CN110619476A

CN110619476A - Method for predicting rice grain heap yellowness index by optimizing SVR (singular value representation) parameters by grid search method

Info

Publication number: CN110619476A
Application number: CN201910890527.9A
Authority: CN
Inventors: 张潇; 李培灵; 王�锋
Original assignee: Henan University of Technology
Current assignee: Henan University of Technology
Priority date: 2019-09-20
Filing date: 2019-09-20
Publication date: 2019-12-27

Abstract

The invention discloses a method for predicting a rice grain heap yellowness index by optimizing SVR parameters by a grid search method. The method comprises the following steps: firstly, dividing and editing the characteristic values of the rice grain piles according to a specific structure, reading the edited rice grain pile data by using a libsvm toolbox in MATLAB, and carrying out normalization processing; then, the read data are transmitted into a grid searching method, and an optimal punishment parameter c and a kernel function parameter g are searched and searched according to a certain step length; and finally, predicting the yellowness index of the rice grain heap by using the screened optimal solution. The method for predicting the yellowness index of the rice grain heap by optimizing the SVR parameters by the grid search method can effectively improve the accuracy of the prediction result of the yellowness index of the rice grain heap through the optimization of the grid search method and the SVR support vector regression prediction.

Description

Method for predicting rice grain heap yellowness index by optimizing SVR (singular value representation) parameters by grid search method

Technical Field

The invention belongs to the technical field of neural networks, and particularly relates to a method for predicting a rice grain heap yellowness index by optimizing SVR parameters by a grid search method.

Background

The annual output of rice all over the world is about 5.2 million tons, about more than 10 million people in 39 countries use rice as main food, particularly Asia has the strongest dependence on rice; wherein, 1.85 hundred million tons of rice are produced in China in year, which accounts for about 35 percent of the world and is the first place; the second are india, indonesia and thailand. China is also a large country for rice consumption, and at present, people directly consume about 100kg of rice annually, and the total consumption of grain is about 1.2 hundred million tons. The rice is one of the most difficult-to-preserve grains, because the rice husk and the husk layer which protect the endosperm are removed in the rice processing process, the endosperm is directly linked with the factors of the external environment, such as temperature and humidity, and the like, and the rice grains are hydrophilic colloids rich in nutrients such as starch, protein and the like, and are extremely easy to deteriorate under the influence of humidity, heat, oxygen, insects, mildew and the like. Especially under the conditions of high temperature and high humidity in summer, the quality of the rice is deteriorated and the mildewing speed is accelerated, so that the acidity of the rice is increased, the viscosity of the rice is reduced, and the edible quality of the rice is reduced and even the edible value of the rice is lost. The yellowness index is an important parameter for the quality of rice. Therefore, the research for predicting the yellowness index of the rice according to the storage condition of the rice has strong theoretical value and practical significance.

The neural network simulates partial functions of human brain, and is an important method for realizing artificial intelligence. The use of neural networks for nonlinear regression prediction processing of data is currently a relatively widespread application. Compared with the traditional empirical model, the neural network prediction fits the relation according to the internal relation between the data, so that the model has higher reliability and practicability. As one of the grains consumed in China, the rice is of great importance in quality. However, the research on the quality of rice in the storage process is less, and at present, the research mostly depends on experience and manual judgment, so that the time and labor are wasted, and the accuracy is low. Support Vector Regression (SVR) is a good and stable neural network algorithm, and the grid search method is to substitute the data of c and g little by little in a certain range and continuously update the optimal c and g. The two are used together to predict the grain bulk yellowness index of rice with very good effect.

The invention content is as follows:

the invention discloses a method for predicting a rice grain heap yellowness index by optimizing SVR parameters by a grid search method. The method comprises the following steps: firstly, dividing and editing the characteristic values of the rice grain piles according to a specific structure, reading the edited rice grain pile data by using a libsvm toolbox in MATLAB, and carrying out normalization processing; then, the read data are transmitted into a grid searching method, and an optimal punishment parameter c and a kernel function parameter g are searched and searched according to a certain step length; and finally, predicting the yellowness index of the rice grain heap by using the screened optimal solution. The grid searching method for optimizing the SVR parameter to predict the yellowness index of the rice grain heap, disclosed by the invention, can effectively improve the accuracy of the prediction result of the yellowness index of the rice grain heap through the grid searching method and the SVR support vector regression prediction. The specific flow of our method is as follows:

a method for predicting the yellowness index of rice grain heap by optimizing SVR parameters by a grid search method is characterized by comprising the following steps: comprises the following steps of (a) carrying out,

preprocessing original rice grain heap data;

reading and normalizing the rice grain pile data;

optimizing by utilizing grid search algorithm parameters;

and (D) setting SVR parameters by using the optimized optimal solution and predicting the yellowness index of the rice grain heap.

1. And (A) preprocessing the original rice grain pile data, and firstly, performing characteristic classification editing on the original rice grain pile data in order to read rice characteristic data by a tool box.

2. And (B) reading and normalizing the rice grain pile data, reading the rice grain pile data edited in the step (A) by using a libsvm tool box, and normalizing the read data to a [0,1] interval according to the same characteristic class by adopting an MATLAB self-carrying function mapminmax.

3. Step (C), optimizing parameters by using a grid search method, transmitting the normalized feature data obtained in the step (B) into the grid search method to search an optimal punishment parameter C and a kernel function parameter g, and initializing data including the optimal parameters C and g, maximum and minimum C and g values and a minimum mean square error; then dividing c and g into grids and starting searching; the data of the rice are roughly and averagely divided into K groups by adopting a K-CV method, wherein the K-1 group is used as a training set, and the last group is used as a verification set.

4. The method for predicting the yellowness index of rice grain heap by optimizing SVR parameters according to claim 1, wherein: and (D) setting an SVR parameter by using the optimal solution obtained by optimization and predicting the yellowness index of the rice grain pile, setting an optimal punishment parameter c and a kernel function parameter g obtained by the grid searching method as parameters of the SVR, inputting the characteristic data of the rice grain pile into a model of the SVR, and predicting the yellowness index of the rice.

SVR support vector regression is an important branch of the SVM support vector machine, the SVM itself is directed to the two-class problem, and SVR is directed to the intrinsic relationship to find a pile of data. The support vector machine is to make the distance of the nearest sample point of the obtained hyperplane maximum; the support vector regression is to make the "minimum distance" of the farthest sample point of the hyperplane as shown in fig. 1;

regression is as an intrinsic relationship to find a pile of data. No matter the data of the pile is composed of several categories, a formula is obtained, the data are fitted, and when a new coordinate value is given, a new value can be obtained; therefore, for SVR, a hyperplane or a function is obtained, and all data can be fitted (i.e., all data points, regardless of which class they belong to, are closest to the hyperplane or function); the statistical understanding is: minimizing the intra-class variance of all the data, and regarding all the data of the classes as one class;

traditional regression methods consider the prediction to be correct if and only if the regression f (x) is completely equal to y, and their losses need to be calculated; support Vector Regression (SVR) considers that prediction is correct without calculating loss as long as the degree of deviation between f (x) and y is not too great.

Description of the drawings:

as shown in the figures, the SVM/SVR diagram of FIG. 1 is shown.

The specific implementation mode is as follows:

to verify the performance of our proposed model, we used existing data for prediction and comparison with real data. A total of 28 sets of data, temperature, relative humidity and days of storage of the grain bulk containing rice.

Table 1 shows the comparison between the performance of our grid search method optimization SVR algorithm and the SVR algorithm, BP neural network algorithm.

TABLE 1 comparison between grid search optimization SVR and SVR, BP Performance

Algorithm	Mean square error
		Grid search optimization SVR	0.0438
SVR	0.1219
		BP	0.1322

Claims

1. The invention discloses a method for predicting a rice grain heap yellowness index by optimizing SVR parameters by a grid search method, which is characterized by comprising the following steps of: comprises the following steps of (a) carrying out,

preprocessing original rice grain heap data;

reading and normalizing the rice grain pile data;

optimizing by utilizing grid search algorithm parameters;

2. The method for predicting the yellowness index of rice grain heap by optimizing SVR parameters according to claim 1, wherein: and (A) preprocessing the original rice grain pile data, and firstly, performing characteristic classification editing on the original rice grain pile data in order to read rice characteristic data by a tool box.

3. The method for predicting the yellowness index of rice grain heap by optimizing SVR parameters according to claim 1, wherein: and (B) reading and normalizing the rice grain pile data, reading the rice grain pile data edited in the step (A) by using a libsvm tool box, and normalizing the read data to a [0,1] interval according to the same characteristic class by adopting an MATLAB self-carrying function mapminmax.

4. The method for predicting the yellowness index of rice grain heap by optimizing SVR parameters according to claim 1, wherein: step (C), optimizing parameters by using a grid search method, transmitting the normalized feature data obtained in the step (B) into the grid search method to search an optimal punishment parameter C and a kernel function parameter g, and initializing data including the optimal parameters C and g, maximum and minimum C and g values and a minimum mean square error; then dividing c and g into grids and starting searching; the data of the rice are roughly and averagely divided into K groups by adopting a K-CV method, wherein the K-1 group is used as a training set, and the last group is used as a verification set.

5. The method for predicting the yellowness index of rice grain heap by optimizing SVR parameters according to claim 1, wherein: setting an SVR parameter by using the optimal solution obtained by optimization and predicting the yellowness index of the rice grain pile, setting an optimal punishment parameter c and a kernel function parameter g obtained by optimizing the grid searching method as parameters of the SVR, inputting the characteristic data of the rice grain pile into a model of the SVR, and predicting the yellowness index of the rice;

SVR support vector regression is an important branch of an SVM support vector machine, the SVM is specific to the problem of two classes, and the SVR is specific to the internal relation of searching a pile of data; the support vector machine is to make the distance of the nearest sample point of the obtained hyperplane maximum; the support vector regression is to make the "minimum distance" of the farthest sample point of the hyperplane as shown in fig. 1;

regression is as if finding the intrinsic relationship of a pile of data; no matter the data of the pile is composed of several categories, a formula is obtained, the data are fitted, and when a new coordinate value is given, a new value can be obtained; therefore, for SVR, a hyperplane or a function is obtained, and all data can be fitted (i.e., all data points, regardless of which class they belong to, are closest to the hyperplane or function); the statistical understanding is: minimizing the intra-class variance of all the data, and regarding all the data of the classes as one class;