CN107121407A - The method that near-infrared spectrum analysis based on PSO RICAELM differentiates Cuiguan pear maturity - Google Patents

Abstract

The invention discloses a method for identifying the maturity of Cuiguan pear based on PSO-RICAELM near-infrared spectrum analysis, comprising the following steps: (1) collecting samples, collecting sample spectra, obtaining sample near-infrared diffuse reflection spectrum data, and analyzing the spectral data Perform second-order derivative processing, and use discrete wavelet transform to compress the original near-infrared spectral data; (2) decompose the compressed spectral data with robust independent component analysis, and extract the basic component spectral matrix and the corresponding approximate concentration matrix; (3) The approximate concentration matrix is used as the model input, and the corresponding maturity is used as the output, and the PSO-RICAELM model is established by using the particle swarm optimization algorithm; (4) The quality evaluation of the model identifies the maturity of the sample. The invention can quickly identify the maturity of the Cuiguan pear, enriches the chemometric method, and has good application prospects.

Description

Identification of Cuiguan pear maturity based on near-infrared spectroscopy analysis of PSO-RICAELM method

technical field

The invention relates to the technical field of infrared non-destructive testing, in particular to a method for identifying the maturity of Cuiguan pears through PSO-RICAELM near-infrared spectrum analysis.

Background technique

Modern near-infrared spectroscopy is a fast and non-destructive detection method. Its principle is to absorb the double frequency combination of hydrogen-containing groups X-H in organic matter. The physical and chemical indicators of organic matter are measured by stoichiometric methods. Effective mathematics The algorithm establishes a functional relationship between physical and chemical indicators and spectra, which is faster and more convenient than traditional methods, and saves time, and has been widely used in many aspects of analysis. However, in the currently existing fruit ripeness identification method based on near-infrared spectrum analysis, such as BP neural network, the convergence speed will be slow during the training process, and the training will fail when entering the local minimum; due to the support vector machine (SVM) Large-scale training is difficult to implement, so it is very difficult to solve multi-classification, and has disadvantages such as low robustness. Therefore, in order to solve the problems of large sample base required for modeling, long training time, instability, etc., a universal and strong model is urgently needed, which can effectively identify the maturity of Cuiguan pear.

Traditional decomposition and dimensionality reduction methods such as PCA and SVD can only decompose irrelevant components, but cannot guarantee that these components are independent of each other, which lacks practical significance and reduces the typicality of the extracted features.

Extreme learning machine (ELM) is a new type of neural network, which randomly gives weights to the input layer and biases to the hidden layer, but this will easily lead to poor generalization ability and insufficient precision of the model.

Contents of the invention

The object of the present invention is to provide a method for identifying the maturity of Cuiguan pear based on PSO-RICAELM near-infrared spectrum analysis for the deficiencies in the prior art. The output of ICA is a non-Gaussian signal, which improves the accuracy of the model. The robust independent component analysis algorithm proposed in this paper is a kind of ICA, which reduces the number of iterations and has a faster convergence speed, and can extract all non-zero Basic composition spectral matrix. The spectrum processed by the robust independent component analysis algorithm is used as input to the extreme learning machine network to obtain the RICAELM model. The invention proposes a particle swarm algorithm to optimize the input layer weight and the hidden layer deviation of the RICAELM network to improve the accuracy of the model.

The present invention is achieved through the following technical solutions: a method for identifying the maturity of Cuiguan pear based on PSO-RICAELM near-infrared spectrum analysis, comprising the following steps:

(1) Collect samples; start picking samples one week before the sample maturity period, and pick samples every n (4<n<10) days, at least three times; scan the samples multiple times with a near-infrared spectrometer, and after multiple scans Taking the average value to obtain the near-infrared spectrum of samples with different maturity; and performing second derivative and discrete wavelet transform processing on the near-infrared spectrum to obtain the compressed near-infrared spectrum data matrix;

(2) Decompose the compressed near-infrared spectrum data matrix obtained in step (1) using robust independent component analysis to obtain an approximate concentration matrix; establish a calibration set model, and use the actual known maturity and approximate concentration matrix as the initial model input , get the RICAELM model; optimize the RICAELM model with the particle swarm optimization algorithm, and get the best Cuiguan pear maturity identification model;

(3) Input the approximate concentration matrix of the prediction set samples into the best Cuiguan pear maturity identification model to obtain the maturity of the prediction set samples.

Further, the specific implementation of the step (1) is as follows:

Samples were picked one week before the sample maturity period, and samples were picked every n (4<n<10) days, at least three times; samples of different maturity were scanned with a near-infrared spectrometer, and the near-infrared spectra of the samples were obtained. Samples of different maturity are randomly divided into calibration set samples and prediction set samples, among which, the calibration set samples are used for model training, and the prediction set samples are used for model quality evaluation; the near-infrared spectra of samples are subjected to second derivative and discrete wavelet transform Processing, the discrete wavelet transform selects the wavelet mother function as dbn, n as the vanishing moment, and obtains the compressed near-infrared spectrum data matrix.

Further, the specific implementation of the step (2) is as follows: the processed data is used as an input to perform a robust independent component analysis operation, and under the premise of independence of the spectral components, the spectrum is decomposed into different components, and the output is an approximate concentration matrix . The robust independent component analysis algorithm takes the kurtosis as the control function, and optimizes the kurtosis through the search method of the optimal step size, and obtains all basic component spectral matrices and corresponding approximate concentration matrices with kurtosis not equal to 0. Input the output non-Gaussian approximate concentration matrix containing most of the spectral features into the extreme learning machine network to obtain the RICAELM model.

Further, the specific implementation of the step (3) is as follows: the connection weight between the input layer and the hidden layer of the Cuiguan pear maturity identification model and the bias value of the hidden layer nodes are used as one of the particle swarm search space Particles, use the particle swarm optimization algorithm to iteratively optimize the particles, set the number of particles m, generate a population, initialize the particle velocity v, inertia weight w, learning factors c ₁ and c ₂ , and the maximum number of iterations k of the PSO, according to the adaptation The degree function calculates the fitness value of the particle, and then updates the position and velocity of the particle. The specific update formula is as follows:

v _i (a+1)＝v _i (a)+c ₁ r ₁ [pbest _i -x _i (a)]+c ₂ r ₂ [gbest-x _i (a)]

x _i (a+1)= _xi (a)+v _i (a+1)

where i is the particle (i=1, 2...), x _i (a+1), v _i (a+1) is the position and velocity of the i-th particle at the a+1-th iteration, c ₁ and c ₂ learning factor, r ₁ and r ₂ are random numbers uniformly distributed in the interval [0,1], v _i , x _i are the speed and position of the i-th particle in the a-th iteration, pbest _i is the i-th particle The individual best position of gbest is the global best position.

When the optimization satisfies the optimization and reaches the maximum number of iterations k or the evaluation value is less than the specified accuracy b (b>0), the search ends, and the optimal input weight and hidden layer bias are obtained and assigned to the RICAELM network. The best identification model of Cuiguan pear maturity was obtained.

Further, in the step (3), the extreme learning machine ELM hidden layer activation function selects the Sigmoidal function.

The beneficial effects of the invention are: the second-order derivative is used to eliminate the baseline drift of the spectrum and the background light interference, and improve the spectral resolution. The use of discrete wavelet transform can compress data, save cost and space, and improve the efficiency and accuracy of modeling. The robust independent component analysis algorithm is used to select the optimal number of independent components to establish a correction model; the PSO-RICAELM algorithm is used to establish a model, which is a model with higher classification accuracy, more stability, and better generalization ability. practical meaning. The whole measurement process does not consume chemical reagents, the test is fast, saves manpower and material resources, the result is accurate in batch testing, and the detection efficiency is greatly improved. The method can be popularized and used in links such as picking, storing, transporting and selling.

Description of drawings

Fig. 1 is a flowchart of the maturity method of the present invention;

Fig. 2 is the sample original spectrogram of four kinds of different maturities of the present invention;

detailed description

Since the near-infrared spectrum can be regarded as a combination of multiple main component spectra, in recent years, some scholars have introduced the problem of "blind source separation (BBS)" into the near-infrared spectral analysis method, and the spectra of the main components are separated from the complex mixed spectrum. seperate. The robust independent component analysis algorithm is an algorithm based on kurtosis and optimal step size, which provides a high-efficiency method for signal blind source separation. The main advantage is that the optimal step size and any non-zero independent component Component, suitable for spectrum decomposition, can extract the typical features of the spectrum. At the same time, ELM is a feed-forward neural network, with fast learning and training speed and few setting parameters, which can greatly reduce human errors. The spectrum processed by the robust independent component analysis algorithm is input into the ELM network, and a RICAELM model with higher precision and stability is obtained. After adopting PSO optimization, RICAELM input layer weights and hidden layer deviations select the optimal input, and obtain a model with higher classification accuracy, more stability, and better generalization ability.

The present invention adopts the PSO-RICAELM combined method, so that the near-infrared spectral analysis method can effectively identify the maturity of Cuiguan pear, enriches the stoichiometric method and provides a theoretical premise and technology for the identification of crop maturity in the field of near-infrared spectral analysis support.

The invention provides a near-infrared spectroscopic analysis method for accurately identifying the maturity of Cuiguan pears. The present invention will be further described below in conjunction with accompanying drawing 1, accompanying drawing 2 and embodiment. The examples are examples of the present invention and are not intended to limit the present invention.

The technical scheme of the present invention is to collect samples and collect spectra, and to preprocess the original spectrum, first process the near-infrared spectrum data with the second order derivative, and then use the discrete wavelet transform to effectively compress the near-infrared spectrum data to obtain a suitable amount of data. The matrix is decomposed by the robust independent component analysis algorithm, the basic component spectral matrix and the corresponding approximate concentration matrix are extracted, the approximate concentration matrix is used as the model input, and the particle swarm optimization RICAELM model is used. The flow chart of the whole program is shown in Figure 1.

Attached Figure 2 is the original spectrum directly acquired by the near-infrared spectrometer. There are differences in the spectra of Cuiguan pears at different maturity levels, and some spectral regions overlap seriously. There is a difference in the absorption of infrared spectrum, which is the basis for identifying the maturity, and it also shows that it is meaningful to identify the maturity of Cuiguan pear. Using the second order derivative can effectively eliminate the influence of baseline drift and other disturbances. Using the discrete wavelet transform method to compress the spectral data and remove the spectral noise, the main information of the spectrum is preserved and the amount of data is greatly reduced. This step needs to choose a reasonable wavelet function and wavelet decomposition layers.

A robust independent component analysis algorithm is used to decompose the wavelet-compressed data. Robust independent component analysis algorithm is a kind of blind source separation algorithm, which provides a high-efficiency method for signal blind source separation, the purpose is to separate the independent components in the mixed signal, the main advantage of this type of algorithm is that it can choose The optimal step size and any non-zero independent components solve the problem of blind source separation signal ordering and improve robustness. When used for spectral data matrix decomposition, the independent component matrix and the corresponding approximate concentration matrix can be obtained. Each row of the independent component matrix is equivalent to the spectral information of a statistical independent component, and the pure substance information in the mixture sample, which is the basic component matrix of the spectrum. The corresponding approximate concentration matrix reflects the proportion of the substance in the original data, that is, the contribution of the independent component to the near-infrared spectrum of the entire sample.

PSO-RICAELM is a particle swarm optimization extreme learning machine network. ELM is a feed-forward neural network with fast learning and training speed and few setting parameters, which can greatly reduce human errors. The spectrum processed by the robust independent component analysis algorithm is input into the ELM network, and a RICAELM model with higher precision and stability is obtained. RICAELM input layer weights and hidden layer deviations are randomly input, and PSO is used to optimize, and a model with higher classification accuracy, more stability, and better generalization ability is obtained. In the implementation, the PSO-RICAELM algorithm is applied, the number of neurons in the hidden layer is initially set to 10, and the number of neurons in the hidden layer is increased to 24 in turn with d=2 as the step size, and the training is repeated 15 times under the value of each hidden neuron. The model parameters of the best model are obtained to obtain the best PSO-RICAELM model.

Example:

1. Sample Collection and Spectrum Acquisition

The test samples came from a standardized orchard. The samples were picked one week before the normal maturity period of the samples. 36 Cuiguan pears were selected with medium size, similar color and no pests and mechanical damage. The samples were picked every 6 days. Second sample; the picked sample completed the experiment within the same day, and the sample was scanned three times with a spectrometer to get the average value; after the spectrum collection was completed, a total of 144 spectral data of four samples with different maturity levels were obtained, and all the data were calculated using the K/S algorithm According to the ratio of 3:1, it is divided into calibration set samples and prediction set samples. The spectrum acquisition instrument is a German Zeiss spectrometer model MCS600. Spectrum area collection range: 800nm-1700nm, experiments are carried out at room temperature. In the experiment, each sample was scanned three times at different positions, and the average spectrum was taken. The modeling software is based on Matlab2015a.

2. Spectral data processing

The second derivative processing is performed on the collected spectrum to eliminate the interference of baseline drift and background light.

Perform discrete wavelet transform on the processed spectrum to compress the data. The wavelet basis function chooses db2 and decomposes to 4 layers. The compressed data accounts for about 7% of the original data, and the compressed data contains most of the information of the original data.

The compressed matrix is decomposed using a robust independent component analysis algorithm. In the robust independent component analysis algorithm, the number of independent components needs to be set, and the initial value of the number of independent components is set to 5. In order to find the optimal number of independent components, the number of independent components is increased to 20 in steps of 1, and the analysis is carried out. The optimal value of the independent component is 12. According to the independent component values, the compressed matrix is decomposed using the robust independent component analysis algorithm to obtain the basic component spectral matrix and the corresponding approximate concentration matrix

3. Establish PSO-RICAELM model

The actual maturity of the four Cuiguan pears and their approximate concentration matrix are used as the analysis model, and the analysis model is established by using the PSO-RICAELM method; in the process of establishing the model, the parameters of the PSO are as follows: set the number of particles m=20, learn The factors c ₁ =1.5 and c ₂ =1.5, the maximum number of iterations k=200, and the minimum precision b is 0.01. Applying the PSO-RICAELM algorithm, the number of neurons in the hidden layer is initially set to 10, and the number of neurons in the hidden layer is increased to 24 in turn with d=2 as the step size, and the training is repeated 15 times under the value of each hidden neuron to obtain the best The model parameters of the model, so as to obtain the best PSO-RICAELM analysis model.

4. Results and Analysis

The correct recognition rate (CAR) is used as the evaluation index of the qualitative model:

Correct recognition rate = number of samples correctly recognized/total number of samples that should be recognized.

It has been verified that after particle optimization, the prediction ability of the model increases, and the accuracy rate increases from 88.9% to 94.4%, thus verifying the correctness of the PSO-RICAELM model.

Claims (5)

1. a method based on PSO-RICAELM's near-infrared spectrum analysis to differentiate the maturity of Cuiguan pear, is characterized in that, comprises the following steps: (1) Collect samples; start picking samples one week before the sample maturity period, and pick samples every n (4<n<10) days, at least three times; scan the samples multiple times with a near-infrared spectrometer, and after multiple scans Taking the average value to obtain the near-infrared spectrum of samples with different maturity; and performing second derivative and discrete wavelet transform processing on the near-infrared spectrum to obtain the compressed near-infrared spectrum data matrix; (2) Decompose the compressed near-infrared spectrum data matrix obtained in step (1) using robust independent component analysis to obtain an approximate concentration matrix; establish a calibration set model, and use the actual known maturity and approximate concentration matrix as the initial model input , get the RICAELM model; optimize the RICAELM model with the particle swarm optimization algorithm, and get the best Cuiguan pear maturity identification model; (3) Input the approximate concentration matrix of the prediction set samples into the best Cuiguan pear maturity identification model to obtain the maturity of the prediction set samples. 2. the method for the identification of Cuiguan pear maturity based on the near-infrared spectrum analysis of PSO-RICAELM according to claim 1, is characterized in that, described step (1) is concretely realized as follows: Samples were picked one week before the sample maturity period, and samples were picked every n (4<n<10) days, at least three times; samples of different maturity were scanned with a near-infrared spectrometer, and the near-infrared spectra of the samples were obtained. Samples of different maturity are randomly divided into calibration set samples and prediction set samples, among which, the calibration set samples are used for model training, and the prediction set samples are used for model quality evaluation; the near-infrared spectra of samples are subjected to second derivative and discrete wavelet transform Processing, the discrete wavelet transform selects the wavelet mother function as dbn, n as the vanishing moment, and obtains the compressed near-infrared spectrum data matrix. 3. the method for distinguishing the maturity of Cuiguan pear according to the near-infrared spectrum analysis based on PSO-RICAELM according to claim 1, it is characterized in that, described step (2) is concretely realized as follows: the data after processing is carried out robustly as input The independent component analysis operation decomposes the spectrum into different components under the premise of the independence of the spectral components, and the output is an approximate concentration matrix. The robust independent component analysis algorithm takes the kurtosis as the control function, and optimizes the kurtosis through the search method of the optimal step size, and obtains all basic component spectral matrices and corresponding approximate concentration matrices with kurtosis not equal to 0. Input the output non-Gaussian approximate concentration matrix containing most of the spectral features into the extreme learning machine network to obtain the RICAELM model. 4. the method for distinguishing the maturity of Cuiguan pear according to the near-infrared spectrum analysis based on PSO-RICAELM according to claim 1, is characterized in that, described step (2) is concretely realized as follows: the Cuiguan pear maturity discrimination model input The connection weight between the layer and the hidden layer and the bias value of the hidden layer node are used as a particle in the particle swarm search space, and the particle swarm optimization algorithm is used to iteratively optimize the particle, and the particle number m is set to generate a population , initialize the particle velocity v, inertia weight w, learning factors c ₁ and c ₂ , and the maximum number of iterations k of PSO, calculate the fitness value of the particle according to the fitness function, and then update the position and velocity of the particle. The specific update formula is as follows: v _i (a+1)＝v _i (a)+c ₁ r ₁ [pdest _i -x _i (a)]+c ₂ r ₂ [gbest-x _i (a)] x _i (a+1)= _xi (a)+v _i (a+1) where i is the particle (i=1, 2...), x _i (a+1), v _i (a+1) is the position and velocity of the i-th particle at the a+1-th iteration, c ₁ and c ₂ learning factor, r ₁ and r ₂ are random numbers uniformly distributed in the interval [0,1], v _i , x _i are the speed and position of the i-th particle in the a-th iteration, pbest _i is the i-th particle The individual best position of gbest is the global best position. When the optimization satisfies the optimization and reaches the maximum number of iterations k or the evaluation value is less than the specified accuracy b (b>0), the search ends, and the optimal input weight and hidden layer bias are obtained and assigned to the RICAELM network. The best identification model of Cuiguan pear maturity was obtained. 5. according to claim 1 or 4 described method based on the near-infrared spectrum analysis of PSO-RICAELM to differentiate Cuiguan pear maturity, it is characterized in that, in described step (3), RICAELM hidden layer activation function selects Sigmoidal function .