CN103810396A - Fermentation process quality prediction method based on dicaryon multiway partial least squares of characteristic space - Google Patents

Abstract

The invention discloses a fermentation process quality prediction method based on dicaryon multiway partial least squares of characteristic space, belongs to both the field of biological fermentation and the field of information science, and particularly relates to a method for predicting quality variables which are difficultly acquired in time when online measurement is conducted in the fermentation process. The method includes the specific steps that training data are acquired, unfolded and standardized, characteristic extraction is conducted on the standardized measurement data and quality data in nuclear space, and the extracted characteristic space is used as the reference for high-dimensional projection; a partial least square model is built, real-time data are acquired to predict projections of the quality data in the high-dimensional space, and the quality data in the high-dimensional space are processed into real prediction values. The method reduces the degree of nonlinearity of the whole model, improves prediction accuracy of the model, and has significance in achieving quality prediction, control and optimal control of the fermentation process.

Description

A kind of sweat qualitative forecasting method of the multidirectional offset minimum binary of double-core based on feature space

Technical field

The invention belongs to biofermentation field, belong to again information science field, be specifically related to apply during the fermentation forecast model that the multidirectional offset minimum binary of double-core based on feature space the sets up sweat quality variable for predicting that on-line measurement is difficult to obtain in time.

Background technology

Sweat is to adopt modern biotechnology, the activity in production that utilizes useful microorganism to carry out, and the microorganism of application or microbial product can produce a large amount of economic benefits.Sweat is the important component part of biotechnology, in the fields such as pharmacy, medical treatment, food, chemical industry, environment, has very important society and economic worth." food industry " 12 " development plan " that National Development and Reform Committee, Ministry of Industry and Information issue write, and by 2015, the fermentation industry gross output value will reach more than 4,600 hundred million yuan, and average growth rate per annum reaches more than 15%.Biotechnology, medicine are also confirmed as the major fields of development in " medical industry " 12 " development plan ".

In sweat, process belongs to batch process greatly, and modern industry batch process is tended to miniaturization, high-speeding, and high added valueization development, the data that can measure in this process are also more and more.But some data, for example cell concentration, production concentrations etc., have two features, and one is that these data and product quality have very strong correlativity, even directly affects the qualification rate of product; Second is these data and is not easy on-line measurement, often traditional measuring method relative data that other easily obtains in measuring these data have certain hysteresis, can not react timely production status, this can affect supervision and the operation of operating personnel to whole production run, even likely incurs loss through delay crucial controller meeting because can not judge in time.For this problem, research emphasis both domestic and external is divided into two large classes, and a class is to analyze qualitatively, such as expert system, although this mode can be in conjunction with knowhow in the past, predicts the outcome often fuzzyyer; Another kind of is to determine quantitative analysis, such as mechanism model modeling, although this mode can explain more accurately production run and data are predicted, due to production run complicated mechanism, in earlier stage when Analysis on Mechanism difficulty, and practical application, still need to adjust quantity of parameters.Analytical approach based on data-driven belongs to the one of quantitative analysis tech, and wherein offset minimum binary is less for amount of training data requirement, and computational complexity is lower, explains that effect is better widely used.What but it was processed is linear problem, has very important shortcoming for the higher object handles of this nonlinear degree of production run.

In order to solve the nonlinear problem of process, core skill is applied in partial least squares algorithm.Core skill can process for producing data nonlinear problem, by by the data projection collecting to the nonlinearity that reduces data itself in higher dimensional space, and then can in nuclear space, can apply partial least squares algorithm to data analysis, this method can improve overall precision of prediction more significantly, but still has problems.In the time of offset minimum binary modeling, data are divided into two groups, and one group is measurement data, can measure in real time, and another group is for qualitative data, these data can not or very difficult Real-time Collection.So although can obtain measurement data and the qualitative data in identical moment in production run, often there is certain hysteresis in obtaining of qualitative data result.Partial least squares algorithm is exactly the relation of calculating between these two groups of data, finds the method with measurement data calculated mass data, plays the effect of real-time estimate qualitative data.Core skill projects to measurement data in higher dimensional space, data after projection still can be calculated, but qualitative data is projected in higher dimensional space, reducing the problem of the nonlinearity of qualitative data is not studied, and how to projecting to qualitative data in higher dimensional space, to calculate real qualitative data be also problem.

Summary of the invention

The present invention is directed to the nonlinear problem of batch process, propose a kind of double-core based on feature space and process multidirectional partial least squares algorithm, project to the qualitative data of nuclear space with nuclear space vector forecasting.Solve the nonlinear problem that traditional core skill is difficult to considering quality data Y, improved the precision of prediction of model.

The present invention has adopted following technical scheme and performing step:

1, a sweat qualitative forecasting method for the multidirectional offset minimum binary of double-core based on feature space, is characterized in that comprising following steps:

Step (1): obtain training data:

The quality variable that gathers in sweat measurand that can on-line measurement and be difficult to on-line measurement forms training data, and a concrete data sampling representation is as follows:

X=(x ₁,x ₂,...,x _nx) (1)

Y=(y ₁,y ₂,...,y _ny) (2)

Wherein, x ₁, x ₂..., x _nxrepresent nx measurand can measuring in real time online in sweat, the different indexs such as such as stir speed (S.S.), temperature, ventilation, i.e. measurement point; In like manner, y ₁, y ₂..., y _nyrepresent ny quality variable can not measuring in real time online in sweat, form data X through repeatedly sampling _{p, nx, r}and Y _{p, ny, r}, wherein p represents a batch algebraically, in order to distinguish batch process different batch, i.e. batch number; R represents hits.For example a batch process is prepared certain product, and this process need just can complete production task through 400 hours, and whole process need is measured speed of agitator, substrate temperature and new 3 variablees of wind air quantity, measurement point nx=3; Output product has cell concentration, 2 variablees of production concentration, measurement point ny=2; Whole production run needs measure once for 1 hour, hits r=400/1=400; This batch process has as above been carried out 10 times, batch number p=10.Can form three-dimensional matrice data: X through the data of measuring _3Dand Y _3D, matrix specification is batch count × hits of number × measurement;

Step (2): by X _3Dand Y _3Dcarry out data expansion and standardization according to AT method, the variance that record is obtained and average, data after treatment are X _2Dand Y _2D;

Step (3): respectively to X _2Dand Y _2Dcarry out the feature extraction in nuclear space, obtain successively feature base S _xand S _y;

Step (4): computation and measurement data X _2Dwith qualitative data Y _2Dprojection in nuclear space:

Choose gaussian kernel function, utilize core skill by X _2D, Y _2Dwith S _xand S _yfor base projects to respectively in nuclear space, obtain K _x=K (X _2D, S _x) and K _y=K (Y _2D, S _y);

Step (5): utilize the data Kx and the K that obtain _yset up offset minimum binary forecast model:

The partial least squares algorithm utilizing is set up data K _xto K _ymapping, and obtain PLS mapping coefficient B, concrete manifestation form is:

K _Y=K _XB (6)

Step (6): on-line measurement and deal with data, concrete steps are:

Step (6.1): the measurement data X that online acquisition is new _new, its data layout is identical with the X in formula (1);

Step (6.2): by the new data X collecting _newaT disposal route according to the modeling moment is processed and standardization;

Step (6.3): utilize core skill by X _newwith S _xfor base projects in nuclear space, obtain K _xnew=K (X _new, S _x);

Step (7): by K _xnewbe input in the PLS model having established, utilize formula (6) to obtain K _ynew, i.e. K _ynew=K _xnewb;

Step (8): by K _ynewcarry out inverse mapping reconstruct, and contrary standardization, predicted value Y obtained _new, concrete steps are:

Step (8.1): utilize S _yform new matrix data, wherein S _ydata layout be:

Wherein, V _mnrepresent S _yin vector matrix, a total m row vector, and the element number of each row vector is n.Wherein m represents the feature base S through feature extraction _ythe number of middle row vector, n=ny;

Through the K calculating _ynewdata layout be:

K _Ynew=(Ky ₁,Ky ₂,...,Ky _m) (8)

Order:

L=K _Ynew (9)

For each the element L in L _i(i is natural number and i ∈ [1, m]), L _i=Ky _i;

Order again:

$\mathrm{Lmi}=\left[\begin{array}{c}{L}_1-L_m\\ L_2-L_m\\ .\\ .\\ .\\ L_{m-1}-L_m\end{array}\right]---\left(11\right)$

Step (8.2): computing formula

Y _new=-0.5(Vmi ^TVmi) ^-1Vmi ^T（Lmi-VmisI） (13)

Wherein V _mi ^tfor the transposition of Vmi, I is that all elements is 1 n dimensional vector;

Step (8.3): according to the reverse Y that calculates of the data processing method of original AT method _newactual prediction data Y before standardization _pre;

Step (9): if production run does not finish, repeating step (6) is to (8).

Beneficial effect

Compared with other prior aries, the present invention has realized the processing of qualitative data nonlinear characteristic, qualitative data is projected to higher dimensional space by the core skill based on feature space, in forecasting process, the qualitative data reduction of nuclear space will be projected to, reach reduction block mold nonlinearity, improved the ability of the precision of prediction of model.

Accompanying drawing explanation

Fig. 1 the inventive method block diagram;

Fig. 2 the inventive method process flow diagram;

Fig. 3 is the AT method of deploying schematic diagram of batch process data;

Fig. 4 is the present invention (FS Dual-KMPLS), multidirectional offset minimum binary (MPLS), the multidirectional offset minimum binary of core (FSKMPLS) based on feature extraction, root-mean-square error (RMSE) the track schematic diagram of the multidirectional offset minimum binary of core (KMPLS), and wherein horizontal ordinate is batch number that ordinate is corresponding;

Fig. 5 is the prediction effect figure of 1 crowd of 4 kinds of methods;

Embodiment

Provide following embodiment in conjunction with content of the present invention:

In bio-pharmaceuticals industry, medical protein utilizes the colibacillary sweat of transgenosis to be prepared mostly, and this preparation process often adopts sequencing batch type fed-batch fermentation mode.Escherichia coli fermentation process is a series of very complicated biochemical reaction process, and this process exists a large amount of Multivariable Couplings, sample information nonlinearity, time variation and uncertainty etc.In this class process, recombination bacillus coli is prepared proleulzin production run and is mainly comprised three phases, is respectively without the bacterial classification cultivation stage of feed supplement, bacterial classification Fast Growth stage and the bacterial classification induced product synthesis phase of feed supplement.Wherein the first stage continues 6h left and right, is that bacterial classification shaking table is cultivated the postvaccinal laundering period; Subordinate phase continues 3 to 4h left and right, and the sugared concentration in this stage fermentation tank need keep higher level, requires to need to continue to supplement glycogen, to meet colibacillary growth demand for reaching this; In phase III, the concentration of sugar need remain on medium level, is beneficial to the expression of foreign protein proleulzin in Escherichia coli.Experiment is carried out in Yizhuang Development District, Beijing pharmaceutical factory, and sweat adopts Sartorius BIOSTAT BDL50L fermentation tank.

In the middle of process of the test, the Escherichia coli fermentation cycle is 19h left and right, sampling interval is 0.5h, choose 8 main process variable and 1 quality variable, process variable is: PH, oxyty, tank pressure, temperature, speed of agitator, benefit sugar amount, benefit nitrogen amount, throughput, quality variable is: OD value, this value has very strong representativeness to overall ferment effect, and can not directly pass through sensor measurement.28 batches of normal data are collected in experiment, and these data are combined into three-dimensional data (variable number × 35,28 batches of number × 9 hits) after arranging.

It is as follows by the concrete implementation step of MATLAB program realization on computers that double-core based on feature space is processed multidirectional partial least squares algorithm:

Step (1): obtain training data:

The quality variable that gathers in sweat measurand that can on-line measurement and be difficult to on-line measurement forms training data, and a concrete data sampling representation is as follows:

X=(x ₁,x ₂,...,x _nx) (1)

Y=(y ₁,y ₂,...,y _ny) (2)

Wherein, x ₁, x ₂..., x _nxrepresent nx measurand can measuring in real time online in sweat, the different indexs such as such as stir speed (S.S.), temperature, ventilation, i.e. measurement point; In like manner, y ₁, y ₂..., y _nyrepresent ny quality variable can not measuring in real time online in sweat, form data X through repeatedly sampling _{p, nx, r}and Y _{p, ny, r}, wherein p represents a batch algebraically, in order to distinguish batch process different batch, i.e. batch number; R represents hits.For example a batch process is prepared certain product, and this process need just can complete production task through 400 hours, and whole process need is measured speed of agitator, substrate temperature and new 3 variablees of wind air quantity, measurement point nx=3; Output product has cell concentration, 2 variablees of production concentration, measurement point ny=2; Whole production run needs measure once for 1 hour, hits r=400/1=400; This batch process has as above been carried out 10 times, batch number p=10.Can form three-dimensional matrice data: X through the data of measuring _3Dand Y _3D, matrix specification is batch count × hits of number × measurement; Nx=8 in this process, ny=1, m=28, o=35.

Step (2): by X _3Dand Y _3Dcarry out data expansion and standardization according to AT method, the variance that record is obtained and average, data after treatment are X _2Dand Y _2D;

Step (3): respectively to X _2Dand Y _2Dcarry out the feature extraction in nuclear space, obtain successively feature base S _xand S _y; S _xand S _ythe middle application characteristic that comprises respectively extracts skill from a plurality of proper vectors that extracting data goes out separately, the S obtaining in this process _xthere are 55 vectors, S _yin have 3 vectors;

Step (4): computation and measurement data X _2Dwith qualitative data Y _2Dprojection in nuclear space:

Choose gaussian kernel function, utilize core skill by X _2D, Y _2Dwith S _xand S _yfor base projects to respectively in nuclear space, obtain K _x=K (X _2D, S _x) and K _y=K (Y _2D, S _y);

Step (5): utilize the data Kx and the K that obtain _yset up offset minimum binary forecast model:

The partial least squares algorithm utilizing is set up data K _xto K _ymapping, and obtain PLS mapping coefficient B, concrete manifestation form is:

K _Y=K _XB (6)

Step (6): on-line measurement and deal with data, concrete steps are:

Step (6.1): the measurement data X that online acquisition is new _new, its data layout is identical with the X in formula (1);

Step (6.2): by the new data X collecting _newaT disposal route according to the modeling moment is processed and standardization;

Step (6.3): utilize core skill by X _newwith S _xfor base projects in nuclear space, obtain K _xnew=K (X _new, S _x);

Step (7): by K _xnewbe input in the PLS model having established, utilize formula (6) to obtain K _ynew, i.e. K _ynew=K _xnewb;

Step (8): by K _ynewcarry out inverse mapping reconstruct, and contrary standardization, predicted value Y obtained _new, concrete steps are:

Step (8.1): utilize S _yform new matrix data, wherein S _ydata layout be:

Wherein, V _mnrepresent S _yin vector matrix, a total m row vector, and the element number of each row vector is n.Wherein m=3, represents the feature base S through feature extraction _ythe number of middle row vector, n=ny=1;

Through the K calculating _ynewdata layout be:

K _Ynew=(Ky ₁,Ky ₂,...,Ky _m) (8)

Order:

L=K _Ynew (9)

For each the element L in L _i(i is natural number and i ∈ [1, m]), L _i=Ky _i;

Order again:

$\mathrm{Lmi}=\left[\begin{array}{c}{L}_1-L_m\\ L_2-L_m\\ .\\ .\\ .\\ L_{m-1}-L_m\end{array}\right]---\left(11\right)$

Step (8.2): computing formula

Y _new=-0.5(Vmi ^TVmi) ^-1Vmi ^T（Lmi-VmisI） (13)

Wherein V _mi ^tfor the transposition of Vmi, I is that all elements is 1 n dimensional vector;

Step (8.3): according to the reverse Y that calculates of the data processing method of original AT method _newactual prediction data Y before standardization _pre;

Step (9): if production run does not finish, repeating step (6) is to (8).

Realize by MATLAB program on computers according to above step, model prediction root-mean-square error and the predicted time of setting up four kinds of methods are as shown in table 1, i.e. the present invention (FS Dual-KMPLS), multidirectional offset minimum binary (MPLS), the multidirectional offset minimum binary of core (FSKMPLS) based on feature extraction, the multidirectional offset minimum binary of core (KMPLS):

Table 1

Claims (1)

1. a sweat qualitative forecasting method for the multidirectional offset minimum binary of double-core based on feature space, is characterized in that comprising following steps:

Step (1): obtain training data:

The quality variable that gathers in sweat measurand that can on-line measurement and be difficult to on-line measurement forms training data, and a concrete data sampling representation is as follows:

X=(x ₁,x ₂,...,x _nx) (1)

Y=(y ₁,y ₂,...,y _ny) (2)

Wherein, x ₁, x ₂..., x _nxrepresent nx measurand, i.e. measurement point in sweat, can measuring in real time online; In like manner, y ₁, y ₂..., y _nyrepresent ny quality variable can not measuring in real time online in sweat, form data X through repeatedly sampling _{p, nx, r}and Y _{p, ny, r}, wherein p represents a batch algebraically, in order to distinguish batch process different batch, i.e. batch number; R represents hits; Data composition three-dimensional matrice data through measuring: X _3Dand Y _3D, matrix specification is batch count × hits of number × measurement;

Step (2): by X _3Dand Y _3Dcarry out data expansion and standardization according to AT method, the variance that record is obtained and average, data after treatment are X _2Dand Y _2D;

Step (3): respectively to X _2Dand Y _2Dcarry out the feature extraction in nuclear space, obtain successively feature base S _xand S _y;

Step (4): computation and measurement data X _2Dwith qualitative data Y _2Dprojection in nuclear space:

Choose gaussian kernel function, utilize core skill by X _2D, Y _2Dwith S _xand S _yfor base projects to respectively in nuclear space, obtain K _x=K (X _2D, S _x) and K _y=K (Y _2D, S _y);

Step (5): utilize the data Kx and the K that obtain _yset up offset minimum binary forecast model:

The partial least squares algorithm utilizing is set up data K _xto K _ymapping, and obtain PLS mapping coefficient B, concrete manifestation form is:

K _Y=K _XB (6)

Step (6): on-line measurement and deal with data, concrete steps are:

Step (6.1): the measurement data X that online acquisition is new _new, its data layout is identical with the X in formula (1);

Step (6.2): by the new data X collecting _newaT disposal route according to the modeling moment is processed and standardization;

Step (6.3): utilize core skill by X _newwith S _xfor base projects in nuclear space, obtain K _xnew=K (X _new, S _x);

Step (7): by K _xnewbe input in the PLS model having established, utilize formula (6) to obtain K _ynew, i.e. K _ynew=K _xnewb;

Step (8): by K _ynewcarry out inverse mapping reconstruct, and contrary standardization, predicted value Y obtained _new, concrete steps are:

Step (8.1): utilize S _yform new matrix data, wherein S _ydata layout be:

Wherein, V _mnrepresent S _yin vector matrix, a total m row vector, and the element number of each row vector is n; Wherein m represents the feature base S through feature extraction _ythe number of middle row vector, n=ny;

Through the K calculating _ynewdata layout be:

K _Ynew=(Ky ₁,Ky ₂,...,Ky _m) (8)

Order:

L=K _Ynew (9)

For each the element L in L _i, i is natural number and i ∈ [1, m], L _i=Ky _i;

Order again:

$\mathrm{Lmi}=\left[\begin{array}{c}{L}_1-L_m\\ L_2-L_m\\ .\\ .\\ .\\ L_{m-1}-L_m\end{array}\right]---\left(11\right)$

Step (8.2): computing formula

Y _new=-0.5(Vmi ^TVmi) ^-1Vmi ^T（Lmi-VmisI） (13)

Wherein V _mi ^tfor the transposition of Vmi, I is that all elements is 1 n dimensional vector;

Step (8.3): according to the reverse Y that calculates of the data processing method of AT method _newactual prediction data Y before standardization _pre;

Step (9): if production run does not finish, repeating step (6) is to (8).

Images