WO2011109922A1

WO2011109922A1 - Near-infrared spectrum characteristic subinterval selection method based on simulated annealing-genetic algorithm

Info

Publication number: WO2011109922A1
Application number: PCT/CN2010/000530
Authority: WO
Inventors: 邹小波; 石吉勇; 赵杰文; 殷晓平; 陈正伟; 黄星奕; 蔡健荣; 陈全胜
Original assignee: 江苏大学
Priority date: 2010-03-12
Filing date: 2010-04-19
Publication date: 2011-09-15
Also published as: CN101832909B; CN101832909A

Abstract

A near-infrared spectrum characteristic subinterval selection method based on a simulated annealing-genetic algorithm comprises pre-treating a near-infrared spectrum first, and then dividing the pretreated near-infrared spectrum into subintervals dynamically, introducing the Metropolis rule of the simulated annealing algorithm into the gene exchange operator and the gene selection operator of the genetic algorithm, selecting an optimal characteristic subinterval using the simulated annealing-genetic algorithm, finally judging an optimal subinterval division manner and an optimal characteristic subinterval combination, establishing a partial least squares (PLS) model for the selected optimal characteristic subinterval.

Description

Sub-interval selection method for near-infrared spectroscopy based on simulated annealing-genetic algorithm

The invention relates to a method for selecting a sub-interval of sub-infrared spectroscopy for analyzing the quality of agricultural products and foods, and particularly relates to a method for selecting sub-intervals of near-infrared spectroscopy based on simulated annealing-genetic algorithm.

Background technique

Near-infrared spectroscopy is widely used in agricultural products and food quality analysis due to its fast analysis speed and high efficiency. However, there are certain deficiencies in near-infrared spectroscopy, such as complex background, low information intensity, overlapping peaks, etc. Conventional spectral analysis methods are analyzed. Therefore, how to effectively extract feature information from a large number of near-infrared spectral data has become the focus of research in this field.

The characteristic absorption of the sample in one or several bands of the near-infrared spectrum determines that the wavenumber points adjacent to the high-information wavenumber point have a relatively large amount of information, that is, the near-infrared spectrum data has a certain continuous correlation. According to the characteristics of the near-infrared spectroscopy data, the calculation of the wavelength selection algorithm is reduced, and the efficiency of the algorithm is improved. Generally, the near-infrared full spectrum is divided into several sub-intervals, and wavelength selection is performed in intervals. The classical spectral interval selection algorithm has interval partial least squares method. The algorithm divides the whole spectrum into sub-interval sub-intervals, and calculates the RMSECV (Root Mean Square of Cross Validation) for each sub-interval. An interval with the smallest square root error is used as the modeling interval. The derivative algorithms of the interval partial least squares algorithm are joint interval partial least squares method, forward/backward interval partial least squares algorithm, moving window partial least squares method, etc., compared with classical interval partial least squares algorithm, derivative algorithm Not only the single interval but also the combination of several intervals. Although these algorithms can extract the characteristic information of the spectrum, the process of dividing the subintervals has certain subjectivity.

Genetic algorithm is a new subject emerging in the 1970s. It is based on the simulation of the natural selection and natural genetic mechanism of the biological world to solve practical problems. It is a highly parallel, random and adaptive search algorithm. In recent years, some scholars have combined genetic algorithm with classical interval partial least squares algorithm to select the characteristic subinterval of near-infrared spectroscopy, simulate natural evolution processes such as natural genetic variation, and solve the optimal combination of feature subintervals, but still exist. Some shortcomings, such as sub-intervals, often rely on experience and have certain subjectivity; genetic algorithms are prone to premature convergence and fall into local optimal solutions, and cannot guarantee global optimal approximation solutions.

The simulated annealing algorithm is a stochastic optimization algorithm based on the Mote Carlo iterative solution strategy. The starting point is based on the similarity between the physical annealing process and the combined optimization. The simulated annealing algorithm starts from a higher initial temperature and uses the Metropolis sampling strategy with probability jump to perform a random search in the candidate solution combination. After the repeated sampling process, the global optimal solution of the problem is finally obtained, which is suitable for solving the large-scale combinatorial optimization problem.

Summary of the invention

In order to overcome the deficiencies of the sub-intervals of the near-infrared spectroscopy in the prior art, and to ensure the global optimal approximate solution, the present invention proposes a sub-interval selection method based on simulated annealing-genetic algorithm for near-infrared spectroscopy, which will simulate The core Metropolis acceptance criterion in the annealing algorithm introduces the genetic algorithm, and prevents the premature fall into the local optimal solution on the basis of ensuring the efficiency of the genetic algorithm, so as to obtain the optimal combination of the sub-intervals of the near-infrared spectrum.

The technical scheme adopted by the invention is: pre-processing the near-infrared spectrum, dynamically dividing the sub-intervals of the pre-processed near-infrared spectrum, and introducing the Metropolis criterion in the simulated annealing algorithm into the gene exchange and gene selection calculation in the genetic algorithm. Son, using the simulated annealing-genetic algorithm to select the optimal feature subinterval, and finally judging the best subinterval partitioning method and the optimal feature subinterval combination, and establishing the PLS model for the selected optimal feature subinterval.

The present invention achieves the following effects by using the above technical solutions:

1. Introduce the Metropolis criterion in the simulated annealing algorithm into the exchange operator and mutation operator, and generate high-quality offspring individuals through the improved mutation and exchange operator, which not only improves the overall fitness level of the population, but also the evolution of the population. Provides enough power.

2. The introduction of Metropolis criterion effectively solves the problem of premature convergence and local optimal solution of traditional genetic algorithm; dynamically dividing the spectral subinterval, effectively avoiding the artificially specified total number of spectral subintervals in the modeling process. insufficient.

The sub-interval selection method of near-infrared spectroscopy based on simulated annealing-genetic algorithm lays a solid foundation for quickly obtaining spectral models with high precision and strong predictive ability.

DRAWINGS

The present invention will be further described in detail below in conjunction with the drawings and specific embodiments.

Figure 1 is a flow chart of the present invention;

Figure 2 is a schematic diagram of Metropolis acceptance criteria;

Figure 3 is a schematic diagram of an exchange operator introducing the Metropolis criterion;

Figure 4 is a schematic diagram of a mutation operator introducing the Metropolis criterion;

Figure 5 is a graph showing the results of sub-interval selection of simulated annealing-genetic algorithm;

Figure 6 is a comparison result of the simulated annealing-genetic algorithm and the traditional genetic algorithm modeling effect;

Figure 7 is a near-infrared spectrum of the flavonoids of cucumber leaves pretreated by standard orthogonal transformation.

Detailed ways

The invention firstly pretreats the near-infrared spectrum, and processes the agricultural product and the food raw near red with an appropriate de-noising method. After the spectrum, the denoising method includes standard orthogonal variation, multivariate scatter correction, centralization, first-order/second-order derivative preprocessing methods, etc. The spectral pre-processing process also includes the division of the correction set and the prediction set sample. The pre-processed near-infrared spectrum is dynamically divided into sub-intervals. When sub-intervals are divided, the number of sub-intervals changes dynamically within a range [m, n]. The subsequent processing of the algorithm will select the number of optimal feature subintervals in the range of [m, n]. When the number of spectral subintervals is k≡ [m, n], the full spectrum is equally divided into subintervals, if the total number of points is divided by i is equal to p, and there is a remainder q, then the number of wave points in each subinterval of the first q subintervals is p+l, and the number of wavenumber points in each subinterval in the remaining subintervals is p.

The Metropol is criterion in the simulated annealing algorithm refers to a judgment rule used in the simulated annealing algorithm to judge the importance of the new solution and the old solution. The Metropol is criterion judges which solution in the old solution and the new solution is an important solution according to the objective function value corresponding to the old solution and the new solution. If the new solution is considered to be an important solution, replace the old solution with the new solution into the next iteration; Then keep the old solution unchanged. For the optimal feature subinterval problem, assume that the objective function values corresponding to the old solution x and the new solution y are f and θλ respectively, then judge the importance of the old solution x and the new solution y based on the following Metropolis criteria: When f (y)>f (x), the new solution is an important solution, otherwise it is judged whether the following formula Pt = _ejr p( ^^ ) > _r e [0, l] holds, where is the new solution transition probability, r is the range of 0~1 The probability density function is randomly generated; if the above formula holds, the new solution y is considered to be an important solution, otherwise the old solution Jf is considered to be an important solution.

The invention introduces the Metropol is criterion in the above simulated annealing algorithm into the gene exchange and gene selection operator in the genetic algorithm, which is called "simulated annealing-genetic algorithm", that is, in the traditional gene exchange operator and the gene mutation operator, the parent The chromosome generates the progeny chromosome through gene exchange or gene mutation, and introduces the Metropolis criterion to judge the parent chromosome (corresponding to the old solution X) and the progeny chromosome (corresponding to the importance of the new solution, if the progeny chromosome is more important than the parent chromosome, then accept the progeny chromosome, Otherwise the child chromosome is rejected.

Simulated annealing-genetic algorithm was used to select the spectral optimal feature subinterval, combined with the gene exchange, gene mutation operator and other operators of traditional genetic algorithm introduced by Metropolis criterion, and the optimal feature subinterval was selected for the near-infrared spectrum after subinterval. Intelligently judge the optimal sub-interval division method and the optimal feature sub-interval combination, establish the PLS model of the correction set and the prediction set for the selected optimal feature subinterval, and calculate the correction set rms error and the prediction set rms error. , modeling parameters such as correction set correlation coefficient and prediction set correlation coefficient.

The present invention divides the subinterval and selects the optimal feature subinterval, and needs to set the following parameters:

(1) The minimum number of subintervals I. : Refers to dividing at least the full spectrum into I. Sub-interval.

(2) The maximum number of sub-intervals I _f: refers to dividing the full spectrum into If sub-intervals at most.

(3) The objective function ί χ λ The role of the objective function is to judge the quality of the current solution X. In general, f (the higher the value of 3⁄4, the better the quality of the current solution X. The goal of the simulated annealing-genetic algorithm of the present invention is the preferred feature. Subinterval, The currently selected all feature subintervals are considered to be the current solution x, and the objective function is defined as f(3⁄4=l/(l+RMSECV), where RMSECV is the interactive verification root mean square error corresponding to all selected interval PLS models. value.

(4) Gene coding: Since the genetic algorithm cannot directly process the two-bit near-infrared spectroscopy data, it needs to represent them as genotype string structure data of the genetic space by binary coding, 1 indicates that the corresponding position of the gene is selected, and 0 indicates the corresponding position. The genes were not selected. For example, 0011001101, it means that there are 10 genes in the chromosome, and the corresponding genes in the 3rd, 4th, 7th, 8th and 10th positions are selected.

(5) Population size: refers to the number of chromosomes in the population and the number of genes in each chromosome. The number of genes is generally determined according to the parameters of the actual problem. For the feature sub-interval selection problem, the number of chromosomes is generally selected from 30 to 100, and the number of genes is equal to the number of sub-intervals.

(6) Gene exchange probability p. : In the process of gene exchange, the proportion of chromosome individuals involved in gene exchange accounts for the total number of chromosomes, and the probability of gene exchange is generally set to 0.65~0.9.

(7) Probability of gene mutation p _{D: In the} process of gene mutation, the chromosome individual involved in the gene mutation accounts for the ratio of the chromosome always, and the probability of gene mutation is generally set to 0.001 to 0.1.

(8) Initialization temperature t. : Corresponding to the initial temperature during solid annealing, the initial temperature is usually set to 200 to 1000 degrees.

(9) Temperature decay function g(d) _: used to control the temperature cooling rate during solid annealing. Usually, + ₁ = t _k g(a )-α , α is usually in the range of 0.5~0.99.

(10) End temperature t _f: When the annealing temperature reaches the end temperature, the solid will reach a certain stable state, and the solid annealing process is finished. Generally, the annealing temperature t _f is about 0 degree.

The following processing steps are adopted to select the optimal feature subinterval for the near-infrared spectrum after dividing the feature subinterval:

(1) When the number of subintervals is i, the near-infrared spectrum is divided into i sub-intervals, and binary gene coding is performed, and the number of genes is the number of sub-intervals i.

(2) Chromosome initialization, randomly generating an initial population of a given size.

(3) When the temperature is t, calculate the objective function of the chromosome in the population / · 6 , use the selection operator to select the individual with fitness ,, and eliminate the individuals with low fitness to achieve the survival of the fittest.

(4) The improved gene crossover and gene mutation operator for introducing the Metropolis criterion in the simulated annealing algorithm into the genetic algorithm for gene exchange and gene mutation operations.

(5) reducing the temperature t of the temperature decay function g (ci), if t is not equal to the end temperature t _F, repeating steps (3) to (4), if the temperature t equal to the end _F, the step (6).

(6) The number of subintervals 2 ' increases by 1. If i is not equal to the maximum number of subintervals I _f , then repeat (1) ~ (5), If i is equal to the maximum number of subintervals I _f , then step (7) is performed.

(7) Determine the total number of best subintervals and the selected optimal subintervals.

The specific steps of the sub-interval selection method of the near-infrared spectroscopy characteristic of the simulated annealing-genetic algorithm are shown in Fig. 1. After the pre-processing of the near-infrared spectrum, when the total number of sub-intervals i=Io, the characteristic sub-interval of the full spectrum is performed. Divide, and perform binary gene coding to randomly initialize the population. After determining the initial population, the annealing temperature ί is from the starting temperature t. Initially, according to the temperature decay function g ( a ) slowly decreases, whenever the temperature t decreases, the fitness of each individual in the population is calculated, the parent chromosome is selected by the chromosome selection operator, and the gene exchange is performed according to the improved gene exchange operator. Gene mutation according to the improved gene mutation operator, and repeatedly performing the above process until the annealing temperature reaches the end temperature, and the corresponding optimal solution is saved when the total number of sub-intervals i = 10, and is incremented by i = i+1. The same solution is used to calculate the optimal solution corresponding to the total number of new subintervals, and the above process is repeated until the total number of subintervals i is greater than the end window width If. At this point, the optimal solutions corresponding to the total number of sub-intervals i' e [I0, If] have been obtained. From these optimization solutions, the solution with the largest value of the objective function is selected as _Xi , X, which is the global optimal feature of the near-infrared spectrum. Sub-interval set, subscript i is the total number of sub-intervals corresponding to the optimal solution. Finally, the correction set and the prediction set model are established according to the selected global optimal solution.

Figure 2 shows the process by which the Metropolis Code judges the importance of the new solution. The new solution transition probability pt is compared with the random probability density function re [0, 1]. If pt>r holds, the new solution is accepted, otherwise the old solution remains unchanged. The specific judgment process is as follows: (1) The Metropol is criterion first calculates the objective function value f (x) f (y) corresponding to the old solution X new solution y; (2) Produces the random probability density function value ί"; (3) Calculates the new (4) Compare the new solution transition probability Pi with the value of the random probability function value r. If it is greater than or equal to r, replace the old solution with a new solution. Otherwise, the old solution remains unchanged. According to this criterion, As follows: The Metropol is criterion can not only accept the optimization solution, but also accept the deterioration solution with a certain probability, which provides a guarantee for avoiding the algorithm falling into the local optimal solution.

Figure 3 Figure 4 shows the improved gene exchange operator and gene mutation operator flow chart, because the improved gene exchange operator is similar to the improved gene mutation operator, taking the improved gene exchange operator as an example. Details the workflow of this operator. Based on the traditional genetic operation, the Metropolis acceptance criterion in simulated annealing algorithm is introduced. The probability of positive mutation is increased on the basis of the original, the probability of negative mutation is reduced, and the algorithm can jump out of the local optimal solution to the global optimal solution. . The exchange operator randomly selects the parents of the parents from the parent group (denoted as Pi), and generates new generations of children by gene exchange (reported as Ci to calculate their fitness values and f(Ci) respectively, and judge whether to accept the new generation according to the Metropolis criterion. The individual judgment process is shown in Figure 3.

Example

Figure 7 shows the near-infrared spectrum of 100 pieces of cucumber leaves pretreated by standard orthogonal change, spectral range lOOOC^^OOcm— ¹ , the number of scans is 32; the wave number interval is 7. 712cm—'; the resolution is 16cm- ¹ . The spectrum of 70 leaves was used as a correction set, and the near infrared spectrum of the remaining 30 leaves was used as a prediction set. The minimum and maximum number of subintervals are 30, 60, the number of populations is 60, the probability of gene exchange is 0.9, the probability of gene mutation is 0.01, the initial temperature is 200, the end temperature is 0.1, and the temperature attenuation coefficient is 0.95. The simulated annealing-genetic algorithm is used to select the feature subinterval. The specific process is as follows:

(1) When the number of subintervals is 30, the full spectrum is divided into 30 sub-areas and binary coded;

(2) The number of chromosomes in the population is 60, and the number of genes per chromosome is 30, and the population is initialized;

(3) When the temperature is 200, the chromosome fitness in the population is calculated, and the selection operator is used to select individuals with high fitness, and the individuals with low fitness are eliminated to achieve the survival of the fittest.

(4) Using the improved gene exchange operator and gene mutation operator for gene exchange and gene mutation operations.

(5) Decrease the temperature according to the temperature decay function. If the temperature is not equal to the end temperature, repeat steps (3) to (4). If it is equal to the end temperature, perform step (6).

(6) The number of subintervals is increased by 1. If the number of subintervals is not equal to the maximum number of subintervals of 60, repeat (1:)~(5). If the number of subintervals is equal to the maximum number of subintervals of 60, execute step (7). .

(7) When it is judged that the number of subintervals is 40, seven subintervals are selected, which are sub-intervals of 3, 5, 14, 18, 21, 32, and 33, and the established model is optimal.

Figure 5 shows the results of sub-interval selection of near-infrared spectroscopy of cucumber leaf lutein using simulated annealing-genetic algorithm. Figure 6 is a comparison of simulated annealing-genetic algorithm and traditional genetic algorithm modeling of cucumber leaf lutein model. In Figure 6, the abscissa is the number of modeling times, the ordinate is the calibration set correlation coefficient of the spectral model, and the curve with the 厶 mark. The calibration set correlation coefficient obtained by modeling the near-infrared spectrum of cucumber leaf lutein by simulated annealing-genetic algorithm, and the curve with mouth mark is the correction set correlation of the near-infrared model of cucumber leaf lutein obtained by traditional genetic algorithm. coefficient. It can be seen from Figure 6 that the spectral model obtained by simulated annealing-genetic algorithm is superior to the spectral model established by traditional genetic algorithm.

Claims

Claims, a near-infrared spectrum feature sub-interval selection method based on simulated annealing-genetic algorithm, which is characterized by: first preprocessing the near-infrared spectrum, and then dynamically dividing the pre-processed near-infrared spectrum into sub-intervals, and then simulated annealing The Metropolis criterion in the algorithm introduces the gene exchange and gene selection operators in the genetic algorithm, uses simulated annealing-genetic algorithm to select the optimal characteristic sub-interval, and finally determines the optimal sub-interval division method and optimal characteristic sub-interval combination. The optimal characteristic sub-interval is used to establish a PLS model. . The near-infrared spectrum characteristic sub-interval selection method based on simulated annealing-genetic algorithm according to claim 1, which is characterized by: dividing the sub-intervals and using the simulated annealing-genetic algorithm to select the optimal characteristic sub-interval. The parameters that need to be set are: The most The number of small subintervals I. , the maximum number of subintervals If, the objective function / ^ enters the gene encoding, the population size, and the gene exchange probability P. , gene mutation probability ρ„, initialization temperature t., temperature attenuation function g(a) and end temperature, the near-infrared spectrum feature sub-interval selection method based on simulated annealing-genetic algorithm according to claim 2, which is characterized by: selection The optimal characteristic sub-interval adopts the following steps:

(1) When the number of subintervals is i, the near-infrared spectrum is divided into i subintervals for binary gene encoding, and the number of genes is the number of subintervals.

(2) Chromosome initialization, randomly generate an initial population of a given size;

(3) When the temperature is t, calculate the objective function ffe) of the chromosomes in the population, use the selection operator to select individuals with high fitness, and eliminate individuals with low fitness to achieve the survival of the fittest in the population;

(4) Use the improved gene exchange operator and gene mutation operator that introduce the Metropolis criterion in the simulated annealing algorithm into the genetic algorithm to perform gene exchange and gene mutation operations;

(5) Reduce the temperature t according to the temperature attenuation function g (a). If t is not equal to the end temperature t _f , repeat steps (3) to (4). If it is equal to the end temperature t _f , perform step (6);

(6) The number of subintervals 2' is increased by 1. If i is not equal to the maximum number of subintervals If, then (1) to (5) are repeated. If _i is equal to the maximum number of subintervals _If , then step (7) is performed;

(7) Determine the total number of optimal subintervals and the selected optimal characteristic subinterval.

. The near-infrared spectrum feature sub-interval selection method based on simulated annealing-genetic algorithm according to claim 3, which is characterized by: the improved gene exchange operator and gene mutation operator described in step (4) perform gene exchange and gene The method of mutation operation is: increase the probability of positive mutation and reduce the probability of negative mutation. The exchange operator randomly selects parent individuals from the parent population, generates new offspring individuals through genetic exchange, and calculates their fitness values respectively. Determine whether to accept newly created individuals according to Metropolis guidelines.

. The near-infrared spectrum characteristic sub-interval selection method based on simulated annealing-genetic algorithm according to claim 1, characterized by: The Metropolis criterion in the simulated annealing algorithm is to judge the old solution based on the objective function values corresponding to the new solution. Which solution among the solution and the new solution is the important solution? If the new solution is considered to be the important solution, the new solution will replace the old solution and enter the next iteration; otherwise, the old solution will remain unchanged.

. The near-infrared spectrum feature sub-interval selection method based on simulated annealing-genetic algorithm according to claim 1, characterized by: the number of divided sub-intervals changes dynamically within a range [m, n], using simulated annealing-genetic algorithm Select the optimal characteristic sub-interval within the range [m, n]. When the number of sub-intervals is ce [ _m , _n ], the full spectrum is equally divided into /c sub-intervals. If the total number of wavenumber points divided by / is equal to p, there is The remainder is 9, then the number of wave number points in each sub-interval in the first q sub-intervals is P+l, and the number of wave number points in each sub-interval in the remaining sub-intervals is P.