Method for predicting brittleness of tilapia meat

Abstract

In order to solve the problem that an obtained prediction result is not accurate enough due to insufficient parameters of an existing fish meat crispness prediction model, the tilapia mossambica meat crispness prediction method comprises the following steps: setting a crispness stage for tilapia mossambica meat, respectively taking samples according to the crispness stage, and setting an accompanying mark; detecting the sample, generating a Raman spectrum and a visible near-infrared spectrum with the wave band of 400-3000 nm, matching the corresponding visible near-infrared spectrum and Raman spectrum, and performing data fusion to generate a fused spectrum; processing the sample by adopting a texture analyzer to obtain corresponding texture data; extracting a characteristic wave band from the fusion spectrum corresponding to the sample; and matching the characteristic wave bands and the texture data according to the accompanying marks, establishing a mapping relation, summarizing to obtain a brittleness prediction model, and obtaining brittleness prediction values of the tilapia meat at different brittleness stages. The fusion of the multi-source data and the time parameters improves the richness of the brittleness prediction model parameter information and the accuracy of the prediction result.

Description

Technical Field

[0001] The invention relates to the technical field of meat crispness prediction, and in particular to a method for predicting the crispness of tilapia meat. Background Art

[0002] Fish occupy a key position in aquatic products, and during the crisping process of fish meat, the fiber structural protein changes, the muscle fibers become thicker and the collagen increases, which not only improves the muscle elasticity, but also changes the crispness of fish meat. Therefore, when testing fish meat, we should focus on its crispness index. In actual testing work, not all fish meat will be tested for crispness. Usually, the test data of some fish meat is obtained by sampling, and a prediction model is established for the tested data. Based on the prediction model, the crispness of the remaining fish meat is predicted.

[0003] At present, the crispness of fish meat is usually detected and predicted by combining vibrational spectroscopy with texture analysis. However, when only spectroscopy is used to analyze the composition and structure of fish meat without limiting the time parameters of the sample, the obtained spectral data may have limitations, making it difficult to fully explore the composition and structure information of fish meat under different time parameters, causing the test results to lose accuracy to a certain extent. This further leads to insufficient basic parameters of the established prediction model, and the final crispness prediction results are not accurate enough.

[0004] For example, a "non-destructive rapid detection method for the texture of fish meat with different freshness" disclosed in Chinese patent literature, with the announcement number CN118758892A, includes the following steps: preparing fish meat samples with different freshness, so that the modeled fish meat samples contain primary, secondary and tertiary freshness; using a texture analyzer to test the elasticity, hardness and chewiness indexes of fish meat with different freshness; using near infrared and Raman spectroscopy to collect near infrared and Raman spectroscopy information of fish meat respectively; using CARS and SiPLS methods to extract features from the spectral information of fish meat, and then using the OSC method for preprocessing on this basis, performing intermediate data fusion on the processed spectra, and establishing PLSR prediction models for elasticity, hardness and chewiness respectively; finally, using the established prediction model to test the texture characteristics of fish meat with unknown freshness. This scheme can predict the crispness of fish meat to a certain extent. Only using the freshness of the fish meat when it is stored to limit the sample will still limit the data collected, making the parameters of the established crispness prediction model insufficient, and making the final crispness prediction result inaccurate. Summary of the invention

[0005] The invention mainly solves the problem that the parameters of the existing fish meat crispness prediction model are not sufficient, resulting in inaccurate prediction results.

[0006] In order to solve the above technical problems, a method for predicting the crispness of tilapia meat of the present invention comprises: a sampling process: setting a crispness stage for tilapia meat, taking samples respectively according to the crispness stage, and setting accompanying marks; a spectral fusion process: detecting the sample and generating a Raman spectrum and a visible near-infrared spectrum with a band of 400nm to 3000nm, matching the corresponding visible near-infrared spectrum and the Raman spectrum and performing data fusion to generate a fused spectrum; a texture data collection process: using a texture analyzer to process the sample to obtain corresponding texture data; a crispness prediction model establishment process: extracting characteristic bands from the fused spectrum corresponding to the sample; matching the characteristic bands and the texture data according to the accompanying marks, establishing a mapping relationship, and obtaining a crispness prediction model after summarizing to obtain the crispness prediction value of the tilapia meat at different crispness stages.

[0007] Preferably, the embrittlement stage includes: for tilapia with a growth time of 0 to 360 days, a embrittlement stage is set every 20 days.

[0008] Preferably, the sampling process further comprises: processing the sample into a cube with a side length of any value between 1 cm and 2 cm.

[0009] Preferably, the sampling process further comprises: maintaining the sample at any temperature between 0 and 4°C.

[0010] Preferably, the method further comprises restoring the sample to any temperature between 15 and 20° C. before the spectral fusion process.

[0011] Preferably, the spectral fusion process further comprises: detecting the sample and generating a near infrared spectrum with a wavelength range of 780nm to 2526nm and a near infrared spectrum with a wavelength range of 400cm -1 ~3500cm -1 Raman spectrum.

[0012] Preferably, the spectral fusion process also includes: using low-level data fusion to directly splice the visible near-infrared spectrum and the Raman spectrum corresponding to the sample into the fused spectrum according to spectral parameters; or using intermediate data fusion to respectively extract characteristic variables from the visible near-infrared spectrum and the Raman spectrum corresponding to the sample, and splicing the corresponding characteristic variables into the fused spectrum.

[0013] As a further preferred embodiment, the process of establishing a brittleness prediction model includes: obtaining a regression coefficient and an intercept based on the fused spectrum corresponding to the sample and the texture data; establishing a weight coefficient corresponding to the texture data; establishing a mapping relationship between the texture data and the characteristic wavelength under the corresponding intercept in combination with the weight coefficient, and obtaining a texture prediction value corresponding to the texture data; the brittleness prediction model is: summing the product of the texture prediction value and the corresponding weight coefficient, and then dividing it by the sum of the weight coefficients, and the result obtained is equal to the brittleness prediction value.

[0014] As a further preference, the process of establishing a brittleness prediction model also includes: extracting the characteristic band from the fused spectrum by a partial least squares regression method, establishing a regression relationship between the characteristic band and the texture data, and obtaining the regression coefficient and the intercept.

[0015] As a further preference, the process of establishing the brittleness prediction model also includes: establishing the weight coefficient according to a cross-validation method or a principal component analysis method.

[0016] The beneficial effects of the present invention are as follows:

[0017] First, by combining the two spectra, we can capture information on different chemical compositions and structural characteristics in fish samples, and simultaneously capture the spectral absorption characteristics and microstructural change characteristics related to most chemical and physical properties in fish. The fusion of the above multi-source data improves the richness of parameter information in the crispness prediction model and the accuracy of the prediction results.

[0018] Second, by extracting characteristic bands from the fused spectrum, we can focus on those spectral regions that have the greatest impact on brittleness. This not only reduces the data dimension and model complexity, but also enhances the model's sensitivity to key factors, thereby improving prediction accuracy.

[0019] Third, combining the texture data measured by the texture analyzer with the spectral characteristic bands ensures a direct connection between the physical measurement results and the chemical composition, which helps to more accurately map the relationship between the physical properties of fish crispness and its internal chemical structure, making the crispness prediction model parameters richer and increasing the reliability of the crispness prediction model.

[0020] Fourth, different crispness stages are set for the fish meat from which the samples are taken, and accompanying marks are set for the samples at each crispness stage to facilitate tracking of the spectral and texture characteristics of the samples, thereby realizing the prediction of the crispness value of tilapia fish meat at different crispness stages. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow chart of a method for predicting the crispness of tilapia meat according to one embodiment of the present invention;

[0022] Figure 2 The present invention is a flowchart of a spectral fusion process of a method for predicting the crispness of tilapia meat according to an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The present invention will be described in more detail below in conjunction with the accompanying drawings. It should be noted that the following description of the present invention with reference to the accompanying drawings is only illustrative and not restrictive.

[0024] Where possible, the various embodiments described below may be recombined with each other to form other embodiments not shown in the following description; the various technical features described below may also be recombined with each other to form other embodiments not shown in the following description.

[0025] Embodiment 1:

[0026] Please refer to Figure 1 and Figure 2 .

[0027] In order to solve the problem that the parameters of the existing fish meat crispness prediction model are not sufficient, resulting in inaccurate prediction results, the present embodiment provides a method for predicting the crispness of tilapia meat, including sampling, spectral fusion, collecting texture data, establishing a crispness prediction model, and finally accurately obtaining the crispness prediction value of tilapia meat through the crispness prediction model.

[0028] The sampling process includes: setting the crispness stage for tilapia meat, taking samples for each crispness stage, and setting accompanying marks. It is known that the number of samples will affect the accuracy of the crispness prediction model, so different numbers of samples can be taken as the basis for detecting the crispness of tilapia meat. Single or multiple samples can be taken, and the higher the number of samples, the richer the parameters obtained by the test.

[0029] Moreover, the parameters of tilapia meat, such as protein, collagen, amino acids, fat content, and moisture, will vary at different growth times. Therefore, by setting different brittle stages according to the growth time of tilapia meat, the samples obtained can cover various changes from the initial state to the complete brittle process. In addition, the introduction of the brittle stage as a time parameter can not only intuitively reflect the state changes of tilapia meat at different storage times, but also establish a connection with parameters such as protein, collagen, amino acids, fat content, and moisture, helping to deeply explore the changing patterns of these parameters during the brittle process and provide a comprehensive sample basis for subsequent precise testing.

[0030] Setting the accompanying mark corresponding to the brittleness stage can make the data more orderly and easier to compare in data management and analysis. At the same time, setting the accompanying mark can accurately associate the characteristic bands with the texture data, which is conducive to establishing a more accurate, generalized and easy to verify and optimize brittleness prediction model. It is equivalent to adding a time parameter synchronized with the sample to the original brittleness prediction model. When the brittleness prediction model is established, the corresponding brittleness prediction value can be obtained by obtaining the brittleness stage information of the fish to be predicted. At the same time, for the brittleness prediction value obtained by the brittleness prediction model, the brittleness stage of the corresponding sample can be obtained by obtaining its accompanying mark.

[0031] The spectrum fusion process includes: detecting samples and generating Raman spectra and visible near-infrared spectra with a band of 400nm to 3000nm, matching the corresponding visible near-infrared spectra and Raman spectra and fusing the data to generate a fused spectrum. In practical applications, visible near-infrared spectra of any of the above bands can be detected and generated for spectrum fusion with Raman spectra. For example, spectra with bands of 400nm, 1700nm, 3000nm, etc. can be used to match the corresponding Raman spectra.

[0032] The Raman spectra and visible near-infrared spectra in the above-mentioned bands corresponding to all samples are detected, and the two spectra corresponding to each sample are fused to obtain more comprehensive and rich information. The visible near-infrared spectrum in the 400nm to 3000nm band covers the spectral absorption characteristics related to most chemical and physical properties in fish meat. Within this band, rich information can be obtained, including spectral information of indicators such as protein, moisture, and fat. Raman spectroscopy is based on inelastic scattering of light and can provide information on molecular vibration and rotation energy levels. For example, it can show the vibration modes of specific chemical bonds in components such as protein and fat.

[0033] The two spectral technologies each have their own advantages, and there will be information loss when used alone. Therefore, the spectral data of the two are fused to avoid the problem of incomplete information obtained by relying on a single spectral detection, thereby improving the accuracy of the detection and the richness of the parameters, providing an accurate and rich parameter basis for the subsequent establishment of a brittleness prediction model.

[0034] The process of collecting texture data includes: using a texture analyzer to process the sample to obtain corresponding texture data. The TPA texture analyzer can be used in a measurement mode to associate the evaluation of tilapia meat texture data with sensory evaluation indicators, and efficiently and non-destructively obtain the texture data of tilapia meat. Texture data generally reflects the physical properties of tilapia meat in multiple dimensions and is closely related to the crispness of the fish meat.

[0035] The process of establishing a crispness prediction model includes: extracting characteristic bands from the fused spectrum corresponding to the sample; matching the characteristic bands and texture data of the sample according to the accompanying markers, and establishing a mapping relationship, and then summarizing to obtain a crispness prediction model. The accompanying markers and the crispness prediction model obtain the crispness prediction values ​​of tilapia meat at different crisping stages.

[0036] The characteristic bands extracted from all fused spectra are mapped to the corresponding texture data. This group of mapping relationships usually contains a single mapping relationship between characteristic bands and multi-dimensional texture data. This process combines the molecular structure and composition information reflected by the spectral data with the physical properties reflected by the texture data, so that the spectral data and texture data can complement and verify each other, and can more comprehensively understand the relationship between the crispness of fish meat and various factors.

[0037] The results of all mapping relationships are summarized into a crispness prediction model, which integrates the molecular level information contained in the spectral data and the physical property information presented by the texture data to form a comprehensive and systematic prediction system. Through this model, the crispness prediction value of tilapia meat can be accurately output.

[0038] like Figure 1 As shown, according to the above content, the steps of using the prediction method of tilapia meat crispness to generate the predicted crispness values ​​at different crispening stages in this embodiment are as follows:

[0039] Step 1: Take tilapia meat samples at different brittle stages and set accompanying marks for the samples respectively;

[0040] Step 2: Detect and generate the Raman spectrum and visible near-infrared spectrum with a wavelength range of 400nm to 3000nm corresponding to the sample; use a texture analyzer to process the sample to obtain the corresponding texture data;

[0041] Step 3: Match two spectra of the same sample based on accompanying markers;

[0042] Step 4: Determine whether the two spectra belong to the same sample. If yes, proceed to step 5; otherwise, proceed to step 3.

[0043] Step 5: Fuse the two corresponding spectra to obtain all samples and their corresponding fused spectra, and extract characteristic bands from the fused spectra;

[0044] Step 6: Match the characteristic bands and texture data corresponding to the sample according to the accompanying markers;

[0045] Step 7: Determine whether the characteristic band and texture data belong to the same sample. If so, proceed to step 8; otherwise, proceed to step 6.

[0046] Step 8: Establish a mapping relationship between characteristic bands and texture data, summarize to obtain a crispness prediction model, and output the crispness prediction values ​​of tilapia meat at different crispness stages.

[0047] In a further preferred embodiment, when a sample is taken from the tilapia meat at each brittle stage, the tracking of samples at different brittle stages can be completed by only setting an accompanying mark on the sample; and when several samples are taken from the tilapia meat at each brittle stage, it is necessary to set a two-dimensional accompanying mark, that is, each brittle mark includes a brittle stage mark and a sample mark, so as to facilitate the tracking of all samples at different brittle stages.

[0048] In a further preferred embodiment, the embrittlement stage includes: for tilapia with a growth time of 0 to 360 days, a embrittlement stage is set every 20 days.

[0049] Clearly dividing the time stages helps to establish the relationship between time and fish crispness. Dividing the stages within the time range of 0 to 360 days can ensure the diversity of sample data. It is known that the cultivation time of tilapia meat is generally 360 days, and the fish meat at different stages may have significant differences in composition, structure and texture characteristics. In the early growth stage, fish meat contains more water and less fibrous tissue. Over time, the fibrous tissue gradually increases, the water content decreases relatively, and the crispness will also change. Therefore, taking 20 days as a crispening stage can accurately track the dynamic changes of tilapia meat during the crispening process. 20 days is a more appropriate stage unit, which can avoid the fish meat crispening significantly due to too long intervals, and can also avoid the fish meat crispening not being obvious due to too short intervals. Moreover, 20 days can divide 360 ​​days into 18 even stages. For example, 0 to 19 days corresponds to stage 0, 20 to 39 days corresponds to stage 1, 40 to 59 days corresponds to stage 2, 60 to 79 days corresponds to stage 3, 80 to 99 days corresponds to stage 4, and so on.

[0050] By analyzing and fusing the two spectra and texture data of samples at different stages, the mapping relationship between low-level fusion data, intermediate fusion data and texture data can be continuously improved. The crispness prediction model established in this way can more accurately identify the characteristics of fish meat at different time stages, thereby improving its ability to predict the crispness of tilapia meat, combining the crispness of tilapia meat with the brittleness time, making it more accurate and reliable in actual detection applications. Therefore, through this staged sampling method, samples of fish meat at various stages of gradual brittleness during growth can be obtained, providing a rich data basis for establishing an accurate detection model, so that the model can better adapt to the detection of fish meat with different degrees of crispness.

[0051] In a further preferred embodiment, the 60-79-day-old tilapia meat is sampled. During the growth of tilapia meat, less than 60 days is in the early stage of rapid growth, with the main components accumulating and the changes being gentle. From 60 to 79 days, the growth slows down and the physiological changes occur. The changes in water content, protein, fat and related enzymes become the turning point of the composition change. After more than 79 days, the growth is stable and the composition changes tend to be stable. These three stages constitute the complete process of the composition changing with the growth time. Therefore, 60-79-day-old tilapia can be selected as the sampling raw material. For example, 60-day-old, 65-day-old, 77-day-old or 79-day-old tilapia can be selected as the sampling raw material.

[0052] In another preferred embodiment, the sample taken is maintained at any temperature within the range of 0 to 4°C. The above setting can prevent the fish meat from freezing in a storage environment below 0°C, thereby destroying its cell structure and causing deviations in subsequent test data; and can prevent the fish meat from undergoing biochemical reactions or internal moisture changes in a storage environment above 4°C, thereby interfering with subsequent test data. Therefore, the temperature of the sample can be set to 0°C, 1°C, 3.5°C, 4 degrees, etc., to ensure that its storage environment is reasonable. It should be noted that if there are multiple samples, the storage temperature of all samples must be unified.

[0053] In another preferred embodiment, the sample is processed into a cube with a side length of any value between 1 cm and 2 cm. The above setting can prevent samples below 1 cm from being too small and affecting the operability of subsequent tests; it can also avoid the problem that samples above 2 cm are too large, resulting in internal unevenness, unsuitability for testing instruments, or waste of testing materials. Therefore, the sample can be set to a cube with a side length of 1 cm, 1.3 cm, 2 cm, etc. It should be noted that if there are multiple samples, the specifications of all samples must be unified.

[0054] In another preferred embodiment, the back muscle area of ​​tilapia meat is selected as the area for sampling. The meat quality of the back muscle of tilapia is relatively stable compared to other parts. The back muscle of tilapia plays a major supporting role for the fish body during the movement of the body, so the changes in its tissue structure and composition are relatively small, and the meat quality of the back of tilapia will not be too different due to factors such as the shape and growth rate of the fish body. The fibers of the back muscles of different tilapia individuals are arranged more closely and regularly, and the individual differences are small, and the muscle development and accumulation of nutrients are relatively stable. This allows the back of the fish meat to provide more consistent test data when testing the crispness of the fish meat, which is more conducive to accurately detecting and analyzing characteristics related to crispness compared to the fish meat in other parts.

[0055] In another preferred embodiment, the sample is restored to any temperature between 15 and 20°C before the spectral fusion process. Within this temperature range, the physical and chemical properties of the sample are stable, which can ensure the accuracy of the test results and avoid sample deterioration due to temperatures above 20°C and the normal physiological structure of the sample being affected by temperatures below 15°C. For example, the sample can be restored to temperatures such as 15°C, 17°C, 19.5°C, and 20°C to reduce the error in the test results caused by temperature differences. It should be noted that if there are multiple samples, the temperatures of the multiple samples need to be kept uniform to ensure the comparability and consistency of the data.

[0056] In another preferred embodiment, a near infrared spectrum with a wavelength range of 780 nm to 2526 nm and a near infrared spectrum with a wavelength range of 400 cm -1 ~3500cm -1 For example, near-infrared spectra at 780nm, 1653nm, 2526nm and the Raman spectrum at 400cm -1 、2450cm -1 、3500cm -1 The Raman spectra of different wavelength bands are used in combination.

[0057] Detecting the two spectra according to the above-mentioned band range can cover the information of the near-infrared region, molecular vibration, and rotational energy level, and can give full play to the complementarity of near-infrared spectroscopy and Raman spectroscopy. For example, when detecting protein in fish, near-infrared spectroscopy can provide information on hydrogen-containing groups of amino acid residues in proteins, while Raman spectroscopy can display the vibration modes of chemical bonds such as protein peptide bonds. The combination of the two spectra can more completely describe the state of proteins, thereby more accurately correlating the composition and crispness of fish. The effective band of near-infrared spectroscopy is in the range of 780nm to 2526nm, and within this range, the complementary detection effect with Raman spectroscopy can be better achieved.

[0058] Similarly, the Raman spectrum is less than 400 cm -1 When entering the far Raman spectral region, the signals of most chemical bond vibration information closely related to the composition and structure of the material are weak; and the band greater than 3500cm -1 When the frequency of the Raman spectrum is above 400 cm, it exceeds the vibration frequency range of most chemical bonds in common molecules, and the effective information about the composition of the substance that can be provided is greatly reduced. -1 ~3500cm -1 When it is outside, it is also impossible to achieve the complementary effect with the near-infrared spectrum.

[0059] In another preferred embodiment, the spectral data fusion process also includes: using low-level data fusion to directly splice the two spectra corresponding to the sample into a fused spectrum according to spectral parameters; or using intermediate data fusion to respectively extract characteristic variables from the two spectra corresponding to the sample, and splice the corresponding characteristic variables into a fused spectrum.

[0060] The low-level data fusion method simply and directly splices the data of the two spectra into a matrix, namely the fused spectrum. This fused spectrum retains the spectral parameters of the two spectra, such as wavelength position, absorbance or scattering intensity. This allows subsequent data processing and analysis to use the complete information of the two spectra at the same time without losing any details that may be valuable for judging the crispness of fish meat. When establishing a relationship model between the crispness of fish meat and spectral data, a large number of data points in the fused spectrum can be used as variable inputs, increasing the complexity and accuracy of the model.

[0061] The intermediate data fusion method further optimizes the data structure, extracts features from the two spectra separately, and then splices the extracted feature variables into a new matrix, namely the fused spectrum. This fused spectrum can focus on the key parts of the two spectra that best reflect the information related to the crispness of fish meat. Therefore, the variables in the generated fused spectrum are more strongly correlated with the crispness of fish meat. When establishing a fish crispness prediction model, the influence of irrelevant or interfering data can be reduced, thereby improving the efficiency and accuracy of the model. For example, by performing data analysis methods such as principal component analysis or multivariate linear regression on the feature matrix, the quantitative relationship between the crispness of fish meat and the spectral characteristics can be more accurately determined, thereby achieving a more accurate crispness prediction.

[0062] like Figure 2 As shown, the spectrum fusion process of this embodiment includes the following steps:

[0063] Step 1: Match the two spectra corresponding to the sample according to the accompanying markers;

[0064] Step 2: Determine whether the two spectra belong to the same sample, if yes, proceed to step 3, otherwise proceed to step 6;

[0065] Step 3: Perform low-level data fusion or intermediate data fusion on the two spectra. The low-level data fusion is performed in step 4, and the intermediate data fusion is performed in step 5;

[0066] Step 4: Directly concatenate the spectral parameters of the two spectra by matching the accompanying markers according to the sample complexity;

[0067] Step 5: According to the sample complexity, extract the feature vectors of the two spectra through the accompanying markers and merge the feature vectors;

[0068] Step 6: Generate a fusion spectrum of the two spectra corresponding to all samples, and output the fusion spectrum corresponding to the sample according to the accompanying label.

[0069] In this embodiment, the process of low-level data fusion is specifically described as follows:

[0070] Assume that the number of samples is m, and the matrix formed by the above visible near infrared spectrum data is X IR , whose dimension is m×n i The matrix formed by the Raman spectrum data is X R , whose dimension is m×n j .

[0071] In the low-level data fusion process, the visible near-infrared spectrum and the Raman spectrum are directly spliced ​​to form a fusion spectrum of low-level data X low , the splicing rule is And at this time, the fusion spectrum of the low-level data X low The dimension is m×(n i +n j ), which directly integrates the raw data of the two spectra.

[0072] In this embodiment, the process of intermediate data fusion is specifically described as follows:

[0073] In the intermediate data fusion process, the visible near infrared spectrum data matrix X IR For feature extraction, assuming that the principal component analysis method (PCA) is used to extract the first p principal components, the above-mentioned visible near-infrared spectrum feature matrix is ​​obtained as F IR (dimension is m×p), the extraction process is: F IR =X IR W IR . Where W IR is the PCA transformation matrix, whose dimension is n i ×p, by IR The covariance matrix of PCA is obtained by eigenvalue decomposition and other operations. IR Let the above visible near infrared spectrum feature matrix F IR Keep the aforementioned visible near infrared spectrum data matrix X as much as possible IR Similarly, for the Raman spectroscopy data matrix X R Perform feature extraction to obtain the characteristic matrix F of the Raman spectrum R , whose dimension is m×q. The extracted visible near-infrared spectral feature matrix F IR and the Raman spectrum feature matrix F R Fusion spectra stitched into intermediate data X mid , expressed as Its dimension is m×(p+q). This matrix integrates the key information of the two spectra after feature extraction and removes some redundant data.

[0074] In a further preferred embodiment, in order to improve the accuracy of the fused spectrum obtained after the low-level data fusion, the feature matrix of the low-level data fusion is usually preprocessed or normalized. low The element x in the i-th row and j-th column of ij (i=1,2,...,m; j=1,2,...,n i +n j ), the normalized formula is Among them, min k (x kj ) and max k (x kj ) are the minimum and maximum values ​​in the jth column element respectively. After normalization, the fused spectrum of the processed low-level data is obtained

[0075] In a further preferred embodiment, the spectral parameters include: wavelength position, absorbance and scattering intensity; the characteristic variables include: wavelength point, peak intensity, peak area and low-dimensional characteristic vector. Among them, the wavelength position reflects the molecular energy level information, which helps to identify the corresponding spectral position of the component; the absorbance is related to the concentration and structure of the substance, and can establish a preliminary connection between the two spectra; the scattering intensity supplements the sample structure and other information to make the fused data more complete and accurate. The wavelength point, peak intensity, and peak area respectively mine the spectral features from the perspectives of key identification, quantitative and qualitative, and substance content to enhance the analysis ability; the low-dimensional characteristic vector reduces the dimension of the high-dimensional spectrum, retains the main information, fuses the data at the abstract level, and provides effective input for subsequent analysis.

[0076] In a further preferred embodiment, the process of establishing a brittleness prediction model includes: obtaining regression coefficients and intercepts based on the fused spectrum and texture data corresponding to the sample; establishing weight coefficients corresponding to the texture data; establishing a mapping relationship between the texture data and the characteristic wavelength under the corresponding intercept in combination with the weight coefficients, and obtaining a texture prediction value corresponding to the texture data; the brittleness prediction model is: summing the product of the texture prediction value and the corresponding weight coefficient, and then dividing it by the sum of the weight coefficients, and the result obtained is equal to the brittleness prediction value.

[0077] Based on the content of this embodiment, the process of obtaining characteristic bands, regression coefficients and intercepts is described by taking the texture data as examples of elasticity, hardness, chewiness and cohesion. These four properties are important indicators for measuring the crispness of fish meat, and are related to the changes in the internal structure and composition of the fish meat when it is crispy, and are reflected in the spectral data. Combining texture data with spectral data can evaluate the crispness in multiple dimensions, and using machine learning training can build an accurate crispness prediction model to assist in fish meat quality detection and prediction. However, it should be noted that in the actual detection and prediction of the crispness of tilapia fish meat, in addition to the four items of elasticity, hardness, chewiness and cohesion, there are other factors such as cohesion and resilience that may be related to the crispness of fish meat and can be reflected through spectral data.

[0078] The process of using partial least squares regression (PLSR) to obtain the characteristic bands, regression coefficients and intercepts of the fused spectrum includes: setting the maximum number of iterations maxIter, for example, it can be set to 100 times to ensure that the algorithm can converge. Setting the convergence threshold ε, for example, it can be set to 10 -6 , used to determine whether the algorithm converges. Initialize the regression coefficient vector β to an all-zero vector with a dimension of (n i +n j )×1, and the intercept β0 is initialized to 0.

[0079] In each iteration, the predicted value is calculated Where X represents the fused spectrum, which can represent X low , or X mid Calculate the error vector Calculate the weight matrix W of X, usually using Update the regression coefficient vector β new =β+λW, where λ is the step size parameter, which is usually determined by experience or debugging, and can be set to 0.01, for example. Update intercept Among them, NN is the texture parameter. When NN is E, it represents elasticity; when it is H, it represents hardness; when it is C, it represents chewiness; when it is C0, it represents cohesiveness.

[0080] Calculate convergence indicators, such as and If Δβ<ε and Δβ0<ε or the number of iterations reaches maxIter, stop the iteration; otherwise, update the regression coefficient vector β to β new , update the intercept vector β0 to Continue to the next iteration.

[0081] The importance score of each feature band can be calculated by calculating the absolute value of each element in the regression coefficient vector β. The spectral band corresponding to the element with the larger absolute value is more important. The importance scores are sorted, and the first k features with higher scores are selected as the feature bands of the fused spectrum. The selection of k is determined according to the actual situation and requirements. For example, the optimal k value can be selected by cross-validation method.

[0082] In a further preferred embodiment, the characteristic band is extracted from the fused spectrum by using the partial least squares regression method, and a regression relationship is established between the characteristic band and the texture data to obtain the regression coefficient and the intercept.

[0083] The parameters of the brittleness prediction model are as follows:

[0084] The intercept and regression coefficients include: In the elastic model, the intercept is β 0E , the regression coefficient is β 1E , β 2E , ..., β mE ; In the hardness model, the intercept is β 0H , the regression coefficient is β 1H , β 2H , ..., β mH ; In the chewiness model, the intercept is β 0C , the regression coefficient is β 1C , β 2C , ..., β mC ; In the cohesive model, the intercept is β 0C0 , the regression coefficient is β 1C0 , β 2C0 , ..., β mC0 .

[0085] Characteristic wavelengths include: X low , or X mid (X will be used instead in the following sections) indirectly reflects the information of the characteristic wavelength X(:, k:l) of the spectral data. These characteristic wavelength information, after operations such as feature extraction and matrix concatenation, are the basis for calculating various texture prediction values ​​and play a key role in accurately predicting texture data and the final brittleness prediction value.

[0086] Weight parameter: set to (ω E ,ω H ,ω C ,ω C0 ), these weight coefficients are used to measure the importance of various texture data such as elasticity, hardness, chewiness and cohesion in the prediction of crispness.

[0087] Based on the above parameters, the mapping relationship between elasticity and characteristic band is established: E = β 0E +β 1EX(:,1:p)+β 2E X(:,p+1:2p)+…+β mE X(:, m-1: mp), and calculate the elasticity prediction value E. Here X(:, k: l) represents the submatrix from the kth column to the lth column in the fusion feature X of the intermediate data. Establish the mapping relationship between hardness and feature band H = β 0H +β 1H X(:,1:p)+β 2H X(:,p+1:2p)+…+β mH X(:, m-1: mp), and calculate the hardness prediction value H. Establish the mapping relationship between chewiness and characteristic band C = β 0C +β 1C X(:,1:p)+β 2C X(:,p+1:2p)+…+β mC X(:, m-1: mp), and calculate the chewing prediction value C. Establish the mapping relationship between cohesion and characteristic band C0 = β 0C0 +β 1C X(:,1:p)+β 2C0 X(:,p+1:2p)+…+β mC0 X(:, m-1:mp), and calculate the predicted cohesion value C0.

[0088] Finally, combining various mapping relationships, a brittleness prediction model is constructed: That is, the predicted value of each texture is multiplied by the corresponding weight coefficient, the sum is then divided by the sum of the weight coefficients to obtain the predicted value of brittleness D.

[0089] In a further preferred embodiment, the process of establishing the brittleness prediction model further includes: establishing weight coefficients according to a cross-validation method or a principal component analysis method.

[0090] Assume that the combination of weight parameters is (ω E ,ω H ,ω C ,ω C0 ).

[0091] If the cross-validation method is used to determine the weight parameters, the data set is divided into a training set and a validation set. E ,ω H ,ω C ,ω C0 ) to build a model and evaluate the model performance on the validation set. By continuously adjusting the weight coefficient combination, the weight coefficient combination that can achieve the best prediction performance on the validation set is selected. In mathematical expression, it is to find a set of weight coefficients Make a certain performance indicator on the validation set (such as minimum mean square error, highest accuracy, etc., assuming the performance indicator is P) meet

[0092] If the principal component analysis method is used to determine the weight parameters, the texture data (including elasticity, hardness, chewiness, cohesion and other related data) are subjected to principal component analysis. First, the original texture data is converted into mutually orthogonal principal components, and then the weight coefficient is determined based on the variance contribution rate of each principal component. The larger the variance contribution rate, the more important the information contained in the principal component, and the higher the corresponding texture data weight. Assume that the variance contribution rates obtained by principal component analysis are (λ E , H , C , C0 ), the weight coefficient is calculated as:

[0093] It should be understood that the embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope limited by the appended claims of the application.

Claims

1. A method for predicting the crispness of tilapia meat, characterized in that: include: Sampling process: setting crisping stages for tilapia meat, taking samples according to the crisping stages, and setting accompanying marks; Spectral fusion process: detecting the sample and generating a Raman spectrum and a visible near-infrared spectrum with a wavelength range of 400nm to 3000nm, matching the corresponding visible near-infrared spectrum and the Raman spectrum and performing data fusion to generate a fused spectrum; Process of collecting texture data: using a texture analyzer to process the sample to obtain corresponding texture data; The process of establishing a crispness prediction model: extracting characteristic bands from the fused spectrum corresponding to the sample; matching the characteristic bands and the texture data according to the accompanying tags, and establishing a mapping relationship, and obtaining a crispness prediction model after summarizing, and obtaining the crispness prediction values ​​of the tilapia meat at different crisping stages.

2. The method for predicting the crispness of tilapia meat according to claim 1, wherein: The embrittlement stage includes: For tilapia with a growth period of 0 to 360 days, a brittle stage is set every 20 days.

3. The method for predicting the crispness of tilapia meat according to claim 1, wherein: The sampling process also includes: The sample is processed into a cube with a side length ranging from 1 cm to 2 cm.

4. The method for predicting the crispness of tilapia meat according to claim 1, wherein: The sampling process also includes: The sample was maintained at any temperature between 0 and 4°C.

5. The prediction method according to claim 1, characterized in that: The method further comprises restoring the sample to any temperature between 15° C. and 20° C. before the spectrum fusion process.

6. The method for predicting the crispness of tilapia meat according to claim 1, wherein: The spectral fusion process further includes: The sample is detected and a near infrared spectrum with a wavelength range of 780nm to 2526nm and a near infrared spectrum with a wavelength range of 400cm -1 ~3500cm -1 Raman spectrum.

7. The method for predicting the crispness of tilapia meat according to claim 1, wherein: The spectral fusion process further includes: Using low-level data fusion, the visible near-infrared spectrum and the Raman spectrum corresponding to the sample are directly spliced ​​into the fused spectrum according to spectral parameters; or By adopting intermediate data fusion, characteristic variables are respectively extracted from the visible near-infrared spectrum and the Raman spectrum corresponding to the sample, and the corresponding characteristic variables are spliced ​​into the fused spectrum.

8. The method for predicting the crispness of tilapia meat according to claims 1 to 7, characterized in that: The process of establishing the brittleness prediction model comprises: Obtaining a regression coefficient and an intercept according to the fused spectrum corresponding to the sample and the texture data; Establishing a weight coefficient corresponding to the texture data; Establishing a mapping relationship between the texture data and the characteristic wavelength at a corresponding intercept in combination with the weight coefficient, and obtaining a texture prediction value corresponding to the texture data; The crispness prediction model is: summing the products of the texture prediction value and the corresponding weight coefficients, and then dividing the sum by the sum of the weight coefficients, and the result obtained is equal to the crispness prediction value.

9. The method for predicting the crispness of tilapia meat according to claim 8, wherein: The process of establishing the brittleness prediction model also includes: extracting the characteristic band from the fused spectrum by a partial least squares regression method, establishing a regression relationship between the characteristic band and the texture data, and obtaining the regression coefficient and the intercept.

10. The method for predicting the crispness of tilapia meat according to claim 8, characterized in that: The process of establishing the brittleness prediction model also includes: establishing the weight coefficient according to a cross-validation method or a principal component analysis method.

Images