CN105568732A - Disc mill control method - Google Patents

Abstract

The invention discloses a disc mill control method. The disc mill control method comprises the following steps: a neural network is established, and a BP neural network and an additional momentum learning rule are used to train the neural network according to existing historical monitoring data; the optimal input parameter of the neural network is solved with particle swarm optimization; through estimated and actually measured error data of the neural network, the additional momentum learning rule is used to update the trained neural network. The disc mill control method solves the problems that the beating degree of a pulp mill is mainly determined by manual assay and far lags behind a beating process and a disc mill cannot be controlled precisely and effectively, soft measurement of the beating degree is realized, and on-line detection signals are provided for real-time control of the disc mill.

Description

Disc mill control method

Technical field

The present invention relates to a kind of disc mill control method, belong to paper equipment technical field processed.

Background technology

Disc mill is a kind of paper equipment processed, just can control beating degree by the advance and retreat cutter of control panel grinding machine.Making beating is the indispensable link of paper production, by measure beating degree can grasp fiber be cut off, moisten rise, the degree of sub-wire, fibrillating.Paper manufacturing commonly uses weight in wet base, beating degree two indices weighs pulp quality.But also beating degree is not carried out to the equipment of on-line monitoring at present, due to not to the signal that beating degree detects in real time, be also difficult to carry out effectively controlling accurately to disc mill.

Summary of the invention

The object of the present invention is to provide a kind of disc mill control method, according to carrying out on-line monitoring to beating degree, carrying out effectively controlling accurately to disc mill.

Object of the present invention is achieved by the following technical programs:

A kind of disc mill control method, comprises the following steps:

1) according to the record to the creation data of pulp mill, count into slurry flow, the electrical power of entering to starch concentration, disc mill consumption, the data of the beating degree in corresponding moment; Using entering to starch flow, enter to starch concentration, electrical power that disc mill consumes as input parameter, using beating degree as output parameter, set up neutral net, according to existing Historical Monitoring data, use BP neutral net, additional momentum learning rules, neural network training;

2) according to the setting of pulp mill's beating degree, by particle cluster algorithm, solve neutral net optimum input parameter, namely enter to starch flow, enter to starch concentration, disc mill consume electrical power;

3) judge whether to need artificial sample according to a upper Recognition with Recurrent Neural Network evaluated error, as needs, by artificial sample then off-line analysis, contrast draws the error of the beating degree that the beating degree of actual measurement and neutral net estimate, then by the beating degree data of this group actual measurement, and neutral net is estimated, together with the error information of actual measurement, to use additional momentum learning rules, upgrades neural network training; If do not needed artificial sample, then return step 2).

Object of the present invention can also be realized further by following technical measures:

Above-mentioned disc grinding machine control method, wherein particle cluster algorithm, step is as follows:

1) initialize population: determine population size NP, particle cluster algorithm iterations NG, initialize particle position, calculate the fitness of each particle and initialize globally optimal solution and individual optimal solution;

The function calculating particle fitness is:

$Fitness=\underset{i}{\Σ}{(O_i-O_i^{\′})}^2$

Wherein, O _irepresent i-th element of neutral net output vector, O ' _ifor i-th element of the output vector of theoretical expectation;

2) population is upgraded: the equation of motion of population is as follows:

v(t)＝ω·v(t-1)+c ₁·(lbest-x(t))+c ₂·(gbest-x(t))

x(t+1)＝x(t)+c ₃·v(t)

Wherein ω is taken as i is the current iteration number of times of particle cluster algorithm, c ₁, c ₂, c ₃for constant, c ₁, c ₂value is 2.8, c ₃value is 0.3, lbest is the individual optimal solution that each particle search is crossed, and gbest is the globally optimal solution that all particle search are crossed;

3) the particle fitness of current iteration is calculated, upgrade individual optimal solution and globally optimal solution: namely to each particle, by the fitness that current iteration produces, compared with current individual optimal solution, get fitness less for individual optimal solution, compared with the globally optimal solution crossed with all particle search, get fitness less for globally optimal solution;

4) judge whether to reach iteration NG time, if so, then export globally optimal solution, if not, then return step 2).

Above-mentioned disc grinding machine control method, wherein additional momentum learning method, update rule as shown in the formula:

$\mathrm{\ω}(t+1)=\mathrm{\ω}\left(t\right)-(1-a)\mathrm{\η}\frac{\∂E_T}{\∂\mathrm{\ω}\left(t\right)}+a\mathrm{\Δ}\mathrm{\ω}\left(t\right)$

Wherein Δ ω (t)=ω (t)-ω (t-1), E _tfor the training error of neutral net, η is weight, and a is factor of momentum, gets 0.9.

Above-mentioned disc grinding machine control method can also be achieved by another kind of technical scheme:

A kind of disc mill control method, comprises the following steps:

1) according to the record to the manufacturing parameter of pulp mill, count into slurry flow, the electrical power of entering to starch concentration, disc mill consumption, the data of the beating degree in corresponding moment; Using entering to starch flow, enter to starch concentration, electrical power that disc mill consumes as input parameter, using beating degree as output parameter, set up neutral net, according to existing Historical Monitoring data, use BP neutral net, additional momentum learning rules, neural network training;

2) according to the setting of pulp mill's beating degree, by genetic algorithm, solve neutral net optimum input parameter, namely enter to starch flow, enter to starch concentration, disc mill consume electrical power;

Described genetic algorithm comprises the following steps:

1. adopt real coding, initialize chromosome, form initial population;

2. each chromosome in fitness function evaluation each generation is utilized;

3. genetic manipulation is carried out;

4. the adaptive value of each individuality is recalculated;

5. after choosing new population, the optimum individual in new population is retained, replace the poorest individuality in this generation with the optimum individual of the previous generation;

6. judge whether to reach evolutionary generation, if do not have, then return the and 2. walk, otherwise terminate;

7. using the value of the optimum individual in new population as with, remain unchanged, adopt BP algorithm to learn, until meet performance indications;

3) judge whether to need artificial sample according to a upper Recognition with Recurrent Neural Network evaluated error, as needs, by artificial sample then off-line analysis, contrast draws the error of the beating degree that the beating degree of actual measurement and neutral net estimate, then by the beating degree data of this group actual measurement, and neutral net is estimated, together with the error information of actual measurement, to use additional momentum learning rules, upgrades neural network training; If do not needed artificial sample, then return step 2).

Above-mentioned disc grinding machine control method, carries out data prediction to the image data of entering to starch flow, and preprocess method is first adopt the Pauta criterion of statistical energy method to reject the abnormal data containing appreciable error, and then carry out filtering, filtering method is as follows:

1) carry out filtering to measured parameter, namely to measured parameter continuous sampling repeatedly, sampled value sorted, choosing median is this efficiently sampling value;

2) finite impulse response filter is carried out to measured parameter, first the frequency characteristic H of given ideal filter _d(e ^jw);

3) unit sample respo of ideal filter is calculated,

4) filter form is set, window function type, length of window N parameter be: sample frequency fs=150Hz, cut-off frequecy of passband fp=5Hz, stopband initial frequency fst=15Hz, stopband attenuation is not less than-50dB, window function type adopts Hamming window, filter order N=30;

5) Calling MATLAB function calculating filter coefficient w (n);

6) unit sample respo h (the n)=hd (n) w (n) of filter is calculated;

7) by N number of h (n) sequence of designing stored in corresponding stored district;

8) using median-filtered result x1 as x (n) stored in corresponding stored district;

9) circulation reading h (n), x (n) value carry out convolution algorithm, try to achieve online filter result $y\left(n\right)=\underset{m=0}{\overset{N-1}{\Σ}}x\left(m\right)h(n-m).$

Compared with prior art, the invention has the beneficial effects as follows: the beating degree that the invention solves pulp mill is mainly determined by artificial chemical examination, greatly lag behind pulping process, the problem of precisely effectively control cannot be carried out to disc mill, achieve the hard measurement to beating degree, provide line detection signal to the real-time control of disc mill.

Accompanying drawing explanation

Fig. 1 is Artificial Neural Network Structures figure of the present invention.

Detailed description of the invention

Below in conjunction with the drawings and specific embodiments, the invention will be further described.

As shown in Figure 1, the present invention set up into slurry flow, enter to starch concentration, disc mill consume electrical power and the beating degree in corresponding moment between mapping relations, set up neutral net.This neutral net, after training, carries out hard measurement to beating degree.

For realizing this purpose, specifically comprise the following steps:

1) according to the detailed record of the history of the manufacturing parameter of pulp mill, count into slurry flow, the electrical power of entering to starch concentration, disc mill consumption, the data of the beating degree in corresponding moment; Using entering to starch flow, enter to starch concentration, electrical power that disc mill consumes as input parameter, using beating degree as output parameter, set up neutral net, according to existing Historical Monitoring data, use BP neutral net, additional momentum learning rules, neural network training; Additional momentum learning method update rule as shown in the formula:

$\mathrm{\ω}(t+1)=\mathrm{\ω}\left(t\right)-(1-a)\mathrm{\η}\frac{\∂E_T}{\∂\mathrm{\ω}\left(t\right)}+a\mathrm{\Δ}\mathrm{\ω}\left(t\right)$

Wherein Δ ω (t)=ω (t)-ω (t-1), E _tfor the training error of neutral net, η is weight, and a is factor of momentum, gets 0.9.

2) according to the setting of pulp mill's beating degree, by particle cluster algorithm, solve neutral net optimum input parameter, namely enter to starch flow, enter to starch concentration, disc mill consume electrical power; Particle cluster algorithm step is as follows:

(1) initialize population: determine population size NP, particle cluster algorithm iterations NG, initialize particle position, calculate the fitness of each particle and initialize globally optimal solution and individual optimal solution;

The function calculating particle fitness is:

$Fitness=\underset{r}{\Σ}{(O_i-O_i^{\′})}^2$

Wherein, O _irepresent i-th element of neutral net output vector, O ' _ifor i-th element of the output vector of theoretical expectation;

(2) population is upgraded: the equation of motion of population is as follows:

v(t)＝ω·v(t-1)+c ₁·(lbest-x(t))+c ₂·(gbest-x(t))

x(t+1)＝x(t)+c ₃·v(t)

Wherein ω is taken as i is the current iteration number of times of particle cluster algorithm, c ₁, c ₂, c ₃for constant, c ₁, c ₂value is 2.8, c ₃value is 0.3, lbest is the individual optimal solution that each particle search is crossed, and gbest is the globally optimal solution that all particle search are crossed;

(3) the particle fitness of current iteration is calculated, upgrade individual optimal solution and globally optimal solution: namely to each particle, by the fitness that current iteration produces, compared with current individual optimal solution, get fitness less for individual optimal solution, compared with the globally optimal solution crossed with all particle search, get fitness less for globally optimal solution;

(4) judge whether to reach iteration NG time, if so, then export globally optimal solution, if not, then return step (2).

3) enter to starch flow due to neural network, enter to starch concentration, electrical power that disc mill consumes, mapping relations between the beating degree in corresponding moment, neutral net is after training, can predict beating degree and control, in order to obtain better observing and controlling effect, reduce neutral net evaluated error, need to upgrade neural network training.

Specific practice is for judge whether to need artificial sample according to a upper Recognition with Recurrent Neural Network evaluated error, its judgment rule is: if last time sampled data and Neural Network Data error less, then extend next artificial sample and this interval time of sampling, if these artificial sample data and Neural Network Data error are comparatively large, then reduce next artificial sample and this interval time of sampling; Concrete sampling interval duration then needs to determine according to the requirement of working control.As needs, by artificial sample then off-line analysis, contrast draw actual measurement with the error of data estimated, then by this group measured data, and neutral net estimate is together with the error information between measured value, use additional momentum learning rules, upgrade neural network training; If do not needed artificial sample, then return step 2).

Above-mentioned steps 2) solve neutral net optimum input parameter, namely solve corresponding with beating degree to enter to starch flow, enter to starch concentration, electrical power that disc mill consumes, then by equipment to entering to starch flow, enter to starch concentration and carry out regulable control, carried out the electrical power of control panel grinding machine consumption by the variable-frequency motor of control panel grinding machine, thus obtain desirable beating degree.

Object of the present invention can also be achieved by another kind of technical scheme, namely same based on neutral net, but uses genetic algorithm, solves the optimum input parameter of neutral net.

The method comprises the following steps:

1) according to the record to the manufacturing parameter of pulp mill, count into slurry flow, the electrical power of entering to starch concentration, disc mill consumption, the data of the beating degree in corresponding moment; Using entering to starch flow, enter to starch concentration, electrical power that disc mill consumes as input parameter, using beating degree as output parameter, set up neutral net, according to existing Historical Monitoring data, use BP neutral net, additional momentum learning rules, neural network training;

2) according to the setting of pulp mill's beating degree, by genetic algorithm, solve neutral net optimum input parameter, namely enter to starch flow, enter to starch concentration, disc mill consume electrical power;

Described genetic algorithm comprises the following steps:

1. adopt real coding, initialize chromosome, form initial population;

2. each chromosome in fitness function evaluation each generation is utilized;

3. genetic manipulation is carried out;

4. the adaptive value of each individuality is recalculated;

5. after choosing new population, the optimum individual in new population is retained, replace the poorest individuality in this generation with the optimum individual of the previous generation;

6. judge whether to reach evolutionary generation, if do not have, then return the and 2. walk, otherwise terminate;

7. using the value of the optimum individual in new population as with, remain unchanged, adopt BP algorithm to learn, until meet performance indications;

3) judge whether to need artificial sample according to a upper Recognition with Recurrent Neural Network evaluated error, as needs, by artificial sample then off-line analysis, contrast draws the error of the beating degree that the beating degree of actual measurement and neutral net estimate, then by the beating degree data of this group actual measurement, and neutral net is estimated, together with the error information of actual measurement, to use additional momentum learning rules, upgrades neural network training; If do not needed artificial sample, then return step 2).

Genetic algorithm is a kind of global optimization method of the random search based on biological evolution process, and it greatly reduces the impact of original state by crossover and mutation, makes search obtain optimal result, and does not rest on Local Minimum place.Therefore, in order to play genetic algorithm and BP algorithm strong point separately, regulating with BP algorithm and optimizing the parameter with locality, with genetic algorithm optimization, there is parameter of overall importance.

Owing to being vulnerable to interference to the detection of entering to starch flow, in order to obtain better technique effect, data prediction is carried out to the image data of entering to starch flow, preprocess method is first adopt the Pauta criterion of statistical energy method to reject the abnormal data containing appreciable error, then carry out filtering, filtering method is as follows:

1) carry out filtering to measured parameter, namely to measured parameter continuous sampling repeatedly, sampled value sorted, choosing median is this efficiently sampling value;

2) finite impulse response filter is carried out to measured parameter, first the frequency characteristic H of given ideal filter _d(e ^jw);

3) unit sample respo of ideal filter is calculated,

4) filter form is set, window function type, length of window N parameter be: sample frequency fs=150Hz, cut-off frequecy of passband fp=5Hz, stopband initial frequency fst=15Hz, stopband attenuation is not less than-50dB, window function type adopts Hamming window, filter order N=30;

5) Calling MATLAB function calculating filter coefficient w (n);

6) unit sample respo h (the n)=hd (n) w (n) of filter is calculated;

7) by N number of h (n) sequence of designing stored in corresponding stored district;

8) using median-filtered result x1 as x (n) stored in corresponding stored district;

9) circulation reading h (n), x (n) value carry out convolution algorithm, try to achieve online filter result $y\left(n\right)=\underset{m=0}{\overset{N-1}{\Σ}}x\left(m\right)h(n-m).$

In addition to the implementation, the present invention can also have other embodiments, and all employings are equal to the technical scheme of replacement or equivalent transformation formation, all drop in the protection domain of application claims.

Claims (5)

1. a disc mill control method, is characterized in that, comprises the following steps:

1) according to the record to the creation data of pulp mill, count into slurry flow, the electrical power of entering to starch concentration, disc mill consumption, the data of the beating degree in corresponding moment; Using entering to starch flow, enter to starch concentration, electrical power that disc mill consumes as input parameter, using beating degree as output parameter, set up neutral net, according to existing Historical Monitoring data, use BP neutral net, additional momentum learning rules, neural network training;

2) according to the setting of pulp mill's beating degree, by particle cluster algorithm, solve neutral net optimum input parameter, namely enter to starch flow, enter to starch concentration, disc mill consume electrical power;

3) judge whether to need artificial sample according to a upper Recognition with Recurrent Neural Network evaluated error, as needs, by artificial sample then off-line analysis, contrast draws the error of the beating degree that the beating degree of actual measurement and neutral net estimate, then by the beating degree data of this group actual measurement, and neutral net is estimated, together with the error information of actual measurement, to use additional momentum learning rules, upgrades neural network training; If do not needed artificial sample, then return step 2).

2. disc mill control method as claimed in claim 1, it is characterized in that, described particle cluster algorithm, step is as follows:

1) initialize population: determine population size NP, particle cluster algorithm iterations NG, initialize particle position, calculate the fitness of each particle and initialize globally optimal solution and individual optimal solution;

The function calculating particle fitness is:

$Fitness=\underset{i}{\Σ}{(O_i-O_i^{\′})}^2$

Wherein, O _irepresent i-th element of neutral net output vector, O _i' be i-th element of the output vector of theoretical expectation;

2) population is upgraded: the equation of motion of population is as follows:

v(t)＝ω·v(t-1)+c ₁·(lbest-x(t))+c ₂·(gbest-x(t))

x(t+1)＝x(t)+c ₃·v(t)

Wherein ω is taken as i is the current iteration number of times of particle cluster algorithm, c ₁, c ₂, c ₃for constant, c ₁, c ₂value is 2.8, c ₃value is 0.3, lbest is the individual optimal solution that each particle search is crossed, and gbest is the globally optimal solution that all particle search are crossed;

3) the particle fitness of current iteration is calculated, upgrade individual optimal solution and globally optimal solution: namely to each particle, by the fitness that current iteration produces, compared with current individual optimal solution, get fitness less for individual optimal solution, compared with the globally optimal solution crossed with all particle search, get fitness less for globally optimal solution;

4) judge whether to reach iteration NG time, if so, then export globally optimal solution, if not, then return step 2).

3. disc mill control method as claimed in claim 1 or 2, is characterized in that, described additional momentum learning method, update rule as shown in the formula:

$\mathrm{\ω}(t+1)=\mathrm{\ω}\left(t\right)-(1-a)\mathrm{\η}\frac{\∂E_T}{\∂\mathrm{\ω}\left(t\right)}+a\mathrm{\Δ}\mathrm{\ω}\left(t\right)$

Wherein Δ ω (t)=ω (t)-ω (t-1), E _tfor the training error of neutral net, η is weight, and a is factor of momentum, gets 0.9.

4. a disc mill control method, is characterized in that, comprises the following steps:

1) according to the record to the manufacturing parameter of pulp mill, count into slurry flow, the electrical power of entering to starch concentration, disc mill consumption, the data of the beating degree in corresponding moment; Using entering to starch flow, enter to starch concentration, electrical power that disc mill consumes as input parameter, using beating degree as output parameter, set up neutral net, according to existing Historical Monitoring data, use BP neutral net, additional momentum learning rules, neural network training;

2) according to the setting of pulp mill's beating degree, by genetic algorithm, solve neutral net optimum input parameter, namely enter to starch flow, enter to starch concentration, disc mill consume electrical power;

Described genetic algorithm comprises the following steps:

1. adopt real coding, initialize chromosome, form initial population;

2. each chromosome in fitness function evaluation each generation is utilized;

3. genetic manipulation is carried out;

4. the adaptive value of each individuality is recalculated;

5. after choosing new population, the optimum individual in new population is retained, replace the poorest individuality in this generation with the optimum individual of the previous generation;

6. judge whether to reach evolutionary generation, if do not have, then return the and 2. walk, otherwise terminate;

7. using the value of the optimum individual in new population as with, remain unchanged, adopt BP algorithm to learn, until meet performance indications;

3) judge whether to need artificial sample according to a upper Recognition with Recurrent Neural Network evaluated error, as needs, by artificial sample then off-line analysis, contrast draws the error of the beating degree that the beating degree of actual measurement and neutral net estimate, then by the beating degree data of this group actual measurement, and neutral net is estimated, together with the error information of actual measurement, to use additional momentum learning rules, upgrades neural network training; If do not needed artificial sample, then return step 2).

5. the disc mill control method as described in claim 1 or 4, it is characterized in that, carry out data prediction to the image data of entering to starch flow, preprocess method is first adopt the Pauta criterion of statistical energy method to reject the abnormal data containing appreciable error, then carry out filtering, filtering method is as follows:

1) carry out filtering to measured parameter, namely to measured parameter continuous sampling repeatedly, sampled value sorted, choosing median is this efficiently sampling value;

2) finite impulse response filter is carried out to measured parameter, first the frequency characteristic H of given ideal filter _d(e ^jw);

3) unit sample respo of ideal filter is calculated,

4) filter form is set, window function type, length of window N parameter be: sample frequency fs=150Hz, cut-off frequecy of passband fp=5Hz, stopband initial frequency fst=15Hz, stopband attenuation is not less than-50dB, window function type adopts Hamming window, filter order N=30;

5) Calling MATLAB function calculating filter coefficient w (n);

6) unit sample respo h (the n)=hd (n) w (n) of filter is calculated;

7) by N number of h (n) sequence of designing stored in corresponding stored district;

8) using median-filtered result x1 as x (n) stored in corresponding stored district;

9) circulation reading h (n), x (n) value carry out convolution algorithm, try to achieve online filter result $y\left(n\right)=\underset{m=0}{\overset{N-1}{\Σ}}x\left(m\right)h(n-m).$