Method for detecting crispness of crisp tilapia based on UDSHelm algorithm

Abstract

A UDSHelm algorithm-based crisp tilapia mossambica brittleness detection method comprises the following steps: acquiring Raman spectrum characteristic data and brittleness values of crisp tilapia mossambica with different brittleness by adopting a confocal Raman spectrometer and a texture analyzer, and establishing a data set; a UDSHelm algorithm model is constructed, the UDSHelm algorithm model comprises a UMAP algorithm module, an SSDA-HELM prediction module and a dung beetle algorithm DBO module, the UMAP algorithm module is used for carrying out dimensionality reduction on Raman spectrum characteristic data of a full-spectrum band, sparse self-encoding SAE is used as a front end for pre-training to provide initialization parameters of a multilayer extreme learning machine HELM, and the HELM prediction module is used for carrying out dimensionality reduction on the Raman spectrum characteristic data of the full-spectrum band; a dung beetle algorithm constructs a fitness function to optimize the number of neurons and the number of layers of HELM; after the UDSHelm algorithm model is trained, the Raman spectrum of the crisp tilapia meat to be detected can be input for brittleness prediction. Compared with a conventional brittleness detection method, the method provided by the invention is more convenient and faster, takes UDSHelm as a prediction model, and has a better prediction effect and high accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of meat product detection, and particularly relates to a method for detecting the crispness of crispy tilapia based on the UDSHelm algorithm. Background Art

[0002] Crispy tilapia is very popular in the market because of its Q - elastic, long - boiling - resistant, and crispy taste. The crispness of these fish is an important index, and the crispness can be improved by adding special feed during the feeding process. However, excessive crisping of the fish meat during the feeding process will lead to symptoms such as hemolysis, hypoxia, and organ functional lesions, and even directly lead to death due to blood circulation disorders. Therefore, the detection of the crispness of crispy tilapia species can not only evaluate the grade of the meat quality and better distinguish the market, but also control the types and dosages of feed by detecting the crispness of the fish meat to achieve the purpose of scientific feeding. The existing methods for evaluating crispness are mainly sensory evaluation and TPA texture analysis. The results of sensory evaluation are greatly affected by human factors. When TPA texture analysis is used for the crispness detection of fish meat, generally, an American FTC texture analyzer with a force sensor of 500N is required. Appropriate test parameters are set, multiple samples are cut, the test samples need to be stored frozen before testing, each sample needs to be measured multiple times and the average value is taken, and then the test data is obtained through specific analysis software. This operation has problems such as cumbersome operation, long time, and great damage to the samples. At the same time, the meat components of different fish species are different, and the applicable evaluation methods are also different. Summary of the Invention

[0003] The purpose of the present invention is to overcome the above - mentioned disadvantages of the prior art and provide a method for detecting the crispness of crispy tilapia based on the UDSHelm algorithm, which is fast, convenient, and has a high accuracy rate.

[0004] The present invention is realized through the following technical solutions:

[0005] A method for detecting the crispness of crispy tilapia based on the UDSHelm algorithm includes the following steps:

[0006] S1. Use a confocal Raman spectrometer to scan the Raman spectra of crispy tilapia with different crispness levels, and extract the Raman spectral characteristic data of the Raman spectra; use a texture analyzer to test the crispness values of crispy tilapia with different crispness levels at a set compression speed and deformation rate. Establish a data set according to the Raman spectral characteristic data and the corresponding crispness values. After pre - processing the data in the data set, randomly divide it into a training set and a validation set according to a certain proportion;

[0007] S2. Construct a UDSHelm algorithm model. The UDSHelm algorithm model includes a UMAP algorithm module, an SSDA-HELM prediction module, and a dung beetle optimization (DBO) algorithm module. The UMAP algorithm module is used to reduce the dimensionality of the Raman spectral feature data in the full spectral band, find the key wavenumbers with greater influence, and eliminate the non-critical wavenumbers with less influence. The SSDA-HELM prediction module is used to predict the brittleness. It uses sparse autoencoder (SAE) as the front end for pre-training to provide initial weights, and the trained parameters are used as the initial parameters of the multi-layer extreme learning machine (HELM) to obtain the optimal solution. The dung beetle optimization (DBO) algorithm module is used to construct a fitness function to optimize SSDA-HELM and find the optimal number of neurons and layers of SSDA-HELM.

[0008] S3. Use the training set to train the UDSHelm algorithm model, use the test set to test and score the UDSHelm algorithm model and optimize it, and then the Raman spectrum of the to-be-tested brittle tilapia fish can be input to predict the brittleness value.

[0009] Further, in step S3, the method for training the UDSHelm algorithm model is as follows: Using brittle tilapia fish with different brittleening times as samples, obtain Raman spectral feature data and brittleness values. First, use UMAP to reduce the dimensionality of the Raman spectral feature data in the full spectral band to determine the dimensionality of the input data. Then, input the data into the input layer of the SSDA autoencoder (stacked sparse denoising autoencoder). The SSDA hidden layer extracts relevant features from the complex input data, and uses an unsupervised learning method to perform layer-by-layer pre-training and fine-tuning to obtain the initial weights. Then, remove the decoding part of the SDAE, connect the ELM network, and use the obtained initial weights as the initial values of the multi-layer extreme learning machine HELM for assignment to output the brittleness value. Softmax can also be used in the output layer to classify the brittleness value. Use the dung beetle optimization (DBO) algorithm to construct a fitness function to optimize the number of neurons and layers of SSDA-HELM and find the optimal number of neurons and layers.

[0010] Further, the method for using the dung beetle optimization (DBO) algorithm module to construct a fitness function to optimize the number of neurons and layers of SSDA-HELM is as follows:

[0011] Initialize the dung beetle population in the DBO algorithm. Each individual represents a combination of a learning rate and the number of neurons. Construct a fitness function, using MSE, RMSE, and accuracy as evaluation indicators, and train the SSDA-HELM model.

[0012] Update the population individuals according to the rules of behaviors such as rolling and searching in DBO, calculate the fitness of the new population, and select the optimal individual.

[0013] If the maximum number of iterations is reached or the error converges, stop the search; otherwise, continue the optimization until the optimal learning rate and the number of neurons are obtained.

[0014] Furthermore, the specific method of the dung beetle optimization (DBO) module is as follows:

[0015] S2-11. If the current individual is a rolling dung beetle, let the random number δ = rand(1). If δ < 0.9, the updated position of the dung beetle is: x i (t + 1) = x i (t) + α × k × x i (t - 1) + b × |x i (t) - X w |,

[0016] Otherwise, the updated position of the dung beetle is: x i (t + 1) = x i (t) + tan(θ)|x i (t) - x i (t - 1)|,

[0017] where t represents the current iteration number; x i (t) represents the position of the i-th dung beetle at the t-th iteration; α is the natural coefficient, taking 1 or -1, -1 indicating deviation from the original direction, and 1 indicating no deviation; k represents the deflection coefficient, k ∈ (0, 0.2]; b is a fixed value in (0, 1); X w is the global worst position; θ ∈ [0, π], which is the deflection angle;

[0018] S2-12. If the current individual is an egg ball, the position of the dung beetle is updated by the following function:

[0019] B i (t + 1) = X * + b1 × (B i (t) - L b * ) + b2 × (B i (t) - U b * ),

[0020] L b * = max(X * × (1 - R), L b ),

[0021] U b * = max(X * × (1 + R), U b ),

[0022] where B i(t) represents the position of the i-th oosphere at the t-th iteration; b1 and b2 represent two independent random vectors of size 1×D, where D is the dimension of the optimization problem; X * is the current local optimal position; L b * and U b * represent the lower and upper bounds of the spawning area respectively; R = 1 - t / T max , T max represents the maximum number of iterations; L b and U b represent the lower and upper bounds of the optimization problem respectively;

[0023] S2 - 13. If the current individual is a small dung beetle, the position of the dung beetle is updated by the following function:

[0024] x i (t + 1) = x i (t) + c1×(x i (t) - L b b ) + c2×(x i (t) - U b b ),

[0025] L b b = max(X b ×(1 - R), L b ),

[0026] U b b = max(X b ×(1 + R), U b ),

[0027] where x i (t) represents the position information of the i-th small dung beetle at the t-th iteration; c1 represents a random number following a normal distribution; c2 represents a random vector belonging to (0, 1); X b represents the global best position, L b b and U b b represent the lower and upper bounds of the best foraging area respectively;

[0028] S2 - 14. If the current individual is a thief dung beetle, the position of the dung beetle is updated by the following function:

[0029] x i (t + 1) = X b + S×g×(|x i (t) - X * | + |xi (t)-X b |),

[0030] Among them, xi(t) represents the position information of the i-th dung beetle at the t-th iteration; g represents a random vector of size 1×D that obeys the normal distribution; S represents a constant value.

[0031] Furthermore, in step S2, the method used by the UMAP algorithm module to reduce the dimension of the Raman spectrum feature data of the full spectrum band is:

[0032] The local density relationship between data points is calculated by the k-nearest neighbor algorithm (k-NN), and the neighborhood size is adjusted using the smoothness parameter;

[0033] After selecting the target to reduce the dimension, the weighted graph is calculated, the connection probability between data points is converted into a weight matrix, and the local manifold structure is optimized using the Riemannian metric;

[0034] Gradient descent is used to optimize the objective function so that the low-dimensional representation maintains the topological structure of the original high-dimensional space as much as possible and minimizes the topological differences between data points in the high-dimensional space and the low-dimensional space.

[0035] Furthermore, in step S2, the algorithm of the SSDA-HELM prediction module is:

[0036] S2-21. Select the number of hidden layers of the SSDA network and initialize the network depth k, X1 = X; X = [x (1) , x (2) , ..., x (m) ] T , is the number of nodes in the hidden layer. The decoding part of the SSDA network is deleted and connected with the HELM network to build the SSDA-HELM network;

[0037] S2-22, starting from the first hidden layer, train the input weights W of each hidden layer i and hidden layer bias b i , and the weight W i and hidden layer bias b i Initialize the SSDA-HELM network as input weights;

[0038] S2-23, input weight W obtained by pre-training i and hidden layer bias b i , calculate the hidden layer output matrix A: A i =H i- 1W li , where A i is the output of the i-th layer node; H i-1 is the input of the i-th layer node; Wli is the weight matrix;

[0039] S2-24. According to the ELM theory:

[0040]

[0041] where H is the output of the hidden layer nodes, β is the output weight, T is the expected output, g(x) is the activation function, and W i =[w i，1 , w 1，2 ,..., w i,n T is the weight between the input node and the i-th hidden node, β i is the weight between the i-th hidden node and the output node, and b i is the bias of the i-th hidden layer node; W1·X j represents the inner product of W i and X j ;

[0042] Calculate the output weight matrix of the neural network where is the generalized inverse matrix of matrix A;

[0043] S2-25. Calculate the output result: where is the output of the i-th layer, H i-1 is the input of the i-th layer, and g(·) is the activation function of the hidden layer;

[0044] S2-26. Repeat the above steps S2-22 to S2-25 until the output calculation of the last hidden layer is completed to obtain the eigenvalues;

[0045] S2-27. Use the extracted features as input values and output the brittleness values, which can be sent to the SOFTMAX classifier for classification prediction.

[0046] Furthermore, the method for training the input weight W i of the hidden layer and the hidden layer bias b i in the S2-22 step is as follows:

[0047] Construct multiple autoencoders, each corresponding to a hidden layer, for pre-training θ i ∈{W i , b i ​} Parameters; the input layer of the hidden layer of each autoencoder is the output of the hidden layer of the previous autoencoder, and the output layer is the reconstruction of the previous hidden layer; the greedy layer-by-layer training method is used to perform unsupervised training of each autoencoder layer by layer to obtain the weights of each layer of the SSDA network, and then the weights are fine-tuned layer by layer as a whole through the backpropagation algorithm to train the optimal weights of the SSDA network;

[0048] The encoding process of the autoencoder is as follows:

[0049] Among them, W1 is the weight matrix from the input layer to the hidden layer and from the hidden layer to the output layer, b1 is the unit bias coefficient of the hidden layer and the output layer; σ(·) represents the activation function, and the logsig function is selected; θ represents the parameter matrix of the network, θ i ∈{W i , b i}.

[0050] Furthermore, the network parameters for training the SSDA network are set as follows: the learning rate is 0.1, the maximum number of pre-training iterations is 400, the maximum number of fine-tuning iterations is 300, the sparse parameter is 0.5, the sparse penalty term parameter is 3, the activation function uses the sigmoid function, and the fine-tuning loss function is:

[0051]

[0052] Among them, θ represents the set of all parameters involved in this loss function, N represents the number of samples, and i represents the sample index; y (i) is the predicted value of the i-th sample, which is the result predicted by the model; x (i) is the true value of the i-th sample, which is the actual observed value of the sample; ||·|| F 2 is the Frobenius norm, which represents the square of the Euclidean norm (L2 norm) and is used to measure the distance between the predicted value y (i) and the true value x (i) . By averaging this distance for all samples (dividing by N), it reflects the overall error degree between the model's predicted value and the true value and prevents overfitting; λ is the regularization parameter, which is used to balance the weights of the fitting error (the first term) and the regularization term (the second term) in the loss function, balance the prediction error and the model complexity, and control the penalty strength for the model parameters; l represents the number of layers of the neural network, and here the relevant calculations are performed from layer 1 to layer l; w (l) represents the weight matrix of the l-th layer of the neural network; w′ (l) represents the bias term correlation matrix of the l-th layer of the neural network.

[0053] When using the test set to test and score the UDSHelm algorithm model, the K-fold cross-validation method is used.

[0054] Furthermore, in step S1, the method for using a texture analyzer to test the crispness values of crispy tilapia with different crispness levels at a set compression speed and deformation rate is as follows: Use a 6 mm cylindrical stainless steel probe of the FTC texture analyzer. In the TPA test mode, the trigger force is set to 0.75 N, the pre-test speed, mid-test speed, and post-test speed are 1 mm / s, 0.5 mm / s, and 1 mm / s in sequence. Each sample is pressed down twice, and the interval time between the two presses is 2 s.

[0055] Furthermore, in step S1, the method for using a confocal Raman spectrometer to scan and obtain the Raman spectra of crispy tilapia with different crispness levels is as follows: The scanning range of the confocal Raman spectrometer is 500 - 2000 cm -1 , with a 50× objective lens, a laser of 532 nm, an integration time of 5 s, a grating of 1200 I / mm. Each sample is tested three times at different positions, and the average spectral line is taken.

[0056] Furthermore, in step S1, the method for extracting the Raman spectral characteristic data of the Raman spectrum is as follows: First, perform polynomial fitting on the Raman spectrum to achieve baseline correction, then use the Savitz - Golay filter for smoothing to remove noise, and then perform Min - Max normalization processing; then obtain the wavenumber, intensity, and peak area of each characteristic peak in the full wavelength range of the Raman spectrum through Origin; the Raman spectral characteristic data includes the wavenumber and peak intensity of each wave peak in the full wavelength range.

[0057] The present invention uses the Raman spectrum of crispy tilapia as the input to detect its crispness. Compared with conventional crispness detection methods, it is more convenient and fast. Using UDSHelm as the prediction model, the UMAP algorithm is used to reduce the dimension of the spectral map in the full spectral band. The sparse auto - encoder SAE is used as the front - end for pre - training to provide optimal initial parameters for the multi - layer ELM. The dung beetle optimization (DBO) algorithm is used to construct a fitness function to optimize the number of neurons and layers of the multi - layer ELM, so that the UDSHelm prediction model has better performance in the crispness prediction of crispy tilapia compared with other algorithm models, with high accuracy, which is of great significance for the quality assessment and scientific breeding of meat products. Description of the Drawings

[0058] Figure 1 It is the overall flowchart of the prediction model in the embodiment of the present invention.

[0059] Figure 2 It is the original Raman spectrogram in the embodiment of the present invention.

[0060] Figure 3 It is the pre - processed Raman spectrogram in the embodiment of the present invention.

[0061] Figure 4These are the Raman spectra after preprocessing by four preprocessing methods in the embodiments of the present invention, where Figure 4 (a) is Savitzky-Golay (S-G) preprocessing, Figure 4 (b) is multiplicative scatter correction (MSC) preprocessing, Figure 4 (c) is standard normal variate (SNV) preprocessing, Figure 4 (d) is Normalize preprocessing.

[0062] Figure 5 This is the structural diagram of SSDA in the embodiments of the present invention.

[0063] Figure 6 This is the structural diagram of the extreme learning machine (ELM) algorithm in the embodiments of the present invention. Detailed implementation manners

[0064] A method for detecting the crispness of crispy tilapia based on the UDSHelm algorithm, as Figure 1 shown, includes the following steps:

[0065] S1. Use a confocal Raman spectrometer to scan the Raman spectra of crispy tilapia with different crispness levels, and extract the Raman spectral feature data of the Raman spectra; use a texture analyzer to measure the crispness values of crispy tilapia with different crispness levels at a set compression speed and deformation rate. Establish a data set based on the Raman spectral feature data and the corresponding crispness values. After preprocessing the data in the data set, randomly divide it into a training set and a validation set according to a certain proportion.

[0066] The crispness of crispy tilapia is closely related to its protein content and structure. The protein content and structure can be characterized by Raman spectra, which can reflect the secondary structure information of proteins. Moreover, Raman spectra can quickly and non-destructively detect fish meat with high sensitivity and high resolution.

[0067] To more accurately explore the characteristic relationship between the crispness value and the Raman spectral data, when collecting crispy tilapia with various different crispness levels, the data range should be as wide as possible to cover various different crispness ranges. Therefore, crispy tilapia with different crisping times are selected as data samples, and the crisping times are 0 days, 30 days, 60 days, 90 days, and 120 days respectively. Take the fish meat in the middle of the back of the crispy tilapia, cut it into pieces of a certain size, first collect the Raman spectral data, and then collect its crispness data through a texture analyzer.

[0068] In this embodiment, a French HORIBA confocal Raman spectrometer is used. During the test, the set scanning range is 500 - 2000 cm -1 , the microscopic lens uses a 50× objective lens, the laser is 532 nm, the integration time is 5 s, the grating is 1200 I / mm, and each sample is tested three times at different positions, and the average spectral line is taken.

[0069] The raw data of Raman spectra often contain noise and background signals due to reasons such as the test environment, samples, and the instrument itself. Therefore, in order to improve the signal-to-noise ratio and accuracy of the data, it is necessary to preprocess the raw data. In this embodiment, the method of data preprocessing is as follows: First, perform polynomial fitting on the Raman spectrum to achieve baseline correction and eliminate the influence of background signals. Then, use the Savitz-Golay filter for smoothing to remove noise and improve the clarity and stability of the signal. Next, perform Min-Max normalization to eliminate the difference in signal amplitude. The preprocessed data uses Origin to obtain the wavenumber, intensity, and peak area of each characteristic peak in the full wavelength range of the Raman spectrum, thereby obtaining Raman spectrum characteristic data. The Raman spectrum characteristic data includes the wavenumber and peak intensity of each peak in the full wavelength range.

[0070] When preprocessing the raw Raman spectrum, PeakFit software and NGSLabSpec5 software can be used for peak area fitting, Origin software can be used to draw the model diagram, and matlab software and Unscrambler X software can be used for modeling analysis.

[0071] From Figure 2 the raw Raman spectrum, it can be seen that due to the large number of spectra and the reasons of the instrument or the uneven surface of the meat sample, baseline drift and noise are generated, making the peak position information not obvious. It can be roughly seen that in the wavenumber range of 500 - 2000 cm -1 there are 7 - 10 fingerprint peaks. After preprocessing the raw Raman spectrum and marking the peaks with Origin, as Figure 3 shown, many detailed peaks in the spectrogram can be seen.

[0072] In order to compare the effects of each preprocessing step and method, the raw Raman spectrum is processed by methods such as smoothing filtering (S-G), standard normal variate (SNV), multiplicative scatter correction (MSC), and normalization (Normalize) respectively. The processing effects are as Figure 4 shown. Through Figure 4 it can be seen that after S-G (as Figure 4 a) preprocessing, the noise of the Raman spectrum decreases, but the baseline drift of the spectrum is still not eliminated, and the characteristic peaks in each band are not obvious; after MSC (as Figure 4 b) and SNV (as Figure 4 c) preprocessing, the effects are relatively similar. The spectrum becomes more compact, the characteristic peaks are clear, and light scattering is eliminated. However, the noise in the 500 - 800 cm -1 band range seems to increase, and the spectrum discrimination is not high, indicating that these two methods eliminate the interference of baseline drift and light scattering to a certain extent; after Normalize (as Figure 4d) The preprocessed Raman spectra become more tightly ordered, baseline interference is eliminated, and the characteristic peak signals are strong but not as strong as those of MSC and SNV.

[0073] By processing the Raman spectrogram, the peak wavenumber, peak intensity, and peak area can be obtained. Using deep learning methods, the peak value at each wavenumber can be used as input to extract features (the peak wavenumber and peak area are composed of the peak values at each wavenumber), so that deeper differential point features can be found. Figure 3 It can be seen that different embrittlement times (0 days, 30 days, 60 days, etc.) represent different brittleness, and the spectrograms of different brittleness have obvious differences.

[0074] The settings for the texture analyzer test are as follows: a 6 mm cylindrical stainless steel probe of the FTC texture analyzer is used. In the TPA test mode, the trigger force is set to 0.75 N, and the pre-test speed, in-test speed, and post-test speed are 1 mm / s, 0.5 mm / s, and 1 mm / s in sequence; the samples are stored in a 4°C refrigerator for a certain period of time (at least 4 h) before the test. Each sample is pressed down twice, and the interval time between the two presses is 2 s, and the average value is taken. The brittleness value is read and calculated by the TMS-Pro physical property analysis system supporting the texture analyzer.

[0075] S2. Construct a UDSHelm algorithm model. The UDSHelm algorithm model includes a UMAP algorithm module, an SSDA-HELM prediction module, and a dung beetle algorithm DBO module. The UMAP algorithm module is used to reduce the dimension of the Raman spectral feature data in the full spectral band, find the key wavenumbers with greater influence, and eliminate the non-key wavenumbers with less influence, thereby reducing the data volume; the SSDA-HELM prediction module is used to predict the brittleness value. It uses sparse autoencoder SAE as the front end for pre-training to provide initial weights, and the trained parameters are used as the initial weights of the multi-layer extreme learning machine HELM to obtain the optimal solution; the dung beetle algorithm DBO module is used to construct a fitness function to optimize SSDA-HELM and find the optimal number of neurons and layers of SSDA-HELM.

[0076] Stacked Sparse Denoising Autoencoders (SSDA) is an architecture formed by stacking multiple SAE stacks. It combines sparse coding and deep network training, and extracts low-level structural features from the original image through layer-by-layer unsupervised learning. The model first injects noise into the input image, and then through a series of encoder and decoder networks, layer by layer learns to remove noise and reconstruct the original image. During the training process, the model continuously optimizes the weight parameters to minimize the reconstruction error.

[0077] The ELM model (extreme learning machine) directly calculates the output based on the generalized inverse matrix theory, without multiple leading and reverse training optimal values, and is less disturbed by training samples. Therefore, the extreme learning machine has the advantages of fewer training parameters, fast learning speed, and strong generalization ability. However, since its initial parameters are all randomly generated, the quality of the random initial parameters directly affects the output results. The multi-layer extreme learning machine HELM (Hierarchical Extreme Learning Machine) is an algorithm based on multiple layers. It fully utilizes the generalization approximation ability of ELM based on random feature mapping, but at the same time, its initial parameters, the number of neurons and the number of layers directly affect the output results. In order to optimize the random initial parameters, the present invention uses sparse autoencoder SAE as the front end for pre-training to provide initialization weights, and the trained parameters are used as the initialization parameters of the multi-layer ELM model to obtain the optimal solution. At the same time, DBO (dung beetle algorithm) is used to construct the fitness function to optimize the number of neurons and the number of layers.

[0078] S3. The UDSHelm algorithm model is trained using a training set, and the UDSHelm algorithm model is tested, scored and optimized using a test set, and then the optimal parameter combination is output. After the optimal parameter combination is loaded into the UDSHelm algorithm model, the Raman spectrum of the crispy tilapia meat to be tested can be input for crispness prediction, and finally the crispness value is output.

[0079] The UMAP (Uniform Manifold Approximation and Projection) algorithm is an innovative nonlinear dimensionality reduction technology. It is based on topological data analysis and manifold learning. By constructing a topological graph of the data and optimizing the low-dimensional representation, it can map high-dimensional data to a low-dimensional space while retaining the data structure and relative distance. In this embodiment, the UMAP algorithm module is used to reduce the dimensionality of the Raman spectral feature data of the full spectrum band as follows:

[0080] The local density relationship between data points is calculated by the k-nearest neighbor algorithm (k-NN), and the neighborhood size is adjusted using the smoothness parameter. The weighted graph is calculated after selecting the target dimensionality reduction, the connection probability between data points is converted into a weight matrix, and the local manifold structure is optimized using the Riemannian metric. The objective function is optimized using gradient descent so that the low-dimensional representation maintains the topological structure of the original high-dimensional space as much as possible and minimizes the topological differences between data points in the high-dimensional space and the low-dimensional space.

[0081] The method for training the UDSHelm algorithm model is as follows: Using tilapia with different embrittlement times as samples, Raman spectral feature data and brittleness values are obtained. First, UMAP is used to reduce the dimensionality of the Raman spectral feature data in the full spectral band to determine the dimensionality of the input data. Then, the data is input into the input layer of the SSDA autoencoder (stacked sparse denoising autoencoder). The SSDA hidden layer extracts relevant features from the complex input data. Using unsupervised learning methods, layer-by-layer pre-training and fine-tuning are performed to obtain the initial weights. Then, the decoding part of the SDAE is removed, and the ELM network is connected. The obtained initial weights are used as the initial values of the multi-layer extreme learning machine HELM for assignment, and the brittleness value is output. The dung beetle algorithm DBO is used to construct a fitness function to optimize the number and number of layers of neurons in SSDA-HELM, and the optimal number and number of layers of neurons are found. For the case where the brittleness needs to be graded, in order to distinguish the quality grade or feeding measures of the product according to the brittleness, the softmax can also be used to classify the brittleness value according to the output brittleness value in the output layer of the model. The categories are represented by the catalysis time and can be divided into five categories: embrittlement time of 0 days, 30 days, 60 days, 90 days, and 120 days.

[0082] The network parameter settings for training the SSDA network are as follows: the learning rate is 0.1, the maximum number of pre-training iterations is 400, the maximum number of fine-tuning iterations is 300, the sparse parameter is 0.5, the sparse penalty term parameter is 3, the activation function uses the sigmoid function, and the fine-tuning loss function is:

[0083]

[0084] where θ represents the set of all parameters involved in this loss function, N represents the number of samples, and i represents the sample index; y (i) is the predicted value of the i-th sample, which is the result predicted by the model; x (i) is the true value of the i-th sample, which is the actual observed value of the sample; is the Frobenius norm, which represents the square of the Euclidean norm (L2 norm) and is used to measure the distance between the predicted value y (i) and the true value x (i) The distance between them. By averaging this distance for all samples (dividing by N), it reflects the overall error degree between the model's predicted value and the true value, preventing overfitting; λ is the regularization parameter, which is used to balance the weights of the fitting error (the first term) and the regularization term (the second term) in the loss function, balance the prediction error and the model complexity, and control the penalty strength for the model parameters; l represents the number of layers of the neural network. Here, relevant calculations are performed from layer 1 to layer l; w (l) represents the weight matrix of the l-th layer of the neural network; w′ (l) represents the bias term correlation matrix of the l-th layer of the neural network.

[0085] When using the test set to test and score the UDSHelm algorithm model, the K-fold cross-validation method is adopted.

[0086] SSDA-HELM uses SSDA to pre-train the initial values of HELM and optimize the initial values. SSDA is docked with HELM, and the number of iteration layers and neurons of the two are the same. The number of iteration layers and neurons will affect the final prediction effect. Therefore, in order to improve the prediction effect, the method for constructing a fitness function by using the dung beetle algorithm DBO module to optimize the number of neurons and layers of SSDA-HELM is as follows:

[0087] Initialize the dung beetle population in the DBO algorithm, including the positions and velocities of dung beetle individuals. Each individual represents a combination of a learning rate and the number of neurons; construct a fitness function, using MSE, RMSE, and accuracy as evaluation indicators, and train the SSDA-HELM model; update the population individuals according to the rules of behaviors such as rolling and searching in DBO, calculate the fitness of the new population, including the fitness of each individual and update the optimal position, and select the optimal individual; if the maximum number of iterations or error convergence is reached, stop the search, otherwise continue to optimize until the best learning rate and the number of neurons are obtained.

[0088] The present invention adopts the UDSHelm (UMAP-DBO-SSDA-HELM) model algorithm, which belongs to an optimization algorithm in deep learning. Generally, deep learning algorithms obtain network models by repeatedly training parameters. Extreme learning machine (ELM) does not require training parameters and is directly calculated, with high speed. The accuracy is related to the initial parameters. Therefore, the SSDA-HELM algorithm pre-trains the initial parameters through SSDA to ensure that the initial parameters are close to the optimal, thereby ensuring the accuracy of the ELM algorithm. The advantage is to try to ensure that the accuracy and calculation speed reach the optimal.

[0089] The original spectral data is first processed by dimensionality reduction using the UMAP algorithm, and then the dimensionality-reduced data is normalized to ensure that the contribution values of the respective feature values after dimensionality reduction are the same. Then, it is divided into a training set and a validation set according to 8:2; the best learning rate and the number of neurons obtained by the DBO algorithm are used to train the final SSDA-HELM model. Finally, the performance of the final model is evaluated, and actual application or deployment is carried out.

[0090] The Dung Beetle Optimization Algorithm (DBO) constructs a fitness function to simulate the behaviors of dung beetles in nature, including rolling balls, dancing, foraging, reproduction, stealing, etc. It can make the solutions in the solution space move rapidly towards better solutions, has strong global search ability, can better avoid falling into local optimal solutions, and has certain advantages for complex multi-variable function optimization. The present invention adopts the DBO algorithm, and according to the optimization rules of DBO, continuously obtains the number and layers of the optimized SSDA-HELM neural network model, and calculates the model evaluation index. The specific method of the dung beetle algorithm is as follows:

[0091] S2-11. If the current individual is a ball-rolling dung beetle, let the random number δ = rand(1). If δ < 0.9, the updated position of the dung beetle is: x i (t + 1) = x i (t) + α × k × x i (t - 1) + b × |x i (t) - X w |,

[0092] Otherwise, the updated position of the dung beetle is: x i (t + 1) = x i (t) + tan(θ)|x i (t) - x i (t - 1)|,

[0093] Wherein, t represents the current iteration number; x i (t) represents the position of the i-th dung beetle at the t-th iteration; α is the natural coefficient, taking 1 or -1, -1 indicating deviation from the original direction, and 1 indicating no deviation; k represents the deflection coefficient, k ∈ (0, 0.2]; b is a fixed value in (0, 1); X w is the global worst position; θ ∈ [0, π], which is the deflection angle;

[0094] S2-12. If the current individual is an egg ball, the position of the dung beetle is updated by the following function:

[0095] B i (t + 1) = X * + b1 × (B i (t) - L b * ) + b2 × (B i (t) - U b * ),

[0096] L b * = max(X * × (1 - R), L b ),

[0097] Ub * = max(X * ×(1 + R), U b ),

[0098] where B i (t) represents the position of the i-th oosphere at the t-th iteration; b1 and b2 represent two independent random vectors of size 1×D, where D is the dimension of the optimization problem; X * is the current local optimal position; L b * and U b * represent the lower and upper bounds of the spawning area respectively; R = 1 - t / T max , T max represents the maximum number of iterations; L b and U b represent the lower and upper bounds of the optimization problem respectively;

[0099] S2 - 13. If the current individual is a small dung beetle, the position of the dung beetle is updated by the following function:

[0100] x i (t + 1) = x i (t) + c1×(x i (t) - L b b ) + c2×(x i (t) - U b b ),

[0101] L b b = max(X b ×(1 - R), L b ),

[0102] U b b = max(X b ×(1 + R), U b ),

[0103] where x i (t) represents the position information of the i-th small dung beetle at the t-th iteration; c1 represents a random number following a normal distribution; c2 represents a random vector belonging to (0, 1); X b represents the global best position, L b b and U b b represent the lower and upper limits of the best foraging area respectively;

[0104] S2-14. If the current individual is a thief dung beetle, the position of the dung beetle is updated by the following function:

[0105] x i (t + 1)= X b + S × g ×(|x i (t)- X * |+|x i (t)- X b |),

[0106] where x i (t) represents the position information of the i-th thief dung beetle at the t-th iteration; g represents a random vector of size 1×D that follows a normal distribution; S represents a constant value.

[0107] In the SSDA-HELM prediction module, the SSDA model architecture is as Figure 5 shown, and the structure diagram and algorithm description of the ELM algorithm are as Figure 6 , and the algorithm of the SSDA-HELM prediction module is as follows:

[0108] S2-21. Select the number of hidden layers of the SSDA network, initialize the network depth k, X1 = X; X = [x (1) , x (2) ,..., x (m) T , is the number of nodes in the hidden layer. Delete the decoding part of the SSDA network and connect it to the HELM network to construct the SSDA-HELM network;

[0109] S2-22. Starting from the first hidden layer, train to obtain the input weights W i and the hidden layer bias b i , and use the weights W i and the hidden layer bias b i as the input weights to initialize the SSDA-HELM network;

[0110] Among them, the method for the SSDA network to train and obtain the input weights W i and the hidden layer bias b i is as follows:

[0111] Construct multiple autoencoders, each autoencoder corresponding to a hidden layer, for pre-training θ i ∈{W i ,b i ​} Parameters; the input layer of each hidden layer of the autoencoder is the output of the hidden layer of the previous autoencoder, and the output layer is the reconstruction of the previous hidden layer; the greedy layer-by-layer training method is used to perform layer-by-layer unsupervised training on each autoencoder to obtain the weights of each layer of the SSDA network, and then the weights are fine-tuned layer by layer as a whole through the backpropagation algorithm to train the optimal weights of the SSDA network;

[0112] The encoding process of the autoencoder is:

[0113] Among them, W1 is the weight matrix from the input layer to the hidden layer and from the hidden layer to the output layer, b1 is the unit bias coefficient of the hidden layer and the output layer; σ(·) represents the activation function, and the logsig function is selected; θ represents the parameter matrix of the network, θ i ∈{W i , b i}.

[0114] S2-23. Calculate the output matrix A of the hidden layer from the input weight W i and the hidden layer bias b i : A i = H i- 1W li , where A i is the output of the i-th layer node; H i-1 is the input of the i-th layer node; W li is the weight matrix;

[0115] S2-24. According to the ELM theory:

[0116]

[0117] Among them, H is the output of the hidden layer node, β is the output weight, T is the expected output, g(x) is the activation function, W i = [w i，1 , w i，2 ,..., w i，n T is the weight between the input node and the i-th hidden node, β i is the weight between the i-th hidden node and the output node, b i is the bias of the i-th hidden layer node; W i ·X j represents the inner product of W i and X j ;

[0118] Calculate the output weight matrix of the neural network Among them, is the generalized inverse matrix of matrix A;

[0119] S2-25. Calculate the output result: Among them is the output of the i-th layer, and H i-1 is the input of the i-th layer, and g(·) is the activation function of the hidden layer;

[0120] S2-26. Repeat the above steps S2-22 to S2-25 until the output calculation of the last hidden layer is completed to obtain the eigenvalue;

[0121] S2-27. Take the extracted feature as the input value and output the brittleness value, which can be sent to the SOFTMAX classifier for classification prediction.

[0122] In this embodiment, during model training, 80% of the data is randomly selected as the training set, and the remaining data is used as the validation set. The K-fold cross-validation method is used to increase the credibility of the training results. Through the adjustment of the value of K, it is found that the model with K = 10 has better generalization performance, so K = 10 is selected.

[0123] To prove the prediction performance of the present invention's UDSHelm, the support vector regression model SVR and the UMAP-SSDA-HELM model are used for comparison respectively. UMAP-SSDA-HELM is the UMAP algorithm combined with the SSDA-HELM model, that is, the DBO algorithm in the present invention's UDSHelm model is not added to optimize the number of layers of SSDA-HELM. The detection results of each model are shown in Table 1.

[0124] Table 1 Model detection results based on Raman spectroscopy

[0125]

[0126] As can be seen from Table 1, compared with other model algorithms, the root mean square error RMSE, mean squared error MSE, mean absolute error MAE, and determination coefficient R of the present invention's UDSHelm model 2 all have excellent performances, indicating that the prediction accuracy and precision of this model are high, the effect is good, and it is more suitable for the prediction of the brittleness of crispy tilapia with Raman spectroscopy as the input, and it also has positive significance for the detection of the brittleness and other properties of other meats.

[0127] The above detailed description is a specific description of the feasible embodiments of the present invention. This embodiment is not intended to limit the patent scope of the present invention. Any equivalent implementation or modification without departing from the present invention shall be included in the patent scope of this case.

Claims

1. A method for detecting the crispness of crispy tilapia based on the UDSHelm algorithm, characterized in that, The steps are as follows: S1. Use a confocal Raman spectrometer to scan the Raman spectra of crispy tilapia with different crispness levels, and extract the Raman spectral feature data of the Raman spectra. Use a texture analyzer to measure the crispness values of crispy tilapia with different crispness levels at a set compression speed and deformation rate. Establish a data set based on the Raman spectral feature data and the corresponding crispness values. After preprocessing the data in the data set, randomly divide it into a training set and a validation set according to a certain proportion; S2. Construct a UDSHelm algorithm model. The UDSHelm algorithm model includes a UMAP algorithm module, an SSDA-HELM prediction module, and a dung beetle algorithm DBO module. The UMAP algorithm module is used to reduce the dimension of the Raman spectral feature data in the full spectral band, find the key wavenumbers with greater influence, and eliminate the non-key wavenumbers with less influence. The SSDA-HELM prediction module is used to predict the crispness value. It uses a sparse autoencoder SAE as the front end for pre-training to provide initial weights, and the trained parameters are used as the initial weights of a multi-layer extreme learning machine HELM to obtain the optimal solution. The dung beetle algorithm DBO module is used to construct a fitness function to optimize SSDA-HELM and find the optimal number of neurons and layers of SSDA-HELM; S3. Use the training set to train the UDSHelm algorithm model. After using the test set to test and score the UDSHelm algorithm model and optimize it, input the Raman spectrum of the crispy tilapia meat to be tested for predicting the crispness value.

2. The method for detecting the crispness of crispy tilapia based on the UDSHelm algorithm according to claim 1, wherein In step S3, the method for training the UDSHelm algorithm model is as follows: Use crispy tilapia with different crisping times as samples to obtain Raman spectral feature data and crispness values. First, use UMAP to reduce the dimension of the Raman spectral feature data in the full spectral band to determine the dimension of the input data. Then input the data into the input layer of the SSDA autoencoder. The SSDA hidden layer extracts relevant features from the complex input data, and uses an unsupervised learning method to perform layer-by-layer pre-training and fine-tuning to obtain the initial weights. Then remove the decoding part of the SDAE, connect the ELM network, and assign the obtained initial weights as the initial values of the multi-layer extreme learning machine HELM to output the crispness value. Use the dung beetle algorithm DBO to construct a fitness function to optimize the number of neurons and layers of SSDA-HELM and find the optimal number of neurons and layers.

3. A method for detecting the crispness of crispy tilapia based on the UDSHelm algorithm according to claim 2, characterized in that, The method for using the dung beetle algorithm DBO module to construct a fitness function to optimize the number of neurons and layers of SSDA-HELM is as follows: Initialize the dung beetle population in the DBO algorithm. Each individual represents a combination of a learning rate and the number of neurons. Construct a fitness function, use MSE, RMSE, and accuracy as evaluation indicators, and train the SSDA-HELM model; Update the population individuals according to the behavior rules of DBO, calculate the fitness of the new population, and select the optimal individual; If the maximum number of iterations or error convergence is reached, stop the search; otherwise, continue to optimize until the optimal learning rate and the number of neurons are obtained.

4. A crispness detection method for crispy tilapia based on the UDSHelm algorithm according to claim 1, characterized in that, The specific method of the dung beetle algorithm DBO module is as follows: S2-11. If the current individual is a rolling dung beetle, let the random number δ = rand(1). If δ < 0.9, the updated position of the dung beetle is: x i (t + 1) = x i (t) + α × k × x i (t - 1) + b × |x i (t) - X w |, Otherwise, the position of the dung beetle is updated as: x i (t + 1) = x i (t) + tan(θ)|x i (t) - x i (t - 1)|, where \(t\) represents the current iteration number; \(x\) i (t) represents the position of the \(i\)-th dung beetle at the \(t\)-th iteration; \(\alpha\) is the natural coefficient, taking 1 or -1, -1 indicates deviation from the original direction, and 1 indicates no deviation; \(k\) represents the deflection coefficient, \(k\in(0,0.2]\); \(b\) is a fixed value in \((0,1)\); \(X\) w is the globally worst position; \(\theta\in[0,\pi]\), which is the deflection angle; S2-12. If the current individual is an egg ball, the position of the dung beetle is updated by the following function: B i (t + 1)= X * + b1×(B i (t)- L b * )+ b2×(B i (t)- U b * ), L b * = max(X * × (1 - R), L b ) U b * = max(X * ×(1 + R), U b ), Among them, B i (t) represents the position of the i-th oosphere at the t-th iteration; b1 and b2 represent two independent random vectors of size 1×D, where D is the dimension of the optimization problem; X * is the current local optimal position; L b * and U b * represent the lower and upper bounds of the spawning area respectively; R = 1 - t / T max , T max represents the maximum number of iterations; L b and U b represent the lower and upper bounds of the optimization problem respectively; S2-13. If the current individual is a small dung beetle, the position of the dung beetle is updated by the following function: x i (t + 1)=x i (t)+c1×(x i (t)-L b b )+c2×(x i (t)-U b b ), L b b = max(X b × (1 - R), L b ) U b b = max(X b × (1 + R), U b ), where x i (t) represents the position information of the i-th dung beetle at the t-th iteration; c1 represents a random number following a normal distribution; c2 represents a random vector belonging to (0, 1); X b represents the global best position, L b b and U b b represent the lower and upper limits of the best foraging area, respectively; S2-14. If the current individual is a thief dung beetle, the position of the dung beetle is updated by the following function: x i (t + 1)= X b + S×g×(|x i (t)- X * | + |x i (t)- X b |), where x i (t) represents the position information of the i-th thief dung beetle at the t-th iteration; g represents a random vector of size 1×D that follows a normal distribution; S represents a constant value.

5. A method for detecting the crispness of crispy tilapia based on the UDSHelm algorithm according to claim 1, characterized in that, In step S2, the UMAP algorithm module is used to reduce the dimension of the Raman spectrum feature data of the full spectrum band as follows: The local density relationship between data points is calculated by the k-nearest neighbor algorithm k-NN, and the neighborhood size is adjusted using the smoothness parameter; After selecting the target to reduce the dimension, the weighted graph is calculated, the connection probability between data points is converted into a weight matrix, and the local manifold structure is optimized using the Riemannian metric; Gradient descent is used to optimize the objective function so that the low-dimensional representation maintains the topological structure of the original high-dimensional space as much as possible and minimizes the topological differences between data points in the high-dimensional space and the low-dimensional space.

6. A method for detecting the crispness of crispy tilapia based on the UDSHelm algorithm according to claim 1, characterized in that, In step S2, the algorithm of the SSDA-HELM prediction module is: S2-21. Select the number of hidden layers of the SSDA network, initialize the network depth k, X1 = X; X = [x (1) , x (2) ,..., x (m) T , where [] is the number of nodes in the hidden layer. Delete the decoding part of the SSDA network and interface it with the HELM network to construct the SSDA-HELM network;​ S2-22. Starting from the first hidden layer, train to obtain the input weights W of each hidden layer i and the hidden layer biases b i , and use the weights W i and the hidden layer biases b i to initialize the SSDA-HELM network as input weight pairs; S2-23. The input weight W obtained by pre-training i and the hidden layer bias b i , calculate the hidden layer output matrix A: A i = H i-1 W li , where A i is the output of the nodes in the i-th layer; H i-1 is the input of the nodes in the i-th layer; W li is the weight matrix; S2-24. According to ELM theory: Among them, H is the output of the hidden layer node, β is the output weight, T is the expected output, g(x) is the activation function, and W i = [w i，1 , w i，2 ,..., w i，n T is the weight between the input node and the i-th hidden node, β i is the weight between the i-th hidden node and the output node, and b i is the bias of the i-th hidden layer node; W i ·X j represents the inner product of W i and X j ;​ Calculating the output weight matrix of the neural network wherein, is the generalized inverse matrix of matrix A; S2-25. Calculate the output result: where is the output of the i-th layer, H i-1 is the input of the i-th layer, and g(·) is the activation function of the hidden layer; S2-26, repeat the above steps S2-22 to S2-25 until the output calculation of the last hidden layer is completed to obtain the eigenvalue; S2-27, using the extracted features as input values, and outputting a predicted value of brittleness; Among them, the method for obtaining the input weight W of the hidden layer and the hidden layer bias b through the training of the SSDA network in the step S2-22 is as follows: i and the hidden layer bias b i is as follows: Construct multiple autoencoders, each corresponding to a hidden layer, for pre-training θ i ∈{W i ,b i} parameters; the input layer of the hidden layer of each autoencoder is the output of the hidden layer of the previous autoencoder, and the output layer is the reconstruction of the previous hidden layer; use the greedy layer-by-layer training method to perform unsupervised training on each autoencoder layer by layer, obtain the weights of each layer of the SSDA network, and then fine-tune the weights layer by layer as a whole through the backpropagation algorithm to train the optimal weights of the SSDA network; The encoding process of the autoencoder is as follows: Among them, W1 is the weight matrix from the input layer to the hidden layer and from the hidden layer to the output layer, and b1 is the unit bias coefficient of the hidden layer and the output layer; σ(·) represents the activation function, and the logsig function is selected; θ represents the parameter matrix of the network, θ i ∈{W i ,b i}.

7. A method for detecting the crispness of crispy tilapia based on the UDSHelm algorithm according to claim 6, characterized in that, The network parameters of the SSDA network training are set as follows: learning rate is 0.1, the maximum number of pre-training iterations is 400, the maximum number of fine-tuning iterations is 300, the sparsity parameter is 0.5, the sparsity penalty parameter is 3, the activation function uses the sigmoid function, and the fine-tuning loss function is: Among them, θ represents the set of all parameters involved in this loss function, N represents the number of samples, and i represents the sample index; y (i) is the predicted value of the i-th sample, which is the result predicted by the model; x (i) is the true value of the i-th sample, which is the actual observed value of the sample; ||·|| F 2 is the Frobenius norm, which represents the square of the Euclidean norm and is used to measure the distance between the predicted value y (i) and the true value x (i) . By averaging this distance for all samples, it reflects the overall error degree between the predicted value and the true value of the model; λ is the regularization parameter, which is used to balance the weights of the fitting error and the regularization term in the loss function and control the penalty strength on the model parameters; l represents the number of layers of the neural network. Here, relevant calculations are performed from layer 1 to layer l; w (l) represents the weight matrix of the l-th layer of the neural network; w′ (l) represents the matrix related to the bias term of the l-th layer of the neural network; When using the test set to test and score the UDSHelm algorithm model, the K-fold cross-validation method is used.

8. A method for detecting the crispness of crispy tilapia based on the UDSHelm algorithm according to claim 1, characterized in that, In step S1, the method of using a texture analyzer to test the crispness value of crispy tilapia with different crispness at a set compression speed and deformation rate is as follows: using a 6mm cylindrical stainless steel probe of an FTC texture analyzer, in TPA test mode, the trigger force is set to 0.75N, the pre-test speed, the test speed, and the post-test speed are 1mm / s, 0.5mm / s, and 1mm / s respectively, each sample is pressed twice, and the interval between the two presses is 2s.

9. A method for detecting the crispness of crispy tilapia based on the UDSHelm algorithm according to claim 1, characterized in that, In step S1, the method for obtaining the Raman spectra of crispy tilapia with different crispness by scanning with a confocal Raman spectrometer is as follows: the scanning range of the confocal Raman spectrometer is 500-2000 cm -1 , 50× objective lens, 532 nm laser, integration time of 5 s, 1200 I / mm grating, each sample is tested three times at different positions, and its average spectrum is taken.

10. A method for detecting the crispness of crispy tilapia based on the UDSHelm algorithm according to claim 1, characterized in that, In step S1, the method for extracting the Raman spectrum characteristic data of the Raman spectrum is as follows: firstly, the Raman spectrum is subjected to polynomial fitting to realize baseline correction, then a Savitz-Golay filter is used for smoothing to remove noise, and then a Min-Max normalization process is performed; Then, the wave number, intensity and peak area of each characteristic peak in the full band of the Raman spectrum are obtained through Origin; the Raman spectrum characteristic data includes the wave number and peak intensity of each peak in the full band.

Images