Data-driven FNN-ARX modeling method for tobacco shred drying and head drying process
阅读说明：本技术 基于数据驱动的烟草烘丝干头过程fnn-arx建模方法 (Data-driven FNN-ARX modeling method for tobacco shred drying and head drying process ) 是由 周锋 曹丽华 朱培栋 郭文明 于 2019-08-14 设计创作，主要内容包括：本发明公开了一种基于数据驱动的烟草烘丝干头过程FNN-ARX建模方法,针对HAUNI牌滚筒式薄板烘丝机烘丝干头过程,提出一种基于数据驱动的辨识建模及参数优化方法。该方法首先采用一种基于模糊神经网络来拟合自回归模型的函数型系数构造的FNN-ARX模型对烟草烘丝过程干头段进行建模,然后采用一种改进的变量投影算法对建立的FNN-ARX模型参数进行优化。本发明可有效的与变量投影算法相结合进行参数分离优化,大大提高了算法的优化过程；且设计的FNN-ARX模型具有伪线性模型的结构形式,可方便后续基于模型的先进控制算法的设计。(The invention discloses a data-driven FNN-ARX modeling method for a tobacco cut-tobacco drying head process, and provides a data-driven identification modeling and parameter optimization method for a tobacco drying head process of a HAUNI-brand roller-type sheet tobacco dryer. The method comprises the steps of firstly, modeling a dry end section of a tobacco shred drying process by adopting a FNN-ARX model constructed by fitting a function coefficient of an autoregressive model based on a fuzzy neural network, and then optimizing the parameters of the established FNN-ARX model by adopting an improved variable projection algorithm. The method can be effectively combined with a variable projection algorithm to carry out parameter separation optimization, thereby greatly improving the optimization process of the algorithm; and the designed FNN-ARX model has the structural form of a pseudo linear model, so that the design of a subsequent advanced control algorithm based on the model can be facilitated.)
1. A data-driven FNN-ARX modeling method for a tobacco shred drying head process is characterized by comprising the following steps:
1) Collecting the dry end of the cut tobacco, i.e. T1calendar of process variables over a period of timeHistory data;
2) Establishing a FNN-ARX model structure in a tobacco shred drying head process:
Wherein: y istThe moisture content of the outlet cut tobacco in the head drying process is shown,The water content of the cut tobacco at the inlet is shown,The flow rate of the tobacco shreds at the inlet is shown,The temperature of the drum is shown,indicating the opening degree, xi, of the moisture exhausting dampertRepresenting white gaussian noise;Is the state quantity of the FNN-ARX model; phi is a0(xt-1)、φy,1(xt-1)、φy,2(xt-1)、 Andare all about model state quantity xt-1the FNN network of (a); the parameters of the FNN-ARX model are divided into a linear parameter set thetaLAnd a non-linear parameter set thetaN；
3) Parameter theta of FNN-ARX model in cut tobacco drying processLAnd thetaNAnd optimizing to obtain an optimized FNN-ARX model of the tobacco shred drying process.
2. the FNN-ARX modeling method for tobacco cut-drying process based on data drive of claim 1, wherein the FNN network is as follows:
Wherein the fuzzy membership function And isAndTo be the center of the fuzzy membership function,AndIs the width of the fuzzy membership function.
3. The FNN-ARX modeling method based on data driven tobacco dry end process according to claim 1, wherein the FNN-ARX model is transformed into the following standard form: y ist＝μ(θN,xt-1)TθL+ξtWherein, in the step (A), And is
4. The data-driven-based FNN-ARX modeling method for tobacco dry-end processing according to claim 1, wherein the parameter θ of FNN-ARX model for dry-end processingLAnd thetaNthe specific implementation process for optimization comprises the following steps:
1) Defining an objective function of parameter optimization of a FNN-ARX model in a dry head process aswherein, r (theta)L,θN)＝(rn(θL,θN),rn-1(θL,θN),...,r3(θL,θN))TAnd r when t is 3t(θL,θN)＝yt-μ(θN,xt-1)TθL，The parameter optimization problem for the dry head process FNN-ARX model is expressed as
2) Setting the non-linearity parameter thetaNKnowing that the linear parameter of the FNN-ARX model to be optimized is represented by thetaL＝Φ(θN)+y, where Φ (θ)N)+is a matrix phi (theta)N) Moore-Penrose inverse of (g), and matrix phi (theta)N)＝(μ(θN,xn-1),μ(θN,xn-2),...μ(θN,x2))T，y＝(yn,yn-1,...y3)TThen the optimization problem is transformed into
3) The jacobian matrix form of the objective function is chosen as: j ═ D (Φ (θ)N))Φ(θN)-y, where Φ (θ)N)-Is phi (theta)N) Inverse of minus sign of D (phi (theta))N))＝[D1(Φ(θN)),D2(Φ(θN))...D8(Φ(θN))]，Dg(Φ(θN) Represents a matrix phi (theta)N) Partial derivatives of the g-th nonlinear parameter; obtaining a Jacobian matrix J and then obtaining a nonlinear parameter thetaNUpdate direction d ofk＝-((Jk)TJk)-1(Jk)TrkWherein JkDenotes the value of k at time J, rkIs the value of r at time k, and r ═ I- Φ (θ)N)Φ(θN)+) y; nonlinear parameter is updated toWherein eta is the search step length and is determined by adopting a mixed quadratic polynomial interpolation method; when the objective function R (theta) of the optimization is finishedN) When the subtraction deviation of the linear parameter set of the model and the value of the previous optimization is less than the optimization termination condition delta or the optimization times reaches the maximum iteration times mu, the whole optimization process is ended, and the linear parameter set of the model passes through thetaL＝Φ(θN)+y is obtained by calculation; wherein the non-linear parameter set θNIs designed to be at an initial value ofData set of water content of cut tobacco with alpha as inletaverage value of (a).
5. The FNN-ARX modeling method based on data-driven tobacco shred drying head process according to claim 4, wherein β is1＝0.2，β2＝0.8，β3＝0.2，β4＝0.8。
Technical Field
The invention relates to the technical field of tobacco shred processing in the tobacco industry, in particular to a data-drive-based modeling and parameter optimization method for a tobacco shred drying processing dry end process of a drum-type tobacco shred dryer.
Background
The cut tobacco drying processing of the cigarette is an important process in the cigarette production process, and mainly reduces the moisture content of the cut tobacco by baking the cut tobacco, so that the moisture content of the baked cut tobacco is uniform and consistent to meet the cut tobacco process standard of the brand cigarette. At present, the main equipment for realizing the cut tobacco drying process in the tobacco plant cut tobacco manufacturing workshop is a cut tobacco drying machine, which comprises: a drum-type cut-tobacco drier and an air-flow type cut-tobacco drier. The invention relates to a HAUNI-brand drum-type thin-plate cut tobacco dryer, which mainly uses steam as a heat source to heat a drum body so that the drum wall reaches high temperature, cut tobacco contacts with the high-temperature drum wall in a rotating drum body to generate heat exchange and curling, and moisture is evaporated and discharged out of the drum.
Over the past few decades, some mechanistic models have been built to simulate the cut-tobacco drying process of a cut-tobacco dryer. However, the tobacco shred drying process is an extremely complex physical and chemical process and has the characteristics of multivariable, large time lag, strong coupling and lack of important variables (lack of detection values of the moisture content of the tobacco shreds at the outlet or the inlet in the process of drying the head or the tail). Therefore, it is difficult to establish an accurate mechanism model of the tobacco cut-tobacco drying process. At present, the most common solution is to simplify the complex physicochemical process of the cut tobacco drying process, design the assumed conditions, and then establish a simplified mechanism model of each stage of the cut tobacco drying process. Such as: research on influence of cut tobacco drying process parameters on quality of cut tobacco after drying [ J ] Chinese tobacco science, 2007,13(6):24-29.) of Hujian army et al (Hujian army, Zhou Ji Heng, panda, etc..) researches a functional relation between cut tobacco drying process parameters and moisture content of outlet cut tobacco through an orthogonal test. Xin et al (Xin Y N, Zhang J W, Li B. drying kinetics of tobaco strands at differential properties and relative humidity [ J ]. Journal of Thermal Analysis and calibration, 2018,132(2):1347-1358.) developed the effect of different air temperatures and relative humidities on the drying kinetics of a wire dryer, and proposed an improved Arrhenius type diffusion coefficient equation. Even so, due to the limitations of the complexity and structure of the actual tobacco cut-tobacco production process, certain key parameters in the resulting simplified mechanistic model are still not determinable or difficult to determine by their physical structure. Especially, in the process of drying cut tobacco by a roller type cut tobacco dryer, due to the complex production process requirement, a mechanism model capable of effectively describing the process is more difficult to obtain. In the dry end section of the cut tobacco drying process of the HAUNI-brand drum-type thin plate cut tobacco drying machine, the flow of the cut tobacco at the inlet is continuously increased, but the outlet does not have the detection value of the moisture content of the cut tobacco, so that the problems of too low moisture content of the cut tobacco in the dry end process, too much dry material and too high waste material rate are easily caused. At present, the moisture content of the tobacco shred drying process mainly adopts an input variable setting control method based on a simplified mechanism model. Generally, due to inaccuracy of the mechanism model, the control performance is not good, and a "dry head" phenomenon may be caused. Therefore, the method has important significance for modeling research on the dry end section of the tobacco shred drying process by adopting a complex system identification modeling and optimization technology.
Disclosure of Invention
the invention aims to solve the technical problem that aiming at the defects of the prior art, the FNN-ARX modeling method based on the data-driven tobacco drying head process is provided, the parameter separation optimization is effectively carried out by combining with a variable projection algorithm, and the optimization process of the algorithm is improved.
In order to solve the technical problems, the technical scheme adopted by the invention is as follows: a data-driven FNN-ARX modeling method for a tobacco shred drying head process comprises the following steps:
1) collecting the dry end of the cut tobacco, i.e. T1Historical data of process variables in a time period;
2) Establishing a FNN-ARX model structure in a tobacco shred drying head process:
Wherein: y istthe moisture content of the outlet cut tobacco in the head drying process is shown,the water content of the cut tobacco at the inlet is shown,The flow rate of the tobacco shreds at the inlet is shown,The temperature of the drum is shown,Indicating the opening degree, xi, of the moisture exhausting dampertRepresenting white gaussian noise;Is the state quantity of the FNN-ARX model; phi is a0(xt-1)、φy,1(xt-1)、φy,2(xt-1)、 AndAre all about model state quantity xt-1the FNN network of (a); the parameters of the FNN-ARX model are divided into a linear parameter set thetaLAnd a non-linear parameter set thetaN(ii) a The subscript t denotes time t;
3) parameter theta of FNN-ARX model in cut tobacco drying processLAnd thetaNAnd optimizing to obtain an optimized FNN-ARX model of the tobacco shred drying process.
The FNN network is as follows:
Wherein the fuzzy membership function And isAndto be the center of the fuzzy membership function,AndIs the width of the fuzzy membership function.
The FNN-ARX model was converted to the following standard form: y ist＝μ(θN,xt-1)TθL+ξtWherein, in the step (A), And isParameter theta of FNN-ARX model in cut tobacco drying processLAnd thetaNthe specific implementation process for optimization comprises the following steps:
1) Defining an objective function of parameter optimization of a FNN-ARX model in a dry head process asWherein, r (theta)L,θN)＝(rn(θL,θN),rn-1(θL,θN),...,r3(θL,θN))Tand r when t is 3t(θL,θN)＝yt-μ(θN,xt-1)TθL，The parameter optimization problem for the dry head process FNN-ARX model is expressed as
2) Setting the non-linearity parameter thetaNknowing that the linear parameter of the FNN-ARX model to be optimized is represented by thetaL＝Φ(θN)+y, where Φ (θ)N)+Is a matrix phi (theta)N) Moore-Penrose inverse of (g), and matrix phi (theta)N)＝(μ(θN,xn-1),μ(θN,xn-2),...μ(θN,x2))T，y＝(yn,yn-1,...y3)TThen the optimization problem is transformed into
3) The jacobian matrix form of the objective function is chosen as: j ═ D (Φ (θ)N))Φ(θN)-y, where Φ (θ)N)-is phi (theta)N) Inverse of minus sign of D (phi (theta))N))＝[D1(Φ(θN)),D2(Φ(θN))...D8(Φ(θN))]，Dg(Φ(θN) Represents a matrix phi (theta)N) Partial derivatives of the g-th nonlinear parameter; obtaining a Jacobian matrix J and then obtaining a nonlinear parameter thetaNUpdate direction d ofk＝-((Jk)TJk)-1(Jk)TrkWherein JkDenotes the value of k at time J, rkIs the value of r at time k, and r ═ I- Φ (θ)N)Φ(θN)+) y; nonlinear parameter is updated toWherein eta is the search step length and is determined by adopting a mixed quadratic polynomial interpolation method; when the objective function R (theta) of the optimization is finishedN) When the subtraction deviation of the linear parameter set of the model and the value of the previous optimization is less than the optimization termination condition delta or the optimization times reaches the maximum iteration times mu, the whole optimization process is ended, and the linear parameter set of the model passes through thetaL＝Φ(θN)+y is obtained by calculation; wherein the non-linear parameter set θNIs designed to be at an initial value ofdata set of water content of cut tobacco with alpha as inletaverage value of (a).
Compared with the prior art, the invention has the beneficial effects that: aiming at the problem that a mechanism model cannot be directly established in the tobacco drying head process of the drum-type tobacco dryer, the invention adopts a Fuzzy Neural Network (FNN) with stronger learning and generalization capability than a general neural network to approximate to a function coefficient of an autoregressive model to construct a FNN-ARX model in the tobacco drying head process, and adopts an efficient variable projection algorithm to optimize the established FNN-ARX model parameters. The designed FNN-ARX model for the cut tobacco drying process has the characteristic of separable parameters, so that the parameter separation optimization can be effectively carried out by combining with a variable projection algorithm, and the optimization process of the algorithm is greatly improved; and the designed FNN-ARX model has the structural form of a pseudo linear model, so that the design of a subsequent advanced control algorithm based on the model can be facilitated.
Drawings
FIG. 1 is a schematic view of a process for drying tobacco shreds in a drum-type tobacco shred drying machine according to the present invention.
Detailed Description
The invention provides an identification modeling and parameter optimization method based on data driving. The method comprises the steps of firstly adopting a FNN-ARX model which is constructed based on a Fuzzy Neural Network (FNN) fitting auto-regression model with extrinsic input, modeling a dry end section of a tobacco shred drying process, and then adopting an improved variable projection algorithm to optimize parameters of the established FNN-ARX model.
The technological schematic diagram of the cut tobacco drying head process of the HAUNI-brand roller-type thin plate cut tobacco dryer is shown in figure 1. Firstly, inlet cut tobacco (generally with the water content of 14-20%) needing cut tobacco drying processing enters a blue rotary drum (with fixed frequency) through a conveyor belt, the cut tobacco contacts with a high-temperature drum wall in the rotary drum to generate heat exchange and curling, so that water is evaporated and is discharged out of the drum through a moisture exhaust air door, and the purpose of reducing the water content of the cut tobacco is achieved. As shown in figure 1, 1 is an inlet cut tobacco moisture content detection sensor with the model of LM-100, which is arranged at the inlet of a roller; 2 is an inlet cut tobacco flow detection sensor with model MF3000 installed at the inlet of the roller; 3 is a drum temperature detecting sensor with the model PT100 installed in the middle of the drum; 4 is an electric adjusting valve of a moisture exhaust air door with model number DN220, which is arranged at the upper end of the tail part of the roller; and 5, an outlet cut tobacco moisture content detection sensor which is installed at the outlet of the roller and has the model of LM-100.
in the tobacco drying process of the drum-type tobacco dryer shown in fig. 1, when the inlet tobacco shred flow is detected at the position 2, the drying process of the tobacco shred aimed by the invention is started, and when the water content of the outlet tobacco shred is detected at the position 5 to be higher than 12.5%, the drying process of the tobacco shred aimed by the invention is ended. Since the blue drum rotates at a fixed frequency, the duration T of the whole process of drying the cut tobacco1(in seconds) can be obtained by measurement. Aiming at the process of drying the cut tobacco, the FNN-ARX modeling method based on data driving specifically comprises the following steps:
Step 1: collecting tobacco cut-drying process (T)1Time period) historical data for each process variable.
After the tobacco shred drying process begins, T is collected at a sampling rate of 2 (seconds/piece)1Data set of water content of outlet cut tobacco in time period (y)1…ynData set of moisture content of inlet cut tobaccodata set of inlet cut tobacco flowData set of drum temperatureAnd moisture damper opening data setwherein Representing a rounding down operation.
Step 2: the FNN-ARX model structure of the tobacco shred drying head process shown in figure 1 is established as follows:
Wherein: y istthe moisture content of the outlet cut tobacco in the head drying process is shown,the water content of the cut tobacco at the inlet is shown,The flow rate of the tobacco shreds at the inlet is shown,the temperature of the drum is shown,Indicating the opening degree, xi, of the moisture exhausting dampertrepresenting white gaussian noise;Is the state quantity of the FNN-ARX model; phi is a0(xt-1)、φy,1(xt-1)、φy,2(xt-1)、 andare all about model state quantity xt-1the FNN network of (1) has the following specific structure:
fuzzy membership function in the above formula and is Andto be the center of the fuzzy membership function,AndIs the width of the fuzzy membership function. As described abovethe FNN-ARX model may be regarded as a special case of a parameter separable model and may be converted into the following standard form: y ist＝μ(θN,xt-1)TθL+ξtWherein, in the step (A), And is The parameters of the FNN-ARX model of the dry end process of the cut tobacco can be divided into a linear parameter set thetaLAnd a non-linear parameter set thetaN。
And step 3: parameter theta of FNN-ARX model in cut tobacco drying processLAnd thetaNAnd (6) optimizing.
Firstly, defining an objective function of parameter optimization of a FNN-ARX model in a dry head process as follows:
Wherein, r (theta)L,θN)＝(rn(θL,θN),rn-1(θL,θN),...,r3(θL,θN))Tand r when t is 3t(θL,θN)＝yt-μ(θN,xt-1)TθL，then pairthe parameter optimization problem of the dry head process FNN-ARX model can be expressed asAn improved variable projection algorithm is used to perform parameter optimization on the objective function.
first, assume a non-linearity parameter θNKnowing that, the linear parameter of the FNN-ARX model to be optimized can be represented as thetaL＝Φ(θN)+y, where Φ (θ)N)+Is a matrix phi (theta)N) Moore-Penrose inverse of (g), and matrix phi (theta)N)＝(μ(θN,xn-1),μ(θN,xn-2),...μ(θN,x2))T，y＝(yn,yn-1,...y3)T. The optimization problem is transformed into
To solve the problem, the present invention selects the jacobian matrix form of the objective function as: j ═ D (Φ (θ)N))Φ(θN)-y, where Φ (θ)N)-Is phi (theta)N) Inverse of minus sign of D (phi (theta))N))＝[D1(Φ(θN)),D2(Φ(θN))...D8(Φ(θN))]，Dg(Φ(θN) Represents a matrix phi (theta)N) Partial derivatives of the g-th non-linear parameter, e.g.after a Jacobian matrix J is obtained, a nonlinear parameter theta can be obtainedNupdate direction d ofk＝-((Jk)TJk)-1(Jk)TrkWherein JkDenotes the value of k at time J, rkIs the value of r at time k, and r ═ I- Φ (θ)N)Φ(θN)+) y. In order to satisfy the objective function V (theta) in the search processL,θN) Is decreasing, the next step of the invention is to update the non-linear parameter toWherein eta is the search step length and is determined by a mixed quadratic polynomial interpolation method. When the objective function R (theta) of the optimization is finishedN) When the subtraction deviation from the optimized value in the previous step is less than an optimization termination condition delta (a termination error set value) or the optimization times reaches a maximum iteration time M, the whole optimization process is ended, and at the moment, the linear parameter set of the model passes through thetaL＝Φ(θN)+and y is calculated. The nonlinear parameter set theta of the optimization methodNis designed to be at an initial value ofwherein alpha is the data set of the water content of the cut tobacco at the inlet collected in the step 1average value of, and β1＝0.2，β2＝0.8，β3＝0.2，β4＝0.8。
By adopting the modeling and parameter optimization strategies, the FNN-ARX model of the tobacco shred drying and head drying process can be finally obtained. Therefore, under the condition of not analyzing the complex physical and chemical relationship of the tobacco cut-tobacco drying process, the mathematical model of the process is established only by using the actually sampled field data, and the established FNN-ARX model has a structure of a pseudo linear model, and the structure is very suitable for the design of a subsequent advanced controller.
- 上一篇：一种医用注射器针头装配设备
- 下一篇：一种烟垛分层松散设备