Medlar quality stability evaluation method and system based on neural network

Abstract

The invention discloses a wolfberry quality stability evaluation method and system based on a neural network, and the method comprises the steps: collecting wolfberry samples from different producing areas, obtaining the growth environment parameters and harvesting and processing information of the wolfberry samples, and building a correlation database of the producing area environment and the wolfberry quality, analyzing the influence rule of the producing area factors on the quality stability of the Chinese wolfberry through a data mining algorithm to obtain a quality stability static prediction model of the Chinese wolfberry in different producing areas; a long-short-term memory neural network is adopted, a multi-level feature extraction structure is designed, network parameters are optimized through an adaptive learning algorithm, generalization and robustness constraints are introduced at the same time, a quality stability dynamic prediction model is obtained, and the prediction capacity and the anti-interference capacity of the model for new samples are enhanced. According to the method, the problems of unstable quality caused by large climate and soil difference in the wolfberry planting area and high subjectivity of wolfberry quality evaluation are solved, and accurate prediction of the wolfberry quality stability is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of quality assessment, and in particular relates to a method and system for assessing wolfberry quality stability based on a neural network. Background Art

[0002] As a common Chinese medicinal material and health food, the quality stability of wolfberry directly affects the efficacy of the product and the safety of consumers' medication. However, due to the wide range of wolfberry planting areas, the climate, soil, harvesting and processing methods of each production area vary greatly, resulting in uneven quality of wolfberry and difficult to ensure stability. At the same time, traditional wolfberry quality evaluation mainly relies on manual experience, and the evaluation criteria are highly subjective, making it difficult to quantify and standardize. The neural network-based wolfberry quality stability evaluation can establish a complex nonlinear relationship between the quality indicators and stability of wolfberry through the learning of a large amount of real sample data, and automatically extract the key features implicit in the data, thereby obtaining a stability prediction model that comprehensively reflects the quality of wolfberry. However, since wolfberry quality evaluation involves sensory indicators and intrinsic component indicators, how to select a suitable combination of indicators, organically integrate these heterogeneous data, design a reasonable neural network structure, and take into account the generalization and robustness of the model is a key technical problem that needs to be solved urgently. In addition, the quality and quantity of wolfberry sample data also directly affect the prediction performance of the model. Therefore, how to collect and screen representative sample data in each link of wolfberry planting, harvesting, processing and storage is also a problem worthy of in-depth study. In order to overcome the above problems, it is urgent to establish a scientific, objective and accurate wolfberry quality stability evaluation method. Summary of the invention

[0003] In order to solve the above technical problems, the present invention proposes a method and system for evaluating wolfberry quality stability based on a neural network to solve the problems existing in the above prior art.

[0004] To achieve the above object, the present invention provides a method for evaluating wolfberry quality stability based on a neural network, comprising:

[0005] The growth environment parameters, harvesting and processing information and quality index test values ​​of wolfberry samples in the selected area are obtained and stored in association to obtain an origin-quality association database; based on the origin-quality association database, the association rules between the origin environmental factors and the quality of wolfberry are obtained; based on the association rules and the data in the origin-quality association database, the first key factor of the stability of wolfberry quality is determined through stepwise regression analysis;

[0006] Training a static prediction model for wolfberry quality stability based on the first key factor;

[0007] Select the sensory indexes and active ingredient content indexes of wolfberry, build a wolfberry quality quantitative evaluation model and train it through the index values;

[0008] Obtaining a second key factor according to an origin-quality association database, and training a dynamic prediction model for wolfberry quality stability according to the first key factor and the second key factor;

[0009] The quality of wolfberry is evaluated according to the trained wolfberry quality quantitative evaluation model, and a comprehensive evaluation of the quality stability of wolfberry is obtained according to the trained wolfberry quality stability static prediction model and the wolfberry quality stability dynamic prediction model.

[0010] Optionally, the process of obtaining association rules between origin environmental factors and wolfberry quality based on the origin-quality association database includes:

[0011] The origin environmental factors and wolfberry quality index data in the origin-quality association data were extracted, the text descriptions in the wolfberry quality index data were segmented and mapped into numerical representations; the processed data were mined and analyzed using the Apriori association rule algorithm to obtain the association rules between the origin environmental factors and wolfberry quality.

[0012] Optionally, the process of determining the first key factor of wolfberry quality stability through stepwise regression analysis based on association rules and data in the origin-quality association database includes:

[0013] Through correlation analysis, environmental parameters whose correlation with wolfberry quality is higher than a preset threshold are screened out; the screened environmental parameters are used as independent variables, and the wolfberry quality index is used as the dependent variable, a regression model is established, the regression coefficients of the environmental parameters of each production area are obtained, and the first key factor is determined based on the regression coefficient.

[0014] Optionally, the process of training the static prediction model of wolfberry quality stability based on the first key factor includes: extracting data corresponding to the first key factor, constructing a characteristic vector matrix of the key factor, and using a support vector machine to construct a static prediction model of wolfberry quality stability; training the static prediction model of wolfberry quality stability according to the characteristic vector matrix, and determining the optimal parameter combination of kernel function type, penalty coefficient, and kernel function parameters of the support vector machine model through cross-validation.

[0015] Optionally, the process of selecting wolfberry sensory indicators and active ingredient content indicators, constructing a wolfberry quality quantitative evaluation model, and training through indicator values ​​includes:

[0016] According to the relevant standards for wolfberry quality evaluation, wolfberry sensory indicators and effective ingredient content indicators are selected to construct a wolfberry quality evaluation index system; multi-source data corresponding to the wolfberry quality evaluation index system are obtained, and the multi-source data are fused at the feature layer through a decision tree algorithm to generate a wolfberry quality evaluation fusion feature vector, based on the wolfberry quality evaluation fusion feature vector, a convolutional neural network is used to extract wolfberry high-level features, and a recursive feature elimination method is used to screen a key feature subset that meets the conditions for quality contribution, the wolfberry quality quantitative evaluation model is trained through the key feature subset, and the optimal parameters of the model are determined through 5-fold cross validation.

[0017] Optionally, the process of training the dynamic prediction model for wolfberry quality stability includes:

[0018] The maximum information coefficient was used to screen in the origin-quality association database to obtain the second key factor. The original data corresponding to the first key factor and the second key factor were smoothed by the wavelet threshold denoising method. The dynamic prediction model of wolfberry quality stability was trained based on the processed data. The Adam optimization algorithm was used for training, and L1 and L2 regularization terms were added to the cross entropy loss function.

[0019] Optionally, adversarial samples are generated by randomly adding Gaussian noise to the original data after processing, and the adversarial data is input into the dynamic prediction model of wolfberry quality stability for training.

[0020] Optionally, the dynamic prediction model for wolfberry quality stability includes a long short-term memory neural network and a convolutional network structure, the outputs of the long short-term memory neural network and the convolutional network structure are cascaded, and the results are output through a softmax classifier; wherein the long short-term memory neural network has 3 layers, and each layer contains 128, 64, and 32 neurons respectively; the convolutional network structure includes 5 convolutional layers, 3 pooling layers, and 1 global average pooling layer, and the convolution kernel size is 3×3.

[0021] The present invention also provides a wolfberry quality stability evaluation system based on a neural network, comprising: a wolfberry quality static evaluation module, used to construct a wolfberry quality stability static prediction model and perform training;

[0022] Wolfberry quality scoring module, used to build and train a wolfberry quality quantitative evaluation model;

[0023] The dynamic evaluation module of wolfberry quality is used to build and train the dynamic prediction model of wolfberry quality stability;

[0024] The comprehensive evaluation module is used to evaluate the quality of wolfberry according to the trained wolfberry quality quantitative evaluation model, and to obtain a comprehensive evaluation of the quality stability of wolfberry according to the trained wolfberry quality stability static prediction model and the wolfberry quality stability dynamic prediction model.

[0025] Compared with the prior art, the present invention has the following advantages and technical effects:

[0026] The present invention aims at the problem of unstable quality caused by large differences in climate and soil in wolfberry planting areas, collects wolfberry samples from different origins, obtains growth environment parameters and harvesting and processing information, establishes a database related to the origin environment and wolfberry quality, and analyzes the influence of origin factors on quality stability through data mining. In view of the problem that wolfberry quality assessment is highly subjective, sensory indicators such as color and taste and intrinsic quality indicators such as effective ingredient content are selected, objective data are obtained through sensors such as electronic tongue and electronic nose and analytical methods such as high performance liquid chromatography, a multi-source heterogeneous data fusion model is established, key features are extracted using deep learning, and objective and quantitative quality assessment standards are constructed. The present invention adopts a long short-term memory neural network, designs a multi-level feature extraction structure, optimizes network parameters through an adaptive learning algorithm, improves the model's ability to depict dynamic changes in quality, introduces generalization and robustness constraints, enhances the model's prediction ability and anti-interference ability in the face of new samples, ensures stable operation in a complex and changeable real environment, and realizes accurate prediction of wolfberry quality stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0028] Figure 1 A schematic diagram of a process for constructing a quality assessment standard according to an embodiment of the present invention;

[0029] Figure 2 A schematic diagram of a process for calculating evaluation indicators according to an embodiment of the present invention;

[0030] Figure 3 This is a diagram of the training process of the dynamic prediction model for wolfberry quality stability according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0032] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0033] Embodiment 1

[0034] like Figure 1-3 As shown, this embodiment provides a method and system for evaluating wolfberry quality stability based on a neural network, including:

[0035] Step S101, in view of the large differences in climate, soil, etc. in wolfberry planting areas, wolfberry samples from different origins are collected to obtain their growth environment parameters and harvesting and processing information, and a correlation database between the origin environment and wolfberry quality is established. The influence of origin factors on the quality stability of wolfberry is analyzed through a data mining algorithm to obtain a quality stability prediction model for wolfberries from different origins, that is, a static prediction model for the quality stability of wolfberries.

[0036] The quality index data of the collected wolfberry samples are obtained. If some indicators exceed the national standard range, they are judged as unqualified samples and eliminated. The data mining algorithm is used to obtain the influence of production environment factors and wolfberry quality stability through correlation analysis and regression analysis methods. According to the influence law, a quality stability prediction model of wolfberry from different production areas is constructed, and support vector machine and random forest machine learning algorithms are used for training and optimization to predict the quality stability level of wolfberry produced in new production areas.

[0037] Specifically, we first selected 20 representative production areas in the main wolfberry planting areas such as Ningxia, Inner Mongolia, Qinghai, Xinjiang, and Gansu, and randomly collected 500 wolfberry samples in each production area. At the same time, we recorded the growth environment parameters of the samples, including climate data such as temperature, humidity, precipitation, and sunshine time, as well as soil pH, organic matter content, and total nitrogen content. The collected wolfberry samples were tested for quality, mainly testing the color, dryness, particle size, and active ingredient content of the wolfberry, and samples that did not meet the national standard GB / T 18672-2014 were eliminated. SQL Server was used to build a relational database to associate and store the production environment parameters, harvesting and processing information, and the quality test results of the corresponding wolfberry samples. The Apriori association rule algorithm was used to mine and analyze the database, and the minimum support was set to 10% and the minimum confidence was set to 80%. The association rules between the environmental factors of the origin and the quality of wolfberry were obtained. For example, when the temperature was between 15℃-25℃ and the soil pH value was between 5-5, the color and active ingredient content of wolfberry were better. Stepwise regression analysis determined that the key factors affecting the stability of wolfberry quality were temperature, light duration and soil pH value. The K-means clustering algorithm was used to divide the 20 origins into 3 categories. The average quality qualified rate of wolfberries produced in each category of origin was calculated, and it was found that the quality of wolfberries from category I origin was the most stable, with an average qualified rate of more than 95%. Finally, a wolfberry quality stability prediction model was constructed based on the SVM support vector machine algorithm. The origin environmental factors were used as input variables, and the wolfberry quality qualified rate was used as the output variable. The optimal model parameters were determined by 5-fold cross validation, and the prediction accuracy could reach 90%. When the new origin environmental parameters were input into the model, the quality stability level of wolfberries in the origin could be predicted, providing a reference for the selection of high-quality production areas.

[0038] According to the collected wolfberry sample data, the Apriori association rule algorithm was used to mine the association rules between the production environment factors and the quality of wolfberry, and the influencing rules were obtained.

[0039] Obtain the origin environmental factors in the wolfberry sample data, and extract the numerical and categorical characteristic variables of environmental temperature, humidity, light intensity, and soil fertility. For wolfberry quality indicators, obtain numerical indicators or use natural language processing technology to extract keywords from text descriptions and vectorize them as the target attributes for association rule mining. Combine the environmental factor characteristics and the quality target attributes into a transaction data set, with each transaction corresponding to a wolfberry sample. Through frequent item set mining, obtain the association rules between the environmental factor feature item set and the quality target item set, and calculate the support, confidence, and lift evaluation indicators of the association rules, such as Figure 2 shown.

[0040] Specifically, the wolfberry sample data is first preprocessed to extract numerical features such as environmental temperature, humidity, light intensity, and soil fertility. For example, the temperature range is between 15℃ and 30℃, the humidity is between 50% and 80%, and the light intensity is between 10000 and 30000lux. The soil fertility is graded using NPK, which is divided into three levels: low, medium, and high. For quality indicators, numerical indicators such as the sugar content of wolfberry are directly used as target attributes, while text descriptions such as "bright color" extract the keywords "color" and "bright" and map them to the value 1. Then the environmental characteristics and quality targets are combined into a transaction data set, and each transaction represents a wolfberry sample. The minimum support of the Apriori algorithm is set to 0.5 and the minimum confidence is set to 8, and the algorithm is used to mine frequent item sets for the transaction data set. For example, the frequent item set {temperature = 25℃, humidity = 60%, light = 20000lux} → {sugar content = 15%} is obtained, indicating that the probability of wolfberry having a high sugar content under this environmental condition is relatively high. Then, association rules are generated from the frequent item sets, and the support, confidence, and lift of each rule are calculated. For example, the support of the rule "temperature 25℃ ∧ humidity 60% → sugar content 15%" is 08, the confidence is 85, and the lift is 2, indicating that the rule has strong relevance and practical value. Finally, the rules are sorted and screened by confidence and lift, and the strong correlation between environmental factors and quality indicators is obtained, which is intuitively displayed through the association network diagram. In the figure, the nodes represent environmental factors or quality indicators, the edges represent the correlation between the two, and the thickness of the edges represents the strength of the association. It can be found from the figure that there is a significant positive correlation between the environmental combinations of temperature 25℃, humidity 60%, and light 20000lux and the high sugar content of wolfberry, which provides strong data support and decision-making basis for the optimization of wolfberry planting environment.

[0041] Through stepwise regression analysis, the first key factor affecting the quality stability of wolfberry among the production environment parameters was obtained, and variables such as temperature, light duration and soil pH value were determined.

[0042] The data of various environmental parameters and wolfberry quality indexes are obtained, and the environmental parameters with quality correlation higher than the preset threshold are screened out through correlation analysis. A stepwise regression analysis is adopted, and the screened environmental parameters are used as independent variables, and the wolfberry quality index is used as the dependent variable to establish a regression model to obtain the regression coefficient of each environmental parameter. The key environmental factors are determined according to the regression coefficients, and corresponding planting management measures are formulated for the key environmental factors, including adjusting the temperature of the planting site, extending the light time, and adjusting the soil pH value, so as to ensure the stability of the quality of wolfberry.

[0043] Specifically, in order to study the key environmental factors affecting the quality stability of wolfberry, meteorological stations were first set up in major wolfberry planting areas such as Ningxia, Qinghai, and Xinjiang. Environmental parameters such as temperature, light duration, and air humidity were automatically collected every hour. The monitoring was carried out continuously for one year to obtain the daily environmental parameter data of each production area throughout the year and establish an environmental parameter database. At the same time, 10 plantations were randomly selected in each production area, and fresh wolfberry fruit samples were collected in June, July, August, and September, 2 kg per plantation, and repeated three times per month. High-performance liquid chromatography was used to determine 10 quality indicators such as sugar, vitamin C, carotenoids, calcium, and iron in wolfberry fruits, and the average values ​​were calculated to establish a wolfberry quality database. The two databases were cleaned, and the 3σ principle was used to eliminate outliers, and then the Z-score method was used for data standardization. The Pearson correlation coefficient r between each environmental parameter and each quality indicator was calculated. When |r|≥65, the correlation between the two was considered significant, and 6 environmental parameters with high correlation with quality indicators were screened out. Taking these six environmental parameters as independent variables and a certain quality index as the dependent variable, a regression model was established by multivariate linear stepwise regression analysis. Each time an independent variable was introduced, the F test was performed using the variance analysis method. When p < 0.5, it was considered that the independent variable had a significant effect on the dependent variable. Finally, a regression model containing four independent variables was obtained, and the regression coefficients were from large to small: light time 531, day and night temperature difference -427, soil pH value -318, and average temperature in early July 275. Thus, four key environmental factors affecting the quality index were determined. In view of the key environmental factors, wolfberry planting management measures were formulated: select production areas with light time ≥ 14h for planting; build a sunshade to reduce the temperature difference between day and night; use lime to adjust the soil pH value to 5; spray 2% potassium dihydrogen phosphate on the leaves in early July to improve the plant's ability to resist high temperatures, thereby ensuring the stability of wolfberry quality.

[0044] The SVM (support vector machine) algorithm was used to construct a prediction model for wolfberry quality stability. The model parameters were determined according to the cross-validation results to judge the quality stability level of wolfberry from new production areas.

[0045] The key characteristic indicators reflecting the quality stability of wolfberry are obtained, and a characteristic vector matrix is ​​constructed; a grid search method is used to determine the optimal parameter combination of the kernel function type, penalty coefficient, and kernel function parameters of the support vector machine model through cross-validation; wolfberry samples from new origins are obtained, characteristic indicators identical to those of the model training data are extracted, and a characteristic vector of the sample to be predicted is constructed; the characteristic vector is input into the trained support vector machine model to obtain a quality stability prediction result of the sample.

[0046] Specifically, firstly, the historical data of wolfberry quality are feature extracted, and key indicators that can reflect the quality stability are selected, such as color, drying loss, sugar content, total acid, etc., to construct a feature matrix with n rows and m columns. Then, the grid search method is used to set the parameter search range of the support vector machine model. For example, the kernel function type can select linear kernel, Gaussian kernel, etc., the penalty coefficient C can take 10 values ​​between 1 and 100, and the kernel function gamma parameter can take 10 values ​​between 0.01 and 1. The optimal parameter combination is determined by 5-fold cross validation. For example, the Gaussian kernel is finally determined to be used, C = 10, gamma = 0.1. The support vector machine model is trained on all historical data using the optimal parameter combination obtained by grid search to obtain a wolfberry quality stability prediction model. For wolfberry samples from new origins, the same feature indicators as the model training data are extracted, such as color L* = 32, a* = 11, b* = 18, drying loss 1%, sugar content 42%, total acid 2%, and the feature vector of the sample to be predicted is constructed. The feature vector is input into the trained support vector machine model, and the predicted value of the sample quality stability is calculated by the model. For example, if the predicted value is 8, it means that the probability of the quality stability of this batch of wolfberries is 80%, which is a relatively high level of stability. If the predicted value is lower than 6, it is recommended to strengthen the quality control of raw materials, such as controlling the maturity and purity of picking, and optimizing the processing parameters, such as drying temperature and time, so as to improve the quality stability of wolfberries. In the future, the quality data of wolfberries from new production areas should be continuously collected, such as taking 100 samples in each production season, and updating the support vector machine prediction model every 2 months to continuously improve the prediction accuracy and scope of application of the model.

[0047] Step S102, in view of the high subjectivity of traditional wolfberry quality assessment, sensory indicators such as wolfberry color and taste and intrinsic quality indicators such as effective ingredient content are selected, objective data are obtained through sensors such as electronic tongue and electronic nose, and component content data are obtained by combining analytical methods such as high performance liquid chromatography, a multi-source heterogeneous data fusion model is established, and a deep learning algorithm is used to extract the key characteristics of wolfberry quality, and an objective and quantitative quality assessment standard is constructed.

[0048] According to the relevant standards for wolfberry quality evaluation, the sensory indicators and active ingredient content indicators of wolfberry are determined, and a wolfberry quality evaluation index system is constructed; a data fusion algorithm is used to fuse multi-source heterogeneous data at the feature layer to generate fusion features for wolfberry quality evaluation; based on the fusion features, a feature selection algorithm is used to determine a feature subset whose contribution to quality is greater than a preset threshold; based on the feature subset, a regression algorithm is used to establish a nonlinear mapping relationship between features and quality scores, and a wolfberry quality quantitative evaluation model is constructed to achieve objective evaluation of wolfberry quality.

[0049] Specifically, according to the national standard GB / T 18672-2014 "Wolfberry" and the industry standard NY / T2638-2014 "Wolfberry Sensory Requirements and Inspection Methods", it is determined that the color of wolfberry should be bright red or dark red, the taste is sweet, and the dryness is ≤13%. At the same time, reference literature research determined that the content of wolfberry polysaccharides ≥8% and the content of betaine ≥2% are key indicators for wolfberry quality evaluation. The Astree II electronic tongue system of Alpha MOS Company in the United States was used to detect the taste of wolfberry. The system is equipped with 7 liquid sensors and can comprehensively analyze the taste characteristics of wolfberry, such as sour, sweet, bitter, astringent, and salty. The PEN3 portable electronic nose of AIRSENSE Company in France was used to detect the smell of wolfberry. The device integrates 10 metal oxide sensors and can quickly identify the aromatic components of wolfberry. The color information of wolfberry was collected using a high-speed camera, and the RGB and HSV color features of wolfberry were extracted through image segmentation algorithms. The contents of Lycium barbarum polysaccharides and betaine were determined by Agilent 1260 high performance liquid chromatograph. The chromatographic column was Agilent Zorbax SB-C18 (250mm×6mm, 5μm), the mobile phase was acetonitrile-0.5mol / L potassium dihydrogen phosphate solution, and the detection wavelengths were 203nm and 254nm, respectively. The signal data of the electronic tongue and electronic nose were subjected to wavelet transform denoising and standardization, the color characteristics were subjected to minimum-maximum normalization, and the chromatographic data were baseline corrected and the peak area was integrated. Then, the decision tree algorithm was used to fuse the above heterogeneous data at the feature layer. The maximum depth of the decision tree, the minimum number of leaf node samples and other parameters were optimized by grid search to generate a fusion feature vector for the quality evaluation of Lycium barbarum that takes into account both sensory characteristics and chemical components. Based on the fusion features, the convolutional neural network was used to automatically extract the high-level features such as texture, color, and taste for the quality evaluation of Lycium barbarum, and the recursive feature elimination method was used to screen out the top 10 key feature subsets that contributed to the quality. On this basis, support vector regression was used to construct a quantitative evaluation model for wolfberry quality. The optimal parameters of the model were determined through 5-fold cross validation. The model was trained on 900 wolfberry samples and tested on 100 samples. The mean absolute error (MAE) was 58 and the determination coefficient (R) was 2.3. 2 It reached 93%, indicating that this method can accurately evaluate the quality of wolfberry and provide a new idea for quality control in the wolfberry industry.

[0050] Step S103, in view of the complex nonlinear characteristics of wolfberry quality stability prediction, a long short-term memory neural network is used to design a multi-level feature extraction structure, and the network parameters are optimized through an adaptive learning algorithm to improve the model's ability to describe the dynamic changes in wolfberry quality; at the same time, generalization and robustness constraints are introduced to enhance the model's prediction ability and anti-interference ability for new samples, ensuring stable operation in a complex and changeable real environment. Figure 3 shown.

[0051] The historical data of wolfberry quality are obtained, the second key factor affecting the stability of wolfberry quality is determined, and the second key factor and the first key factor are used as input features of the model; the original data are denoised and feature enhanced by wavelet analysis signal processing technology, highlighting the dynamic change characteristics of the wolfberry quality; a long short-term memory neural network model is designed for the processed data, and the memory of long-term dependencies is realized through gated units to capture the temporal evolution law of the wolfberry quality; during model training, an adaptive learning rate optimization algorithm is introduced to dynamically adjust the network parameters according to the change of the loss function, so as to improve the convergence speed and stability of the model.

[0052] Specifically, the maximum information coefficient and other algorithms were used together with the origin-quality association database to screen out 20 key factors affecting the quality of wolfberry. Then the wavelet threshold denoising method was used to smooth the original data, and the first-order and second-order differences of the key characteristics of wolfberry quality were extracted to highlight the dynamic trend of quality characteristics. On this basis, a three-layer LSTM neural network was built, each layer containing 128, 64, and 32 neurons respectively. The processed key characteristic data of wolfberry quality were input with a time step of 7 days to achieve adaptive learning of the time series evolution law of quality. At the same time, a CNN structure consisting of 5 convolutional layers and 3 pooling layers was designed. Local features were extracted with a 3×3 convolution kernel, and global features were fused through global average pooling. Finally, the output vectors of LSTM and CNN were cascaded, and the quality grade of wolfberry was predicted through a softmax classifier. During the model training process, the Adam optimization algorithm was used, and the initial learning rate was set to 0.01, which was reduced to 5 times the original rate every 10 epochs. At the same time, L1 and L2 regularization terms were added to the cross entropy loss function, and the regularization coefficients were set to 0.1 and 0.01, respectively, to effectively suppress overfitting. In addition, by randomly adding 5% to 10% Gaussian noise to the original data, randomly blocking 10% to 20% of the feature dimensions, constructing adversarial samples and adding them to the training set, the model's anti-interference ability was improved. Finally, the model was tested on 2,000 wolfberry quality inspection data. The classification accuracy of the model was stable at more than 95%, which can accurately warn of quality abnormality risks and provide strong support for quality control in the wolfberry industry.

[0053] Multidimensional data such as meteorology and soil were obtained based on expert knowledge, and the key factors affecting wolfberry quality were screened out through the maximum information coefficient.

[0054] The maximum information coefficient between each dimension data and wolfberry quality is calculated to screen out the key factors with the greatest correlation with wolfberry quality. A dynamic prediction model for wolfberry quality stability is established for the key factors, and the prediction model is a decision tree, support vector machine or neural network model. If the performance of the prediction model does not reach the preset threshold, the hyperparameters of the prediction model are adjusted until the performance of the prediction model reaches the preset threshold. The prediction model is applied to the newly collected meteorological data and soil data to obtain the predicted wolfberry quality grade.

[0055] The maximum information coefficient between each dimension data and wolfberry quality was calculated, and the key factors with the maximum information coefficient greater than 5 were selected as model input. Then the decision tree algorithm was selected to establish the wolfberry quality prediction model. The preprocessed multidimensional data was divided into training set and test set in a ratio of 8:2. The decision tree model was trained using the training set data, and the optimal parameters obtained through grid search were: the maximum tree depth was 5, and the minimum number of leaf node samples was 10. The model performance was evaluated on the test set, and the accuracy reached 85%, meeting the preset threshold, and it was saved as the final prediction model. The model was used to predict the newly collected meteorological and soil data. If the wolfberry quality was predicted to be level one, the planting management measures of watering once every 7 days and applying organic fertilizer were formulated; if the wolfberry quality was predicted to be level two, the management measures of watering once every 10 days and reducing the amount of organic fertilizer by 30% were formulated to ensure that the wolfberry quality reached the expected level. Through the above method, the quality of wolfberry can be accurately predicted according to meteorological and soil data, and targeted planting management measures can be formulated, which is of great significance to improving the quality of wolfberry.

[0056] Specifically,

[0057] Wavelet analysis was used to reduce the noise of the original data, and the dynamic change trend of wolfberry quality was obtained by extracting differential features.

[0058] The original time series data of wolfberry is obtained, and multi-scale decomposition is performed using wavelet transform to remove high-frequency noise and obtain the denoised time series data;

[0059] Performing a differential operation on the denoised time series data to extract the change features between adjacent time points to obtain a differential feature sequence;

[0060] According to the differential feature sequence, a dynamic prediction model for wolfberry quality stability is constructed, and the model is trained using a support vector machine algorithm;

[0061] The trained dynamic prediction model of wolfberry quality stability is used to predict the quality of newly collected wolfberry time series data and obtain the quality change trend.

[0062] If the quality change trend decreases, it is judged that the quality of wolfberry has decreased, and the early warning mechanism is triggered;

[0063] According to the quality change trend and early warning situation, a time series prediction algorithm is used to predict the quality of wolfberry in the future.

[0064] Specifically, the original time series data was decomposed into 5 layers using wavelet transform, and the Daubechies4 wavelet basis function was selected. The noise components in the high-frequency detail coefficients after decomposition were removed by threshold processing, and then the denoised time series data was obtained by wavelet reconstruction. Then, the denoised time series data was subjected to first-order difference operation to extract the change characteristics between adjacent data points, and a differential feature sequence consisting of 4319 differential eigenvalues ​​was obtained. Based on the differential feature sequence, the Gaussian kernel function was selected to construct a dynamic prediction model for the quality stability of wolfberry, and the model was trained using the 10-fold cross-validation method to obtain a dynamic prediction model for the quality stability of wolfberry with good generalization performance. The dynamic prediction model for the quality stability of wolfberry was used to predict the quality of wolfberry time series data collected continuously for 7 days. By analyzing the changing trend of the prediction results, it was found that the quality began to decline from the 5th day, and the decline exceeded 10%, triggering the early warning mechanism. Combining the quality decline trend and early warning situation, a long short-term memory neural network algorithm was used, and the prediction results of the dynamic prediction model of wolfberry quality stability were used as input to predict the quality of wolfberry in the next three days. The prediction results showed that if no measures were taken, the quality would continue to deteriorate, with an average decline rate of 5% / day, which provided an important basis for decision-making on wolfberry quality control.

[0065] Gaussian noise and other interference factors were introduced through adversarial training strategy to determine the robustness and stability of the model in predicting wolfberry quality.

[0066] The method comprises the following steps: obtaining wolfberry image data to be predicted, adding Gaussian noise interference factors to the wolfberry image data, and generating an adversarial sample data set; using the adversarial sample data set to train a prediction model, and obtaining an adversarially trained dynamic prediction model for wolfberry quality stability; obtaining a target wolfberry image, and using the prediction model to perform quality prediction on the target wolfberry image, and obtaining a preliminary prediction result; for the preliminary prediction result, introducing different degrees of Gaussian noise interference, repeating the prediction multiple times, and obtaining a group of prediction results; judging the discreteness of the group of prediction results, and outputting the prediction result if the discreteness is less than a preset threshold.

[0067] The quality of wolfberry is evaluated according to the trained wolfberry quality quantitative evaluation model, and the comprehensive evaluation of wolfberry quality stability is obtained according to the trained wolfberry quality stability static prediction model and wolfberry quality stability dynamic prediction model.

[0068] Through this series of methods, the accuracy and stability of wolfberry quality prediction can be effectively improved, providing reliable technical support for wolfberry quality grading.

[0069] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for evaluating wolfberry quality stability based on neural network, characterized in that: The following steps are involved: The growth environment parameters, harvesting and processing information and quality index test values ​​of wolfberry samples in the selected area are obtained and stored in association to obtain an origin-quality association database; based on the origin-quality association database, the association rules between the origin environmental factors and the quality of wolfberry are obtained; based on the association rules and the data in the origin-quality association database, the first key factor of the stability of wolfberry quality is determined through stepwise regression analysis; Training a static prediction model for wolfberry quality stability based on the first key factor; Select the sensory indexes and active ingredient content indexes of wolfberry, build a wolfberry quality quantitative evaluation model and train it through the index values; Obtaining a second key factor according to an origin-quality association database, and training a dynamic prediction model for wolfberry quality stability according to the first key factor and the second key factor; The quality of wolfberry is evaluated according to the trained wolfberry quality quantitative evaluation model, and a comprehensive evaluation of the quality stability of wolfberry is obtained according to the trained wolfberry quality stability static prediction model and the wolfberry quality stability dynamic prediction model.

2. The method for evaluating wolfberry quality stability based on neural network according to claim 1, characterized in that: The process of obtaining the association rules between origin environmental factors and wolfberry quality based on the origin-quality association database includes: The origin environmental factors and wolfberry quality index data in the origin-quality association data were extracted, the text descriptions in the wolfberry quality index data were segmented and mapped into numerical representations; the processed data were mined and analyzed using the Apriori association rule algorithm to obtain the association rules between the origin environmental factors and wolfberry quality.

3. The method for evaluating wolfberry quality stability based on neural network according to claim 1, characterized in that: Based on the association rules and the data in the origin-quality association database, the process of determining the first key factor of wolfberry quality stability through stepwise regression analysis includes: Through correlation analysis, environmental parameters whose correlation with wolfberry quality is higher than a preset threshold are screened out; the screened environmental parameters are used as independent variables, and the wolfberry quality index is used as the dependent variable, a regression model is established, the regression coefficients of the environmental parameters of each production area are obtained, and the first key factor is determined based on the regression coefficient.

4. The method for evaluating wolfberry quality stability based on neural network according to claim 1, characterized in that: The process of training the static prediction model of wolfberry quality stability based on the first key factor includes: extracting the data corresponding to the first key factor, constructing the characteristic vector matrix of the key factor, and using a support vector machine to construct the static prediction model of wolfberry quality stability; training the static prediction model of wolfberry quality stability according to the characteristic vector matrix, and determining the optimal parameter combination of the kernel function type, penalty coefficient, and kernel function parameters of the support vector machine model through cross-validation.

5. The method for evaluating wolfberry quality stability based on neural network according to claim 1, characterized in that: The process of selecting wolfberry sensory indicators and active ingredient content indicators, constructing a wolfberry quality quantitative evaluation model, and training it through indicator values ​​includes: According to the relevant standards for wolfberry quality evaluation, wolfberry sensory indicators and effective ingredient content indicators are selected to construct a wolfberry quality evaluation index system; multi-source data corresponding to the wolfberry quality evaluation index system are obtained, and the multi-source data are fused at the feature layer through a decision tree algorithm to generate a wolfberry quality evaluation fusion feature vector, based on the wolfberry quality evaluation fusion feature vector, a convolutional neural network is used to extract wolfberry high-level features, and a recursive feature elimination method is used to screen a key feature subset that meets the conditions for quality contribution, the wolfberry quality quantitative evaluation model is trained through the key feature subset, and the optimal parameters of the model are determined through 5-fold cross validation.

6. The method for evaluating wolfberry quality stability based on neural network according to claim 1, characterized in that: The process of training the dynamic prediction model of wolfberry quality stability includes: The maximum information coefficient was used to screen in the origin-quality association database to obtain the second key factor. The original data corresponding to the first key factor and the second key factor were smoothed by the wavelet threshold denoising method. The dynamic prediction model of wolfberry quality stability was trained based on the processed data. The Adam optimization algorithm was used for training, and L1 and L2 regularization terms were added to the cross entropy loss function.

7. The method for evaluating wolfberry quality stability based on neural network according to claim 6, characterized in that: Adversarial samples are generated by randomly adding Gaussian noise to the original data after processing, and the adversarial data is input into the dynamic prediction model of wolfberry quality stability for training.

8. The method for evaluating wolfberry quality stability based on neural network according to claim 1, characterized in that: The dynamic prediction model for wolfberry quality stability includes a long short-term memory neural network and a convolutional network structure. The outputs of the long short-term memory neural network and the convolutional network structure are cascaded, and the results are output through a softmax classifier; wherein the long short-term memory neural network has three layers, and each layer contains 128, 64, and 32 neurons respectively; the convolutional network structure includes five convolutional layers, three pooling layers, and one global average pooling layer, and the convolution kernel size is 3×3.

9. A neural network-based wolfberry quality stability evaluation system, characterized in that: include: The wolfberry quality static evaluation module is used to build and train a static prediction model for wolfberry quality stability; Wolfberry quality scoring module, used to build and train a wolfberry quality quantitative evaluation model; The dynamic evaluation module of wolfberry quality is used to build and train the dynamic prediction model of wolfberry quality stability; The comprehensive evaluation module is used to evaluate the quality of wolfberry according to the trained wolfberry quality quantitative evaluation model, and to obtain a comprehensive evaluation of the quality stability of wolfberry according to the trained wolfberry quality stability static prediction model and the wolfberry quality stability dynamic prediction model.

Images