Near infrared spectrum light distance correction method for box

Abstract

The invention discloses a near infrared spectrum light distance correction method for a box, and the method comprises the steps: S1, carrying out the fitting of a collected near infrared spectrum signal, introducing a penalty factor to obtain the optimal distortion degree and roughness, and obtaining an optimal fitting signal based on the optimal distortion degree and roughness; s2, performing linear regression through nonlinear transformation of a generalized logistic function to predict different weight values, and distributing weights for the acquired near infrared spectrum data so as to obtain an optimal fitting baseline; s3, performing variable screening on the corrected near infrared spectrum data by using a BOSS algorithm to determine an optimal subset of variables; and S4, utilizing a particle swarm algorithm to optimize penalty parameters and kernel function parameters of support vector regression, and establishing a prediction model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of infrared spectrum analysis, and more particularly to a near-infrared spectrum light distance correction method for a box. BACKGROUND

[0002] Stacking is one of the common applications of industrial robots, and the traditional stacking control method is to control the target position according to the absolute displacement or relative displacement instruction of the servo motor. However, this method is too rigid and has no feedback information, and cannot effectively ensure the normal operation of the equipment according to the actual situation. At present, the stacking robot triggers operation by whether the material reaches the station, and the detection of the material is based on infrared reflection. However, for a box, the reflected spectrum data of different colors on the surface of the box and different materials of the box are different, so it is necessary to correct the light distance of the infrared spectrum to enable the stacking robot to operate normally.

[0003] Therefore, a new scheme needs to be proposed to solve the above problems. SUMMARY

[0004] The present application aims to overcome the shortcomings of the prior art and provide a near-infrared spectrum light distance correction method for a box.

[0005] To achieve the above purpose, the present application adopts the following technical scheme:

[0006] A near-infrared spectrum light distance correction method for a box, comprising the following steps:

[0007] Step S1, fitting the collected near-infrared spectrum signal to obtain the best distortion and roughness by introducing a penalty factor, and obtaining the optimal fitting baseline based on the best distortion and roughness;

[0008] Step S2, predicting different weight values by linear regression through nonlinear transformation of the generalized logistic function, and assigning weights to the collected near-infrared spectrum data according to the optimal fitting baseline;

[0009] Step S3, using the BOSS algorithm to perform variable screening on the corrected near-infrared spectrum data to determine the optimal subset of variables;

[0010] Step S4, using the particle swarm algorithm to optimize the penalty parameter and kernel function parameter of support vector regression, and establishing a prediction model.

[0011] Further, step S1 comprises the following steps:

[0012] Step S11, fitting the near-infrared spectrum data y collected by the infrared spectrometer to obtain a fitting signal z, and setting the length value of the fitting signal z and the length value of the near-infrared spectrum data to be N;

[0013] Step S12, based on step S11, the distortion degree F of the fitting signal z and the actual signal, and the roughness R of the fitting signal z are obtained:

[0014] (1)

[0015] In formula (1), i is the number of near-infrared spectrum data;

[0016] Step S13, by introducing a penalty factor λ to seek the optimal value between roughness R and distortion degree F, and let formula (2) minimum to obtain the optimal fitting baseline z:

[0017] (2)

[0018] The matrix of formula (2) is:

[0019] (3)

[0020] Wherein, D is the second order difference matrix coefficient term; the superscript T is the transpose operator of the matrix; is the second order difference equation.

[0021] Further, the optimal fitting baseline z is smoothed, including:

[0022] The partial derivative of the optimal fitting baseline z is solved, and the partial derivative is 0:

[0023] (4)

[0024] After solving, the expression of the fitting signal z can be obtained as:

[0025] (5)

[0026] Wherein, I is the unit matrix, the row and column are the same as the length value N of the collected near-infrared spectrum data y.

[0027] Further, in step S2, the rule of assigning weight to near-infrared spectrum data is: when the near-infrared spectrum data signal is lower than the optimal fitting baseline signal, the baseline of its peak-free region is assigned similar weight; when the near-infrared spectrum data signal is higher than the optimal fitting baseline signal, it is assigned a weight of 0.

[0028] Further, step S2 includes the following steps:

[0029] Step S21, in the iteration process, the weight vector w is updated according to the asymmetric weight rule,

[0030] (6)

[0031] wherein and are the mean and standard deviation of respectively; set , is a part of d and is defined only in the region of ; The function is a generalized function, defined as follows:

[0032] wherein e is the natural constant;

[0033] Step S22, according to the rule of assigning weight to the near infrared spectrum data, the expression of the fitting signal z is transformed:

[0034] (7)

[0035] In formula (7), W is the weight matrix of the weight vector w;

[0036] Solved by Whittaker filter.

[0037] Further, step S3 includes the following steps:

[0038] Step S31, set the near infrared spectrum data as a data matrix;

[0039] Step S32, apply the independent sampling method to generate K variable subsets in the variable space, in each data set, extract the variables selected by the independent sampling method, and remove the data samples with collinearity, the number of sample data replaced by the independent sampling method is the same as the total number of variable subsets, and each variable is assigned to the same weight;

[0040] Step S33, train K-PLS sub-models through the variable subsets in step S32, calculate the prediction error of the sub-models to extract the best model with the minimum prediction error;

[0041] Step S34, calculate the regression coefficients of each variable subset, change each value in the regression vector to a positive number, make each regression vector have the same unit length, and sum the calculated regression vectors to obtain new weights for the sample data;

[0042] Step S35, apply the weighted independent sampling method to generate new subsets according to the new weights;

[0043] Step S36, repeat steps S33-S35 until the number of new subsets is 1, and in the iteration process, the subset with the minimum prediction error is selected as the optimal variable set.

[0044] Further, in step S34, the calculated regression vector is summed up according to the following formula:

[0045] (8)

[0046] In formula (8), K is the number of sub-models; is the absolute value of the normalized regression coefficient of variable i in the Kth sub-model.

[0047] Further, step S4 includes the following steps:

[0048] Step S41, the variable-selected near-infrared spectrum data set is normalized;

[0049] Step S42, the penalty parameter c and the kernel function parameter g of the support vector machine are randomly initialized as the initial particle position and particle velocity, respectively, and then the particle fitness value is calculated according to formula (10) to update the particle velocity and position information:

[0050] (10)

[0051] In formula (10), is the inertia weight; is the velocity of particle x in the jth dimension at the kth iteration; is the velocity of particle x in the jth dimension at the (k+1)th iteration; and is the acceleration constant; and is a random number distributed in the interval [0, 1]; is the jth component of the individual historical optimal position of particle x at the kth iteration; is the jth component of the population historical optimal position of the particle group at the kth iteration; is the position of particle x in the jth dimension at the kth iteration; is the position of particle x in the jth dimension at the (k+1)th iteration; j=1, 2, …, J; x=1, 2, …, X;

[0052] Step S43, find the individual extreme value and the population extreme value of the particle, then update the velocity and the position of the particle; and update the individual extreme value and the population extreme value according to the fitness of the particles in the new population;

[0053] Step S44, steps S402 and S403 are repeated until the optimal particle point in the population is found;

[0054] Step S45, a training prediction model is established according to the optimal penalty parameter c and the kernel function parameter g found, and a test set is predicted;

[0055] Step S46, index analysis is performed according to the prediction result

[0056] The present application has the following advantages:

[0057] 1. The present application can model and analyze near-infrared spectrum data obtained by reflecting infrared light on boxes of different materials and different colors, so as to adjust the photoelectric signal distance to achieve timely sensing of any box and accurate control of stacking.

[0058] 2. The present application performs baseline correction and variable selection on the near-infrared spectrum data to improve the generalization ability and accuracy of the model; the selected characteristic variables are input, and a linear and nonlinear model is established by a support vector machine optimized by a particle swarm. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 It is a change graph of the number of variable selection, root mean square error and regression coefficient in the present embodiment, which changes with the number of iterations;

[0060] Figure 2 It is a change graph of the weight of the near-infrared spectrum wavelength selection in the BOSS algorithm in the present embodiment, which changes with the number of iterations;

[0061] Figure 3 It is a 3D view of the parameter selection result when the particle swarm optimization algorithm is used to seek the support vector regression penalty parameter c and the kernel function parameter g in the present embodiment. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0063] Embodiment: A near-infrared spectrum distance correction method for boxes, as shown in Figure 1 The method comprises the following steps:

[0064] Step S1, fitting the collected near-infrared spectrum signal to obtain the best roughness and roughness by introducing a penalty factor, and obtaining the optimal fitting baseline based on the best roughness and roughness;

[0065] Specifically, step S1 comprises the following steps:

[0066] Step S11, the near infrared spectrum data y obtained by the infrared spectrometer is fitted to obtain a fitting signal z, and the length value of the fitting signal z is the same as that of the near infrared spectrum data, both being N;

[0067] Step S12, based on step S11, the distortion degree F of the fitting signal z and the actual signal and the roughness R of the fitting signal z are obtained:

[0068] (1)

[0069] In formula (1), i is the i-th near infrared spectrum data, i is a natural number starting from 3;

[0070] Step S13, a penalty factor λ is introduced to seek the optimal value between the roughness R and the distortion degree F, and formula (2) is minimized to obtain the optimal fitting baseline z:

[0071] (2)

[0072] The matrix of formula (2) is:

[0073] (3)

[0074] Wherein, D is a second-order difference matrix coefficient term; the superscript T is a transpose operator of the matrix; is a second-order difference equation, and the optimal fitting baseline z is an independent variable of the equation;

[0075] As can be seen from formula (1), the lower the roughness R of the fitting signal z obtained by fitting, the greater the distortion degree F between the fitting signal z and the actual signal, and the roughness R and the distortion degree F are two quantities that restrict each other. Therefore, in order to balance the distortion degree F and the roughness R of the fitting signal z, a penalty factor λ is introduced to the roughness R of the fitting signal z to seek the optimal value between the roughness R and the distortion degree F, and the optimal fitting baseline z is obtained by minimizing formula (2).

[0076] As can be seen from formula (2), the greater the value of the penalty factor λ, the higher the weight of the roughness R in the optimal fitting baseline z. The smaller the value of the optimal fitting baseline z, the smoother the fitting signal z.

[0077] In formula (3), D is a second-order difference matrix coefficient term. When the length of the obtained actual infrared spectrum signal is 6, the second-order difference matrix coefficient term D is represented by a matrix as follows:

[0078] ;

[0079] In order to make the fitting signal z more smooth, the partial derivative of the optimal fitting baseline z in formula (3) is taken, and the partial derivative is set to 0, and the formula after the partial derivative is shown as formula (4):

[0080] (4)

[0081] Further solving formula (4) can obtain the expression of the fitting signal z as formula (5):

[0082] (5)

[0083] Wherein, I is a unit matrix, and the row and column are the same as the length value N of the collected near-infrared spectrum data y.

[0084] In step S2, different weight values are predicted by linear regression through nonlinear transformation of the generalized logistic function, and the collected near-infrared spectrum data is assigned weights according to the optimal fitting baseline. Specifically:

[0085] Formula (5) is a very typical Whittaker filter. In order to apply formula (5) in baseline correction, a weight vector w is introduced on the basis of formula (5). When the collected near-infrared spectrum data signal is lower than the optimal fitting baseline signal , the baseline of the peak-free region is assigned similar weights; when the near-infrared spectrum data signal is higher than the optimal fitting baseline signal , the weight is set to 0, so as to realize the smoothing processing of the collected near-infrared spectrum signal. At this time, formula (5) is converted into formula (7):

[0086] (7)

[0087] In formula (7), W is the weight matrix of the weight vector w, which is solved by Whittaker filtering.

[0088] In the above method, when the collected near-infrared spectrum data signal is lower than the optimal fitting baseline signal , the background noise is equally distributed above and below the baseline for the baseline of the peak-free region, so that the baseline of this region has similar weights, so that the baseline of the peak-free region is not underestimated; when the near-infrared spectrum data signal is higher than the optimal fitting baseline signal , the weight value of the baseline of this region is set to 0, so as to become part of the peak value of this region.

[0089] In order to realize the above method, the weight vector w is updated according to the asymmetric weight rule,

[0090] (6)

[0091] In formula (6), and are the mean and standard deviation of respectively; set , is a part of d, only defined in the region of ; The function is a generalized function, defined as follows:

[0092] , where e is the natural constant.

[0093] Step S3, using BOSS algorithm to correct the near infrared spectrum data variable screening to determine the optimal subset of variables.

[0094] The BOSS algorithm uses the idea of weighted bootstrap sampling (WBS) and model population analysis (MPA) to select variables. This method generates n subsets of variables through bootstrap sampling (BSS), removes duplicate variables in the selected variables, and performs regression through partial least squares. Then, according to the absolute value of the regression coefficient, the weight of each variable is determined. Apply weighted bootstrap sampling to generate a sub-model according to the weight, and then analyze the new weight generated by the sub-model in the iterative variable process according to the model population analysis (MPA). In the process of BOSS feature variable optimization, follow the rules of soft shrinkage, and do not directly remove less important variables, but assign them a smaller weight to preserve the utilization of data and avoid some useful data for the final model prediction result being removed. This variable selection method runs in an iterative manner until the final number of variables reaches 1. According to the cross-validation in the BOSS algorithm, the root mean square error of cross-validation (RMSECV) is calculated and the variable subset with the smallest cross-validation root mean square error (i.e. prediction error) is selected as the optimal subset.

[0095] Specifically, step S3 includes the following steps:

[0096] Step S31, set the near infrared spectrum data as a data matrix, such as data matrix, each row contains n samples, and each column includes p variables. The size of vector y is , representing the attribute of interest of the sample;

[0097] Step S32, applying the bootstrap method to the variable space to generate K subsets of variables, in each data set, extracting the variables selected by the bootstrap method, and removing the data samples with collinearity to ensure the uniqueness of the sample data, the number of sample data replaced by the bootstrap method (BBS) is the same as the total number of variable subsets, and each variable is assigned the same weight;

[0098] Step S33, training K-PLS sub-models through the variable subsets in step S32, calculating the prediction error of the sub-models to extract the best model with the minimum prediction error;

[0099] Step S34, calculating the regression coefficients of each variable subset, making each value in the regression vector positive, making each regression vector have the same unit length, and summing the calculated regression vectors to obtain new weights for the sample data; the summation formula is as follows:

[0100] (8)

[0101] In formula (8), K is the number of sub-models; is the absolute value of the normalized regression coefficient of variable i in the Kth sub-model

[0102] Step S35, applying the weighted bootstrap method to generate new subsets according to the new weights; this process is the same as the bootstrap method (BBS), which uses characteristic variables to generate new subsets. The amount of sample data in the weighted bootstrap method (WBS) is determined according to the average number of variables in step S34. Therefore, the number of variables in the replaced variable subset is about 0.632 times that of the previous one. This step makes the sample data with larger absolute values of regression coefficients in the selected minimum prediction error model more accurate.

[0103] Step S36, repeating steps S33-S35 until the number of new subsets is 1, and in the iteration process, the subset with the minimum prediction error is selected as the optimal variable set.

[0104] In this embodiment, the number of cross-validation in the weighted bootstrap method is 3, and the number of bootstrap is 25; Figure 1 The variable selection number, root mean square error, and regression coefficient change with the number of iterations, and the root mean square error is the smallest at the sixteenth iteration, and the optimal subset is selected at the fifth iteration; Figure 2 In the bootstrap of soft shrinkage (BOSS) method, the weight of the near-infrared spectrum wavelength selected changes with the number of iterations, and the weight selected in the range of 3000-3500 frequency bands is the highest.

[0105] Step S4, using the particle swarm algorithm to optimize the penalty parameter and kernel function parameter of support vector regression, and establishing a prediction model.

[0106] Specifically, step S4 includes the following steps:

[0107] Step S41, normalizing the near-infrared spectrum data set after variable selection;

[0108] Step S42, randomly initializing the penalty parameter c and the kernel function parameter g of the support vector machine as the initial particle position and particle velocity, respectively, and then calculating according to formula (10) to calculate the fitness value of the particle to update the velocity and position information of the particle:

[0109] (10)

[0110] In formula (10), is the inertia weight; is the velocity of the particle x in the jth dimension at the kth iteration; is the velocity of the particle x in the jth dimension at the (k+1)th iteration; and is the acceleration constant; and is a random number distributed in the interval [0, 1]; is the jth component of the individual historical optimal position of the particle x at the kth iteration; is the jth component of the group historical optimal position of the particle group at the kth iteration; is the position of the particle x in the jth dimension at the kth iteration; is the position of the particle x in the jth dimension at the (k+1)th iteration; j=1, 2, …, J; x=1, 2, …, X;

[0111] Step S43, finding the individual extreme value of the particle and the group extreme value , then updating the velocity and the position of the particle; and updating the individual extreme value and the group extreme value according to the fitness of the particles in the new population;

[0112] Step S44, repeating steps S402 and S403 until the optimal particle point in the population is found;

[0113] Step S45, establishing a training prediction model according to the optimal penalty parameter c and the kernel function parameter g found, and predicting the test set;

[0114] Step S46, according to the prediction results, the index analysis, such as mean absolute error, determination coefficient, mean square error and root mean square error.

[0115] Table 1 modeling method modeling effect table

[0116]

[0117] In Table 1, the mean square error and the root mean square error can represent the generalization and accuracy of the modeling, the determination coefficient represents the fitting degree of the regression analysis model and the prediction ability of the regression analysis model. From Table 1, it can be observed that after the particle swarm optimization of the kernel function parameter g and the penalty parameter c, the root mean square error and the mean square error are smaller, the determination coefficient is larger, the generalization and accuracy of the model are higher, and the response to the characteristic variable is more rapid.

[0118] After the optimization of the penalty parameter c and the kernel function parameter g by the particle swarm optimization algorithm, the prediction accuracy is further improved. Table 2 is the values of the two parameters of the support vector machine found by the 5-fold cross-validation method and the particle swarm optimization algorithm respectively.

[0119] Table 2 comparison table of parameters found by two optimization methods

[0120]

[0121] Figure 3 The 3D view of the parameter selection results when the particle swarm optimization algorithm is used to seek the support vector regression penalty parameter c and the kernel function parameter g. The lines of various colors in the figure represent the contour lines, which are the values of the mean square error changing with the penalty parameter c and the kernel function parameter g parameters of the double frequency, and the lines of the same color are the same height.

[0122] The above only describes the preferred embodiments of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solutions falling within the concept of the present application shall be considered as falling within the protection scope of the present application. It should be noted that for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application shall also be considered as falling within the protection scope of the present application.

Claims

1. A method for near infrared spectroscopy distance correction for a box, characterized by, Includes the following steps: Step S1: Fit the acquired near-infrared spectral signal, introduce a penalty factor to obtain the optimal distortion and roughness, and obtain the optimal fitting baseline based on the optimal distortion and roughness. Step S2: Linear regression is performed using the nonlinear transformation of the generalized logistic function to predict different weight values, and weights are assigned to the collected near-infrared spectral data based on the optimal fitting baseline. Step S3: Use the BOSS algorithm to screen variables in the corrected near-infrared spectral data to determine the optimal subset of variables; Step S4: Optimize the penalty parameters and kernel function parameters of support vector regression using the particle swarm optimization algorithm, and establish a prediction model.

2. The method for near infrared spectroscopy distance correction of a box according to claim 1, wherein, Step S1 includes the following steps: Step S11: Fit the near-infrared spectral data y acquired by the infrared spectrometer to obtain the fitted signal z. Let the length of the fitted signal z and the length of the near-infrared spectral data both be N. Step S12, based on step S11, obtain the distortion F between the fitted signal z and the actual signal, and the roughness R of the fitted signal z: （1） In equation (1), i represents the nth near-infrared spectral data. Step S13: By introducing a penalty factor λ, the optimal value between roughness R and distortion F is sought, and the optimal fitting baseline z is obtained by minimizing equation (2): （2） The matrix of equation (2) is: （3） where D is a second-order difference matrix coefficient term, and a superscript T is a transpose operator of a matrix; is a second-order difference equation.

3. The method for near infrared spectroscopy distance correction of a box according to claim 2, wherein, The optimal fitted baseline z is smoothed, including: Take the partial derivative with respect to the optimal fitted baseline z, and set the partial derivative to 0: （4） After solving, the expression for the fitted signal z can be obtained as follows: （5） Where I is the identity matrix, and its rows and columns are the same as the length N of the acquired near-infrared spectral data y.

4. The method for near infrared spectroscopy distance correction of a box according to claim 1, wherein, In step S2, the rule for assigning weights to near-infrared spectral data is as follows: when the near-infrared spectral data signal is lower than the optimal fitted baseline signal, a similar weight is assigned to the baseline of its peakless region; when the near-infrared spectral signal is higher than the optimal fitted baseline signal, a weight of 0 is assigned to it.

5. The method for near infrared spectroscopy distance correction of a box according to claim 3, wherein, Step S2 includes the following steps: Step S21: During the iteration process, update the weight vector w according to the asymmetric weighting rule. （6） In formula (6), and are respectively the mean and standard deviation; set , is a part of d, only in the region defined; the function is a generalized function, defined as follows: ; where e is the natural constant; Step S22: Based on the weighting rules for near-infrared spectral data, the expression for the fitted signal z is transformed: （7） In equation (7), W is the weight matrix of the weight vector w; Solving using Whittaker filtering.

6. The method for near infrared spectroscopy distance correction of a box according to claim 1, wherein, Step S3 includes the following steps: Step S31: Set the near-infrared spectral data into a data matrix; Step S32: Apply the autonomous sampling method to generate K variable subsets in the variable space. In each dataset, extract the variables selected by the autonomous sampling method and remove data samples with collinearity. The number of sample data replaced by the autonomous sampling method is the same as the total number of variable subsets. Each variable is assigned the same weight. Step S33: Train the K-PLS sub-model using the subset of variables from step S32, calculate the prediction error of the sub-model, and extract the optimal model with the minimum prediction error. Step S34: Calculate the regression coefficients for each subset of variables, make each value in the regression vector positive, so that each regression vector has the same unit length, and sum the calculated regression vectors to obtain new weights for the sample data. Step S35, a new subset is generated according to the new weight by applying the weighted self-adopted sampling method; Step S36, steps S33-S35 are repeated until the number of new subsets is 1, and in the iteration process, the subset with the minimum prediction error is selected as the optimal variable set.

7. The method for near infrared spectroscopy distance correction of a box according to claim 6, wherein, In step S34, the calculated regression vector is summed up as follows: （8） In formula (8), K is the number of sub-models; is the absolute value of the normalized regression coefficient of variable i in the Kth sub-model.

8. The method for near infrared spectroscopy distance correction of a box according to claim 1, wherein, Step S4 includes the following steps: Step S41, the variable-selected near-infrared spectrum data set is normalized; Step S42, the penalty parameter c and the kernel function parameter g of the support vector machine are randomly initialized as the initial particle position and particle velocity, and then calculated according to formula (10) to calculate the fitness value of the particle to update the velocity and position information of the particle: （10） in formula (10), is the inertia weight; is the velocity of particle x in the jth dimension at the kth iteration; is the velocity of particle x in the jth dimension at the (k+1)th iteration; and is the acceleration constant; and is a random number distributed in the interval [0, 1]; is the jth component of the individual historical best position of particle x at the kth iteration; is the jth component of the population historical best position of the particle population at the kth iteration; is the position of particle x in the jth dimension at the kth iteration; is the position of particle x in the jth dimension at the (k+1)th iteration; j = 1, 2, …, J; x = 1, 2, …, X; Step S43, find particle individual extremum and population extremum Then update the velocity and position of the particle; update the individual extremum and population extremum according to the fitness of the particles in the new population; Step S44, steps S402 and S403 are repeated until the optimal particle point in the population is found; Step S45, a training prediction model is established according to the optimal penalty parameter c and the kernel function parameter g found, and a test set is predicted; Step S46, index analysis is performed according to the prediction result.

Images