathrm{\γ}-1}{H}_{w-k}^{\mathrm{\α}}---\left(5\right)$

In formula (5), 1≤k≤γ < w；

Step c, obtained the w time cognitive result DL of test by formula (6)_wα Entropy sequence similarity SD_w:

${\mathrm{SD}}_w=\exp \&lsqb;-{\left(\frac{{\max }_{k\&Element;(0,\mathrm{\&gamma;}-1)}\left\{\right|(H_{w-k+1}^{\mathrm{\&alpha;}}-{\stackrel{\&OverBar;}{H}}_w^{\mathrm{\&alpha;}}),...,(H_{w-k}^{\mathrm{\&alpha;}}-{\stackrel{\&OverBar;}{H}}_w^{\mathrm{\&alpha;}})\left|\right\}}{e}\right)}^g\&rsqb;---\left(6\right)$

In formula (6), g and e represents gradient and the width of exponential function respectively；

Step d, judging whether w≤γ sets up, if setting up, then calculating described the w time test cognition result DL_wInformation Granularity error E G_w, thus calculating the Information Granularity δ of the w+1 time cognitive process_w+1；Return step 3.2；Otherwise, step b is performed；

Step e, judge whether γ < w≤σ sets up, if setting up, then by the test cognition result DL of described the w time cognitive process_wα Entropy sequence similarity SD_wWith set reliability thresholds T₂Compare, if SD_wBe more than or equal to reliability thresholds T₂, then α Entropy sequence similarity SD_wRepresent and meet reliability thresholds T₂；Otherwise, step f is performed；

If whether step f w > σ sets up, if setting up, then by the test cognition result DL of the w-1 time cognitive process_w-1Final cognitive result as described flame image P to be identified；Otherwise, the described cognitive result DL of the w time test is calculated_wInformation Granularity error E G_w, thus calculating the Information Granularity δ of the w+1 time cognitive process_w+1；Return step 3.2.

The described cognitive result DL of the w time test_wInformation Granularity error E G_wIt is obtained by formula (7):

${\mathrm{EG}}_w=\left|\frac{\stackrel{\&OverBar;}{S}{\left(w\right)}^{\&prime;\&prime;}}{{(1+{\left(\stackrel{\&OverBar;}{S}{\left(w\right)}^{\&prime;}\right)}^2)}^{\frac{3}{2}}}\right|-\left|\frac{\stackrel{\&OverBar;}{S}{(w-1)}^{\&prime;\&prime;}}{{(1+{\left(\stackrel{\&OverBar;}{S}{(w-1)}^{\&prime;}\right)}^2)}^{\frac{3}{2}}}\right|---\left(7\right)$

In formula (7),Represent all previous α Entropy sequence H_wCubic spline function；Represent cubic spline functionFirst derivative；Represent cubic spline functionSecond dervative；And EG₁≤0；EG₂≤0。

The Information Granularity δ of described the w+1 time cognitive process_w+1It is obtain as follows:

Step a, judge EG_wWhether≤0 set up, if setting up, then utilizes formula (8) to obtain the Information Granularity δ of the w+1 time cognitive process_w+1, otherwise perform step b:

${\mathrm{\&delta;}}_{w+1}={\mathrm{\&delta;}}_w(1-{\left(2^{w+1}-1\right)}^{-1}{H}_w^{\mathrm{\&alpha;}})---\left(8\right)$

Step b, formula (9) is utilized to obtain the Information Granularity δ of the w+1 time cognitive process_w+1:

${\mathrm{\&delta;}}_{w+1}={\mathrm{\&delta;}}_w(1-\frac{{(2^{w+1}-1)}^{-1}{H}_w^{\mathrm{\&alpha;}}}{2})---\left(9\right).$

Compared with the prior art, beneficial effects of the present invention is embodied in:

1, in order to manual imitation unrestricted choice complete cognitive characteristics of many grades when state is burnt till in different flame images differentiations deliberates repeatedly the characteristics of cognition of comparison, the present invention establishes complete flame image cognitive characteristics space by adopting multiple cognitive characteristics extracting method, integrated RVFL grader is utilized to obtain the cognitive result of robust, by the test cognition result of cognitive process is carried out reliability assessment, calculate the Information Granularity of cognitive process next time, cognition is fed back in the complete cognitive characteristics space freely regulating test sample at many levels, thus significantly improving cognitive precision, solve that to there is cognitive characteristics space in existing open loop feedback-less cognitive system incomplete, cognitive characteristics space immobilizes, the problem that grader robustness is not strong and discrimination is not high.

2, the present invention is by adopting multivariate image analysis method, Principal Component Analysis Method, vision word originally in conjunction with latent semantic analysis method, extract the color coefficients of flame image, overall situation shape factor and local shape factor respectively, and merged as cognitive characteristics vector, establish complete cognitive characteristics space, overcome the defect in simple incomplete cognitive characteristics space in traditional cognitive system, thus significantly improving cognitive precision.

3, the present invention is by setting up the control information system of the cognitive result of test, the cognitive characteristics diversity of tolerance test sample and the generic training sample of present cognitive result, reliability assessment for cognitive result provides foundation, thus overcoming conventional open-loop cognitive system directly export the blindness of cognitive result, enhance the reliability of cognitive result.

4, by reliability assessment step, the present invention differentiates whether the present cognitive result of test sample can as final cognitive result output, thus realizing freely regulating test sample cognitive characteristics space, then feed back cognition at many levels, significantly improve the reliability of cognitive result.

5, the present invention is by testing all previous α Entropy sequence of sample, the Information Granularity error of test sample cognition result is calculated according to its variation tendency, thus differentiating the local convergence of cognitive system, overcome the mode that conventional information granularity gradually changes, quickly update for the self adaptation of Information Granularity and provide foundation.

6, the present invention local convergence by cognitive system, adaptively selected Information Granularity renewal function, realize freely regulating of test sample cognitive characteristics space thereby through the Information Granularity updated, then feed back cognition at many levels, significantly improve cognitive precision.

Accompanying drawing explanation

Fig. 1 is the system general flow chart of the present invention.

Detailed description of the invention

Below in conjunction with accompanying drawing, technical scheme is carried out clear, complete description.Obviously described embodiment is only a part for the embodiment of the present invention, based on embodiments of the invention, and the other embodiments that those skilled in the art obtains under the premise not making creative work, broadly fall into the protection domain of this patent.

The embodiment provides a kind of rotary kiln with imitative feed-back regulatory mechanism and burn till state identification method, to solve prior art and also exist that cognitive characteristics space is incomplete, cognitive characteristics space immobilizes, grader robustness is strong and discrimination is not high problem.Specifically, step is as follows:

Step 1, the rotary kiln under running status is shot, obtain rotary kiln and burn till state video, burn till from rotary kiln and state video is selected v width normal condition flame image, be designated as NI, v width low temperature state flame image, it is designated as LI and v panel height temperature state flame image, is designated as HI；Formed 3v width training flame image set TI by v width normal condition flame image NI, v width low temperature state flame image LI and v panel height temperature state flame image HI, and have TI={I₁,I₂,…,I_τ,…,I_3v}；I_τRepresent τ width training flame image；1≤τ≤3v；In the present embodiment, v=50,3v=150；

Step 2, cognitive characteristics extract

Step 2.1 is as it is shown in figure 1, adopt multivariate image analysis method that τ width is trained flame image I_τProcess, it is thus achieved that τ width training flame image I_τD₁Dimension color coefficients f_1,τ；In the present embodiment, d₁=1；The detailed content of multivariate image analysis method can referring to document " Monitoringflamesinanindustrialboilerusingmultivariateima geanalysis "；

Step 2.2 is as it is shown in figure 1, adopt Principal Component Analysis Method that τ width is trained flame image I_τProcess, it is thus achieved that τ width training flame image I_τD₂Dimension overall situation shape factor f_2,τ；In the present embodiment, d₂=6；The detailed content of Principal Component Analysis Method can referring to document " TwodimensionalPCA:anewapproachtoappearance-basedfacerepr esentationandrecognition "；

Step 2.3 is as it is shown in figure 1, adopt vision word originally in conjunction with latent semantic analysis method, τ width to be trained flame image I_τProcess, it is thus achieved that τ width training flame image I_τD₃Dimension overall situation shape factor f_3,τ；In the present embodiment, d₃=5；Vision word originally can referring to document " FlameImage-BasedBurningStateRecognitionforSinteringProce ssofRotaryKilnUsingHeterogeneousFeaturesandFuzzyIntegral " in conjunction with the detailed content of latent semantic analysis method；

Step 2.4, to τ width train flame image I_τD₁Dimension color coefficients f_1,τ、d₂Dimension overall situation shape factor f_2,τAnd d₃Dimension overall situation shape factor f_3,τMerge, it is thus achieved that d=d₁+d₂+d₃Maintain several C '_τ, and to coefficient C '_τIt is normalized, it is thus achieved that τ width training flame image I_τTraining cognitive characteristics vector C_τ；Thus obtaining the training cognitive characteristics vector set of 3v width training flame image, it is designated as C={C₁,C₂,…,C_τ,…,C_3v}；In the present embodiment, d=12；

Step 3, cognitive system design

Step 3.1, defined variable w；And initialize w=1, initialization information granularity δ_w=1；

Step 3.2, in the w time feedback identifying, based on Information Granularity δ_w, adopt rough set theory that training cognitive characteristics vector set C is carried out yojan process, it is thus achieved that the training cognitive characteristics vector set B of the w time yojan_w；The detailed content that Information Granularity calculates can referring to document " Granularcomputing:Basicissuesandpossiblesolutions "；The detailed computational methods of rough set theory yojan input cognitive characteristics vector can referring to document " Roughsets "；

Step 3.3, the w time yojan training cognitive characteristics vector set B_wAs the input of integrated RVFL grader and be trained, obtain the integrated RVFL grader of optimum and the cognitive result of training of the w time cognitive process；Optimum integrated RVFL grader: the integrated size n of base RVFL network_b=5, the basic function number n in a base RVFL network_h=35；

Step 4, burn till from the rotary kiln of online real time collecting and state video obtains flame image to be identified as test image P；

Step 5, cognitive result assessment

Step 5.1, according to step 2, flame image P to be identified is carried out cognitive characteristics extraction, it is thus achieved that the d dimension test cognitive characteristics vector FB of flame image P to be identified；In the present embodiment, extract the cognitive characteristics FB=[0.16 of flame image P to be identified based on step 2；-21.48；-0.02；-0.12；-0.01；0.02；0.02；1.14；1.76；-0.17；2.9e-16；-1.7e-16]；

Step 5.2, in the w time feedback identifying, based on Information Granularity δ_w, adopt rough set theory that d dimension test cognitive characteristics vector FB is carried out yojan process, it is thus achieved that the test cognitive characteristics vector FB of the w time yojan_w；As it is shown in figure 1, extract the cognitive characteristics vector FB of flame image P to be identified based on step 2, then carry out yojan, and by the test cognitive characteristics vector FB of yojan_wSend into optimum integrated RVFL grader and carry out classification cognition；

Step 5.3, utilize the optimum integrated RVFL grader of the w time cognitive process flame image P to be identified is carried out classification cognition, obtain the test cognition result DL of the w time cognitive process_w；

Step 5.4, test cognition result DL according to the w time cognitive process_w, set up the w time control information system SR_w, and the test cognition result of the w time cognitive process is carried out reliability assessment, if meeting reliability thresholds, then by the test cognition result DL of the w time cognitive process_wFinal cognitive result as flame image P to be identified；Otherwise, w+1 is assigned to w；And return step 3.2 and perform；

Specifically, the w time control information system SR_wIt is utilize formula (1) to set up:

${\mathrm{SR}}_w=(I_l,\stackrel{\&OverBar;}{M}(w,l\left)\right)---\left(1\right)$

In formula (1), I_lRepresent l class training flame image number；L ∈ 1,2,3} represents the 1st class normal condition flame image, the 2nd class low temperature state flame image and the 3rd type high temp state flame image respectively, and in the present embodiment, I_l=50；And have:

$\stackrel{\&OverBar;}{M}(w,l)={\&cup;}_{q=1}^{I_l}|{\mathrm{FB}}_w-B_{q,w}|---\left(2\right)$

In formula (2),Represent the w time test cognitive characteristics vector FB_wWith the poor matrix of feature that l class trains flame image, B_q,wRepresent the training cognitive characteristics vector of the w time yojan of q-th training flame image in l class training flame image；1≤q≤I_l。

In the present embodiment, reliability assessment is carry out as follows by the test cognition result of the w time cognitive process:

Step a, formula (3) is utilized to obtain the w time cognitive result DL of test_wError alpha entropyThus obtaining all previous α Entropy sequence $H_w=\{H_1^{\mathrm{\&alpha;}},H_2^{\mathrm{\&alpha;}},...,H_w^{\mathrm{\&alpha;}}\},$ In the present embodiment, α=2:

$H_w^{\mathrm{\&alpha;}}=\frac{{\mathrm{\&Sigma;}}_{s=1}^h\frac{|E_s{|}^{\mathrm{\&alpha;}}}{|I_l{|}^{\mathrm{\&alpha;}}}-1}{1-2^{(\mathrm{\&alpha;}-1)}}---\left(3\right)$

In formula (3), E_sRepresent, based on equivalence relation, l class is trained flame image I_lAbout feature difference matrixThe equivalence class divided after carrying out granulation dividing processing, all equivalence classes composition quotient set { E₁,E₂,…,E_h, h represents the number of equivalence class in quotient set；1≤s≤h；

Step b, formula (4) is utilized to obtain the w time cognitive result DL of test_wγ tie up error alpha Entropy sequenceIn the present embodiment, γ=2:

$H_w^{\mathrm{\&gamma;}}=\{(H_{w-\mathrm{\&gamma;}+1}^{\mathrm{\&alpha;}}-{\stackrel{\&OverBar;}{H}}_w^{\mathrm{\&alpha;}}),...,(H_w^{\mathrm{\&alpha;}}-{\stackrel{\&OverBar;}{H}}_w^{\mathrm{\&alpha;}})\}---\left(4\right)$

In formula (4),Represent the cognitive result DL of the w time test_wγ tie up error alpha entropy vector average；And have:

$H_w^{\mathrm{\&gamma;}}=\frac{1}{\mathrm{\&gamma;}}{\mathrm{\&Sigma;}}_{k=0}^{\mathrm{\&gamma;}-1}{H}_{w-k}^{\mathrm{\&alpha;}}---\left(5\right)$

In formula (5), 1≤k≤γ < w；

Step c, obtained the w time cognitive result DL of test by formula (6)_wα Entropy sequence similarity SD_w:

${\mathrm{SD}}_w=\exp \&lsqb;-{\left(\frac{{\max }_{k\&Element;(0,\mathrm{\&gamma;}-1)}\left\{\right|(H_{w-k+1}^{\mathrm{\&alpha;}}-{\stackrel{\&OverBar;}{H}}_w^{\mathrm{\&alpha;}}),...,(H_{w-k}^{\mathrm{\&alpha;}}-{\stackrel{\&OverBar;}{H}}_w^{\mathrm{\&alpha;}})\left|\right\}}{e}\right)}^g\&rsqb;---\left(6\right)$

In formula (6), g and e represents gradient and the width of exponential function respectively, in the present embodiment, and g=2, e=0.25 × standard deviation { H_w}；

Step d, judging whether w≤γ sets up, if setting up, then calculating described the w time test cognition result DL_wInformation Granularity error E G_w, thus calculating the Information Granularity δ of the w+1 time cognitive process_w+1；Return step 3.2；Otherwise, step b is performed；

Step e, judge whether γ < w≤σ sets up, if setting up, then by the test cognition result DL of described the w time cognitive process_wα Entropy sequence similarity SD_wWith set reliability thresholds T₂Compare, if SD_wBe more than or equal to reliability thresholds T₂, then α Entropy sequence similarity SD_wRepresent and meet reliability thresholds T₂；Otherwise, step f is performed；In the present embodiment, σ=30；

If whether step f w > σ sets up, if setting up, then by the test cognition result DL of the w-1 time cognitive process_w-1Final cognitive result as described flame image P to be identified；Otherwise, the described cognitive result DL of the w time test is calculated_wInformation Granularity error E G_w, thus calculating the Information Granularity δ of the w+1 time cognitive process_w+1；Return step 3.2.

In the present embodiment, the cognitive result DL of the w time test_wInformation Granularity error E G_wIt is obtained by formula (7):

${\mathrm{EG}}_w=\left|\frac{\stackrel{\&OverBar;}{S}{\left(w\right)}^{\&prime;\&prime;}}{{(1+{\left(\stackrel{\&OverBar;}{S}{\left(w\right)}^{\&prime;}\right)}^2)}^{\frac{3}{2}}}\right|-\left|\frac{\stackrel{\&OverBar;}{S}{(w-1)}^{\&prime;\&prime;}}{{(1+{\left(\stackrel{\&OverBar;}{S}{(w-1)}^{\&prime;}\right)}^2)}^{\frac{3}{2}}}\right|---\left(7\right)$

In formula (7),Represent all previous α Entropy sequence H_wCubic spline function；Represent cubic spline functionFirst derivative；Represent cubic spline functionSecond dervative；And EG₁≤0；EG₂≤0。

In the present embodiment, the Information Granularity δ of the w+1 time cognitive process_w+1It is obtain as follows:

Step a, judge EG_wWhether≤0 set up, if setting up, then utilizes formula (8) to obtain the Information Granularity δ of the w+1 time cognitive process_w+1, otherwise perform step b:

${\mathrm{\&delta;}}_{w+1}={\mathrm{\&delta;}}_w(1-{\left(2^{w+1}-1\right)}^{-1}{H}_w^{\mathrm{\&alpha;}})---\left(8\right)$

Step b, formula (9) is utilized to obtain the Information Granularity δ of the w+1 time cognitive process_w+1:

${\mathrm{\&delta;}}_{w+1}={\mathrm{\&delta;}}_w(1-\frac{{(2^{w+1}-1)}^{-1}{H}_w^{\mathrm{\&alpha;}}}{2})---\left(9\right)$

As it is shown in figure 1, the cognitive result DL treating cognitive the 1st cognitive process of flame image P that optimum integrated RVFL grader is provided₁={ high temperature }, according to 1=w≤γ=2, it is necessary to feed back cognition next time, formula (8) calculate the Information Granularity δ of the 2nd cognitive process₂=0.87, update cognitive characteristics space, send into optimum integrated RVFL grader and carry out classification cognition；Cognitive result DL to the 2nd cognitive process₂={ normal }, according to 2=w≤γ=2, it is necessary to feed back cognition next time, is calculated the Information Granularity δ of the 3rd cognitive process by formula (8)₃=0.79, update cognitive characteristics space, send into optimum integrated RVFL grader and carry out classification cognition；To the 3rd cognitive process cognition result DL₃={ normal }, according to 2=γ < w=3≤σ=30, is calculated the 3rd cognitive process cognition result DL by formula (6)₃α Entropy sequence similarity SD₃, and with set reliability thresholds T₂Compare: SD₃=0.45 < T₂=0.8, illustrate that current cognitive characteristics insufficient space treats cognitive flame image P to distinguish this, it is necessary to feed back cognition next time, formula (8) calculate the Information Granularity δ of the 4th cognitive process₄=0.75, update cognitive characteristics space, send into optimum integrated RVFL grader and carry out classification cognition；

Treat the cognitive result DL of the 4th cognitive process of cognitive flame image₄={ normal }, according to 2=γ < w=4≤σ=30, is calculated its α Entropy sequence similarity SD by formula (6)₄, and with set reliability thresholds T₂Compare: SD₄=0.88 >=T₂, meet reliability thresholds, stop feedback cognitive process, and by DL₄={ normal } is as the final cognitive result of this flame image.

Step 6, judges to treat that the state of burning till of cognition flame image P is whether as normal condition flame image according to final cognitive result, if so, then maintain and currently burn till state；And return step 4, obtain another flame image to be identified as test image；Otherwise, after the currently state of burning till is carried out alarm, return again to step 4 and perform.

Above-mentioned imitative feedback cognitive process is a specific implementation process of the present invention, energy Automatic adjusument Information Granularity when in the face of different sample, realize treating the cognitive complete cognitive characteristics space of sample to treat to distinguish from the overall situation to the foundation of the self adaptation of local from macroscopic view to microcosmic, adopt the robustness of integrated study enhancement mode grader, simultaneously according to the reliability assessment of cognitive result, sample repeatedly imitated feedback cognition, manual imitation unrestricted choice complete cognitive characteristics of many grades when state is burnt till in different flame images differentiations deliberates repeatedly the characteristics of cognition of comparison, thus significantly improving cognitive precision and reliability.

Claims (5)

1. the rotary kiln with imitative feed-back regulatory mechanism burns till a state identification method, and its feature is being to carry out as follows:

Step 1, the rotary kiln under running status is shot, obtain rotary kiln and burn till state video, burn till from described rotary kiln and state video is selected v width normal condition flame image, be designated as NI, v width low temperature state flame image, it is designated as LI and v panel height temperature state flame image, is designated as HI；Formed 3v width training flame image set TI by described v width normal condition flame image NI, v width low temperature state flame image LI and v panel height temperature state flame image HI, and have TI={I₁,I₂,…,I_τ,…,I_3v}；I_τRepresent τ width training flame image；1≤τ≤3v；

Step 2, cognitive characteristics extract

Described τ width is trained flame image I by step 2.1, employing multivariate image analysis method_τProcess, it is thus achieved that τ width training flame image I_τD₁Dimension color coefficients f_1,τ；

Described τ width is trained flame image I by step 2.2, employing Principal Component Analysis Method_τProcess, it is thus achieved that τ width training flame image I_τD₂Dimension overall situation shape factor f_2,τ；

Described τ width is originally trained flame image I in conjunction with latent semantic analysis method by step 2.3, employing vision word_τProcess, it is thus achieved that τ width training flame image I_τD₃Dimension overall situation shape factor f_3,τ；

Step 2.4, to described τ width train flame image I_τD₁Dimension color coefficients f_1,τ、d₂Dimension overall situation shape factor f_2,τAnd d₃Dimension overall situation shape factor f_3,τMerge, it is thus achieved that d=d₁+d₂+d₃Maintain several C '_τ, and to described coefficient C '_τIt is normalized, it is thus achieved that described τ width training flame image I_τTraining cognitive characteristics vector C_τ；Thus obtaining the training cognitive characteristics vector set of 3v width training flame image, it is designated as C={C₁,C₂,…,C_τ,…,C_3v}；

Step 3, cognitive system design

Step 3.1, defined variable w；And initialize w=1, initialization information granularity δ_w=1；

Step 3.2, in the w time feedback identifying, based on Information Granularity δ_w, adopt rough set theory that described training cognitive characteristics vector set C is carried out yojan process, it is thus achieved that the training cognitive characteristics vector set B of the w time yojan_w；

Step 3.3, the w time yojan training cognitive characteristics vector set B_wAs the input of integrated RVFL grader and be trained, obtain the integrated RVFL grader of optimum and the cognitive result of training of the w time cognitive process；

Step 4, burn till from the rotary kiln of online real time collecting and state video obtains flame image to be identified as test image P；

Step 5, cognitive result assessment

Step 5.1, according to step 2, described flame image P to be identified is carried out cognitive characteristics extraction, it is thus achieved that the d dimension test cognitive characteristics vector FB of described flame image P to be identified；

Step 5.2, in the w time feedback identifying, based on Information Granularity δ_w, adopt rough set theory that described d dimension test cognitive characteristics vector FB is carried out yojan process, it is thus achieved that the test cognitive characteristics vector FB of the w time yojan_w；

Step 5.3, utilize the optimum integrated RVFL grader of described the w time cognitive process described flame image P to be identified is carried out classification cognition, obtain the test cognition result DL of the w time cognitive process_w；

Step 5.4, test cognition result DL according to described the w time cognitive process_w, set up the w time control information system SR_w, and the test cognition result of the w time cognitive process is carried out reliability assessment, if meeting reliability thresholds, then by the test cognition result DL of the w time cognitive process_wFinal cognitive result as described flame image P to be identified；Otherwise, w+1 is assigned to w；And return step 3.2 and perform；

Step 6, according to described final cognitive result judges described in treat that the state of burning till of cognition flame image P is whether as normal condition flame image, if so, then maintains and currently burns till state；And return step 4, obtain another flame image to be identified as test image；Otherwise, after the currently state of burning till is carried out alarm, return again to step 4 and perform.

2. the rotary kiln with imitative feed-back regulatory mechanism according to claim 1 burns till state identification method, it is characterized in that the w time control information system SR in described step 5.4_wIt is utilize formula (1) to set up:

${\mathrm{SR}}_w=(I_l,\stackrel{\&OverBar;}{M}(w,l\left)\right)---\left(1\right)$

In formula (1), I_lRepresent l class training flame image number；{ 1,2,3} represents the 1st class normal condition flame image, the 2nd class low temperature state flame image and the 3rd type high temp state flame image to l ∈ respectively；And have:

$\stackrel{\&OverBar;}{M}(w,l)={\&cup;}_{q=1}^{I_l}|{\mathrm{FB}}_w-B_{q,w}|---\left(2\right)$

In formula (2),Represent the w time test cognitive characteristics vector FB_wWith the poor matrix of feature that l class trains flame image, B_q,wRepresent the training cognitive characteristics vector of the w time yojan of q-th training flame image in l class training flame image；1≤q≤I_l。

3. the rotary kiln with imitative feed-back regulatory mechanism described in claim 2 burns till state identification method, it is characterized in that reliability assessment is carry out as follows by the test cognition result of described the w time cognitive process:

Step a, formula (3) is utilized to obtain the w time cognitive result DL of test_wError alpha entropyThus obtaining all previous α Entropy sequence $H_w=\{H_1^{\mathrm{\&alpha;}},H_2^{\mathrm{\&alpha;}},...,H_w^{\mathrm{\&alpha;}}\}:$

$H_w^{\mathrm{\&alpha;}}=\frac{{\mathrm{\&Sigma;}}_{s=1}^h\frac{|E_s{|}^{\mathrm{\&alpha;}}}{|I_l{|}^{\mathrm{\&alpha;}}}-1}{1-2^{(\mathrm{\&alpha;}-1)}}---\left(3\right)$

In formula (3), E_sRepresent, based on equivalence relation, l class is trained flame image I_lAbout feature difference matrixThe equivalence class divided after carrying out granulation dividing processing, all equivalence classes composition quotient set { E₁,E₂,…,E_h, h represents the number of equivalence class in quotient set；1≤s≤h；

Step b, formula (4) is utilized to obtain the w time cognitive result DL of test_wγ tie up error alpha Entropy sequence

$H_w^{\mathrm{\&gamma;}}=\{(H_{w-\mathrm{\&gamma;}+1}^{\mathrm{\&alpha;}}-{\stackrel{\&OverBar;}{H}}_w^{\mathrm{\&alpha;}}),...,(H_w^{\mathrm{\&alpha;}}-{\stackrel{\&OverBar;}{H}}_w^{\mathrm{\&alpha;}})\}---\left(4\right)$

In formula (4),Represent the cognitive result DL of the w time test_wγ tie up error alpha entropy vector average；And have:

$H_w^{\mathrm{\&gamma;}}=\frac{1}{\mathrm{\&gamma;}}{\mathrm{\&Sigma;}}_{k=0}^{\mathrm{\&gamma;}-1}{H}_{w-k}^{\mathrm{\&alpha;}}---\left(5\right)$

In formula (5), 1≤k≤γ < w；

Step c, obtained the w time cognitive result DL of test by formula (6)_wα Entropy sequence similarity SD_w:

${\mathrm{SD}}_w=\exp \&lsqb;-{\left(\frac{{\max }_{k\&Element;(0,\mathrm{\&gamma;}-1)}\left\{\right|(H_{w-k+1}^{\mathrm{\&alpha;}}-{\stackrel{\&OverBar;}{H}}_w^{\mathrm{\&alpha;}}),...,(H_{w-k}^{\mathrm{\&alpha;}}-{\stackrel{\&OverBar;}{H}}_w^{\mathrm{\&alpha;}})\left|\right\}}{e}\right)}^g\&rsqb;---\left(6\right)$

In formula (6), g and e represents gradient and the width of exponential function respectively；

Step d, judging whether w≤γ sets up, if setting up, then calculating described the w time test cognition result DL_wInformation Granularity error E G_w, thus calculating the Information Granularity δ of the w+1 time cognitive process_w+1；Return step 3.2；Otherwise, step b is performed；

Step e, judge whether γ < w≤σ sets up, if setting up, then by the test cognition result DL of described the w time cognitive process_wα Entropy sequence similarity SD_wWith set reliability thresholds T₂Compare, if SD_wBe more than or equal to reliability thresholds T₂, then α Entropy sequence similarity SD_wRepresent and meet reliability thresholds T₂；Otherwise, step f is performed；

If whether step f w > σ sets up, if setting up, then by the test cognition result DL of the w-1 time cognitive process_w-1Final cognitive result as described flame image P to be identified；Otherwise, the described cognitive result DL of the w time test is calculated_wInformation Granularity error E G_w, thus calculating the Information Granularity δ of the w+1 time cognitive process_w+1；Return step 3.2.

4. the rotary kiln with imitative feed-back regulatory mechanism according to claim 3 burns till state identification method, it is characterized in that the described cognitive result DL of the w time test_wInformation Granularity error E G_wIt is obtained by formula (7):

${\mathrm{EG}}_w=\left|\frac{\stackrel{\&OverBar;}{S}{\left(w\right)}^{\&prime;\&prime;}}{{(1+{\left(\stackrel{\&OverBar;}{S}{\left(w\right)}^{\&prime;}\right)}^2)}^{\frac{3}{2}}}\right|-\left|\frac{\stackrel{\&OverBar;}{S}{(w-1)}^{\&prime;\&prime;}}{{(1+{\left(\stackrel{\&OverBar;}{S}{(w-1)}^{\&prime;}\right)}^2)}^{\frac{3}{2}}}\right|---\left(7\right)$

In formula (7),Represent all previous α Entropy sequence H_wCubic spline function；Represent cubic spline functionFirst derivative；Represent cubic spline functionSecond dervative；And EG₁≤0；EG₂≤0。

5. the rotary kiln with imitative feed-back regulatory mechanism according to claim 3 or 4 burns till state identification method, it is characterized in that the Information Granularity δ of described the w+1 time cognitive process_w+1It is obtain as follows:

Step a, judge EG_wWhether≤0 set up, if setting up, then utilizes formula (8) to obtain the Information Granularity δ of the w+1 time cognitive process_w+1, otherwise perform step b:

${\mathrm{\&delta;}}_{w+1}={\mathrm{\&delta;}}_w(1-{\left(2^{w+1}-1\right)}^{-1}{H}_w^{\mathrm{\&alpha;}})---\left(8\right)$

Step b, formula (9) is utilized to obtain the Information Granularity δ of the w+1 time cognitive process_w+1:

${\mathrm{\&delta;}}_{w+1}={\mathrm{\&delta;}}_w(1-\frac{{(2^{w+1}-1)}^{-1}{H}_w^{\mathrm{\&alpha;}}}{2})---\left(9\right).$