A risk assessment method for edible oil based on optimized ELM

Abstract

The present invention discloses an edible oil risk assessment method based on optimized ELM, which belongs to the field of food safety. First, the content data of risk factors in the edible oil to be tested is obtained, and then the filter processing is performed using the Bayesian optimized wavelet threshold method. The filtered data is subjected to grey correlation analysis to calculate the weights of each risk indicator to formulate a risk value label. The original safety detection data is used as the input of an extreme learning machine, and the risk value label is used as the expected output. The parameters of the extreme learning machine are optimized using Bayesian to obtain a predicted risk value. Finally, the predicted risk value is divided into risk levels using a fuzzy comprehensive analysis method. For a new edible oil to be evaluated, after risk evaluation is performed using the evaluation model, its corresponding risk level is directly obtained. The present invention improves the accuracy of risk assessment and helps relevant departments strengthen supervision of relevant edible oil production enterprises, thereby improving the safety level of edible oil and reducing edible oil safety risks.

Description

Technical Field

[0001] The present invention relates to the field of food safety, and in particular to an edible oil risk assessment method based on optimized ELM (Extreme Learning Machine). Background Art

[0002] Food safety issues have always been an important factor affecting public health, economic development and the stability of human society. Edible oil, as an indispensable raw material in people's daily diet, has a quality that is closely related to human health. In order to reduce the occurrence of accidents, risk assessment and early warning of edible oil are very important, which can help relevant departments to carry out targeted risk control.

[0003] Establishing a high-precision and robust risk assessment model is crucial for edible oil risk control. When measuring the actual content of edible oil risk factors, measurement noise and errors are inevitable due to issues such as sensor performance and instrument damage. If left unaddressed, these errors can lead to inaccurate risk assessment results. Traditional risk assessment models don't account for this, compromising both accuracy and robustness.

[0004] Currently, the Delphi method or the analytic hierarchy process (AHP) is mainly used in the field of food safety risk assessment. However, these methods rely too much on subjective factors to determine the weights of risk indicators and are unable to determine the weights of various risk indicators based on existing hazard content data, which affects their practicality. Moreover, after the weights are determined, they are unable to perform risk assessment based on a large amount of hazard content data, and it is difficult to form a quantitative, objective, and universal standard. Summary of the Invention

[0005] To address the limitations of the existing technology, the present invention proposes an edible oil risk assessment method based on optimized ELM. Targeting the existing edible oil safety test data, the robustness and fault tolerance of the evaluation model are improved through filtering processing. The Bayesian optimization algorithm is used to optimize the parameters of the filtering module and the evaluation module respectively, thereby improving the accuracy of the evaluation results. The quantitative evaluation results are subjected to fuzzy comprehensive analysis, and the risk levels are divided for reference by relevant departments, thereby effectively controlling the edible oil risks.

[0006] The edible oil risk assessment method based on optimized ELM has the following specific steps:

[0007] Step 1: For the edible oil to be tested, obtain its safety test data through the sensor and extract the content data of the risk factor from it;

[0008] Risk factors include acid value, peroxide value, arsenic, lead, aflatoxin b1 and benzo[a]pyrene.

[0009] Step 2: Use the Bayesian optimized wavelet threshold method to filter the content data of various risk factors;

[0010] The specific process is:

[0011] Step 201: Optimize the threshold value and mother wavelet function type of each layer of wavelet components using the Bayesian optimization method;

[0012] The number of wavelet decomposition layers is determined by empirical methods.

[0013] Step 202: gradually selecting upper and lower limits of each parameter during the optimization process and inputting them into the Bayesian optimization algorithm;

[0014] Step 203: using the optimized mother wavelet function, perform wavelet decomposition on the content data of various risk factors to obtain decomposition results;

[0015] The calculation formula of wavelet decomposition is as follows:

[0016]

[0017]

[0018] υ=λ i γ

[0019]

[0020] Where σ is the noise estimation variance, D i is the high-frequency subsequence with index i after decomposition, γ is the estimated threshold of noise in each subsequence calculated based on the estimated noise variance, T is the length of the denoised sequence, and the threshold υ is obtained by adjusting the estimated threshold γ by the value of the proportional coefficient λ. i is the proportional coefficient of the indirect adjustment threshold, D i,t The high-frequency subsequence index of the decomposition result of the i-th layer is t, and λ is obtained through the Bayesian optimization algorithm.

[0021] Step 204: extract the detail coefficient class of the first layer in the decomposition result and estimate the standard deviation of the noise; use the threshold value of the optimized wavelet component of each layer to filter out the estimated noise, and then perform wavelet reconstruction to obtain the content data of various risk factors of edible oil after filtering.

[0022] The noise in the data is distributed in various high-frequency components, mainly concentrated in the first three layers of decomposition. As the number of decomposition layers increases, the noise contained in the high-frequency components gradually decreases.

[0023] Step 3: Perform dimensionless processing on the filtered data to obtain the relative risk assessment value corresponding to each test result in each type of risk factor;

[0024] Dimensionless processing refers to the ratio of the actual detected content value of the risk factor to the critical limit of the indicator; that is:

[0025] P ij =x ij / a j

[0026] Where, P ij is the i-th test result of the j-th risk factor (x ij ) Relative risk value after dimensionless processing (i=1,2,...,n,j=1,2,...,m); a j Indicates the critical limit of the indicator; there are m types of risk factors, and each type of risk factor has n test results.

[0027] Step 4: Perform grey correlation analysis on the relative risk evaluation value to obtain the weight corresponding to each test result in each type of risk factor, and then combine it with the relative risk evaluation value to obtain a comprehensive risk evaluation value;

[0028] Right now:

[0029]

[0030] Where Y is the comprehensive risk assessment sequence, and P is the matrix of relative risk assessment values. m ] T is the weight vector, and the calculation formula is: γ pq is the correlation coefficient between the p-th risk factor and the q-th risk factor calculated using grey correlation analysis;

[0031] Step 5: Input the edible oil safety test data into the extreme learning machine (ELM), use the comprehensive risk assessment value as the expected output of the ELM, continuously train the ELM, and use Bayesian optimization to optimize the number of hidden layer nodes of the ELM so that the risk value predicted by the ELM is as close to the expected value as possible, thus forming an evaluation model.

[0032] Step 6: Calculate the membership vectors of the five risk levels corresponding to the comprehensive risk evaluation value of each test result in each risk factor category, and classify the test result into the corresponding risk level according to the maximum membership principle;

[0033] The comprehensive risk assessment value y of the i-th item i The membership degree h iq , q=1,2,3,4,5, the calculation method is as follows:

[0034] When q=1,

[0035]

[0036] When q=2,3,4,

[0037]

[0038] When q=5,

[0039]

[0040] In the above formula, Z q Indicates the representative value corresponding to each of the five risk levels, which correspond to the fuzzy assessment levels of micro risk, low risk, medium risk, high risk and extremely high risk respectively.

[0041] Step 7: For the new edible oil to be evaluated, use the evaluation model to conduct risk assessment and classify it into the corresponding risk level.

[0042] The advantages of the present invention are:

[0043] (1) The present invention provides an edible oil risk assessment method based on optimized ELM, which uses wavelet threshold method to filter the original data, thereby improving the robustness and fault tolerance of the evaluation model.

[0044] (2) The present invention provides an edible oil risk assessment method based on optimized ELM, which combines GRA (Grey correlation analysis) with ELM to evaluate the risk of edible oil, and uses the Bayesian algorithm to optimize the filtering module and evaluation module of the model respectively, thereby improving the accuracy of the final evaluation result. In the process of Bayesian parameter optimization, reasonable upper and lower limits of each parameter are gradually selected, reducing the complexity of the algorithm.

[0045] (3) The present invention proposes an edible oil risk assessment method based on optimized ELM, which performs fuzzy comprehensive analysis on the quantitative comprehensive evaluation results and divides the risk levels for reference by relevant departments, thereby carrying out effective risk control. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of an edible oil risk assessment method based on optimized ELM of the present invention;

[0047] Figure 2 This is a schematic diagram of the results of wavelet decomposition of the content data of various risk factors in the present invention;

[0048] Figure 3 This is a flow chart of the edible oil risk assessment method based on optimized ELM in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The present invention will be described in detail below with reference to the accompanying drawings.

[0050] The edible oil risk assessment method based on optimized ELM described in the present invention is as follows: Figure 1 The specific steps are as follows:

[0051] Step 1: For the edible oil to be tested, obtain its safety test data through the sensor and extract the content data of the risk factor from it;

[0052] Risk factors include acid value, peroxide value, arsenic, lead, aflatoxin b1 and benzo[a]pyrene.

[0053] Step 2: Use the Bayesian optimized wavelet threshold method to filter the content data of various risk factors;

[0054] The specific process is:

[0055] Step 201: Optimize the threshold value and mother wavelet function type of each layer of wavelet components using the Bayesian optimization method;

[0056] The wavelet threshold parameters are optimized according to the Bayesian optimization method. The parameters include the threshold of each wavelet layer after wavelet decomposition and the type of mother wavelet function in wavelet decomposition. The number of wavelet decomposition layers is determined by empirical method.

[0057] Step 202: During the optimization process, reasonable upper and lower limits of each parameter are gradually selected and input into the Bayesian optimization algorithm to improve the accuracy of the optimized parameters and reduce the complexity of the algorithm;

[0058] Step 203: Using the optimized selected mother wavelet function, perform wavelet decomposition on the content data of each risk factor to obtain decomposition results; extract the standard deviation of the detail coefficient class estimation noise of the first layer in the decomposition results;

[0059] The calculation formula of wavelet decomposition is as follows:

[0060]

[0061]

[0062] υ=λ i γ

[0063]

[0064] Where σ is the noise estimation variance, median represents the median of the smoothed sequence, 0.6745 is the Gaussian noise standard variance adjustment coefficient, and D iis the high-frequency subsequence with index i after decomposition, γ is the estimated threshold of noise in each subsequence calculated based on the estimated noise variance, T is the length of the denoised sequence, and the threshold υ is obtained by adjusting the estimated threshold γ by the value of the proportional coefficient λ. i Is the proportional coefficient of the indirect adjustment threshold, the value range is λ i ∈(0,1); i=1,...,N; D i,t The high-frequency subsequence index of the i-th layer decomposition result is t, and λ is obtained through the Bayesian optimization algorithm.

[0065] The wavelet decomposition results are as follows Figure 2 As shown in Figure 1, the standard deviation of the noise is estimated by extracting the detail coefficient class of the first layer. Because the noise in the data is distributed in various high-frequency components, it is mainly concentrated in the first three layers of decomposition. As the number of decomposition layers increases, the noise contained in the high-frequency components gradually decreases.

[0066] Traditional wavelet thresholding methods generally use a global threshold to filter out noise. The present invention sets a threshold for filtering each layer of wavelet components, and the threshold of each layer is selected through a Bayesian optimization algorithm, making the result more accurate and more adaptable to data.

[0067] Step 204: For the estimation noise, the threshold value of each layer of optimized wavelet component is used to filter out the estimation noise, and then wavelet reconstruction is performed to obtain the content data of various risk factors of edible oil after filtering.

[0068] Step 3: Perform dimensionless processing on the filtered data to obtain the relative risk assessment value corresponding to each test result in each type of risk factor;

[0069] Dimensionless processing refers to the ratio of the actual detected content value of the risk factor to the critical limit of the indicator; that is:

[0070] P ij =x ij / a j

[0071] Where, P ij is the i-th test result of the j-th risk factor (x ij ) Relative risk value after dimensionless processing (i=1,2,...,n,j=1,2,...,m); a j Indicates the critical limit of the indicator; there are m types of risk factors, and each type of risk factor has n test results.

[0072] Step 4: Perform grey correlation analysis on the relative risk evaluation value to obtain the weight corresponding to each test result in each type of risk factor, and then combine it with the relative risk evaluation value to obtain a comprehensive risk evaluation value;

[0073] Right now:

[0074]

[0075] Where Y is the comprehensive risk assessment sequence, and P is the matrix of relative risk assessment values. m ] T is the weight vector, and the calculation formula is: γ pq is the correlation coefficient between the p-th risk factor and the q-th risk factor calculated using grey correlation analysis;

[0076] Step 5: Input the edible oil safety test data into the extreme learning machine (ELM), use the comprehensive risk assessment value as the expected output of the ELM, continuously train the ELM, and use Bayesian optimization to optimize the number of hidden layer nodes of the ELM so that the risk value predicted by the ELM is as close to the expected value as possible, thus forming an evaluation model to conduct risk assessment on the edible oil to be evaluated;

[0077] Step 6: Calculate the membership vectors of the five risk levels corresponding to the comprehensive risk evaluation value of each test result in each risk factor category, and classify the test result into the corresponding risk level according to the maximum membership principle;

[0078] The risk factor of edible oil belongs to the negative effect type indicator, that is, this risk factor cannot exceed the key limit of the indicator. Therefore, the membership function can adopt the descending half-step distribution model. For the i-th comprehensive risk assessment value y i The membership degree h iq , q=1,2,3,4,5, the calculation method is as follows:

[0079] When q=1,

[0080]

[0081] When q=2,3,4,

[0082]

[0083] When q=5,

[0084]

[0085] In the above formula, Z q Indicates the representative value corresponding to each of the five risk assessment levels, corresponding to the fuzzy comment levels of slight risk, low risk, medium risk, high risk and very high risk respectively;

[0086] Step 7: For the new edible oil to be evaluated, use the evaluation model to conduct risk assessment and classify it into the corresponding risk level.

[0087] Example:

[0088] like Figure 3 As shown, in this embodiment, first, the safety test data of edible oil obtained by means of sensors or the like is input, including the content data of six key risk factors, namely acid value, peroxide value, arsenic, lead, aflatoxin b1 and benzo[a]pyrene;

[0089] Then, the filtering module is used to perform filtering using the Bayesian optimized wavelet threshold method;

[0090] like Figure 3 As shown, the specific process is:

[0091] Step 201: Optimize the wavelet threshold parameters according to the Bayesian optimization method. When the number of wavelet decomposition layers exceeds 5, the parameter optimization effect is negligible. Therefore, in order to reduce the redundancy of the algorithm, the number of layers is set to 5 according to the empirical method.

[0092] Step 202: During the parameter optimization process of step 201, reasonable upper and lower limits of each parameter are gradually selected and input into the Bayesian optimization algorithm to improve the accuracy of the optimized parameters and reduce the complexity of the algorithm.

[0093] Specifically, the parameters of the wavelet threshold method directly determine the quality of the data filtering effect. In this example, a Bayesian optimization method based on MATLAB is selected: Practical Bayesian Optimization (PBO).

[0094] Bayesian optimization requires the definition of an objective function. For the filtering module, the objective function is the root mean square error (RMSE) between the filtered data and the true value.

[0095]

[0096] Among them, m is the number of input samples, y i (w) is the filtered value, y i is the true value.

[0097] Then the objective function is expressed as:

[0098]

[0099] The process of the Bayesian optimization algorithm is:

[0100] The input is: objective function g(w), parameter initial value x0, parameter hard bounds LB and PB, parameter reasonable upper and lower bounds PLB and PUB

[0101] The output is: optimal parameter w*

[0102] First initialize: Preliminary evaluation of the objective function g(w);

[0103] Then, x search Substitute it into the objective function g(w) for evaluation and stop when the requirements are met; otherwise, continue to iterate and reset the hyperparameters if necessary;

[0104] x search ←search oracle

[0105] If the search is unsuccessful, calculate the polling set P k , evaluate the objective function g(w); otherwise, after k successful iterations, update w at this time k+1 ;

[0106] If successful:

[0107] otherwise,

[0108] Until the final output: Finish.

[0109] Step 203: performing wavelet decomposition according to the optimized selected mother wavelet function to obtain a decomposition result, and then extracting the detail coefficient class of the first layer to estimate the standard deviation of the noise.

[0110] Step 204: For the estimated noise in step 203, the noise is filtered out using the wavelet threshold after Bayesian optimization of each layer, and then wavelet reconstruction is performed to obtain the edible oil risk factor content data that meets the filtering requirements.

[0111] Then, the evaluation module is used to perform dimensionless processing on the filtered data to obtain the relative risk evaluation value. Grey correlation analysis is then performed to obtain the weight corresponding to each risk factor. The weight is combined with the relative risk evaluation value to obtain the comprehensive risk evaluation value, which is used as the expected output of the ELM evaluation model, and the edible oil safety test data is used as the input of the evaluation model.

[0112] Specifically:

[0113] Step 301: The dimensionless processing evaluation value refers to the relative risk value represented by the ratio of the actual detection value of the hazard factor to the critical limit of the indicator;

[0114] Step 302: Perform grey relational analysis on the dimensionless data to obtain the risk factor weight vector W = [w1, w2, ..., w m ] T .

[0115] Specifically, a reference vector and a comparison vector are obtained, wherein the reference vector P1 and the comparison vector P i They are:

[0116] P1={P1(1),P1(2),...,P1(n)};

[0117] P j ={P j (1),P j (2),...,P j (n)};

[0118] Where n is the number of samples, j = 1, 2, ..., m, and m is the number of risk indicators.

[0119] This embodiment calculates the grey correlation coefficient, P at time k j The grey correlation coefficients of P(k) and P1(k) are as follows:

[0120]

[0121] Among them, ξ j (k) is the grey relational coefficient, ρ∈(0,1). Adjusting the parameter ρ can enhance the differences between the coefficients. The sequence P1 and the sequence P j The correlation coefficient between them is:

[0122]

[0123] In order to ensure the accuracy of the results, each risk indicator is used as a reference sequence once, and the correlation coefficient matrix of all risk indicators can be obtained:

[0124]

[0125] Then the weight vector W is:

[0126]

[0127] Where i = 1, 2, ..., m, j = 1, 2, ..., m, m is the number of risk factors;

[0128] Multiply the weight vector W obtained by the above formula by the relative risk matrix P of each indicator obtained in step 301, that is:

[0129]

[0130] Where Y=[y1,y2,...,y n ] T is the comprehensive risk assessment sequence, P is the relative risk assessment matrix;

[0131] Step 303: Using the data Y obtained in step 302 as the expected output of the ELM and the edible oil safety test data as the input, the number of hidden layer nodes of the extreme learning machine is optimized by Bayesian optimization to form an edible oil risk assessment model.

[0132] Specifically, if Figure 3 As shown in the evaluation module, the edible oil detection data is input into the ELM. The weights w and bias b on the hidden layer nodes of the ELM are randomly generated according to any continuous probability distribution and have nothing to do with the training data. This is different from common neural networks such as BP and RBF. As a result, ELM has great advantages over traditional neural networks in terms of efficiency, generalization and robustness. Therefore, the Bayesian optimization algorithm only needs to optimize the number of hidden layer nodes of the ELM to make the predicted risk value as close to the expected value as possible. After the evaluation model is formed, the risk evaluation of the edible oil to be evaluated is performed.

[0133] For this embodiment, the objective function of the Bayesian optimization algorithm has also changed accordingly, namely:

[0134]

[0135] Among them, m is the number of input samples, y i (w) is the predicted value of ELM, y i is the expected value.

[0136] Finally, for the i-th comprehensive risk value (y i ) Calculate its membership degree h iq Finally, the risk level of the test result is determined according to the maximum membership principle.

[0137] The representative values ​​corresponding to each level of the 5-scale evaluation level, namely {Z1, Z2, Z3, Z4, Z5} = {0.25, 0.5, 0.75, 1.00, 2.00}, correspond to the fuzzy evaluation levels of slight risk, low risk, medium risk, high risk and extremely high risk respectively.

[0138] For new edible oils to be evaluated, risk assessment is performed using the evaluation model and the oils are classified into corresponding risk levels.

Claims

1. A risk assessment method for edible oil based on optimized ELM, characterized in that: Specifically include First, for the edible oil to be tested, the safety test data is obtained through the sensor, and the content data of risk factors are extracted from it; and the content data of various risk factors are filtered using the Bayesian optimized wavelet threshold method; Then, the filtered data is dimensionless processed, and the risk evaluation of each detection result in each risk factor is performed using the combination of grey relational analysis and extreme learning machine (ELM) to obtain a quantitative comprehensive risk evaluation value. Calculate the membership vector of the five risk levels corresponding to the comprehensive risk evaluation value of each test result, and classify each test result into the corresponding risk level according to the maximum membership principle; The membership function adopts the descending half-ladder distribution model; Next, the edible oil safety test data is input into the ELM, and the comprehensive risk evaluation value of each test result is used as the expected output of the ELM. The ELM is continuously trained, and the number of hidden layer nodes of the ELM is optimized using Bayesian optimization to make the risk value predicted by the ELM as close as possible to the expected value, thus forming an evaluation model. Finally, for the new edible oil to be evaluated, the evaluation model is used to conduct risk assessment, and the comprehensive risk assessment value of its test results is obtained, which is then matched to the risk level to obtain the final risk assessment result.

2. The edible oil risk assessment method based on optimized ELM according to claim 1, characterized in that: The risk factors include six categories: acid value, peroxide value, arsenic, lead, aflatoxin b1 and benzo[a]pyrene.

3. The edible oil risk assessment method based on optimized ELM according to claim 1, characterized in that: The Bayesian optimized wavelet threshold method is used to filter the content data of various risk factors. The specific process is as follows: Step 201: define the objective function of Bayesian optimization as the root mean square error (RMSE) between the filtered data and the true value; use the Bayesian algorithm to optimize the threshold of each wavelet component after wavelet decomposition and adaptively select the mother wavelet function type; and use the empirical method to determine the number of wavelet decomposition layers. Step 202: gradually selecting upper and lower limits of each parameter during the optimization process and inputting them into the Bayesian optimization algorithm; Step 203: using the optimized mother wavelet function, perform wavelet decomposition on the content data of various risk factors to obtain decomposition results; The calculation formula of wavelet decomposition is as follows: y=l i c Where σ is the noise estimation variance, median represents the median of the smoothed sequence, and D i is the high-frequency subsequence with index i after decomposition, γ is the estimated threshold of noise in each subsequence calculated based on the estimated noise variance, T is the length of the denoised sequence, and the threshold υ is calculated by the proportional coefficient λ i The value of the estimated threshold γ is adjusted to obtain λ i is the proportional coefficient of the index i of the indirect adjustment threshold, D i,t represents the data of the high-frequency subsequence indexed as t with index i, λ i Obtained through Bayesian optimization algorithm; Step 204: extracting the standard deviation of the detail coefficient-like estimated noise of the first layer in the decomposition result; The threshold value of each optimized wavelet component is used to filter out the estimated noise, and then wavelet reconstruction is performed to obtain the content data of various risk factors of edible oil after filtering. The noise in the data is distributed in various high-frequency components, mainly concentrated in the first three layers of decomposition. As the number of decomposition layers increases, the noise contained in the high-frequency components gradually decreases.

4. The edible oil risk assessment method based on optimized ELM according to claim 1, characterized in that: The grey relational analysis combined with the extreme learning machine (ELM) is used to perform risk assessment on each test result in each risk factor. The specific process is as follows: First, the filtered edible oil safety test data is dimensionlessly processed to obtain the ratio of the actual detection content value of the risk factor to the critical limit of the risk factor indicator; that is: P ij =x ij / a j Where, P ij The i-th test result x for the j-th risk factor ij The ratio after dimensionless processing is i=1,2,...,n,j=1,2,...,m;a j Indicates the critical limit of the indicator; there are m types of risk factors, and each type of risk factor has n test results; Then: perform grey relational analysis on the dimensionless ratios, with each risk indicator acting as a reference sequence, and obtain the correlation coefficient matrix γ of all risk indicators: Then the weight vector W is: γ pq is the correlation coefficient between the p-th risk factor and the q-th risk factor calculated using grey correlation analysis; Finally, the weight vector W obtained from the above formula is multiplied by the dimensionless ratio P of each indicator to obtain the quantitative comprehensive risk assessment value; Right now: Where Y=[y1,y2,...,y n ] T is a comprehensive risk assessment sequence, W=[w1,w2,...,w m ] T is the weight vector.

5. The edible oil risk assessment method based on optimized ELM according to claim 1, characterized in that: In the membership vector, the comprehensive risk evaluation value y of the i-th item is i The membership degree h iq , q=1,2,3,4,5, the calculation method is as follows: When q=1, When q=2,3,4, When q=5, In the above formula, Z q Indicates the representative value corresponding to each of the five risk levels, which correspond to the fuzzy assessment levels of micro risk, low risk, medium risk, high risk and extremely high risk respectively.

Images