CN103886316A - Combustion monitoring and diagnosis method based on feature extraction and fuzzy C-means cluster - Google Patents

Abstract

The invention discloses a combustion monitoring and diagnosis method based on feature extraction and fuzzy C-means cluster, comprising steps of utilizing time-frequency domain analysis and comentropy analysis technology to extract a flame characteristic value reflecting the combustion condition in flame detection signals, utilizing the fuzzy C-means cluster algorithm to perform FCM cluster analysis on the flame characteristic value and calculating a fuzzy combustion index through the fuzzy recognition. The invention can obtain a quantitative value of a combustion state through the fuzzy mode recognition and calculation of the combustion index, can classify the combustion states to realize quantitative monitoring of the combustion state, can promptly and accurately reflect boiler operation performance attributes, can provide new thoughts and methods for optimizing the performance and monitoring the state of the boiler, and can provide a reference model to the advanced modules of the power plant monitoring information system (boiler operation optimization, state monitoring and fault diagnosis).

Description

A kind of combustion monitoring based on feature extraction and fuzzy C-means clustering and the method for diagnosis

Technical field

The present invention relates to a kind of combustion condition monitoring method, relate in particular to FCM clustering method and Fuzzy Pattern Recognition, belong to machine learning modeling field.

Background technology

Cluster analysis is the one of multivariate statistical analysis, is also the important branch of non-supervised recognition.Cluster is a process that things is distinguished and sorted out, and in this process, not with the relevant priori of classification, the inherent similarity between things is unique criterion that generic is divided.Therefore, as the unsupervised sorting technique of one, it is according to certain criterion a sample set divide into several classes who there is no a classification mark, and making as far as possible similar sample assemble is a class, and dissimilar sample gathers inhomogeneity.Cluster analysis is divided into traditional cluster and the large class of fuzzy clustering two.Traditional cluster analysis is a kind of hard division, and object to be identified is strictly divided into a certain class, has either-or character.But under actual conditions, a lot of objects do not have strict attribute, and they are existing intermediary aspect form and generic, need to carry out soft division.

The quality of power boiler burning situation, directly has influence on safety and the economical operation of boiler, and combustion instability not only can reduce boiler thermal output, also may cause stove chamber fire-extinguishing extreme in the situation that, therefore needs to detect to judge that by flame whether burning is stable.Conventional flame detecting device can only judge " having " of flame, " nothing ", lacks combustion diagnosis function.Therefore, to the extraction of fire inspection signal, and how its vital role in combustion diagnosis of the signature analysis of fire inspection signal.

In recent years, many scholars have also studied from different angles the problem that boiler combustion stability is differentiated, and have obtained certain achievement.But due to complicacy and the polytrope of boiler combustion, also need to carry out further research work for combustion diagnosis.

Summary of the invention

Goal of the invention: in order to overcome the deficiencies in the prior art, the invention provides a kind of combustion monitoring based on feature extraction and fuzzy C-means clustering and the method for diagnosis.

Technical scheme: for solving the problems of the technologies described above, a kind of combustion monitoring based on feature extraction and fuzzy C-means clustering provided by the invention and the method for diagnosis, first while employing, frequency domain statistical study and the Analysis of Entropy technology extract the flame characteristic amount that reflects combustion position in fire inspection signal, recycling Fuzzy C-Means Cluster Algorithm carries out FCM cluster analysis to flame characteristic amount, then calculates fuzzy fire burning index by Fuzzy Pattern Recognition.

Above-mentioned concrete steps are

1, obtain different combustion conditions via Data Input Interface (good, in, poor) lower corresponding fire inspection signal, and calculate its variance (Var), the degree of bias (Ske), kurtosis (Kur), form factor (SF), singular spectrum entropy (H _t), Power Spectral Entropy (H _f), Wavelet Energy Spectrum entropy (H _w).

2, FCM cluster analysis step is: adopt Fuzzy C-Means Cluster Algorithm to determine that each data point belongs to the degree of certain cluster, in FCM clustering algorithm, it is by limited sample set x={x ₁, x ₂... x _n, x _i={ Var, Ske, Kur, SF, H _t, H _f, H _w, combustion conditions } and (2≤l≤n), arbitrary sample point can be not tight to be divided into l class

Lattice are divided into a certain class, but belong to l not same area with certain degree of membership.

$\begin{array}{c}\min J_m(u,v)=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{u}_{\mathrm{ij}}^m\·d_{\mathrm{ij}}^2=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{u}_{\mathrm{ij}}^m\·{\left|\right|x_j-v_i\left|\right|}_A^2\\ s.t.\underset{i=1}{\overset{k}{\mathrm{\Σ}}}{u}_{\mathrm{ij}}=1,\∀j=1,...,n\end{array}---\left(a\right)$

In formula a:

J _m(u, v)---objective function (containing the inter-object distance sum of squares function of fuzzy set theory);

---sample x _jto cluster centre v _idistance,

U _ij---j sample belongs to the degree of membership of i cluster;

M---FUZZY WEIGHTED index (m > 1).

Can solve: $\left\{\begin{array}{c}{U}_{\mathrm{ij}}=\frac{1}{\underset{k=1}{\overset{k}{\mathrm{\Σ}}}{\left(\frac{d_{\mathrm{ij}}}{d_{\mathrm{kj}}}\right)}^{\frac{2}{m-1}}}\\ v_i=\frac{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{u}_{\mathrm{ij}}^m\·x_j}{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{u}_{\mathrm{ij}}^m}\end{array}\right.---\left(b\right)$

Concrete steps are as follows:

Step 1 initialization: set Fuzzy Weighting Exponent m, cluster number k (2≤k≤n), iteration stopping threshold values ε and regulation iterations b, initialization cluster centre v ⁽⁰⁾;

In step dual-purpose (b), first expression formula is calculated degree of membership matrix

In step three-purpose type (b), second expression formula upgraded cluster centre

If step 4 || v ^(b)-v ^(b+1)|| < ε or iterations exceed stipulated number, stop algorithm, otherwise turn second step;

Step 5 Output rusults.

3, Fuzzy Pattern Recognition (based on maximum subjection principle)

If G is all identified the domain that object forms, A ₁, A ₂..., A _nbe n the subset of G, x ∈ G is an object to be identified.If ${u_A}_i\left(x\right)=\max \{{u_A}_1\left(x\right),{u_A}_2\left(x\right),...,{u_A}_n\left(x\right)\}$

In formula ---x is to A _idegree of membership.

Think that x is preferentially under the jurisdiction of A _i, x is under the jurisdiction of A _ithat class of representative.

4, calculate fuzzy fire burning index

Define the computing formula of the fuzzy fire burning index of i sample:

e＝3×u _i1+2×u _i2+1×u _i3 (c)

U in formula _i1---i sample degree of membership to this classification of fired state " good ";

U _i2---i sample to fired state " in " degree of membership of this classification;

U _i3---i sample degree of membership to this classification of fired state " poor ", and have

Beneficial effect: the method for combustion monitoring based on feature extraction and fuzzy C-means clustering provided by the invention and diagnosis in terms of existing technologies, has the following advantages:

1, burning sample data is carried out to FCM cluster, result shows feasibility and the validity of combustion diagnosis method.

2, the data to combustion position undetermined, are transformed to after burning sample, utilize Fuzzy Pattern Recognition and calculate fire burning index the quantitative values that can obtain fired state, thereby to combustion position classification, realize the Quantitative Monitoring of fired state.

3, reflect more timely and accurately boiler operation performance attribute, also for boiler performance optimization and status monitoring provide new thinking and method.

4, for power plant's monitoring information system Premium Features module (as boiler operatiopn optimization, condition monitoring and fault diagnosis etc.), providing can reference model.

Brief description of the drawings

Fig. 1 is process flow diagram of the present invention.

Fig. 2 is the fire burning index trend curve figure in embodiment in the present invention.

Embodiment

Below in conjunction with figure, the present invention is further described.

Carried out battery of tests on May 15th, 2012,16 days and three days on the 17th, totally 28 operating modes, the duration of each experiment condition is 15 minutes (after stable), change respectively the parameters such as air quantity, atomizing pressure, fuel pressure, fuel oil temperature, obtain the fire inspection signal under each operating mode, whole process mainly contains input data pre-service, entropy calculates, the nucleus modules such as PNN neural net model establishing and monitoring management, detailed process as shown in Figure 1:

1, furnace flame obtains fire inspection signal via flame detector, serial line interface, attached data acquisition system (DAS);

2, data enter after input data pre-service link is data acquisition system (DAS) and data storage document, 3B layer 33# fire inspection signal is carried out to feature extraction, obtain its combustion characteristic amount under each operating mode, comprise variance (Var), the degree of bias (Ske), kurtosis (Kur), form factor (SF), singular spectrum entropy (H _t), Power Spectral Entropy (H _f), Wavelet Energy Spectrum entropy (H _w);

3, get n group combustion characteristic data sample, ensure to contain three kinds of fired states in sample, carry out FCM cluster input data;

4, Fuzzy Weighting Exponent m=2 are set, clusters number is 3, and iteration stopping threshold value is ε=10 ^-5, utilize FCM algorithm to carry out cluster to combustion characteristic amount;

5, input need to be identified the sample of combustion position, calculates fuzzy fire burning index based on maximum membership grade principle, according to the scope of fire burning index to burner combustion condition diagnosing;

6, can be to carrying out as above step after all burner Data Integrations wherein for the fired state diagnosis of whole burner hearth.

In this flow process, FCM sorting procedure is: adopt Fuzzy C-Means Cluster Algorithm to determine that each data point belongs to the degree of certain cluster, in FCM clustering algorithm, it is by limited sample set x={x ₁, x ₂... x _n, x _i={ Var, Ske, Kur, SF, H _t, H _f, H _w, combustion conditions } and (2≤l≤n), arbitrary sample point can strictly not be divided into a certain class, but belongs to l not same area with certain degree of membership to be divided into l class.

$\begin{array}{c}\min J_m(u,v)=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m\&CenterDot;d_{\mathrm{ij}}^2=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m\&CenterDot;{\left|\right|x_j-v_i\left|\right|}_A^2\\ s.t.\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}=1,\&ForAll;j=1,...,n\end{array}---\left(a\right)$

In formula a:

J _m(u, v)---objective function (containing the inter-object distance sum of squares function of fuzzy set theory);

---sample x _jto cluster centre v _idistance,

U _ij---j sample belongs to the degree of membership of i cluster;

M---FUZZY WEIGHTED index (m > 1).

Can solve: $\left\{\begin{array}{c}{U}_{\mathrm{ij}}=\frac{1}{\underset{k=1}{\overset{k}{\mathrm{\&Sigma;}}}{\left(\frac{d_{\mathrm{ij}}}{d_{\mathrm{kj}}}\right)}^{\frac{2}{m-1}}}\\ v_i=\frac{\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m\&CenterDot;x_j}{\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m}\end{array}\right.---\left(b\right)$

Concrete steps are as follows:

Step 1 initialization: set Fuzzy Weighting Exponent m, cluster number k (2≤k≤n), iteration stopping threshold values ε and regulation iterations b, initialization cluster centre v ⁽⁰⁾;

In step dual-purpose (b), first expression formula is calculated degree of membership matrix

In step three-purpose type (b), second expression formula upgraded cluster centre

If step 4 || v ^(b)-v ^(b+1)|| < ε or iterations exceed stipulated number, stop algorithm, otherwise turn second step;

Step 5 Output rusults.

Maximum membership grade principle step is:

If G is all identified the domain that object forms, A ₁, A ₂..., A _nbe n the subset of G, x ∈ G is an object to be identified.If ${u_A}_i\left(x\right)=\max \{{u_A}_1\left(x\right),{u_A}_2\left(x\right),...,{u_A}_n\left(x\right)\}$

In formula

---x is to A _idegree of membership.

Think that x is preferentially under the jurisdiction of A _i, x is under the jurisdiction of A _ithat class of representative.

In order to check this diagnostic method validity, adopt FCM Clustering Model above to carry out combustion diagnosis to 12 groups of burning data samples to be identified, table 1 has provided the characteristic quantity for the treatment of diagnostic sample, diagnostic result is as shown in table 2, in table, thickened portion is the maximal value of each sample degree of membership, treat that as can be seen from the table diagnostic sample has all obtained correct fired state classification, thereby verified feasibility and the validity of the combustion diagnosis method proposing herein.

Table 1 is treated the characteristic quantity of diagnostic sample

Table 2 is treated the diagnostic result of diagnostic sample

Fuzzy fire burning index calculation procedure is:

Define the computing formula of the fuzzy fire burning index of i sample:

e＝3×u _i1+2×u _i2+1×u _i3 (c)

U in formula _i1---i sample degree of membership to this classification of fired state " good ";

U _i2---i sample to fired state " in " degree of membership of this classification;

U _i3---i sample degree of membership to this classification of fired state " poor ", and have

Get respectively 4:44:55AM～4:53:55AM time period May 15 (Test05) in 2012,5:31:02AM～5:40:02AM time period May 16 (Test17), examining data with the fire of 3:21:33AM～3:30:33AM time period May 17 (Test25) is research object, calculate corresponding fire burning index, its trend curve as shown in Figure 2.

According to field observation, these three fired states corresponding to time period be respectively " good ", " in ", " poor ".As can be seen from the figure, the fire burning index of the 1st time period fluctuates between 2.5～3, and flameholding is described, fired state is fine; The fire burning index of the 2nd time period is fluctuation up and down near 2, illustrates that burning is more stable, and fired state is medium; The fire burning index of the 3rd time period fluctuates between 1～1.5, and combustion instability is described, fired state is poor.Therefore, the fuzzy fire burning index of proposition can be used as the index of reflection fired state quality, realizes the Quantitative Monitoring of fired state.

The above is only the preferred embodiment of the present invention; be noted that for those skilled in the art; under the premise without departing from the principles of the invention, can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.

Claims (2)

1. a combustion monitoring based on feature extraction and fuzzy C-means clustering and the method for diagnosis, it is characterized in that: while the steps include: first to adopt, frequency domain statistical study and the Analysis of Entropy technology extract the flame characteristic amount that reflects combustion position in fire inspection signal, recycling Fuzzy C-Means Cluster Algorithm carries out FCM cluster analysis to flame characteristic amount, then calculates fuzzy fire burning index by Fuzzy Pattern Recognition.

2. combustion monitoring based on feature extraction and fuzzy C-means clustering according to claim 1 and the method for diagnosis, is characterized in that:

(1), obtain under different combustion conditions corresponding fire inspection signal via Data Input Interface, and calculate the combustion characteristic amount under each operating mode, comprise variance Var, degree of bias Ske, kurtosis Kur, shape factor S F, singular spectrum entropy H _t, Power Spectral Entropy H _f, Wavelet Energy Spectrum entropy H _w;

(2), FCM cluster analysis step is: adopt Fuzzy C-Means Cluster Algorithm to determine that each combustion characteristic amount data point belongs to the degree of certain cluster, in FCM clustering algorithm, by limited sample set x={x ₁, x ₂... x _n, x _i={ Var, Ske, Kur, SF, H _t, H _f, H _w, combustion conditions }, sample set is divided into l class, and (2≤l≤n), arbitrary sample point can strictly not be divided into a certain class, but belongs to l not same area with certain degree of membership;

$\begin{array}{c}\min J_m(u,v)=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m\&CenterDot;d_{\mathrm{ij}}^2=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m\&CenterDot;{\left|\right|x_j-v_i\left|\right|}_A^2\\ s.t.\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}=1,\&ForAll;j=1,...,n\end{array}---\left(a\right)$

In formula (a):

J _m(u, v)---objective function, containing the inter-object distance sum of squares function of fuzzy set theory;

---sample x _jto cluster centre v _idistance,

U _ij---j sample belongs to the degree of membership of i cluster;

M---FUZZY WEIGHTED index (m > 1);

Can solve:

$\left\{\begin{array}{c}{U}_{\mathrm{ij}}=\frac{1}{\underset{k=1}{\overset{k}{\mathrm{\&Sigma;}}}{\left(\frac{d_{\mathrm{ij}}}{d_{\mathrm{kj}}}\right)}^{\frac{2}{m-1}}}\\ v_i=\frac{\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m\&CenterDot;x_j}{\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m}\end{array}\right.$

Concrete steps are as follows:

Step 1 initialization: set Fuzzy Weighting Exponent m, cluster number k (2≤k≤n), iteration stopping threshold values ε and regulation iterations b, initialization cluster centre v ⁽⁰⁾;

In step dual-purpose (b), first expression formula is calculated degree of membership matrix

In step three-purpose type (b), second expression formula upgraded cluster centre

If step 4 || v ^(b)-v ^(b+1)|| < ε or iterations exceed stipulated number, stop algorithm, otherwise turn second step;

Step 5 Output rusults;

(3) Fuzzy Pattern Recognition is based on maximum subjection principle

If G is all identified the domain that object forms, A ₁, A ₂..., A _nbe n the subset of G, x ∈ G is an object to be identified, if ${u_A}_i\left(x\right)=\max \{{u_A}_1\left(x\right),{u_A}_2\left(x\right),...,{u_A}_n\left(x\right)\}$

In formula

---x is to A _idegree of membership; Think that x is preferentially under the jurisdiction of A _i, x is under the jurisdiction of A _ithat class of representative;

(4) calculate fuzzy fire burning index

Define the computing formula of the fuzzy fire burning index of i sample:

e＝3×u _i1+2×u _i2+1×u _i3 (c)

U in formula c _i1---i sample degree of membership to this classification of fired state " good ";

U _i2---i sample to fired state " in " degree of membership of this classification;

U _i3---i sample degree of membership to this classification of fired state " poor ", and have

Images