CN102636624B - Method for soft measurement of alumina concentration in electrolyzer during aluminum electrolysis process - Google Patents

Abstract

The invention discloses a method for soft measurement of alumina concentration in an electrolyzer during an aluminum electrolysis process, comprising the following steps: (1) collecting process production parameter data; (2) preprocessing the process production parameter data; (3) establishing samples of process production parameter data; (4) establishing a multi-alumina concentration soft measurement model based on different electrolyzer statuses and obtaining model parameters; (5) employing the model to estimate alumina concentration. The method disclosed by the invention can ensure fast, accurate and real-time detection of changes in the alumina concentration so as to achieve optimal control of the production process.

Description

The flexible measurement method of alumina concentration in aluminium electrolysis process electrolytic tank

Technical field

The present invention relates to the measuring technique of aluminium electrolysis process manufacturing parameter, relate in particular to a kind of flexible measurement method that is difficult to measure in real time online with physical sensors alumina concentration in electrolytic tank for solving aluminium electrolysis process.

Background technology

In aluminium electrolytic industry, alumina concentration is the important state parameter of reflection aluminium cell production run process, also be an important control parameter that maintains aluminium cell material balance simultaneously, produce the process schedule requirement of " three low height " according to modern aluminum electrolytic industry, alumina concentration must be controlled to a lower concentration range, but because the sensor cost of directly measuring for alumina concentration is at present high, the economy such as easy damage, technical reason, cause lacking the online measurement means of directly measuring alumina concentration always, this has become restriction aluminium electroloysis industry and has further improved control efficiency, reduce a bottleneck of energy consumption.At present, soft-measuring technique provides effective approach for solving problems, and be considered to fruitful and the most attractive new method, add under the condition of investment not increasing or reduce, soft-measuring technique will be used widely, thereby process control and detection system are produced to tremendous influence.Soft measurement is exactly (to be called auxiliary variable according to the process variable that easily survey can be surveyed, as speed, pressure, temperature etc.) (be called leading variable with the variable to be measured that is difficult to direct-detection, as material component, product quality etc.) mathematical relation, according to certain optiaml ciriterion, adopt various computing method, realize measurement or the estimation to variable to be measured with soft-sensing model.

At present, for the soft measurement of alumina concentration, main or cell resistance slope calculation method,, by can extrapolate the concentration of aluminium oxide to the analysis of cell resistance size, its relation curve is as shown in Figure 1.Because cell resistance R calculates gained by tank voltage and potline current intensity, can be the relation between tank voltage and alumina concentration by the relationship conversion shown in Fig. 1, change and be transformed within the scope of a less time interval by alumina concentration, within the scope of this according to tank voltage or cell resistance slope value and corresponding alumina concentration value thereof, adopt the method for the linear theories such as linear regression model (LRM), set up the soft-sensing model of alumina concentration, this method can only effectively detect alumina concentration and change in very little perform region, and the soft measurement problem of the alumina concentration that can not thoroughly solve whole perform region, this is because alumina concentration is except outside the Pass changing with current tank voltage or cell resistance and having, also with the current or residing electrolytic tank state of electrolytic tank in early time, the groove action adopting, relevant with fluctuation of other manufacturing parameters etc.In addition, based on artificial neural network (Artificial Neural Network, brief note is for ANN) the soft measurement of alumina concentration also someone adopted, soft measurement based on artificial neural network can be under the condition of priori that does not possess object, according to the input/output relation Direct Modeling of object, the on-line correction ability of model is strong, and can be applicable to nonlinearity and serious time-dependent system.But, the selection of quality and quantity, learning algorithm, topology of networks and the type etc. of the learning training sample of neural network has significant impact to the soft-sensing model forming, and the selection of training sample proper vector often lacks reliable technological basis, so adopt the alumina concentration soft-sensing model of neural network also to have its limitation.In sum, because aluminium cell is a non-homogeneous unsteady Model, every factor that the alumina concentration relating to changes intercouples, so combined process is chosen the feature input of suitable alumina concentration model, adopting a kind of accurate, reliable non-linear soft-sensing model is vital to setting up alumina concentration soft-sensing model.

Summary of the invention

In view of this, the object of this invention is to provide the flexible measurement method of alumina concentration in a kind of accurate, reliable aluminium electrolysis process electrolytic tank.

The object of the invention is to be achieved through the following technical solutions:

The flexible measurement method of alumina concentration in this aluminium electrolysis process electrolytic tank, comprises the following steps:

Step 1: gatherer process manufacturing parameter data, comprise that the process that gathers tank voltage, potline current, reinforced interval (NB), aluminum yield produces parameter, and set up the historical data base of storage parameters numerical value by hardware device;

Step 2: set one and read time parameter and read time period parameter, read time parameter every one section, just read the data before time period parameter that read in this moment of historical data base, process manufacturing parameter is carried out to data processing, set up training sample set and test sample book set, the form of sample set is { x _i, y _i, wherein

X _i={ F (k), H (k), R (k), S (k), T (k), NB (k), L (k), V _s(k) }, y _i={ C (k) }, x _i∈ R ⁸, y _i∈ R, wherein, in formula, the implication of each parameter is as follows:

K represents residing certain sampling instant of sample i, moment k-moment (k-1)=read time parameter;

F (k) is cell voltage fluctuation number of times, reads time period parameter with interior all tank voltage data-signal V to what collect _kaccording to eight layers of DB8 wavelet decomposition signals to for tree construction, then the signal of (8,0) frequency range is reconstructed and obtains reading the tank voltage signal V of time period parameter with interior (8,0) frequency range _80k, statistics reads the tank voltage signal wave crest trough number of time period parameter with interior (8,0) frequency range, and a Wave crest and wave trough is designated as 1 secondary undulation, and crest or trough are 0.5 secondary undulation;

H (k) is high-frequency energy number percent, reads time period parameter and is designated as V with interior tank voltage collecting _k, calculate tank voltage signal gross energy

calculate (8,0) frequency range tank voltage signal energy

high-frequency energy number percent $H\left(k\right)=\frac{E_h}{E}=\frac{E-E_{80}}{E};$

R (k) is cell resistance, processes the tank voltage V that obtains (8,0) frequency range by the k moment _80kpotline current I (k) divided by this moment obtains, and the mode of utilizing following smoothing algorithm is deleted the exceptional value in the cell resistance sequential value of sampling rate, wherein, e is the deviation of cell resistance value and smoothing computation value, ω represents shift time, ω=5s, R (k-ω) is the cell resistance value of 5s before the k moment, R (k-2 ω) is the cell resistance value of 10s before the k moment.

R(k)＝0.3×R(k-ω)-0.1×R(k-2ω)+0.2×(e(k)+2×e(k-ω)+e(k-2ω))；

S (k) is cell resistance slope, and being defined as the increment that reads recently the filter resistance in time parameter is rate of change; T (k) is accumulation slope, be defined as the cumulative increment that reads recently the filter resistance in time parameter in the recent period, use difference filter slope calculations, and use recursion formula to calculate accumulation slope, specific algorithm is take 30s as the cycle, low-pass filtering resistance to be sampled, press two formulas calculated resistance slope S and accumulation slope T respectively below

S(k)＝(R(k-6ω)-R(k-18ω)+2×(R(k)-R(k-24ω)))/5，

T(k)＝(15/16)×T(k-1)+S(k)/4；

NB (k) is alumina blanking interval, aluminum electrolysis control system actual set blanking spacing value, represent with respect to benchmark blanking interval multiple, NB (k) is between 0.8-1.3, when NB (k)=1, represent with benchmark blanking interval blanking;

L (k) is aluminum yield, the quality of the metallic aluminium that expression aluminium cell extracts carry out out aluminium operation in the time of the k moment time, and the average aluminum yield using L (k) as the k moment in this method, units/kg,

V _s(k) be setting voltage, represent that aluminum electrolytic cell control system is the tank voltage that electrolytic tank is set at moment k, the V of unit;

C (k) is alumina concentration, represent to be dissolved into the mass percent of the aluminium oxide in electrolyte, general oxidation concentration is all between 0%-12%, C (k) value part of only peeking here, and 2% alumina concentration is expressed as C (k)=2;

Step 3: adopt following algorithm to be normalized sample set, attribute is zoomed between [0,1]:

$x_i=\frac{x_i}{\left|\right|x_i\left|\right|}$

$y_i=\frac{y_i}{12},$

In above formula || || get 2-norm;

Step 4: according to all electrolytic tank states of storing in control system historical data base, set up alumina concentration soft-sensing model corresponding under different electrolytic tank states, solving model, computation model parameter;

Step 5: according to current electrolytic tank state and its corresponding alumina concentration soft-sensing model, test sample book, in conjunction with being input in soft-sensing model, is calculated to its corresponding alumina concentration discreet value

utilize the algorithm of following formula by alumina concentration discreet value

renormalization is alumina concentration value, estimates out alumina concentration corresponding to current test sample book set,

${\hat{y}}_i=12{\hat{y}}_i.$

Further, in step 4, groove state comprises health, ill, Very Ill-conditioned totally seven kinds of states of health, inferior health, critical health, slight morbid state, moderate substantially, collect sample set for the training under different slots state, adopt least square method supporting vector machine (LSSVM) training to obtain the alumina concentration soft-sensing model that different slots state is corresponding, totally 7 kinds, model tormulation is as follows:

$\hat{y}\left(x\right)=f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_{i}K(x,x_i)+b,$

In above formula, N is input number of samples, x _ibe i input sample, x is a certain input variable, α _i≠ 0 is corresponding input sample x _ifor support vector, α _ifor support vector machine coefficient,

for model pre-estimating output, K (x, x _i) be kernel function;

Further, kernel function is radial basis function (RBF);

Further, in step 4, to the electrolytic tank under a certain groove state, the algorithm of alumina concentration soft-sensing model parameter is as follows:

Step 1:

Model parameter initialization; (comprise initialization Population n, maximum iteration time k _max, study factor c ₁, study factor c ₂, penalty coefficient pace of change lower bound Vc _min, penalty coefficient pace of change upper bound Vc _max, kernel function spread factor changes lower bound V σ _min, kernel function spread factor change upper bound V σ _max, kernel function spread factor lower bound σ _min, kernel function spread factor upper bound σ _max, penalty coefficient lower bound c _min, penalty coefficient upper bound c _max, inertia weight ω);

Step 2:

The corresponding algorithm parameter c of all particles and σ in initialization colony, i.e. initialization particle population, initial population is shone upon to produce by Logistic and is obtained, and concrete grammar is: definition ξ ₁₁, ξ ₁₂for the random number in (0,1), and ξ ₁₁≠ ξ ₁₂, by ξ ₁₁, ξ ₁₂substitution Logistic mapping, that is:

ξ _i+1,1＝4ξ _i1(1-ξ _i1)

ξ _i+1,2＝4ξ _i2(1-ξ _i2),i＝1,...,m-1，

Above formula iteration is obtained to m group Chaos Variable for m-1 time

Again Chaos Variable is mapped in feasible zone, even:

c _i＝c _min+ξ _i1(c _max-c _min)

σ _i＝σ _min+ξ _i2(σ _max-σ _min)，，

Definition initial population

wherein x _i=(x _i1, x _i2)=(c _i, σ _i) be the position of i particle;

Calculate optimal adaptation value corresponding to all particles in colony, the individual particles optimal location in definition colony is:

m is Population; Definition colony optimal location: the optimal location P of colony _ginitial value be that of adaptive value minimum in personal best particle; By m P _iin the following formula of radial basis kernel function substitution:

Application least square method is obtained m group model parameter alpha _iand b, then by this m group model parameter substitution successively $\hat{y}\left(x\right)=f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_i\mathrm{Kx}(x_i+b)$ With $f=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{(y_i-{\hat{y}}_i)}^2$ In, the f trying to achieve is m adaptive value corresponding to all particles, and the adaptive value of now trying to achieve is initialized as to the optimal adaptation value that now particle is corresponding, optimal adaptation value refers to the difference of two squares of estimating output and actual output of adaptive value under parameter current;

Step 3:

Search colony optimal adaptation value, i.e. optimal adaptation value corresponding to all particles in colony relatively, finding minimum optimal adaptation value is corresponding colony optimal adaptation value, and using algorithm parameter c corresponding particle and σ as optimum model parameter;

Step 4:

Value x and the pace of change v thereof of all particle algorithm parameter c and σ in renewal colony;

Judge and change in colony all particle algorithm parameter c and σ within the scope of feasible zone, that is:

If v _i1> vc _max, make v _i1=vc _max;

If v _i1< vc _min, make v _i1=vc _min;

If v _i2> v σ _max, make v _i2=v σ _max;

If v _i2< v σ _min, make v _i2=v σ _min;

If x _i1< c _min, make x _i1=c _min;

If x _i1> c _max, make x _i1=c _max;

If x _i2< σ _min, make x _i2=σ _min;

If x _i2> σ _max, make x _i2=σ _max, i=1 ..., Population

Least square method is calculated the model parameter α of each particle under its algorithm parameter c and σ _iand b;

Calculate the optimal adaptation value of all particles in colony, the optimal adaptation value value that all particles are corresponding is the minimum value between adaptive value and the optimal adaptation value calculating of last iteration that time iterative computation goes out;

Search colony optimal adaptation value, i.e. optimal adaptation value corresponding to all particles in colony relatively, finding minimum optimal adaptation value is corresponding colony optimal adaptation value, and using algorithm parameter c corresponding particle and σ as optimum model parameter;

Step 5:

Judge whether iterations reaches maximal value, if do not reached, returns to step 4, otherwise execution step 6;

Step 6:

The algorithm parameter c calculating after last iteration and σ are decided to be the algorithm parameter of aluminium oxide soft-sensing model;

Step 7:

Use least square method computational algorithm parameter c and model parameter α corresponding to σ _iand b;

Further, in step 1, reading time parameter is 10min, and reading time period parameter is 2h.

The invention has the beneficial effects as follows:

(1) realize automatically and control monitor data is provided for aluminium electrolysis process, for the optimal control that realizes electrolytic aluminium process lays the foundation with optimization operation;

(2), according to different electrolytic tank running statuses, training obtains the many alumina concentrations soft-sensing model structure under different conditions, makes the adaptive faculty of soft measurement stronger;

(3) detect in real time alumina concentration and change, realize production process monitoring, improve electrolytic aluminium output;

(4) replace artificial assay, reach the object that promptly and accurately detects production status, further Optimization Technology.

Other advantages of the present invention, target and feature will be set forth to a certain extent in the following description, and to a certain extent, based on will be apparent to those skilled in the art to investigating below, or can be instructed from the practice of the present invention.Target of the present invention and other advantages can be realized and be obtained by instructions and claims below.

Accompanying drawing explanation

In order to make the object, technical solutions and advantages of the present invention clearer, below in conjunction with accompanying drawing, the present invention is described in further detail, wherein:

Fig. 1 is alumina concentration-cell resistance relation curve;

Fig. 2 is the implementation framework of soft-sensing model of the present invention;

Fig. 3 is hardware platform structural representation of the present invention;

Fig. 4 is the WAVELET PACKET DECOMPOSITION restructuring tree construction adopting in sample data process of establishing of the present invention;

Fig. 5 is electrolytic bath change in voltage curve.

Embodiment

Hereinafter with reference to accompanying drawing, the preferred embodiments of the present invention are described in detail.Should be appreciated that preferred embodiment is only for the present invention is described, rather than in order to limit the scope of the invention.

As shown in Figure 2, the flexible measurement method of alumina concentration in aluminium electrolysis process electrolytic tank of the present invention, comprises the following steps:

Step 1: gatherer process manufacturing parameter data, comprise that the process that gathers tank voltage, potline current, reinforced interval (NB), aluminum yield produces parameter, and set up the historical data base of storage parameters numerical value by hardware device;

Step 2: set one and read time parameter and read time period parameter, read time parameter every one section, just read the data before time period parameter that read in this moment of historical data base, process manufacturing parameter is carried out to data processing, set up training sample set and test sample book set, the form of sample set is { x _i, y _i, in the present embodiment, reading time parameter is 10min, read time period parameter and get 2h, i.e. and data before 10min reads this moment of historical data base 2h, wherein

X _i={ F (k), H (k), R (k), S (k), T (k), NB (k), L (k), V _s(k) }, y _i={ C (k) }, x _i∈ R ⁸, y _i∈ R, wherein, in formula, the implication of each parameter is as follows:

K represents residing certain sampling instant of sample i, moment k-moment (k-1)=read time parameter;

F (k) is cell voltage fluctuation number of times, reads time period parameter with interior all tank voltage data-signal V to what collect _kaccording to eight layers of DB8 wavelet decomposition signals to for tree construction, then the signal of (8,0) frequency range is reconstructed and obtains reading the tank voltage signal V of time period parameter with interior (8,0) frequency range _80k, statistics reads the tank voltage signal wave crest trough number of time period parameter with interior (8,0) frequency range, and a Wave crest and wave trough is designated as 1 secondary undulation, and crest or trough are 0.5 secondary undulation;

H (k) is high-frequency energy number percent, reads time period parameter and is designated as V with interior tank voltage collecting _k, calculate tank voltage signal gross energy

calculate (8,0) frequency range tank voltage signal energy high-frequency energy number percent $H\left(k\right)=\frac{E_h}{E}=\frac{E-E_{80}}{E};$

R (k) is cell resistance, processes the tank voltage V that obtains (8,0) frequency range by the k moment _80kpotline current I (k) divided by this moment obtains, and the mode of utilizing following smoothing algorithm is deleted the exceptional value in the cell resistance sequential value of sampling rate, wherein, e is the deviation of cell resistance value and smoothing computation value, ω represents shift time, ω=5s, R (k-ω) is the cell resistance value of 5s before the k moment, R (k-2 ω) is the cell resistance value of 10s before the k moment.

R(k)＝0.3×R(k-ω)-0.1×R(k-2ω)+0.2×(e(k)+2×e(k-ω)+e(k-2ω))；

S (k) is cell resistance slope, and being defined as the increment that reads recently the filter resistance in time parameter is rate of change; T (k) is accumulation slope, be defined as the cumulative increment that reads recently the filter resistance in time parameter in the recent period, use difference filter slope calculations, and use recursion formula to calculate accumulation slope, specific algorithm is take 30s as the cycle, low-pass filtering resistance to be sampled, press two formulas calculated resistance slope S and accumulation slope T respectively below

S(k)＝(R(k-6ω)-R(k-18ω)+2×(R(k)-R(k-24ω)))/5，

T(k)＝(15/16)×T(k-1)+S(k)/4；

NB (k) is alumina blanking interval, aluminum electrolysis control system actual set blanking spacing value, represent with respect to benchmark blanking interval multiple, NB (k) is between 0.8-1.3, when NB (k)=1, represent with benchmark blanking interval blanking;

L (k) is aluminum yield, the quality of the metallic aluminium that expression aluminium cell extracts carry out out aluminium operation in the time of the k moment time, because L (k) sampling period with respect to front several model parameters is much bigger, so average aluminum yield using L (k) as the k moment, average aluminum yield using L (k) as the k moment in this method, units/kg,

V _s(k) be setting voltage, represent that aluminum electrolytic cell control system is the tank voltage that electrolytic tank is set at moment k, the V of unit;

C (k) is alumina concentration, represent to be dissolved into the mass percent of the aluminium oxide in electrolyte, general oxidation concentration is all between 0%-12%, C (k) value part of only peeking here, and 2% alumina concentration is expressed as C (k)=2;

Step 3: adopt following algorithm to be normalized sample set, attribute is zoomed between [0,1]:

$x_i=\frac{x_i}{\left|\right|x_i\left|\right|}$

$y_i=\frac{y_i}{12},$

In above formula || || get 2-norm;

Step 4: according to all electrolytic tank states of storing in control system historical data base, set up alumina concentration soft-sensing model corresponding under different electrolytic tank states, solving model, computation model parameter; Groove state is defined as electrolysis ability and the important indicator of moving steady in a long-term that electrolytic tank has, that decision is how many to electrolytic tank input electric weight, keep the reference frame of efficient output, need to drop into more energy for the electrolytic tank of groove state health, to improve output; On the contrary, just need to reduce the input of energy for the electrolytic tank of groove state morbid state, take corresponding control measure to make groove condition improvement, then increase the input of energy, along with the carrying out producing, the physical characteristics of electrolytic tank itself will change, and the electrolysis ability of groove also changes thereupon.

In the present embodiment, groove state definition comprise health, substantially health, inferior health, critical health, slight ill, moderate is ill, Very Ill-conditioned totally seven kinds of states, collect sample set for the training under different slots state, adopt least square method supporting vector machine (LSSVM) training to obtain the alumina concentration soft-sensing model that different slots state is corresponding, totally 7 kinds, model tormulation is as follows:

$\hat{y}\left(x\right)=f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{i}K(x,x_i)+b,$

In above formula, N is input number of samples, x _ibe i input sample, x is a certain input variable, α _i≠ 0 is corresponding input sample x _ifor support vector, α _ifor support vector machine coefficient,

for model pre-estimating output, K (x, x _i) be kernel function, the kernel function that this patent adopts is radial basis function (RBF), the parametric solution process of least square method supporting vector machine (LSSVM) adopts population (PSO) algorithm dynamic optimization, ask for and obtain algorithm parameter penalty coefficient c and kernel function spread factor σ, c and σ can be obtained to optimum model parameter α in substitution LSSVM again _iand b.

In step 4, to the electrolytic tank under a certain groove state, the algorithm of alumina concentration soft-sensing model parameter is as follows:

Step 1:

Model parameter initialization; (comprise initialization Population n, maximum iteration time k _max, study factor c ₁, study factor c ₂, penalty coefficient pace of change lower bound Vc _min, penalty coefficient pace of change upper bound Vc _max, kernel function spread factor changes lower bound V σ _min, kernel function spread factor change upper bound V σ _max, kernel function spread factor lower bound σ _min, kernel function spread factor upper bound σ _max, penalty coefficient lower bound c _min, penalty coefficient upper bound c _max, inertia weight ω);

Step 2:

The corresponding algorithm parameter c of all particles and σ in initialization colony, i.e. initialization particle population, initial population is shone upon to produce by Logistic and is obtained, and concrete grammar is: definition ξ ₁₁, ξ ₁₂for the random number in (0,1), as get ξ ₁₁, ξ ₁₂=0.25,0.5,0.75 and ξ ₁₁≠ ξ ₁₂.By ξ ₁₁, ξ ₁₂substitution Logistic mapping, that is:

ξ _i+1,1＝4ξ _i1(1-ξ _i1)

ξ _i+1,2＝4ξ _i2(1-ξ _i2),i＝1,...,m-1，

Above formula iteration is obtained to m group Chaos Variable for m-1 time

Further, Chaos Variable is mapped in feasible zone, even:

c _i＝c _min+ξ _i1(c _max-c _min)

σ _i＝σ _min+ξ _i2(σ _max-σ _min)，

Definition initial population

wherein x _i=(x _i1, x _i2)=(c _i, σ _i) be the position of i particle.

(illustrate: c is penalty coefficient, the attention degree of representative model to outlier in sample data, penalty coefficient c is larger represents that we more pay attention to the outlier in sample data, and model training error can dullness decline along with the increase of penalty factor c simultaneously, but when c increases to after certain value, this range of decrease degree can become very little, even go to zero, when the slack variable of all outlier with a timing, fixed c is larger, also larger to the loss of objective function, now just implying and be unwilling to abandon these outlier, the most extreme situation is that c is decided to be to infinity, like this as long as slightly a point peels off, the value of objective function becomes infinity at once, allow problem become without separating at once, thereby be degenerated to hard interval problem, kernel function spread factor σ: when spread factor σ is very little, contact between support vector loose (distance be less than between the support vector of σ just have contact), learns machine relative complex, and extensive Generalization Ability is poor, otherwise σ is too large, the impact between support vector is excessively strong, and regression model is difficult to the precision that reaches enough, easily produces and owes matching)

Calculate optimal adaptation value corresponding to all particles in colony, the individual particles optimal location in definition colony is:

m is Population.Definition colony optimal location: the optimal location P of colony _ginitial value be that of adaptive value minimum in personal best particle.By m P _iin the following formula of radial basis kernel function substitution:

Application least square method is obtained m group model parameter alpha _iand b, then by this m group model parameter substitution successively $\hat{y}\left(x\right)=f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{i}K(x,x_i)+b$ With $f=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(y_i-{\hat{y}}_i)}^2$ In, the f trying to achieve is m adaptive value corresponding to all particles, and the adaptive value of now trying to achieve is initialized as to the optimal adaptation value that now particle is corresponding, optimal adaptation value refers to the difference of two squares of estimating output and actual output of adaptive value under parameter current;

Step 3:

Search colony optimal adaptation value, i.e. optimal adaptation value corresponding to all particles in colony relatively, finding minimum optimal adaptation value is corresponding colony optimal adaptation value, and using algorithm parameter c corresponding particle and σ as optimum model parameter;

Step 4:

Value x and the pace of change v thereof of all particle algorithm parameter c and σ in renewal colony;

Judge and change in colony all particle algorithm parameter c and σ within the scope of feasible zone, that is:

If v _i1> vc _max, make v _i1=vc _max;

If v _i1< vc _min, make v _i1=vc _min;

If v _i2> v σ _max, make v _i2=v σ _max;

If v _i2< v σ _min, make v _i2=v σ _min;

If x _i1< c _min, make x _i1=c _min;

If x _i1> c _max, make x _i1=c _max;

If x _i2< σ _min, make x _i2=σ _min;

If x _i2> σ _max, make x _i2=σ _max, i=1 ..., Population

Least square method is calculated the model parameter α of each particle under its algorithm parameter c and σ _iand b;

Calculate the optimal adaptation value of all particles in colony, the optimal adaptation value value that all particles are corresponding is the minimum value between adaptive value and the optimal adaptation value calculating of last iteration that time iterative computation goes out;

Search colony optimal adaptation value, i.e. optimal adaptation value corresponding to all particles in colony relatively, finding minimum optimal adaptation value is corresponding colony optimal adaptation value, and using algorithm parameter c corresponding particle and σ as optimum model parameter;

Step 5:

Judge whether iterations reaches maximal value, if do not reached, returns to step 4, otherwise execution step 6;

Step 6:

The algorithm parameter c calculating after last iteration and σ are decided to be the algorithm parameter of aluminium oxide soft-sensing model;

Step 7:

Use least square method computational algorithm parameter c and model parameter α corresponding to σ _iand b;

Step 5: according to current electrolytic tank state and its corresponding alumina concentration soft-sensing model, test sample book, in conjunction with being input in soft-sensing model, is calculated to its corresponding alumina concentration discreet value utilize the algorithm of following formula by alumina concentration discreet value

renormalization is alumina concentration value, estimates out alumina concentration corresponding to current test sample book set,

${\hat{y}}_i=12{\hat{y}}_i.$

Alumina concentration concentration on-line detecting system is made up of hardware support platform and soft-sensor software, in the process control level of hardware platform (as shown in Figure 2), slot control machine collects a large amount of real-time process manufacturing parameter data, by CAN bus by these real time data write into Databasces, data are offered data processing server by database, in data processing server, complete the processing of data, the computation processes such as the soft measurement diagnosis of alumina concentration, and result of calculation is offered and in management and monitoring machine, carries out man-machine interaction demonstration, management and monitoring machine can offer decision-making management level by soft measurement result simultaneously, so that for the managerial personnel of higher level provide decision references.The process manufacturing parameter data that soft-sensor software collects according to hardware platform, realize flexible measurement method proposed by the invention, and target is exactly to detect the variation of alumina concentration, to realize the optimal control to production run.

Concrete implementation example:

Step 1: collecting sample

According to the record of production in certain aluminium manufacturer workshop, the electrolytic tank operation manufacturing parameter of No. 3326 electrolytic tank 2011-8-1300:00:00-2011-8-1401:00:00 of continuous acquisition electrolysis 3 series, collection period is 10min, totally 150 pairs of samples, wherein front 100 pairs for modeling, rear 50 pairs for prediction.The tank voltage curve of electrolytic tank within this period as shown in Figure 5, go out current groove state for slight morbid state according to cell parameters curve and in conjunction with micro-judgment, the aluminium oxide soft-sensing model that sample data now trains is this series electrolytic tank corresponding electrolytic tank under slight ill groove state.

Setting up sample set is

wherein

x _i＝{F(k),H(k),R(k),S(k),T(k),NB(k),L(k),V _s(k)}

y _i＝{C(k)}

X _i∈ R ⁸, y _i∈ R, k represents residing certain sampling instant of sample i.

Part sample data is as shown in table 1:

Table 1

Step 2: sample data pre-service

Owing to adopting the Euclidean distance of sample data to calculate in LSSVM algorithm, for avoiding the data domination lesser amt range data of larger amt scope, input data and output data acquisition are all normalized with following algorithm, attribute is zoomed to [0,1], between, the pretreated sample data of part is as shown in table 2 below:

Table 2

Step 3: determine best parameter

Selected radial basis kernel function

as the kernel function of aluminium oxide soft-sensing model, after selected kernel function, adopt alumina concentration soft-sensing model parameter to ask for algorithm computation model parameter.Before algorithm operation, initiation parameter arranges as shown in table 3 below:

Table 3

Asking for through alumina concentration soft-sensing model parameter the optimal parameter that algorithm determines is:

Penalty coefficient c=90; Kernel function spread factor σ=0.33.

The alumina concentration soft-sensing model coefficient of this series electrolytic tank that training obtains correspondence under slight ill groove state is as shown in table 4 below:

Table 4

Step 5: model pre-estimating meter

Be input to alumina concentration soft-sensing model for the sample data of predicting by latter 50 pairs, calculate its corresponding alumina concentration discreet value

Step 6: discreet value renormalization

Utilization algorithm is below by alumina concentration discreet value

renormalization is alumina concentration value, and with actual alumina concentration value y _icompare, the result in table 5 represents that this alumina concentration soft-sensing model can estimate alumina concentration preferably.

${\hat{y}}_i=12{\hat{y}}_i$

Table 5

Finally explanation is, above embodiment is only unrestricted in order to technical scheme of the present invention to be described, although the present invention is had been described in detail with reference to preferred embodiment, those of ordinary skill in the art is to be understood that, can modify or be equal to replacement technical scheme of the present invention, and not departing from aim and the scope of the technical program, it all should be encompassed in the middle of claim scope of the present invention.

Claims (5)

1. the flexible measurement method of alumina concentration in aluminium electrolysis process electrolytic tank, is characterized in that: described flexible measurement method comprises the following steps:

Step 1: gatherer process manufacturing parameter data, comprise that the process that gathers tank voltage, potline current, reinforced interval (NB), aluminum yield produces parameter, and set up the historical data base of storage parameters numerical value by hardware device;

Step 2: set one and read time parameter and read time period parameter, read time parameter every one section, just read the data before time period parameter that read in this moment of historical data base, process manufacturing parameter is carried out to data processing, set up training sample set and test sample book set, the form of sample set is { x _i, y _i, wherein

X _i={ F (k), H (k), R (k), S (k), T (k), NB (k), L (k), V _s(k) }, y _i={ C (k) }, x _i∈ R ⁸, y _i∈ R, wherein, in formula, the implication of each parameter is as follows:

K represents residing certain sampling instant of sample i, moment k-moment (k-1)=read time parameter;

F (k) is cell voltage fluctuation number of times, reads time period parameter with interior all tank voltage data-signal V to what collect _kaccording to eight layers of DB8 wavelet decomposition signals to for tree construction, then the signal of (8,0) frequency range is reconstructed and obtains reading the tank voltage signal V of time period parameter with interior (8,0) frequency range _80k, statistics reads the tank voltage signal wave crest trough number of time period parameter with interior (8,0) frequency range, and a Wave crest and wave trough is designated as 1 secondary undulation, and crest or trough are 0.5 secondary undulation;

H (k) is high-frequency energy number percent, reads time period parameter and is designated as V with interior tank voltage collecting _k, calculate tank voltage signal gross energy

calculate (8,0) frequency range tank voltage signal energy high-frequency energy number percent $H\left(k\right)=\frac{E_h}{E}=\frac{E-E_{80}}{E};$

R (k) is cell resistance, processes the tank voltage V that obtains (8,0) frequency range by the k moment _80kpotline current I (k) divided by this moment obtains, and the mode of utilizing following smoothing algorithm is deleted the exceptional value in the cell resistance sequential value of sampling rate, wherein, e is the deviation of cell resistance value and smoothing computation value, ω represents shift time, ω=5s, R (k-ω) is the cell resistance value of 5s before the k moment, R (k-2 ω) is the cell resistance value of 10s before the k moment.

R(k)＝0.3×R(k-ω)-0.1×R(k-2ω)+0.2×(e(k)+2×e(k-ω)+e(k-2ω))；

S (k) is cell resistance slope, and being defined as the increment that reads recently the filter resistance in time parameter is rate of change; T (k) is accumulation slope, be defined as the cumulative increment that reads recently the filter resistance in time parameter in the recent period, use difference filter slope calculations, and use recursion formula to calculate accumulation slope, specific algorithm is take 30s as the cycle, low-pass filtering resistance to be sampled, press two formulas calculated resistance slope S and accumulation slope T respectively below

S(k)＝(R(k-6ω)-R(k-18ω)+2×(R(k)-R(k-24ω)))/5，

T(k)＝(15/16)×T(k-1)+S(k)/4；

NB (k) is alumina blanking interval, aluminum electrolysis control system actual set blanking spacing value, represent with respect to benchmark blanking interval multiple, NB (k) is between 0.8-1.3, when NB (k)=1, represent with benchmark blanking interval blanking;

L (k) is aluminum yield, the quality of the metallic aluminium that expression aluminium cell extracts carry out out aluminium operation in the time of the k moment time, and the average aluminum yield using L (k) as the k moment in this method, units/kg,

V _s(k) be setting voltage, represent that aluminum electrolytic cell control system is the tank voltage that electrolytic tank is set at moment k, the V of unit;

C (k) is alumina concentration, represents to be dissolved into the mass percent of the aluminium oxide in electrolyte;

Step 3: adopt following algorithm to be normalized sample set, attribute is zoomed between [0,1]:

$x_i=\frac{x_i}{\left|\right|x_i\left|\right|}$

$y_i=\frac{y_i}{12},$

In above formula || || get 2-norm;

Step 4: according to all electrolytic tank states of storing in control system historical data base, set up alumina concentration soft-sensing model corresponding under different electrolytic tank states, solving model, computation model parameter;

Step 5: according to current electrolytic tank state and its corresponding alumina concentration soft-sensing model, test sample book, in conjunction with being input in soft-sensing model, is calculated to its corresponding alumina concentration discreet value

utilize the algorithm of following formula by alumina concentration discreet value renormalization is alumina concentration value, estimates out alumina concentration corresponding to current test sample book set,

${\hat{y}}_i=12{\hat{y}}_i.$

2. the flexible measurement method of alumina concentration in aluminium electrolysis process electrolytic tank according to claim 1, it is characterized in that: in step 4, groove state comprises health, ill, Very Ill-conditioned totally seven kinds of states of health, inferior health, critical health, slight morbid state, moderate substantially, collect sample set for the training under different slots state, adopt least square method supporting vector machine (LSSVM) training to obtain the alumina concentration soft-sensing model that different slots state is corresponding, totally 7 kinds, the optimal function of employing is as follows:

$\min J=\frac{1}{2}{\mathrm{\&omega;}}^T\mathrm{\&omega;}+c{R}_{\mathrm{emp}},$

In above formula, ω ^t, ω is used for controlling the complexity of model, c is error penalty coefficient, representative function smoothness and permissible error are greater than the compromise between the numerical value of ε; R _empfor empiric risk, about the insensitive loss function of ε; Least square method supporting vector machine selects quadratic loss function as optimization problem loss function, and quadratic loss function is expressed as error ξ _iquadratic sum, it is used for measuring special parameter and selects the loss bringing; Definition kernel function

for the inner product operation of high-dimensional feature space, K (x _i, x _j) meet the symmetric function of Mercer condition, optimization problem is converted into to solve linear model as follows:

$\hat{y}\left(x\right)=f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{i}K(x,x_i)+b,$

In above formula, N is input number of samples, x _ibe i input sample, x is a certain input variable, α _i≠ 0 is corresponding input sample x _ifor support vector, α _ifor support vector machine coefficient,

for model pre-estimating output, K (x, x _i) be kernel function.

3. the flexible measurement method of alumina concentration in aluminium electrolysis process electrolytic tank according to claim 2, is characterized in that: kernel function is radial basis function (RBF), is expressed as

σ is kernel function spread factor, and being needs definite algorithm parameter.

4. the flexible measurement method of alumina concentration in aluminium electrolysis process electrolytic tank according to claim 1, is characterized in that: in step 4, to the electrolytic tank under a certain groove state, the algorithm of alumina concentration soft-sensing model parameter is as follows:

Step 1:

Model parameter initialization; Comprise initialization Population n, maximum iteration time k _max, study factor c ₁, study factor c ₂, penalty coefficient pace of change lower bound Vc _min, penalty coefficient pace of change upper bound Vc _max, kernel function spread factor changes lower bound V σ _min, kernel function spread factor change upper bound V σ _max, kernel function spread factor lower bound σ _min, kernel function spread factor upper bound σ _max, penalty coefficient lower bound c _min, penalty coefficient upper bound c _max, inertia weight ω;

Step 2: the corresponding algorithm parameter c of all particles and σ in initialization colony, i.e. initialization particle population, initial population is shone upon to produce by Logistic and is obtained, and concrete grammar is: definition ξ ₁₁, ξ ₁₂for the random number in (0,1), and ξ ₁₁≠ ξ ₁₂, by ξ ₁₁, ξ ₁₂substitution Logistic mapping, that is:

ξ _i+1,1＝4ξ _i1(1-ξ _i1)

ξ _i+1,2＝4ξ _i2(1-ξ _i2),i＝1,...,m-1，

Above formula iteration is obtained to m group Chaos Variable for m-1 time

Again Chaos Variable is mapped in feasible zone, even:

c _i＝c _min+ξ _i1(c _max-c _min)

σ _i＝σ _min+ξ _i2(σ _max-σ _min)，，

Definition initial population

wherein x _i=(x _i1, x _i2)=(c _i, σ _i) be the position of i particle;

Calculate optimal adaptation value corresponding to all particles in colony, the individual particles optimal location in definition colony is:

m is Population; Definition colony optimal location: the optimal location P of colony _ginitial value be that of adaptive value minimum in personal best particle; By m P _iin the following formula of radial basis kernel function substitution:

Application least square method is obtained m group model parameter alpha _iand b, then by this m group model parameter substitution successively $\hat{y}\left(x\right)=f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i\mathrm{Kx}(x_i+b)$ With $f=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(y_i-{\hat{y}}_i)}^2$ In, the f trying to achieve is m adaptive value corresponding to all particles, the adaptive value of now trying to achieve is initialized as to the optimal adaptation value that now particle is corresponding, and the optimal adaptation value value that all particles are corresponding is the minimum value between adaptive value and the optimal adaptation value calculating of last iteration that time iterative computation goes out;

Step 3:

Search colony optimal adaptation value, i.e. optimal adaptation value corresponding to all particles in colony relatively, finding minimum optimal adaptation value is corresponding colony optimal adaptation value, and using algorithm parameter c corresponding particle and σ as optimum model parameter;

Step 4:

Value x and the pace of change v thereof of all particle algorithm parameter c and σ in renewal colony;

Judge and change in colony all particle algorithm parameter c and σ within the scope of feasible zone, that is:

If v _i1> vc _max, make v _i1=vc _max;

If v _i1< vc _min, make v _i1=vc _min;

If v _i2> v σ _max, make v _i2=v σ _max;

If v _i2< v σ _min, make v _i2=v σ _min;

If x _i1< c _min, make x _i1=c _min;

If x _i1> c _max, make x _i1=c _max;

If x _i2< σ _min, make x _i2=σ _min;

If x _i2> σ _max, make x _i2=σ _max, i=1 ..., Population,

Utilize least square method to calculate the model parameter α of each particle under its algorithm parameter c and σ _iand b;

Calculate optimal adaptation value corresponding to all particles in colony, the optimal adaptation value value that all particles are corresponding is the minimum value between adaptive value and the optimal adaptation value calculating of last iteration that time iterative computation goes out;

Search colony optimal adaptation value, i.e. optimal adaptation value corresponding to all particles in colony relatively, finding minimum optimal adaptation value is corresponding colony optimal adaptation value, and using algorithm parameter c corresponding particle and σ as optimum model parameter;

Step 5:

Judge whether iterations reaches maximal value, if do not reached, returns to step 4, otherwise execution step 6;

Step 6:

The algorithm parameter c calculating after last iteration and σ are decided to be the algorithm parameter of aluminium oxide soft-sensing model;

Step 7:

Use least square method computational algorithm parameter c and model parameter α corresponding to σ _iand b;

5. the flexible measurement method of alumina concentration in aluminium electrolysis process electrolytic tank according to claim 1, is characterized in that: in step 1, reading time parameter is 10min, reading time period parameter is 2h.

Images