Method and device for predicting quality of rice product raw material based on deep feature fusion

Abstract

The application belongs to the technical field of rice product processing, and discloses a rice product raw material quality prediction method and device based on deep feature fusion. The method comprises the following steps: a new feature fusion method is proposed, deep feature fusion is composed of coarse level feature fusion and fine level feature fusion, and then the deep feature fusion is input into a prediction model for prediction; different feature extraction methods are adopted to obtain different feature data; different weight values are assigned to different feature data according to the importance of the features to obtain feature fusion data; an intelligent optimization algorithm is selected to optimize the structure of the deep feature fusion and the fusion weight value of the deep feature fusion to obtain an optimal prediction model according to the requirement of achieving an optimal prediction effect; rice powder raw material index data are predicted according to the optimal prediction model and rice product index data; and raw material quality information is obtained according to the prediction result of the rice powder raw material index data. In the above manner, the accuracy of the prediction result is improved, the quality evaluation standard is met, and the rice product raw material quality is predicted.

Description

Technical Field

[0001] This invention relates to the field of rice product processing technology, and in particular to a method and apparatus for predicting the quality of rice product raw materials based on deep feature fusion. Background Technology

[0002] With the development of the grain and food industry and the improvement of living standards, selecting suitable raw materials to produce high-quality rice products is one of the problems that needs to be solved. The fundamental condition for producing high-quality products is having high-quality raw materials. However, with an increasing number of raw materials available for making rice products, selecting suitable raw materials has become crucial for producing high-quality rice products.

[0003] The problem of predicting the quality of rice product raw materials involves obtaining information on the raw material and product indicators of rice products, predicting the content of rice flour raw material indicators, providing relevant suggestions for quality control in the production of high-quality products, providing a reference for the selection of raw materials, and providing guidance for the procurement of raw materials for the production of high-quality rice products.

[0004] Current research on food forecasting mainly involves correlation analysis and regression equation establishment between the types, components, or physicochemical properties of raw materials and product characteristics to predict product indicators. However, simply analyzing the data of raw material indicators and product indicators results in insufficient accuracy in predicting product indicators, and the predicted indicator content fails to meet quality assessment standards.

[0005] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0006] The main objective of this invention is to provide a method and apparatus for predicting the quality of rice product raw materials based on deep feature fusion. This aims to solve the technical problem that existing technologies simply analyze raw material indicators and product indicators, resulting in insufficient accuracy in predicting the content of raw material indicators in rice products, and the predicted indicator content failing to meet quality assessment standards.

[0007] To achieve the above objectives, this invention provides a method for predicting the quality of rice product raw materials based on deep feature fusion, the method comprising the following steps:

[0008] Based on the current number of fusions and the preset feature extraction strategy, determine the current feature extraction strategy;

[0009] Based on the current feature extraction strategy and rice product sample data, feature data is obtained;

[0010] Determine the initial weight data based on the importance of the feature data;

[0011] Based on the initial weight data and feature data, feature fusion is performed to obtain feature fusion data, and the current fusion count is updated;

[0012] Based on the feature fusion data and the current fusion count, deep feature fusion data is obtained;

[0013] The deep feature fusion data is input into a preset prediction model, and an intelligent optimization algorithm is used to optimize the structure of the preset prediction model and the initial weight data of the deep feature fusion data to establish the optimal prediction model.

[0014] Based on the optimal prediction model, the indicator data of the target rice product are predicted to obtain the raw material prediction results;

[0015] Based on the raw material prediction results, raw material quality information is obtained.

[0016] Optionally, obtaining deep feature fusion data based on the feature fusion data and the number of fusions includes:

[0017] When the current fusion count is greater than or equal to the preset fusion count, deep feature fusion data is determined based on the feature fusion data.

[0018] Optionally, before determining the deep feature fusion data based on the feature fusion data when the current fusion count is greater than or equal to a preset fusion count, the method further includes:

[0019] When the current number of fusions is less than the preset number of fusions, a new current feature extraction method is determined based on the current number of fusions;

[0020] Based on the aforementioned feature fusion data and the new current feature extraction method, new feature data is obtained;

[0021] Based on the new feature data, return to the step of determining the initial weight data according to the importance of the feature data.

[0022] Optionally, the step of inputting the deep feature fusion data into a preset prediction model and using an intelligent optimization algorithm to optimize the structure of the preset prediction model and the initial weight data of the deep feature fusion data to establish an optimal prediction model includes:

[0023] The deep feature fusion data is input into a preset prediction model to obtain initial prediction data;

[0024] Based on the initial prediction data and raw material sample data, an intelligent optimization algorithm is used to perform internal and external nesting optimization on the initial weight data of the structure and depth feature fusion data of the preset prediction model, so as to obtain an initial combination of optimized weight data and optimized prediction model structure.

[0025] Based on the initial combination of the optimized weight data and the optimized prediction model structure, an optimal prediction model is established.

[0026] Optionally, establishing the optimal prediction model based on the initial combination of the optimized weight data and the optimized prediction model structure includes:

[0027] Based on the initial combination of the optimized weight data and the optimized prediction model structure, as well as the deep feature fusion data, the optimized prediction data is obtained.

[0028] Based on the optimized prediction data and raw material sample data, evaluation data for the initial combination of each optimized weight data and the optimized prediction model structure are obtained;

[0029] Based on the evaluation data, determine the target combination of optimizing weight data and optimizing the prediction model structure;

[0030] Based on the target combination of the optimized weight data and the optimized prediction model structure, an optimal prediction model is established.

[0031] Furthermore, to achieve the above objectives, the present invention also proposes a prediction device for the quality of rice product raw materials based on deep feature fusion, the prediction device for the quality of rice product raw materials based on deep feature fusion comprising:

[0032] The fusion module is used to determine the current feature extraction strategy based on the current number of fusion attempts and the preset feature extraction strategy.

[0033] The fusion module is also used to obtain feature data based on the current feature extraction strategy and rice product sample data;

[0034] The fusion module is also used to determine initial weight data based on the importance of the feature data;

[0035] The fusion module is further configured to perform feature fusion based on the initial weight data and feature data to obtain feature fusion data, and update the current fusion count;

[0036] The fusion module is further configured to obtain deep feature fusion data based on the feature fusion data and the current fusion count;

[0037] The optimization module is used to input the deep feature fusion data into a preset prediction model, and use an intelligent optimization algorithm to optimize the structure of the preset prediction model and the initial weight data of the deep feature fusion data to establish an optimal prediction model.

[0038] The prediction module is used to predict the index data of the target rice product based on the optimal prediction model, and obtain the raw material prediction results.

[0039] The prediction module is also used to obtain raw material quality information based on the raw material prediction results.

[0040] Optionally, the fusion module is further configured to determine deep feature fusion data based on the feature fusion data when the current fusion count is greater than or equal to a preset fusion count;

[0041] Optionally, the fusion module is further configured to determine a new current feature extraction method based on the current fusion count when the current fusion count is less than a preset fusion count;

[0042] Based on the aforementioned feature fusion data and the new current feature extraction method, new feature data is obtained;

[0043] Based on the new feature data, return to the step of determining the initial weight data according to the importance of the feature data.

[0044] Optionally, the optimization module is further configured to input the deep feature fusion data into a preset prediction model to obtain initial prediction data;

[0045] Based on the initial prediction data and raw material sample data, an intelligent optimization algorithm is used to perform internal and external nesting optimization on the initial weight data of the structure and depth feature fusion data of the preset prediction model, so as to obtain an initial combination of optimized weight data and optimized prediction model structure.

[0046] Based on the initial combination of the optimized weight data and the optimized prediction model structure, an optimal prediction model is established.

[0047] Optionally, the optimization module is further configured to obtain optimized prediction data based on the initial combination of the optimized weight data and the optimized prediction model structure, as well as the deep feature fusion data;

[0048] Based on the optimized prediction data and raw material sample data, evaluation data for the initial combination of each optimized weight data and the optimized prediction model structure are obtained;

[0049] Based on the evaluation data, determine the target combination of optimizing weight data and optimizing the prediction model structure;

[0050] Based on the target combination of the optimized weight data and the optimized prediction model structure, an optimal prediction model is established.

[0051] In this invention, a current feature extraction strategy is determined based on the current number of fusions and a preset feature extraction strategy. Feature data is obtained based on the current feature extraction strategy and rice product sample data. Initial weight data is determined based on the importance of the feature data. Feature fusion is then performed using the initial weight data and the feature data to obtain fused feature data, and the current number of fusions is updated. Furthermore, deep feature fusion data is obtained based on the fused feature data and the current number of fusions. This deep feature fusion data is input into a preset prediction model. An intelligent optimization algorithm is used to optimize the structure of the preset prediction model and the initial weight data of the deep feature fusion data to establish an optimal prediction model. The optimal prediction model is then used to predict the indicator data of the target rice product, obtaining the raw material prediction result and thus the raw material quality information. Compared to existing technologies that simply analyze raw material and product indicators, this invention effectively improves the accuracy of prediction results through deep feature fusion of data and optimization of the prediction model and weights. It overcomes the technical problem of insufficient prediction accuracy of raw material indicator content in rice products, resulting in predicted indicator content that does not meet quality assessment standards. This invention can accurately predict the raw material indicator content of rice products, ensuring it meets quality assessment standards, thereby predicting the quality of rice product raw materials and providing a reference for selecting raw materials for producing high-quality rice products. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the structure of a device for predicting the quality of rice product raw materials based on deep feature fusion, which is part of the hardware operating environment involved in the embodiments of the present invention.

[0053] Figure 2 This is a flowchart illustrating the first embodiment of the method for predicting the quality of rice product raw materials based on deep feature fusion according to the present invention.

[0054] Figure 3 This is a schematic diagram of the BPNN structure of an embodiment of the method for predicting the quality of rice product raw materials based on deep feature fusion according to the present invention.

[0055] Figure 4 This is a schematic diagram of the RF structure of an embodiment of the method for predicting the quality of rice product raw materials based on deep feature fusion according to the present invention.

[0056] Figure 5 This is a schematic diagram of the XGBoost structure of an embodiment of the method for predicting the quality of rice product raw materials based on deep feature fusion according to the present invention.

[0057] Figure 6 This is a schematic diagram of the application process for predicting the quality of rice product raw materials according to an embodiment of the method for predicting the quality of rice product raw materials based on deep feature fusion of the present invention, specifically for predicting the quality of rice product raw materials.

[0058] Figure 7This is a flowchart illustrating the second embodiment of the method for predicting the quality of rice product raw materials based on deep feature fusion according to the present invention.

[0059] Figure 8 This is a schematic diagram of single-layer multi-dimensional feature fusion, representing an embodiment of the method for predicting the quality of rice product raw materials based on deep feature fusion according to the present invention.

[0060] Figure 9 This is a schematic diagram of the deep feature fusion process in an embodiment of the method for predicting the quality of rice product raw materials based on deep feature fusion according to the present invention.

[0061] Figure 10 This is a schematic diagram of the rice flour raw material index content prediction process according to an embodiment of the rice product raw material quality prediction method based on deep feature fusion of the present invention.

[0062] Figure 11 This is a structural block diagram of the first embodiment of the rice product raw material quality prediction device based on deep feature fusion of the present invention.

[0063] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0064] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0065] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a device for predicting the quality of rice product raw materials based on deep feature fusion, which is part of the hardware operating environment of the embodiment of the present invention.

[0066] like Figure 1As shown, the rice product raw material quality prediction device based on deep feature fusion may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0067] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the predictive device for the quality of rice product raw materials based on deep feature fusion, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0068] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a prediction program for the quality of rice product raw materials based on deep feature fusion.

[0069] exist Figure 1 In the rice product raw material quality prediction device based on deep feature fusion shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and memory 1005 in the rice product raw material quality prediction device based on deep feature fusion of the present invention can be set in the rice product raw material quality prediction device based on deep feature fusion. The rice product raw material quality prediction device based on deep feature fusion calls the rice product raw material quality prediction program based on deep feature fusion stored in the memory 1005 through the processor 1001, and executes the rice product raw material quality prediction method based on deep feature fusion provided in the embodiment of the present invention.

[0070] This invention provides a method for predicting the quality of rice product raw materials based on deep feature fusion, referring to... Figure 2 , Figure 2This is a flowchart illustrating the first embodiment of a method for predicting the quality of rice product raw materials based on deep feature fusion according to the present invention.

[0071] In this embodiment, the method for predicting the quality of rice product raw materials based on deep feature fusion includes the following steps:

[0072] Step S10: Determine the current feature extraction strategy based on the current number of fusions and the preset feature extraction strategy.

[0073] It should be noted that the execution subject of this embodiment is a computer, which can be any computer capable of running a prediction program for the quality of rice product raw materials based on deep feature fusion. This embodiment does not impose any restrictions. The prediction program for the quality of rice product raw materials based on deep feature fusion, installed in the computer, reverse-predicts the raw material indicators based on the rice product indicators and determines the quality of the rice product raw materials. To produce high-quality rice products with good taste, attractive appearance, and high nutritional value, it is necessary to analyze the indicators of the rice products, accurately reverse-predict the indicators of the processed products, predict the content of various raw material indicators, and analyze and evaluate the predicted raw material indicator content according to relevant evaluation standards. This provides a good reference for the selection of raw materials and quality control in the production of high-quality rice noodles. Therefore, the core of the problem of predicting the content of raw material indicators in rice noodles is to find a feature extraction method and prediction model suitable for rice noodle data, which can help improve the prediction accuracy of raw material indicator content. This embodiment improves the data feature extraction stage and finds a suitable prediction model to optimize and improve the prediction accuracy.

[0074] It should be understood that the current fusion count refers to the number of times the data has undergone feature fusion, used to determine which layer has been reached, with an initial value of 0. The preset feature extraction strategy is a pre-defined feature extraction method used in each layer of feature fusion, which can be a method such as correlation analysis (Pearson) and factor analysis (FA). This embodiment does not limit this and can be flexibly adjusted according to the actual situation. The number of feature extraction methods in each layer of feature fusion is not limited in this embodiment. The current feature extraction strategy is the feature extraction method used in the current layer.

[0075] In practice, based on the current number of fusion iterations, the layer of feature fusion is determined, and the corresponding feature extraction method is found for subsequent feature extraction and fusion.

[0076] Step S20: Obtain feature data based on the current feature extraction strategy and rice product sample data.

[0077] Table 1 Rice Noodle Index Data

[0078] Indicator Name Minimum value Maximum value average value Standard deviation Viscoelasticity (VIS) 2.500 7.199 4.171 1.389 Subtlety (ES) 3.200 6.300 4.605 0.774 Softness 3.511 7.297 5.038 0.893 Chewiness 2.803 8.485 4.446 1.151 Appearance and color (CAL) 14.007 18.000 16.544 0.880 Tasting Score (TS) 30.801 41.069 340804 2.672 Cooking loss (CL) 2.530 10.000 7.119 1.866 Rehydration Time (RTime) 22.756 32.487 27.435 2.282 Taste 5.101 6.474 5.822 0.394 Product moisture content (PWC) 11.801 13.198 12.627 0.318 Natural breakage rate (BR) 1.159 28.494 10.370 5.938

[0079] It is understood that the rice product sample data refers to the content data of rice product indicators with known corresponding raw material indicator content data, used as sample data for feature extraction. This embodiment does not limit the type of rice product; it can be any rice product, such as rice noodles. The rice product sample data may contain content data corresponding to multiple indicators, as shown in Table 1, which includes 11 indicators such as viscoelasticity, fineness, softness, chewiness, appearance and color, total taste score, cooking loss, and rehydration time. This embodiment does not limit this and adjustments can be made according to different rice products. The feature data refers to all data obtained using the feature extraction method in the current layer; different feature extraction methods can yield different feature data.

[0080] Step S30: Determine the initial weight data based on the importance of the feature data.

[0081] It should be understood that the initial weight data is the weight assigned to each feature data. Different feature data correspond to different initial weight data. The greater the importance of the feature data, the greater its corresponding initial weight data.

[0082] Step S40: Perform feature fusion based on the initial weight data and feature data to obtain feature fusion data, and update the current fusion count.

[0083] It should be noted that the feature fusion data is the data obtained by weighted fusion of the feature data of the current layer.

[0084] In the specific implementation, weights are assigned to the obtained feature data for weighted fusion to obtain the feature fusion data of the current layer, which is used as the input of the next layer. At this time, the number of fusions is increased by 1.

[0085] Step S50: Obtain deep feature fusion data based on the feature fusion data and the current fusion count.

[0086] It is understood that the deep feature fusion data is the data obtained after multi-layer feature fusion.

[0087] In practice, multi-layer feature fusion is adopted, and features are extracted and fused at each layer from top to bottom of the sample data to generate deep fusion features, so as to maximize the performance of the sample data and improve the final prediction accuracy.

[0088] Step S60: Input the deep feature fusion data into the preset prediction model, and use an intelligent optimization algorithm to optimize the structure of the preset prediction model and the initial weight data of the deep feature fusion data to establish the optimal prediction model.

[0089] It should be noted that the preset prediction model is a prediction model that needs to be selected, and there can be multiple models. In this embodiment, three machine learning models commonly used for numerical prediction are selected: Back Propagation Neural Network (BPNN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). Other prediction models can also be selected. This embodiment does not limit this and can be flexibly selected according to the actual situation. Each preset prediction model has a corresponding model structure.

[0090] Understandably, the core idea of ​​the BPNN model is to use backpropagation and gradient descent to iteratively update the network weights, minimizing the total error and ensuring that the predicted value for each sample is as close as possible to the true value. Figure 3 The diagram shows a BPNN structure. In the diagram, I represents neurons in the input layer, J represents neurons in the hidden layer, K represents neurons in the output layer, and W represents neurons in the output layer. ij W represents the connection weights between the input layer and the hidden layer. jk The weights represent the connection weights between the hidden and output layers. The input to each neuron in the next layer is calculated by the neurons in the previous layer based on these weights. X i X j X k These represent the inputs to each layer of neurons, O i O j O k These represent the outputs of each neuron in each layer. The structure of the BPNN model can be optimized in several ways: the number of hidden layers, the number of nodes in the hidden layers, the choice of activation function, and the choice of optimizer. The RF model is a supervised machine learning model based on ensemble learning. It is an overall model composed of many decision trees, making predictions by averaging the predictions of each decision tree. Its basic idea is to combine multiple decision trees when determining the final output. Although each decision tree has a high variance, when all decision trees are combined in parallel, because each decision tree has been perfectly trained on specific sample data, the variance of the result is very low. This process can improve the overall predictive ability of the model, such as... Figure 4 The diagram shows the RF structure. The structure of the RF model can be optimized in several aspects, including the number of decision trees, the maximum depth of the tree, the minimum number of samples required for node splits, and the minimum number of samples required for leaf nodes. XGBoost implements an efficient and highly optimized gradient boosting decision tree. In gradient boosting decision trees, the model is trained continuously, with each model iteratively reducing the error of the previous model. Instead of assigning different weights to the classifier after each iteration, this method fits a new model based on the new residuals of previous predictions and then minimizes the loss when adding the latest predictions. Figure 5The diagram shows the XGBoost structure. The expression for calculating the predicted value of the tree ensemble model is as follows:

[0091]

[0092] In the formula, x represents the regression tree space. i For sample input, y i For the predicted output, q represents the tree structure, with T leaf nodes, and each f... k The corresponding independent tree structure q and weight w, w i Let F represent the score of the i-th leaf node, and F be the prediction model used. The predicted value obtained by the model. The structure of the XGBoost model can be optimized in the following aspects: the number of decision trees, the learning rate, the maximum depth of the trees, the minimum sample weight of the leaf nodes, and other parameters.

[0093] Further, step S60 includes: inputting the deep feature fusion data into a preset prediction model to obtain initial prediction data; using an intelligent optimization algorithm to perform internal and external nesting optimization on the structure of the preset prediction model and the initial weight data of the deep feature fusion data based on the initial prediction data and the raw material sample data to obtain an initial combination of optimized weight data and optimized prediction model structure; and establishing an optimal prediction model based on the initial combination of optimized weight data and optimized prediction model structure.

[0094] Understandably, the initial prediction data is the prediction data obtained using a traditional prediction model. The initial combination of optimized weight data and optimized prediction model structure is the combination of weights and prediction model structure that maximizes the accuracy of the overall prediction result after optimization. After optimization, each preset prediction model can find the model structure that maximizes the prediction accuracy. Each optimized model structure has a corresponding optimal weight, thus obtaining the combination of weights and prediction model structure with the highest accuracy. The optimal prediction model is the final prediction model with the highest prediction accuracy.

[0095] It should be understood that in deep feature fusion, the features extracted from each layer are assigned weights during fusion and then input into the prediction model for prediction. Therefore, the optimization of weights will inevitably affect the optimization of the prediction model. Thus, the optimization of both cannot be performed independently. This embodiment employs a nested optimization approach to simultaneously optimize the fusion weights and model structure during the prediction process. This assigns more appropriate weights to feature data that plays a greater role in improving prediction accuracy, while simultaneously optimizing the structure of the prediction model, further improving prediction accuracy. The intelligent optimization algorithm can be the PSO (Particle Swarm Optimization) algorithm or other optimization algorithms; this embodiment does not impose any restrictions and can be flexibly adjusted according to the actual situation.

[0096] Further, the step of establishing the optimal prediction model based on the initial combination of the optimized weight data and the optimized prediction model structure includes: obtaining optimized prediction data based on the initial combination of the optimized weight data and the optimized prediction model structure and deep feature fusion data; obtaining evaluation data of the initial combination of each optimized weight data and the optimized prediction model structure based on the optimized prediction data and raw material sample data; determining the target combination of the optimized weight data and the optimized prediction model structure based on the evaluation data; and establishing the optimal prediction model based on the target combination of the optimized weight data and the optimized prediction model structure.

[0097] It should be noted that the optimized prediction data is the prediction data obtained by combining different weight data with the prediction model structure. The raw material sample data is the raw material index content data corresponding to the rice product sample data. The raw material sample data may contain the content data corresponding to multiple indicators. For example, the rice flour raw material index data shown in Table 2 includes 11 index content data such as moisture content, starch content, amylose content, gel consistency, protein content, fat content, and water solubility. This embodiment does not limit this and can make corresponding adjustments according to the actual situation.

[0098] Table 2. Rice Noodle Raw Material Data

[0099] Indicator Name Minimum value Maximum value average value Standard deviation Moisture (WC) 11.302 14.346 12.602 0.554 Starch content (SC) 69.451 74.789 72.759 1.341 Amylose content (AC) 11.309 25.789 16.555 3.589 Gel consistency (GC) 17.163 77.887 57.476 14.665 Protein content (PC) 6.849 8.538 7.724 0.363 Fat content (FC) 0.541 1.538 0.981 0.210 Fatty acid esters (FAVs) 16.271 196.878 74.559 46.078 Water-soluble (WS) 3.156 5.283 4.247 0.529 Swelling power (SF) 6.916 8.523 7.793 0.424 Final viscosity (FV) 2571.347 5020.148 3194.764 554.827 Gelatinization temperature (GT) 78.339 91.796 84.802 2.709

[0100] It is understood that the evaluation data represents the error between the optimized predicted data and the actual raw material index content data, as shown in Table 3 for XGB-D. new Evaluation data, XGB-D new The optimized fusion weights can be combined with the XGBoost model through the coefficients (R... 2 The evaluation is based on data such as mean absolute error (MAE), root mean square error (RMSE), relative root mean square error (RRMSE), and accuracy. Other evaluation methods can also be used, and this embodiment does not impose any restrictions on them. The target combination of optimized weight data and optimized prediction model structure is the combination with the highest accuracy among all initial combinations of optimized weight data and optimized prediction model structure, and this is determined through evaluation data.

[0101] In the specific implementation, the predicted data obtained by combining all weight data with the prediction model structure is evaluated, the error data between the predicted value and the actual value is calculated, and comparative analysis is performed to find the combination of weight data and prediction model structure with the most stable prediction and the highest prediction accuracy, and to establish the optimal prediction model for raw material quality prediction.

[0102] Table 3XGB-D new Assessment data

[0103] Indicator Name R2 MAE RMSE RRMSE Accuracy Moisture 0.980 0.0328 0.0622 0.495％ 99.505％ Starch content 0.985 0.0864 0.1622 0.223％ 99.777％ amylose content 0.962 0.1399 0.2711 1.242％ 98.345％ gel consistency 0.954 0.6987 1.4557 1.462％ 97.506％ Protein content 0.978 0.0331 0.0600 0.777％ 99.223％ Fat content 0.951 0.0149 0.0302 3.106％ 96.894％ Fatty acid value 0.912 2.8689 4.9860 6.593％ 93.407％ Water-soluble 0.960 0.0421 0.0768 1.804％ 98.196％ swelling power 0.979 0.0278 0.0502 0.664％ 99.356％ Final viscosity 0.958 27.6691 67.0918 2.134％ 97.866％ gelatinization temperature 0.982 0.2082 0.3788 0.447％ 99.553％

[0104] Step S70: Predict the index data of the target rice product according to the optimal prediction model to obtain the raw material prediction results.

[0105] It should be understood that the target rice product is a rice product for which raw material quality prediction is required, wherein the type of rice product is consistent with the sample data of rice products, and the raw material prediction result is the raw material index content data predicted using the optimal prediction model.

[0106] Step S80: Obtain raw material quality information based on the raw material prediction results.

[0107] It should be noted that the raw material quality information is the raw material quality situation obtained based on the prediction results. The raw material prediction results contain the predicted content data of various raw material indicators. Based on these raw material indicator content data, the raw material quality can be evaluated to obtain raw material quality information.

[0108] like Figure 6 The diagram illustrates the application process for predicting rice noodle raw material indicators. It demonstrates how to inversely predict raw material indicators based on the content values ​​of rice noodle product indicators. The process involves optimization through deep feature fusion and a predictive model to obtain the predicted results. Error analysis is then performed between the predicted results and the actual values. Based on the analysis, an optimal predictive model is generated and applied to quality control in the production of high-quality rice noodles. Simultaneously, the predicted raw material content is used to procure relevant raw materials for producing high-quality rice noodles, providing a wider range of more affordable options for raw material procurement.

[0109] In this embodiment, the current feature extraction strategy is determined based on the current number of fusions and the preset feature extraction strategy. Feature data is obtained based on the current feature extraction strategy and rice product sample data. Initial weight data is determined based on the importance of the feature data. Feature fusion is then performed based on the initial weight data and the feature data to obtain feature fusion data, and the current number of fusions is updated. Then, deep feature fusion data is obtained based on the feature fusion data and the current number of fusions. The deep feature fusion data is input into the preset prediction model. The structure of the preset prediction model and the initial weight data of the deep feature fusion data are optimized using an intelligent optimization algorithm to establish the optimal prediction model. The optimal prediction model is used to predict the index data of the target rice product to obtain the raw material prediction result, and thus obtain the raw material quality information. This embodiment proposes a novel feature fusion method, which consists of deep feature fusion composed of coarse-level feature fusion and more refined feature fusion. It extracts and fuses features at each layer from top to bottom, maximizing the performance of the sample data and effectively improving the accuracy of the prediction results. By optimizing the fusion weight allocation and prediction model structure of the multi-dimensional feature fusion, the feature components that contribute more to improving prediction accuracy are assigned more appropriate weights, thus optimizing the prediction model structure. This enables accurate prediction of the raw material index content of rice products, ensuring they meet quality assessment standards, thereby predicting the quality of rice product raw materials and providing a reference for selecting raw materials for producing high-quality rice products.

[0110] refer to Figure 7 , Figure 7 This is a flowchart illustrating a second embodiment of a method for predicting the quality of rice product raw materials based on deep feature fusion according to the present invention.

[0111] Based on the first embodiment described above, step S50 includes:

[0112] Step S501: When the current fusion count is less than the preset fusion count, a new current feature extraction method is determined based on the current fusion count.

[0113] It's important to note that deep feature fusion is an improvement upon multi-dimensional feature fusion. Multi-dimensional feature fusion involves extracting features from data using multiple basic methods, then fusion the feature groups extracted by each method in a weighted manner to generate a new set of fused features. The specific weighted fusion method between different feature groups is as follows: multiple feature groups are weighted such that features within the same group have the same weight, and the sum of the weights of features from different groups is equal to one. The weighted feature groups are then concatenated to generate a new set of features. Single-layer multi-dimensional feature fusion is an abstract module within deep fusion technology and represents the fundamental computational part of deep feature fusion, such as... Figure 8 The diagram illustrates a single-layer, multi-dimensional feature fusion process. The original data is defined as D0 = {X1, X2, X3, ..., X...}.q In multi-dimensional feature fusion, multiple feature extraction methods are set up: f = {f1, f2, ..., f...} p In f, each method extracts features from D0 to obtain d. i d i Weights w are assigned during feature fusion. i If a single-layer multi-dimensional feature fusion is used to predict the original data, the mathematical model for feature fusion is as follows:

[0114]

[0115]

[0116] In the formula, y is the true value. Let F be the predicted value, and w be the prediction model used. i d i Representing feature d i The weighting process aims to fuse the original data D0 with multi-dimensional features to generate new fused feature data D1, which is then input into the prediction model F for training and prediction, thereby improving the model's loss value. The goal is to minimize these differences. Multi-dimensional feature fusion combines features extracted from multiple base methods in a weighted manner, leveraging the informational value of the original data from different dimensions and enhancing its feature representation. This can be seen as an optimization of the base methods. The weights assigned to the fused features obtained through weighted fusion represent the contributions of each base method to the fused features.

[0117] Understandably, deep feature fusion builds upon multi-dimensional feature fusion by adding multiple layers of feature fusion. Data is processed through the first layer and then the next, until the final layer, generating the final deep fused feature. It's a further modification of multi-feature fusion, extracting and fusing features from each layer from top to bottom, and saving the features extracted by each layer's basic method. This allows for weighted feature fusion at each layer when generating deep fused features, ultimately feeding them into the prediction model. This maximizes the performance of the original data, improving the final prediction accuracy. Figure 9 The diagram illustrates the deep feature fusion process. Deep feature fusion is achieved through multiple layers of feature fusion. Each layer can use different basic methods and numbers of features extracted. After data has undergone feature extraction by the previous layer's basic method, it is temporarily fused by direct concatenation. The resulting new feature data serves as the input for the next layer. The features extracted by various basic methods in the current layer are recorded to facilitate weighted fusion in subsequent layers. The lower-level feature fusion modules sequentially perform the above feature extraction and fusion until the final layer. The process of using deep feature fusion for prediction is shown below:

[0118]

[0119]

[0120] In the formula, y is the true value. Let F be the predicted value, F be the prediction model used, and P1 be the objective function to generate the final fused feature data D after deep feature fusion of the original data D0. k The input is given to the prediction model F for training and prediction, resulting in the model's loss value. To reach the minimum.

[0121] In the specific implementation, the data D obtained by fusing the (k+1)th feature is... k For example, in the deep feature fusion, the k-th layer feature fusion is defined with multiple feature extraction methods f. k ={f k1 ,f k2 ,...,f kp}, f ki Each method in D k Feature extraction was performed to obtain d respectively ki Each d ki Generate D by direct splicing k ={d k1 ,d k2 ,...,d kp The basic method extracts {d} for each layer k1 ,d k2 ,...,d km} will be recorded. In deep feature fusion, to reflect the expressive power of various basic methods at each layer, they will be weighted and assigned to {d}. k1 ,d k2 ,...,d km The weighted values ​​are then processed to generate the fused feature D. k In this system, the weighted feature fusion at each layer is independent, with the weighting order proceeding sequentially from the first layer to the last. Let D be the data obtained from the feature fusion at the k-th layer. k-1 Multiple feature extraction methods are set up. k f k Each method in D k Feature extraction yielded {d k1 ,d k2 ,...,d km}, using a single-layer feature fusion weighting method for {d k1 ,d k2 ,...,d km Perform weighted fusion to generate the weighted fusion feature D of this layer. k .

[0122] It should be noted that the preset number of fusions is the number of feature fusion layers required. If the current number of fusions is less than the threshold number of fusions, it means that the feature fusion process has not yet ended and the next layer of feature fusion needs to be performed.

[0123] Step S502: Based on the feature fusion data and the new current feature extraction method, new feature data is obtained.

[0124] Step S503: Based on the new feature data, return to step S30.

[0125] In the specific implementation, if the current fusion count is less than the fusion count threshold, the feature fusion data obtained from the current layer is used as the input data for the next layer. Based on the fusion count, it is determined which layer of feature fusion needs to be performed, and the corresponding current feature extraction method is found to obtain new feature data. Feature fusion is performed again. After the feature fusion of this layer is completed, the current fusion count is increased as the basis for determining the feature extraction method for the next layer, until the required deep feature fusion data is obtained after feature fusion of each layer.

[0126] Step S504: When the current fusion count is greater than or equal to the preset fusion count, determine the deep feature fusion data based on the feature fusion data.

[0127] Understandably, when the current number of fusions is greater than or equal to the preset number of fusions, it means that the feature fusion of each layer has been completed, and the final deep feature fusion data has been generated.

[0128] like Figure 10 The diagram illustrates the prediction process for rice noodle raw material index content. It employs a two-layer deep feature fusion (k=2) and combines three prediction models to predict the content of rice noodle raw material indexes. The architecture of the two-layer deep feature fusion is as follows: The first layer uses correlation analysis (Pearson) and factor analysis (FA) to extract features from the rice noodle product index data, and the weighted features are directly concatenated and input into the second layer. The second layer uses RF and XGBoost algorithms to score the features input from the upper layer and extract relevant features accordingly.

[0129] In this embodiment, when the current fusion count is less than the preset fusion count, a new current feature extraction method is determined based on the current fusion count. New feature data is obtained based on the feature fusion data and the new current feature extraction method. Feature fusion continues based on the new initial feature data. When the current fusion count is greater than or equal to the preset fusion count, deep feature fusion data is obtained. This embodiment employs multi-layer feature fusion, extracting and fusing features at each layer from top to bottom, ultimately generating deep fusion features. This makes the sample data features more diverse and expressive, enhances the multi-dimensional expressive information of the sample data, maximizes the performance of the sample data, and improves the final prediction accuracy.

[0130] Furthermore, this embodiment of the invention also proposes a storage medium storing a prediction program for the quality of rice product raw materials based on deep feature fusion. When the prediction program for the quality of rice product raw materials based on deep feature fusion is executed by a processor, it implements the steps of the prediction method for the quality of rice product raw materials based on deep feature fusion as described above.

[0131] Reference Figure 11 , Figure 11 This is a structural block diagram of the first embodiment of the rice product raw material quality prediction device based on deep feature fusion of the present invention.

[0132] like Figure 11 As shown in the embodiment of the present invention, the predictive device for the quality of rice product raw materials based on deep feature fusion includes:

[0133] The fusion module 10 is used to determine the current feature extraction strategy based on the current number of fusions and the preset feature extraction strategy.

[0134] The fusion module 10 is also used to obtain feature data based on the current feature extraction strategy and rice product sample data.

[0135] The fusion module 10 is also used to determine initial weight data based on the importance of the feature data.

[0136] The fusion module 10 is further configured to perform feature fusion based on the initial weight data and feature data to obtain feature fusion data, and update the current fusion count.

[0137] The fusion module 10 is further configured to obtain deep feature fusion data based on the feature fusion data and the current fusion count.

[0138] The optimization module 20 is used to input the deep feature fusion data into a preset prediction model, and use an intelligent optimization algorithm to optimize the structure of the preset prediction model and the initial weight data of the deep feature fusion data to establish an optimal prediction model.

[0139] The prediction module 30 is used to predict the index data of the target rice product based on the optimal prediction model to obtain the raw material prediction results.

[0140] The prediction module 30 is also used to obtain raw material quality information based on the raw material prediction results.

[0141] In this embodiment, the current feature extraction strategy is determined based on the current number of fusions and the preset feature extraction strategy. Feature data is obtained based on the current feature extraction strategy and rice product sample data. Initial weight data is determined based on the importance of the feature data. Feature fusion is then performed based on the initial weight data and the feature data to obtain feature fusion data, and the current number of fusions is updated. Then, deep feature fusion data is obtained based on the feature fusion data and the current number of fusions. The deep feature fusion data is input into the preset prediction model. The structure of the preset prediction model and the initial weight data of the deep feature fusion data are optimized using an intelligent optimization algorithm to establish the optimal prediction model. The optimal prediction model is used to predict the index data of the target rice product to obtain the raw material prediction result, and thus obtain the raw material quality information. This embodiment proposes a novel feature fusion method, which consists of deep feature fusion composed of coarse-level feature fusion and more refined feature fusion. It extracts and fuses features at each layer from top to bottom, maximizing the performance of the sample data and effectively improving the accuracy of the prediction results. By optimizing the fusion weight allocation and prediction model structure of the multi-dimensional feature fusion, the feature components that contribute more to improving prediction accuracy are assigned more appropriate weights, thus optimizing the prediction model structure. This enables accurate prediction of the raw material index content of rice products, ensuring they meet quality assessment standards, thereby predicting the quality of rice product raw materials and providing a reference for selecting raw materials for producing high-quality rice products.

[0142] In one embodiment, the fusion module 10 is further configured to determine deep feature fusion data based on the feature fusion data when the current fusion count is greater than or equal to a preset fusion count;

[0143] In one embodiment, the fusion module 10 is further configured to determine a new current feature extraction method based on the current fusion count when the current fusion count is less than a preset fusion count;

[0144] Based on the aforementioned feature fusion data and the new current feature extraction method, new feature data is obtained;

[0145] Based on the new feature data, return to the step of determining the initial weight data according to the importance of the feature data.

[0146] In one embodiment, the optimization module 20 is further configured to input the deep feature fusion data into a preset prediction model to obtain initial prediction data;

[0147] Based on the initial prediction data and raw material sample data, an intelligent optimization algorithm is used to perform internal and external nesting optimization on the initial weight data of the structure and depth feature fusion data of the preset prediction model, so as to obtain an initial combination of optimized weight data and optimized prediction model structure.

[0148] Based on the initial combination of the optimized weight data and the optimized prediction model structure, an optimal prediction model is established.

[0149] In one embodiment, the optimization module 20 is further configured to obtain optimized prediction data based on the initial combination of the optimized weight data and the optimized prediction model structure, as well as the deep feature fusion data;

[0150] Based on the optimized prediction data and raw material sample data, evaluation data for the initial combination of each optimized weight data and the optimized prediction model structure are obtained;

[0151] Based on the evaluation data, determine the target combination of optimizing weight data and optimizing the prediction model structure;

[0152] Based on the target combination of the optimized weight data and the optimized prediction model structure, an optimal prediction model is established.

[0153] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0154] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0155] In addition, for technical details not described in detail in this embodiment, please refer to the method for predicting the quality of rice product raw materials based on deep feature fusion provided in any embodiment of the present invention, which will not be repeated here.

[0156] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0157] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0158] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0159] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for predicting the quality of rice product raw materials based on deep feature fusion, characterized by, The rice product raw material quality prediction method based on deep feature fusion comprises the following steps: According to the current fusion number and the preset feature extraction strategy, determine the current feature extraction strategy; According to the current feature extraction strategy and the rice product sample data, obtain feature data; According to the importance of the feature data, determine the initial weight data; According to the initial weight data and the feature data, perform feature fusion to obtain feature fusion data, and update the current fusion number; According to the feature fusion data and the current fusion number, obtain deep feature fusion data; Input the deep feature fusion data into a preset prediction model, use an intelligent optimization algorithm to optimize the structure of the preset prediction model and the initial weight data of the deep feature fusion data, and establish an optimal prediction model; According to the optimal prediction model, predict the index data of the target rice product to obtain a raw material prediction result; According to the raw material prediction result, obtain the raw material quality information; According to the feature fusion data and the fusion number, obtain deep feature fusion data, comprising: When the current fusion number is greater than or equal to the preset fusion number, determine the deep feature fusion data according to the feature fusion data; When the current fusion number is less than the preset fusion number, determine a new current feature extraction method according to the current fusion number, obtain new feature data according to the feature fusion data and the new current feature extraction method, and return to execute the step of determining the initial weight data according to the importance of the feature data.

2. The method of claim 1, wherein, The deep feature fusion data is input into a preset prediction model, and an intelligent optimization algorithm is used to optimize the structure of the preset prediction model and the initial weight data of the deep feature fusion data, and an optimal prediction model is established, comprising: Input the deep feature fusion data into a preset prediction model to obtain initial prediction data; According to the initial prediction data and the raw material sample data, use an intelligent optimization algorithm to optimize the structure of the preset prediction model and the initial weight data of the deep feature fusion data, and obtain the initial combination of the optimized weight data and the optimized prediction model structure; According to the initial combination of the optimized weight data and the optimized prediction model structure, establish an optimal prediction model.

3. The method of claim 2, wherein, The initial combination of the optimized weight data and the optimized prediction model structure is established, comprising: According to the initial combination of the optimized weight data and the optimized prediction model structure and the deep feature fusion data, obtain optimized prediction data; According to the optimized prediction data and the raw material sample data, obtain the evaluation data of each optimized weight data and the initial combination of the optimized prediction model structure; According to the evaluation data, determine the target combination of the optimized weight data and the optimized prediction model structure; According to the target combination of the optimized weight data and the optimized prediction model structure, establish an optimal prediction model.

4. A device for predicting the quality of a rice product raw material based on deep feature fusion, characterized by The rice product raw material quality prediction device based on deep feature fusion comprises: A fusion module for determining the current feature extraction strategy according to the current fusion number and the preset feature extraction strategy; The fusion module is further configured to obtain feature data according to the current feature extraction strategy and the rice product sample data; The fusion module is further configured to determine initial weight data according to the importance of the feature data; The fusion module is further configured to perform feature fusion according to the initial weight data and the feature data, to obtain feature fusion data, and to update a current fusion number; The fusion module is further configured to obtain deep feature fusion data according to the feature fusion data and the current fusion number; The optimization module is configured to input the deep feature fusion data into a preset prediction model, to optimize the structure of the preset prediction model and the initial weight data of the deep feature fusion data using an intelligent optimization algorithm, and to establish an optimal prediction model; The prediction module is configured to predict index data of a target rice product according to the optimal prediction model, to obtain a raw material prediction result; The prediction module is further configured to obtain raw material quality information according to the raw material prediction result; The fusion module is further configured to determine deep feature fusion data according to the feature fusion data when the current fusion number is greater than or equal to a preset fusion number; When the current fusion number is less than the preset fusion number, a new current feature extraction method is determined according to the current fusion number, new feature data is obtained according to the feature fusion data and the new current feature extraction method, and the step of determining initial weight data according to the importance of the feature data is returned to be executed according to the new feature data.

5. The rice product raw material quality prediction device based on deep feature fusion according to Claim 4, characterized by, The optimization module is further configured to input the deep feature fusion data into a preset prediction model to obtain initial prediction data; According to the initial prediction data and the raw material sample data, the structure of the preset prediction model and the initial weight data of the deep feature fusion data are optimized using an intelligent optimization algorithm, to obtain an initial combination of optimization weight data and an optimized prediction model structure; According to the initial combination of the optimization weight data and the optimized prediction model structure, an optimal prediction model is established.

6. The rice product raw material quality prediction device based on deep feature fusion according to Claim 5, characterized by, The optimization module is further configured to obtain optimization prediction data according to the initial combination of the optimization weight data and the optimized prediction model structure and the deep feature fusion data; According to the optimization prediction data and the raw material sample data, evaluation data of each optimization weight data and the initial combination of the optimized prediction model structure are obtained; According to the evaluation data, a target combination of the optimization weight data and the optimized prediction model structure is determined; According to the target combination of the optimization weight data and the optimized prediction model structure, an optimal prediction model is established.

Images