Grass carp muscle unsaturated fatty acid content nondestructive testing method based on scale characteristic wavelength

Abstract

The invention relates to the technical field of grass carp nondestructive testing, in particular to a grass carp muscle unsaturated fatty acid content nondestructive testing method based on scale characteristic wavelength. The method comprises the following steps: constructing a back propagation neural network, and training the network by using a plurality of samples containing scale wavelength information and unsaturated fatty acid information to obtain a prediction model; and the fresh grass carp scale wavelength information is input into the prediction model to obtain the unsaturated fatty acid content of the living grass carp muscle, so that rapid nondestructive detection can be realized, and the detection accuracy is relatively high.

Description

Technical Field

[0001] The present application relates to the technical field of nondestructive detection of grass carp, and in particular to a nondestructive detection method for the unsaturated fatty acid content in grass carp muscle based on characteristic wavelengths of scales. Background Art

[0002] Grass carp (Ctenopharyngodon idella) is distributed in the Yangtze River and Pearl River systems in China and their affiliated lakes, reservoirs and ponds. As an economic aquatic product, grass carp can meet human food needs. The polyunsaturated fatty acids (PUFAs) contained in grass carp are essential nutrients for the normal physiological functions and growth and development of aquatic animals. At the same time, PUFAs can effectively prevent cardiovascular diseases, regulate blood lipids, enhance immunity and promote brain development, and play an extremely important role in human health, nutrition and fishery.

[0003] However, the traditional method for detecting unsaturated fatty acids in grass carp requires scraping off scales and cutting muscles, which may damage the grass carp and even cause its death. In addition, there are problems such as cumbersome sample pretreatment process, long detection cycle, and high cost, making non-destructive testing impossible. Summary of the invention

[0004] Based on this, the application uses the characteristic wavelengths of the scales of a large number of grass carp samples for machine learning, establishes the relationship between the characteristic wavelengths of the scales of fresh grass carp and the content of unsaturated fatty acids in its muscle, and constructs a prediction model. By inputting the scale wavelength information of fresh grass carp into the prediction model, the content of unsaturated fatty acids in its muscle can be quickly known, which can not only achieve rapid non-destructive detection, but also has a high detection accuracy.

[0005] To this end, the embodiments of the present application disclose at least the following technical solutions:

[0006] The embodiment discloses a method for detecting unsaturated fatty acids in fresh grass carp muscle. The method comprises:

[0007] Obtaining a training set, the training set comprising a plurality of training samples, each of the training samples comprising at least one scale wavelength information and measured information of unsaturated fatty acids corresponding to the scale wavelength information;

[0008] Obtaining a back propagation neural network, the back propagation neural network comprising an input layer, a hidden layer and an output layer, the back propagation neural network having an initialized first connection weight between the input layer and the hidden layer, an initialized second connection weight connecting the hidden layer and the output layer, an initialized hidden layer threshold and an initialized output layer threshold;

[0009] Performing a forward propagation step, the forward propagation step comprising inputting the scale wavelength information in each of the training samples by the input layer, processing the scale wavelength information by the hidden layer to obtain a hidden layer output vector, and obtaining the expected output information of unsaturated fatty acids by the output layer according to the hidden layer output vector;

[0010] Obtaining a network prediction error of the output layer according to the measured information of the unsaturated fatty acid and the expected output information;

[0011] Back-propagating the network prediction error to the input layer, updating the first connection weight, the second connection weight, the hidden layer threshold, and the output layer threshold;

[0012] Utilizing all of the training samples to perform the forward propagation, obtain the network prediction error, and the back-propagation steps, training is performed according to a preset learning rate, and if it is detected that the network prediction error is less than or equal to an error threshold and / or the number of training times is greater than or equal to a preset number of training times, the training of the back-propagation neural network is completed;

[0013] Determine a prediction model according to the first connection weight, the second connection weight, the hidden layer threshold and the output layer threshold of the back propagation neural network when the training is completed;

[0014] Obtaining wavelength information of scales of fresh grass carp samples to be tested;

[0015] The scale wavelength information of the fresh grass carp sample to be tested is input into the prediction model to obtain the prediction information of unsaturated fatty acids. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic flow chart of a method for detecting unsaturated fatty acids in fresh grass carp muscle provided in an embodiment.

[0017] Figure 2 A schematic flow chart of the method for obtaining scale wavelength information provided in an embodiment.

[0018] Figure 3 A flowchart diagram of a method for obtaining flake wavelength information provided in some other embodiments.

[0019] Figure 4 This is a schematic diagram of the method flow of step S805 provided in the embodiment.

[0020] Figure 5 A schematic diagram of the structure of a back-propagation neural network provided in an embodiment.

[0021] Figure 6 The original spectral image of grass carp scales provided in the embodiment.

[0022] Figure 7 Provided for the embodiment Figure 6 The image shown is the image obtained through step S800.

[0023] Figure 8 The graph shows the predicted content (Predicted PUFAs content (mg / g)) and the observed content (Observed PUFAs content (mg / g)) of saturated fatty acids provided in the examples. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present application more clear, the present application is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described here are only used to explain the present application and are not used to limit the present application.

[0025] In order to solve the problems of long cycle, high cost and lethal sampling in the prior art, this application establishes the relationship between the characteristic wavelength of the scales of fresh grass carp and the content of unsaturated fatty acids in its muscle through machine learning and constructs a prediction model. By inputting the characteristic wavelength information of the scales of fresh grass carp into the prediction model, the content of unsaturated fatty acids in its muscle can be quickly known, which can not only realize rapid non-destructive detection, but also has high detection accuracy.

[0026] For this reason, Figure 1 As shown, the embodiment discloses a method for detecting unsaturated fatty acids in fresh grass carp muscle. The method comprises:

[0027] S10, obtaining a training set, the training set comprising a plurality of training samples, each training sample comprising at least one scale wavelength information and measured information of unsaturated fatty acids corresponding to the scale wavelength information;

[0028] S20, obtaining a back propagation neural network, the back propagation neural network comprising an input layer, a hidden layer and an output layer, the back propagation neural network having an initialized first connection weight connecting the input layer and the hidden layer, an initialized second connection weight connecting the hidden layer and the output layer, an initialized hidden layer threshold and an initialized output layer threshold;

[0029] S30, performing a forward propagation step, the forward propagation step comprising inputting the scale wavelength information in each training sample by the input layer, processing the scale wavelength information by the hidden layer to obtain a hidden layer output vector, and obtaining the expected output information of unsaturated fatty acids by the output layer according to the hidden layer output vector;

[0030] S40, obtaining the network prediction error of the output layer according to the measured information of unsaturated fatty acids and the expected output information;

[0031] S50, back-propagating the network prediction error to the input layer, updating the first connection weight, the second connection weight, the hidden layer threshold and the output layer threshold;

[0032] S60: According to a preset learning rate, all training samples are used to perform forward propagation, obtain network prediction error and back propagation steps. If it is detected that the network prediction error is less than or equal to the error threshold and / or the number of training times is greater than or equal to the preset number of training times, the training of the back propagation neural network is completed;

[0033] S70: Determine a prediction model according to the first connection weight, the second connection weight, the hidden layer threshold and the output layer threshold of the back propagation neural network when the training is completed;

[0034] S80, obtaining scale wavelength information of the fresh grass carp sample to be tested;

[0035] S90, inputting the scale wavelength information of the fresh grass carp sample to be tested into the prediction model to obtain the prediction information of unsaturated fatty acids.

[0036] In some embodiments, the flake wavelength information consists of hyperspectral data of 402nm, 416nm, 418nm, 434nm, 479nm, 499nm, 982nm, 1006nm, 1012nm, 1015nm, 1019nm, 1028nm, 1032nm and 1034nm.

[0037] In some embodiments, Figure 2 As shown, in the method, the step of acquiring the wavelength information of the flakes includes:

[0038] S801, obtaining a hyperspectral image of a fresh grass carp, wherein the hyperspectral image is a pixel image of a spectrum of scales of the fresh grass carp, and the wavelength range of the spectrum of the scales is 400-1100 nm;

[0039] S802, obtaining a target area in a hyperspectral image;

[0040] S803, calculating the average reflectivity of all wavelengths within the wavelength range of 400-1100 nm in the current area;

[0041] S804, average spectrum at wavelength corresponding to the average reflectivity value;

[0042] S805: extracting the scale wavelength information from the average spectrum.

[0043] In some embodiments, the hyperspectral image is a 2558×960 pixel image.

[0044] In some embodiments, the target area is a 200×200 pixel area.

[0045] In some embodiments, Figure 3 The step of acquiring the flake wavelength information also includes S800 before step S801: preprocessing the acquired original spectral image to obtain the hyperspectral image.

[0046] In order to eliminate the influence of equipment, samples and environment on the original spectral data collection process of the hyperspectral imaging system, the original spectral image obtained by the hyperspectral imaging camera is subjected to convolution smoothing. Convolution smoothing is a preprocessing method for smoothing spectral data and reducing noise. It smoothes the data by fitting a polynomial curve in the local area, thereby eliminating or reducing random noise in the data and making the data curve smoother. Figure 6 The original spectrum of a fresh grass carp scale is shown. Due to the uneven surface flatness of the grass carp scale and noise, the hyperspectral line is irregular. Figure 7 As shown, after step S800, the hyperspectral image obtained by convolution smoothing the original spectrum image reduces interference such as noise, which is more conducive to machine learning.

[0047] Table 1 shows the comparison between the root mean square error of the original spectral image without step S800 and the root mean square error of the hyperspectral image after step S800. As can be seen from Table 1, the prediction model obtained after step S800 has better prediction performance.

[0048] Table 1

[0049] Whether it passes S800 <![CDATA[R 2 ]]> RMSE no 0.8115 0.2021 yes 0.8397 0.1359

[0050] In addition, since the hyperspectral image contains a lot of variable information and there is a lot of redundant information between the variables, the modeling efficiency and accuracy of the model will be seriously affected. In order to improve the prediction accuracy of the model and improve the efficiency of the model prediction, in some embodiments, such as Figure 4 As shown, step S805 specifically includes:

[0051] S811, obtaining an average spectrum of each sample;

[0052] S812, obtaining Monte Carlo sampling from the average spectrum of each sample;

[0053] S813, obtaining at least two variable subsets from the Monte Carlo sampling;

[0054] S814, cross-validating each variable subset to obtain a root mean square error of each variable subset;

[0055] S815. If it is detected that a certain subset of the variables has the minimum root mean square error, extract the variables in the subset of the variables as the extracted flake wavelength information.

[0056] In step S812, the Monte Carlo sampling is run 50 times.

[0057] This shows that after the Monte Carlo sampling is run 50 times, step S705 eliminates useless information and can more accurately extract the flake wavelength information, which provides help for model training and improving the accuracy of the prediction model.

[0058] In some embodiments, the unsaturated fatty acid information is detected by the method recommended by GB / T 28769-2012 standard.

[0059] In some embodiments, 200 fresh grass carps were randomly sampled using the Kennard-Stone (KS) method. Among them, the scale wavelength information and the corresponding unsaturated fatty acid information of 160 fish were extracted as a training set for training the constructed back propagation neural network. In addition, the scale wavelength information and the corresponding unsaturated fatty acid information of 40 fish were extracted as a validation set for validating the trained back propagation neural network to obtain a prediction model. Table 2 shows the statistical results of the sum of the absolute content of UFA (mg / g) in the muscles of fresh grass carps providing the training set and the validation set.

[0060] Table 2

[0061]

[0062] In some embodiments, the unsaturated fatty acid information is a normalized vector of the unsaturated fatty acid content in the fresh grass carp muscle.

[0063] In some embodiments, Figure 5 As shown, the back propagation neural network includes 1 input layer, 1 hidden layer and 1 output layer. According to the input and output sequence (X, Y), the number of neurons i in the network input layer, the number of neurons j in the hidden layer, the number of neurons k (k = 1) in the output layer, the first connection weight between the input layer and the hidden layer is ωij, the second connection weight between the hidden layer and the output layer is ωjk, the hidden layer threshold a, the output layer threshold b, the preset learning rate and the neuron excitation function are determined.

[0064] Among them, the calculation formula of the hidden layer output H is as follows:

[0065]

[0066] Where j is the number of neurons in the hidden layer, and f is the activation function of the hidden layer. When the activation function is set to purelin function, it is equivalent to having no activation function, and the hidden layer function selects the Logistic function, and its expression is:

[0067]

[0068] Among them, the hidden layer output calculation formula is as follows:

[0069]

[0070] In some embodiments, the network prediction error is calculated as follows:

[0071] ek=Y K -Ok k=1,2,…,m

[0072] In some embodiments, the first connection weight and the second connection weight are updated according to the network prediction error, and the updated calculation formula is as follows:

[0073]

[0074] ωjk=ωjk+ηHjekj=1,2,…,l;k=1,2,…,m

[0075] In the formula, η is the learning efficiency.

[0076] In some embodiments, the hidden layer threshold a and the output layer threshold b are updated according to the network prediction error, and the updated calculation formula is as follows:

[0077]

[0078] bk=bk+ek k=1,2,…,m

[0079] In S60 provided in some embodiments, the input layer of the back propagation neural network has 200 neurons, the number of neurons in the hidden layer is 5, and the number of neurons in the output layer is 1. The preset learning rate is 0.01, and the preset training error threshold is 1×10 -6 , the preset number of training times is 1000.

[0080] In S90 provided in some embodiments, the prediction information is a predicted value of the unsaturated fatty acid content. Figure 8 As shown in the figure, the coefficient of determination (R 2 P ) is 0.868, and the root mean square error (RMSE) is 0.4322, indicating that the results of predicting the unsaturated fatty acid content in grass carp muscle using the method provided in the present application are more accurate.

[0081] Table 3 shows the prediction results of the prediction model provided by the present application for the samples in the training set and the validation set. The performance of the prediction model is evaluated by four statistical parameters, including the training set determination coefficient (R 2 C ), the validation set determination coefficient (R 2 P ), root mean square error (RMSE) and mean absolute error (MAE). 2 The higher the value, the lower the RMSE and MAE, the better the prediction effect of the prediction model. 2 , RMSE and MAE are calculated as shown in formulas (1)-(3):

[0082]

[0083] Where n represents the number of samples, xi and xi represent the estimated value and the measured value respectively. and denote the average of the estimated and measured values, respectively.

[0084] Wherein, obtain BP model (back propagation neural network) and RBF model (radial basis function neural network) to the determination coefficient of training set respectively 0.9467 and 0.8238, to the determination coefficient of verification set respectively 0.868 and 0.8016, to the root mean square error of training set respectively 0.3784 and 0.2215, to the root mean square error of verification set respectively 0.4322 and 0.2657, to the mean absolute error of training set respectively 0.3038 and 0.3152, to the mean absolute error of verification set respectively 0.3423 and 0.3154. Thus illustrate, the method that embodiment provides can carry out rapid nondestructive detection to unsaturated fatty acid content in fresh grass carp muscle. As shown in Table 3, adopt the back propagation neural network provided by the application to be trained to obtain prediction model, and its error is minimum.

[0085] Table 3

[0086]

[0087] The method for detecting unsaturated fatty acids in fresh grass carp muscle provided in the embodiment of the present application can be performed by a device for detecting unsaturated fatty acids in fresh grass carp muscle, or a control unit in the device for detecting unsaturated fatty acids in fresh grass carp muscle for performing the method for detecting unsaturated fatty acids in fresh grass carp muscle. In the embodiment of the present application, the device for detecting unsaturated fatty acids in fresh grass carp muscle is taken as an example to illustrate the device for detecting unsaturated fatty acids in fresh grass carp muscle provided in the embodiment of the present application.

[0088] Optionally, an embodiment of the present application also provides an electronic device, including a processor, a memory, and a program or instruction stored in the memory and executable on the processor. When the program or instruction is executed by the processor, each process of the above-mentioned method embodiment for detecting unsaturated fatty acids in fresh grass carp muscle is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be described here.

[0089] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed in the present application should be covered within the protection scope of the present application.

Claims

1. The method for detecting unsaturated fatty acids in fresh grass carp muscle includes: Obtaining a training set, the training set comprising a plurality of training samples, each of the training samples comprising at least one scale wavelength information and measured information of unsaturated fatty acids corresponding to the scale wavelength information; Obtaining a back propagation neural network, the back propagation neural network comprising an input layer, a hidden layer and an output layer, the back propagation neural network having an initialized first connection weight connecting the input layer with the hidden layer, an initialized second connection weight connecting the hidden layer with the output layer, an initialized hidden layer threshold and an initialized output layer threshold; Performing a forward propagation step, the forward propagation step comprising inputting the scale wavelength information in each training sample by the input layer, processing the scale wavelength information by the hidden layer to obtain a hidden layer output vector, and obtaining the expected output information of unsaturated fatty acids by the output layer according to the hidden layer output vector; Obtaining a network prediction error of the output layer according to the measured information of unsaturated fatty acids and the expected output information; Back-propagating the network prediction error to the input layer, updating the first connection weight, the second connection weight, the hidden layer threshold, and the output layer threshold; According to a preset learning rate, all training samples are used to perform the steps of forward propagation, obtaining the network prediction error and the back propagation, and if it is detected that the network prediction error is less than or equal to an error threshold and / or the number of training times is greater than or equal to a preset number of training times, the training of the back propagation neural network is completed; Determine a prediction model according to the first connection weight, the second connection weight, the hidden layer threshold and the output layer threshold of the back propagation neural network when the training is completed; Obtaining wavelength information of scales of fresh grass carp samples to be tested; The scale wavelength information of the fresh grass carp sample to be tested is input into a prediction model to obtain prediction information of unsaturated fatty acids.

2. According to the method according to claim 1, the flake wavelength information is composed of hyperspectral data of 402nm, 416nm, 418nm, 434nm, 479nm, 499nm, 982nm, 1006nm, 1012nm, 1015nm, 1019nm, 1028nm, 1032nm and 1034nm.

3. The method according to claim 1, wherein the step of acquiring the wavelength information of the flakes comprises: Obtaining a hyperspectral image of a fresh grass carp, wherein the hyperspectral image is a pixel map of a spectrum of scales of the fresh grass carp, and the wavelength range of the spectrum of the scales is 400-1100 nm; Obtaining a target area in a hyperspectral image; Calculate the average reflectivity of all wavelengths in the wavelength range of 400-1100nm in the current area; According to the average spectrum at the corresponding wavelength according to the average reflectance; The flake wavelength information is extracted from the average spectrum.

4. The method according to claim 3, further comprising: The acquired original spectral image is preprocessed to obtain the hyperspectral image.

5. According to the method of claim 3, the step of "extracting the flake wavelength information from the average spectrum" comprises: Obtain the average spectrum of each sample; Obtaining Monte Carlo sampling from the average spectrum of each of the samples; obtaining at least two subsets of variables from the Monte Carlo sampling; Perform cross validation on each variable subset to obtain the root mean square error of each variable subset; If it is detected that a certain subset of the variables has the minimum root mean square error, the variables in the subset of the variables are extracted as the extracted flake wavelength information.

6. According to the method of any one of claims 1 to 5, the input layer of the back propagation neural network has 200 neurons, the number of neurons in the hidden layer is 5, the number of neurons in the output layer is 1, the preset learning rate is 0.01, and the preset training error threshold is 1×10 -6 , the preset number of training times is 1000.

Images