CN103886405A - Boiler combustion condition identification method based on information entropy characteristics and probability nerve network - Google Patents

Abstract

The invention discloses a boiler combustion identification method based on information entropy characteristics and a probability nerve network. The method comprises steps of entering a data pretreatment procedure and obtaining typical load points and a characteristic sampling collection of corresponding exhaust smoke oxygen volume and furnace pressure signals through a data input interface, entering a sampling data entropy analysis process and calculating singular spectral entropy and power spectral entropy of the exhaust smoke oxygen volume and furnace pressure signals under the corresponding working condition, using the obtained entropy value signals and the corresponding load working condition point as a training data collection to construct a PNN boiler combustion working condition identification model and outputting the result to a client terminal to join the optimization operation guide and the condition detection. The invention not only solves procedure state characterization problem in the furnace but also reflects the attributes of the furnace operation performance timely and accurately, avoids fault guidance for the operation personnel caused by falsity data and wrong data, and provides a reference model to the boiler operation optimization, state monitor and failure diagnosis of a power plant monitor information system.

Description

A kind of Combustion Operation of Boilers recognition methods based on Information Entropy Features and probabilistic neural network

Technical field

The present invention relates to a kind of Combustion Operation of Boilers recognition methods based on boiler Information Entropy Features and probabilistic neural network, belong to machine learning modeling field.

Background technology

Machine learning is means and the mechanism of obtaining knowledge from known sample data or information by methods such as excavation, conclusion, deduction, analogies, its object be exactly according to design someway or algorithm, prior given training sample is learnt, then ask for the estimation to dependence between certain system input and output, and make this estimation preferably the unknown output be made prediction as far as possible accurately or its character is judged.Probabilistic neural network (Probabilistic Neural Networks, PNN) first proposed in 1989 by doctor D.F.Specht, the parallel algorithm that a kind of probability density function method of estimation based on Bayes classification rule and Parzen window develops, be also a class formation simple, train succinct, widely used artificial neural network.In actual applications, especially solving in the application of classification problem, complete the work that nonlinear learning algorithm is done with linear learning algorithm, keep the high precision of nonlinear algorithm simultaneously.The weights that PNN network is corresponding are exactly the distribution of pattern sample, thereby can meet the requirement of real-time processing.

The combustion process of large-scale power station pulverized coal firing boiler is a complicated nonlinear time-varying process, when actual motion, can be subject to boiler form, use coal, the impact of pulverized coal preparation system form and the factor such as the method for operation and air distribution mode, therefore be often difficult to set up accurate process mechanism model, thereby bring difficulty for boiler combustion monitoring and performance optimization, usually need to find the effective ways that characterize real-time status in stove, optimize with the Real-Time Monitoring, trend judgement and the operation that complete stove internal procedure.Therefore hearth combustion operating mode accurate modeling is significant to boiler combustion monitoring and performance optimization.

Boiler combustion process is complicated and changeable, chooses the emphasis that the parameter relevant with Combustion Operation of Boilers cross cutting is modeling.When Combustion Operation of Boilers changes, between the signals such as main vapour pressure, oxygen content in exhaust gas and furnace pressure, as mode input, will deviate from larger with actual condition.Boiler combustion coherent signal under typical condition is carried out to computational analysis with the characteristic parameter of descriptive system process in information entropy theory, can acquired information entropy feature with the tendency rule of Combustion Operation of Boilers variation.

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 Operation of Boilers recognition methods based on boiler Information Entropy Features and probabilistic neural network.

Technical scheme: for solving the problems of the technologies described above, a kind of Combustion Operation of Boilers recognition methods based on boiler Information Entropy Features and probabilistic neural network provided by the invention, comprises the following steps:

(1) enter " Data Input Interface machine " by " network switch " from the field data of DCS, enter input data pre-service link, the property samples collection that obtains typical load point and corresponding oxygen content in exhaust gas and furnace pressure signal by Data Input Interface machine, every kind of operating mode is got n sample and is analyzed:

D={x ₁, x ₂, x ₃... x _l; y ₁, y ₂, y ₃... y _lsubscript L is number of samples, and using property samples collection as calculating sample set;

(2) enter sample data entropy and analyze link, calculate oxygen content in exhaust gas and furnace pressure signal singular spectrum entropy and Power Spectral Entropy under corresponding operating mode:

1. the data of above-mentioned each load point of sample are solved to singular spectrum entropy:

For given discrete-time series signal { x _iory _i| i=1,2 ..., n}, given nested dimension m (m < n/2), the delay matrix A of structure embedded space,

$A=\left[\begin{array}{cccc}{x}_1& x_2& ...& x_{n-m+1}\\ x_2& x_3& ...& x_{n-m+2}\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ x_m& x_{m+1}& ...& x_n\end{array}\right]---\left(1\right)$

Matrix A is carried out to svd, and calculating singular value is λ ₁>=λ ₂>=...>=λ _m, establish the number that k is non-zero singular value, k value has reflected the number of the different mode comprising in delay matrix;

2. the data of above-mentioned each load point of sample are solved to Power Spectral Entropy:

Obtain frequency spectrum X (k) and power spectrum S by discrete Fourier transformation _k(k=1,2 ... n).S _kcan be regarded as a kind of energy binned of original signal in frequency domain space, the Power Spectral Entropy H of definable signal _f,

$H_f=-\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{p}_k\log p_k---\left(2\right)$

P in formula _k---k power spectrum shared ratio in whole power spectrum,

represent the component probability of k frequency energy,

for time domain discrete signal { x ₁, x ₂..., x _nfourier transform be the definition of the power Spectral Estimation of X (ω);

(3) the entropy signal of step (2) being tried to achieve and corresponding load condition point are constructed PNN Combustion Operation of Boilers model of cognition as training dataset, the singular spectrum entropy of oxygen content in exhaust gas and furnace pressure and the Power Spectral Entropy of furnace pressure are as the input of PNN neural network, corresponding operating mode kind, as output, is carried out training and testing to PNN model:

1. after training sample and sample to be identified being normalized, send into network input layer;

2. training sample W later normalization and sample X to be identified are sent into mode layer, complete dot-product operation z _i=(X-W _i) ^Τ(X-W _i),

Z in formula _i---the output of i mode layer node,

W _i---i training sample vector;

3. in summation layer, calculate probability density of all categories summation according to Pazen window function method, obtain the output of summation layer $y_k=\frac{1}{N_k}\underset{j=1}{\overset{N_k}{\mathrm{\&Sigma;}}}\exp [-\frac{z_j}{{2\mathrm{\&sigma;}}^2}],$

N in formula _k---be under the jurisdiction of the number of training of k class,

σ---smoothing factor;

4. adopt competitive function at output layer, according to Bayes classification rule discriminant, export maximum pattern and be the affiliated classification of X;

The final PNN model that obtains in real time, can be used for boiler operatiopn state recognition;

(4) result is outputed to client and participate in optimizing operation instruction and state-detection.

Further, singular value spectrum { λ in described step 2 _ican regard that the one in time domain is divided to signal as, and can calculate according to the thought of information entropy thus the singular spectrum entropy of time-domain signal, calculation procedure is as follows:

(1) choose nested dimension m, by time-domain signal sequence { x _i| i=1,2 ..., n} order intercepts, and obtains the delay matrix shown in formula (1);

(2) singular value of compute matrix A spectrum λ _i(i=1,2 ... k);

(3) singular value is composed to λ _i(i=1,2 ... k) substitution formula (2.24) obtains singular spectrum entropy H _t,

$H_t=-\underset{k=1}{\overset{k}{\mathrm{\&Sigma;}}}{p}_i\log p_i,---\left(3\right)$

P in formula _i---i singular value shared ratio in whole singular value spectrum, represent the component probability of i pattern.

Inventive principle: the invention discloses a kind of Combustion Operation of Boilers recognition methods based on boiler Information Entropy Features and probabilistic neural network of the present invention, for generating plant pulverized coal boiler running status characteristic, using oxygen signal and furnace pressure signal as parameter, application message entropy theory is studied the characteristic of boiler combustion process.Application singular spectrum entropy and Power Spectral Entropy characteristic parameter characterize the stove internal procedure under the various operating modes of boiler, obtain the Information Entropy Features parameter of stove internal procedure with the tendency rule of Combustion Operation of Boilers variation, and to adopt probabilistic neural network be PNN network, set up Combustion Operation of Boilers model of cognition.

Beneficial effect: the present invention not only can solve the problem of stove internal procedure state representation, and can reflect more timely and accurately boiler operation performance attribute, avoid the mistake that false data or misdata bring to operations staff to instruct, simultaneously can for the senior functional module such as boiler operatiopn optimization, condition monitoring and fault diagnosis of power plant's monitoring information system provides can reference model.

Brief description of the drawings

Fig. 1 is schematic flow sheet of the present invention.

Fig. 2 is training result and the error schematic diagram of PNN network.

Embodiment

Below in conjunction with drawings and Examples, the present invention is further described.

Embodiment: as shown in Figure 1, a kind of Combustion Operation of Boilers recognition methods based on boiler Information Entropy Features and probabilistic neural network, described step comprises:

(1) enter " Data Input Interface machine " by " network switch " from the field data of DCS, enter input data pre-service link, the property samples collection that obtains typical load point and corresponding oxygen content in exhaust gas and furnace pressure signal by Data Input Interface machine, every kind of operating mode is got n sample and is analyzed:

D={x ₁, x ₂, x ₃... x _l; y ₁, y ₂, y ₃... y _lsubscript L is number of samples, and using property samples collection as calculating sample set;

(2) enter sample data entropy and analyze link, calculate oxygen content in exhaust gas and furnace pressure signal singular spectrum entropy and Power Spectral Entropy under corresponding operating mode:

1. the data of above-mentioned each load point of sample are solved to singular spectrum entropy:

For given discrete-time series signal { x _iory _i| i=1,2 ..., n}, given nested dimension m (m < n/2), the delay matrix A of structure embedded space,

$A=\left[\begin{array}{cccc}{x}_1& x_2& ...& x_{n-m+1}\\ x_2& x_3& ...& x_{n-m+2}\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ x_m& x_{m+1}& ...& x_n\end{array}\right]---\left(1\right)$

Matrix A is carried out to svd, and calculating singular value is λ ₁>=λ ₂>=...>=λ _m, establish the number that k is non-zero singular value, k value has reflected the number of the different mode comprising in delay matrix, singular value spectrum { λ _ican regard that the one in time domain is divided to signal as, and can calculate according to the thought of information entropy thus the singular spectrum entropy of time-domain signal, calculation procedure is as follows:

(1) choose nested dimension m, by time-domain signal sequence { x _i| i=1,2 ..., n} order intercepts, and obtains the delay matrix shown in formula (1);

(2) singular value of compute matrix A spectrum λ _i(i=1,2 ... k);

(3) singular value is composed to λ _i(i=1,2 ... k) substitution formula (2.24) obtains singular spectrum entropy H _t,

$H_t=-\underset{k=1}{\overset{k}{\mathrm{\&Sigma;}}}{p}_i\log p_i,---\left(3\right)$

P in formula _i---i singular value shared ratio in whole singular value spectrum,

represent the component probability of i pattern.

2. the data of above-mentioned each load point of sample are solved to Power Spectral Entropy:

Obtain frequency spectrum X (k) and power spectrum S by discrete Fourier transformation _k(k=1,2 ... n).S _kcan be regarded as a kind of energy binned of original signal in frequency domain space, the Power Spectral Entropy Hf of definable signal,

$H_f=-\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{p}_k\log p_k---\left(2\right)$

P in formula _k---k power spectrum shared ratio in whole power spectrum, represent the component probability of k frequency energy,

for time domain discrete signal { x ₁, x ₂..., x _nfourier transform be the definition of the power Spectral Estimation of X (ω);

(3) the entropy signal of step (2) being tried to achieve and corresponding load condition point are constructed PNN Combustion Operation of Boilers model of cognition as training dataset, the singular spectrum entropy of oxygen content in exhaust gas and furnace pressure and the Power Spectral Entropy of furnace pressure are as the input of PNN neural network, corresponding operating mode kind, as output, is carried out training and testing to PNN model:

1. after training sample and sample to be identified being normalized, send into network input layer;

2. training sample W later normalization and sample X to be identified are sent into mode layer, complete dot-product operation z _i=(X-W _i) ^Τ(X-W _i),

Z in formula _i---the output of i mode layer node,

W _i---i training sample vector;

3. in summation layer, calculate probability density of all categories summation according to Pazen window function method, obtain the output of summation layer $y_k=\frac{1}{N_k}\underset{j=1}{\overset{N_k}{\mathrm{\&Sigma;}}}\exp [-\frac{z_j}{{2\mathrm{\&sigma;}}^2}],$

N in formula _k---be under the jurisdiction of the number of training of k class,

σ---smoothing factor;

4. adopt competitive function at output layer, according to Bayes classification rule discriminant, export maximum pattern and be the affiliated classification of X;

The final PNN model that obtains in real time, can be used for boiler operatiopn state recognition;

(4) result is outputed to client and participate in optimizing operation instruction and state-detection.

In the present embodiment, from studied 1000MW unit, get the training sample of 40 groups of data samples as PNN neural network, each 5 groups of the sample that wherein 8 kinds of combustion conditions are corresponding.The input layer of the PNN disaggregated model of setting up has 3 nodes, corresponding 3 dimension Information Entropy Features vectors, and mode layer has 40 nodes, corresponding 40 number of training, summation layer and output layer have 8 nodes, corresponding 8 kinds of combustion conditions patterns.Table 1 has provided 8 kinds of combustion conditions patterns and corresponding description thereof, and table 2 item has been listed part training sample.

Table 1 operating mode pattern and description

The Information Entropy Features of table 2 training sample

As shown in Figure 2, the training that adopts above-mentioned data as training sample, PNN network to be carried out, the training result demonstration PNN network of building can correctly be classified to all samples.

Test result shows, PNN disaggregated model has very high recognition accuracy to steady load operating mode, load fluctuation operating mode and igniting, blowing out operating mode.

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. the Combustion Operation of Boilers recognition methods based on Information Entropy Features and probabilistic neural network, is characterized in that: comprise the following steps:

(1) enter Data Input Interface machine from the field data of DCS by the network switch, enter input data pre-service link, the property samples collection that obtains typical load point and corresponding oxygen content in exhaust gas and furnace pressure signal by Data Input Interface machine, every kind of operating mode is got n sample and is analyzed:

D={x ₁, x ₂, x ₃... x _l; y ₁, y ₂, y ₃... y _lsubscript L is number of samples, and using property samples collection as calculating sample set;

(2) enter sample data entropy and analyze link, calculate oxygen content in exhaust gas and furnace pressure signal singular spectrum entropy and Power Spectral Entropy under corresponding operating mode:

1. the data of above-mentioned each load point of sample are solved to singular spectrum entropy:

For given discrete-time series signal { x _iory _i| i=1,2 ..., n}, given nested dimension m (m < n/2), the delay matrix A of structure embedded space,

$A=\left[\begin{array}{cccc}{x}_1& x_2& ...& x_{n-m+1}\\ x_2& x_3& ...& x_{n-m+2}\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ x_m& x_{m+1}& ...& x_n\end{array}\right]---\left(1\right)$ Matrix A is carried out to svd, and calculating singular value is λ ₁>=λ ₂>=...>=λ _m, establish the number that k is non-zero singular value, k value has reflected the number of the different mode comprising in delay matrix;

2. the data of above-mentioned each load point of sample are solved to Power Spectral Entropy:

Obtain frequency spectrum X (k) and power spectrum S by discrete Fourier transformation _k(k=1,2 ... n).S _kcan be regarded as a kind of energy binned of original signal in frequency domain space, the Power Spectral Entropy H of definable signal _f,

$H_f=-\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{p}_k\log p_k---\left(2\right)$ P in formula _k---k power spectrum shared ratio in whole power spectrum,

represent the component probability of k frequency energy,

for time domain discrete signal { x ₁, x ₂..., x _nfourier transform be the definition of the power Spectral Estimation of X (ω);

(3) the entropy signal of step (2) being tried to achieve and corresponding load condition point are constructed PNN Combustion Operation of Boilers model of cognition as training dataset, the singular spectrum entropy of oxygen content in exhaust gas and furnace pressure and the Power Spectral Entropy of furnace pressure are as the input of PNN neural network, corresponding operating mode kind, as output, is carried out training and testing to PNN model:

1. after training sample and sample to be identified being normalized, send into network input layer;

2. training sample W later normalization and sample X to be identified are sent into mode layer, complete dot-product operation z _i=(X-W _i) ^Τ(X-W _i),

Z in formula _i---the output of i mode layer node,

W _i---i training sample vector;

3. in summation layer, calculate probability density of all categories summation according to Pazen window function method, obtain the output of summation layer $y_k=\frac{1}{N_k}\underset{j=1}{\overset{N_k}{\mathrm{\&Sigma;}}}\exp [-\frac{z_j}{{2\mathrm{\&sigma;}}^2}],$

N in formula _k---be under the jurisdiction of the number of training of k class,

σ---smoothing factor;

4. adopt competitive function at output layer, according to Bayes classification rule discriminant, export maximum pattern and be the affiliated classification of X;

The final PNN model that obtains in real time, can be used for boiler operatiopn state recognition;

(4) result is outputed to client and participate in optimizing operation instruction and state-detection.

2. a kind of Combustion Operation of Boilers recognition methods based on Information Entropy Features and probabilistic neural network according to claim 1, is characterized in that: singular value spectrum { λ in described step 2 _ican regard that the one in time domain is divided to signal as, and can calculate according to the thought of information entropy thus the singular spectrum entropy of time-domain signal, calculation procedure is as follows:

(1) choose nested dimension m, by time-domain signal sequence { x _i| i=1,2 ..., n} order intercepts, and obtains the delay matrix shown in formula (1);

(2) singular value of compute matrix A spectrum λ _i(i=1,2 ... k);

(3) singular value is composed to λ _i(i=12...k) substitution formula (2.24) obtains singular spectrum entropy H _t,

$H_t=-\underset{k=1}{\overset{k}{\mathrm{\&Sigma;}}}{p}_i\log p_i,---\left(3\right)$

P in formula _i---i singular value shared ratio in whole singular value spectrum, represent the component probability of i pattern.

Images