Agricultural product marketing data summarization analysis method

Abstract

The invention provides a summarizing and analyzing method based on agricultural product marketing data, and belongs to the technical field of electric digital data processing, and the method comprises the steps: collecting marketing data of an agricultural product sales terminal, building a first neural network model and a second neural network model, inputting the marketing data into the first neural network model to obtain a first feature vector, and obtaining a second feature vector; establishing a sales contribution degree matrix, inputting a sales trend feature and a price fluctuation feature in the first feature vector into a second neural network model to obtain a second feature vector, and establishing a kernel function mapping model to map the first feature vector and the second feature vector to a high-dimensional feature space, and establishing an agricultural product market competitiveness evaluation model based on the sales contribution degree matrix, generating sales strategy suggestions by using the agricultural product market competitiveness evaluation model, and performing optimization training on the first neural network model by using a neural network distillation method. The problem that in the prior art, dynamic feature extraction of agricultural product sales data lacks consideration on market competitiveness factors is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of electronic digital data processing, and more particularly, relates to a method for summarizing and analyzing agricultural product marketing data. Background Art

[0002] The analysis of agricultural product marketing data is an important link to support agricultural production and sales decisions. Traditional analysis of agricultural product marketing data mainly uses statistical analysis methods and simple machine learning models, such as linear regression, decision trees, etc. These methods analyze historical sales data, price data, and inventory data to provide a basis for formulating agricultural product sales strategies. In the specific implementation process, steps such as data preprocessing, feature extraction, model training, and predictive analysis are usually adopted. Among them, feature extraction mainly relies on statistically designed features by humans, and predictive analysis targets a single indicator, lacking an overall grasp of the market competition environment.

[0003] However, traditional technologies have significant defects in practical applications. First, the feature extraction process overly relies on human experience, making it difficult to adapt to the rapidly changing market environment and unable to effectively capture the dynamic features of agricultural product sales. Second, existing analysis methods often separate tasks such as sales prediction, price analysis, and competitiveness evaluation, with each module operating independently and lacking organic integration. Third, traditional methods are inefficient in computing when dealing with high-dimensional and non-linear market data and are difficult to mine the deep correlation relationships between agricultural products. Fourth, existing technologies do not make sufficient use of the information on the display positions of agricultural products and do not fully consider the impact of spatial layout on sales. Fifth, the modeling of the market competition environment by traditional models is too simplistic and cannot accurately reflect the competition and cooperation relationships between agricultural products.

[0004] Facing the increasingly complex competition environment in the agricultural product market, existing technologies are difficult to achieve in-depth mining of sales data and accurate evaluation of market competitiveness. Especially in scenarios where it is necessary to simultaneously consider time series features, spatial distribution features, and market competition relationships, traditional methods are difficult to provide a systematic solution. This fragmented analysis method limits the optimization effect of agricultural product marketing strategies and cannot meet the demand for precision marketing in the modern agricultural market. That is to say, there is a technical problem in the existing technology that the extraction of dynamic features of agricultural product sales data lacks consideration of market competitiveness factors. Summary of the Invention

[0005] In view of this, the present invention provides a method for summarizing and analyzing agricultural product marketing data, which can solve the technical problem in the existing technology that the extraction of dynamic features of agricultural product sales data lacks consideration of market competitiveness factors.

[0006] The present invention is implemented as follows: The present invention provides a method for aggregating and analyzing agricultural product marketing data, including the following steps: collecting marketing data of agricultural product sales terminals, establishing a first neural network model and a second neural network model, where the first neural network model is a fully connected neural network for extracting basic features of agricultural product marketing data, and the second neural network model is a convolutional neural network for processing features with spatial correlation. The weight distribution of the first neural network model is a Laplace distribution, and the weight distribution of the second neural network model is a Gaussian distribution. Input the marketing data into the first neural network model to obtain a first feature vector, where the first feature vector includes sales trend features, price fluctuation features, and regional distribution features. Establish a sales contribution matrix, where the sales contribution matrix includes self - contribution values and interaction contribution values. The self - contribution value represents the sales ability of a single agricultural product, and the interaction contribution value represents the sales correlation between different agricultural products. Input the sales trend features and price fluctuation features in the first feature vector into the second neural network model to obtain a second feature vector, where the second feature vector includes price elasticity coefficients and market saturation indicators. Establish a kernel function mapping model, where the kernel function mapping model uses an improved radial basis kernel function. The improved radial basis kernel function considers the contribution weight coefficient based on the sales contribution matrix and the Euclidean distance between the first feature vector and the second feature vector. Map the first feature vector and the second feature vector to a high - dimensional feature space. Establish an agricultural product market competitiveness evaluation model based on the sales contribution matrix. The agricultural product market competitiveness evaluation model is optimized using a loss function, where the loss function includes a mean square error term and a relative entropy term. Generate sales strategy suggestions using the agricultural product market competitiveness evaluation model, and optimize and train the first neural network model using the neural network distillation method, where the neural network distillation method uses the second neural network model as a teacher model.

[0007] Among them, the step of collecting marketing data of agricultural product sales terminals is specifically to collect the sales quantity of agricultural products, the sales amount of agricultural products, the sales time of agricultural products, the agricultural product category number, the origin information of agricultural products, and the display position information of agricultural products. Assign a unique product identification code to each agricultural product through a distributed data collection architecture, and attach it to the agricultural product packaging in the form of a barcode or a QR code.

[0008] Among them, the construction steps of the first neural network model and the second neural network model are specifically as follows: the first neural network model includes an input layer, four hidden layers, and an output layer. The number of neurons in the four hidden layers is 1024, 512, 256, and 128 respectively, and the number of neurons in the output layer is 64. The ReLU activation function is used in the hidden layers, and the Sigmoid activation function is used in the output layer. The second neural network model includes three convolutional layers and two fully connected layers. The convolutional kernel sizes of the three convolutional layers are 5×5, 3×3, and 3×3 respectively, and the number of convolutional kernels in the three convolutional layers is 32, 64, and 128 respectively. The number of neurons in the two fully connected layers is 256 and 128 respectively. A max pooling layer is used after the three convolutional layers, and the pooling kernel size of the max pooling layer is 2×2.

[0009] Among them, the first feature vector includes sales trend features, price fluctuation features, and regional distribution features. Fourier transform is performed on the sales trend features in the first feature vector to obtain a frequency domain feature matrix, which is used to characterize the periodic law of agricultural product sales.

[0010] Among them, the main frequency components are extracted according to the frequency domain feature matrix to establish an agricultural product sales cycle prediction model. The agricultural product sales cycle prediction model is based on the natural growth cycle and market supply and demand relationship. The agricultural product sales cycle prediction model adopts a bidirectional long short-term memory network structure, considering both historical and future time series information.

[0011] Among them, the sales contribution degree matrix includes self-contribution values and interaction contribution values. The self-contribution value characterizes the sales ability of a single agricultural product, and the interaction contribution value characterizes the sales correlation between different agricultural products.

[0012] Among them, the kernel function mapping model adopts an improved radial basis kernel function, which considers the contribution weight coefficient based on the sales contribution degree matrix and the Euclidean distance between the first feature vector and the second feature vector.

[0013] Among them, the agricultural product market competitiveness evaluation model constructs a comprehensive evaluation model based on a graph neural network and a multi-task learning framework. The agricultural product market competitiveness evaluation model represents the competition relationship between different agricultural products as a weighted graph structure, and the edge weights in the graph structure are jointly determined by the sales contribution degree matrix and the feature distance matrix.

[0014] Among them, the loss function of the agricultural product market competitiveness evaluation model includes three terms: the first term is the two-norm of the weight vector, which is used to control the model complexity; the second term is the slack variable term with a penalty factor, which is used to balance the error tolerance of the model; the third term is the relative entropy term with a balance factor, which is used to measure the difference between the predicted distribution and the true distribution.

[0015] Among them, the trained first neural network model is used for real-time analysis of agricultural product marketing data, generating market dynamic warning information, and outputting the sales strategy suggestions and the market dynamic warning information.

[0016] Compared with the prior art, a method for aggregating and analyzing agricultural product marketing data provided by the present invention. The method for aggregating and analyzing agricultural product marketing data proposed by the present invention realizes automatic feature extraction and in-depth analysis of agricultural product sales data by constructing a dual neural network model architecture. This method organically integrates sales trend features, price fluctuation features, and regional distribution features, and realizes non-linear transformation of the feature space through an improved kernel function mapping, thereby establishing a unified market competitiveness evaluation framework.

[0017] The technical solution adopted by the present invention effectively overcomes the limitations of the traditional technology. The first neural network model automatically extracts basic features, avoiding the problem of over-reliance on manual experience; the second neural network model specifically processes spatial correlation features, improving the utilization efficiency of display position information; the introduction of the sales contribution degree matrix realizes accurate modeling of the competition relationship between agricultural products; the application of the knowledge distillation mechanism improves the generalization ability and real-time processing efficiency of the model. This solution realizes comprehensive perception and analysis of market dynamics through multi-level feature extraction and fusion.

[0018] The present invention successfully solves the technical problem in the prior art that the dynamic feature extraction of agricultural product sales data lacks consideration of market competitiveness factors. This is because the method establishes an end-to-end analysis framework, organically combines tasks such as feature extraction, market modeling, and competitiveness evaluation, and realizes the full-process optimization from data collection to strategy generation. At the same time, by introducing an improved radial basis kernel function and a multi-task learning framework, this method can accurately depict the market association relationship between agricultural products, providing reliable technical support for formulating precise marketing strategies. Brief Description of the Drawings

[0019] Figure 1 It is a flowchart of the method of the present invention.

[0020] Figure 2 It is a sales trend feature diagram of a certain vegetable in 30 days in Embodiment 2.

[0021] Figure 3It is a frequency-domain analysis result diagram of the sales data in Example 2.

[0022] Figure 4 It is a heat map of the position contribution value in Example 2

[0023] Figure 5 It is a competitiveness evaluation result diagram in Example 2. Detailed implementation manners

[0024] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] As Figure 1 shown, it is a flowchart of a method for summarizing and analyzing agricultural product marketing data provided by the present invention. This method includes the following steps:

[0026] S01. Collect marketing data of agricultural product sales terminals. The marketing data includes the quantity of agricultural product sales, the amount of agricultural product sales, the time of agricultural product sales, the agricultural product category number, the agricultural product origin information, and the agricultural product display position information;

[0027] S02. Establish a first neural network model and a second neural network model. The first neural network model is a fully connected neural network, and the second neural network model is a convolutional neural network. The weight distribution of the first neural network model is a Laplace distribution, and the weight distribution of the second neural network model is a Gaussian distribution;

[0028] S03. Input the marketing data into the first neural network model for preprocessing to obtain a first feature vector. The first feature vector includes sales trend features, price fluctuation features, and regional distribution features;

[0029] S04. Perform Fourier transform on the sales trend features in the first feature vector to obtain a frequency-domain feature matrix, which is used to characterize the periodic law of agricultural product sales;

[0030] S05. Extract the main frequency components according to the frequency-domain feature matrix and establish an agricultural product sales cycle prediction model. The agricultural product sales cycle prediction model is based on the natural growth cycle and the market supply and demand relationship;

[0031] S06. Calculate the display contribution value of each agricultural product. The display contribution value is based on the agricultural product display position information and the sales conversion rate, and the display contribution value is used to characterize the influence degree of agricultural product display on sales;

[0032] S07. Establish a sales contribution matrix, where the sales contribution matrix includes self - contribution values and interaction contribution values. The self - contribution value represents the sales ability of a single agricultural product, and the interaction contribution value represents the sales correlation between different agricultural products;

[0033] S08. Input the price fluctuation characteristics and the display contribution value into the second neural network model to obtain a second feature vector, where the second feature vector includes a price elasticity coefficient and a market saturation index;

[0034] S09. Establish a kernel function mapping model. The kernel function mapping model uses an improved radial basis kernel function to map the first feature vector and the second feature vector into a high - dimensional feature space;

[0035] S10. Calculate a feature distance matrix in the high - dimensional feature space, where the feature distance matrix is used to represent the market correlation between different agricultural products;

[0036] S11. Establish an agricultural product market competitiveness evaluation model based on the sales contribution matrix and the feature distance matrix. The agricultural product market competitiveness evaluation model is optimized using a loss function, and the loss function includes a mean square error term and a relative entropy term;

[0037] S12. Generate sales strategy suggestions using the agricultural product sales cycle prediction model and the agricultural product market competitiveness evaluation model. The sales strategy suggestions include a display position optimization plan and price adjustment suggestions;

[0038] S13. Optimize and train the first neural network model using the neural network distillation method. The neural network distillation method uses the second neural network model as a teacher model;

[0039] S14. Use the trained first neural network model for real - time agricultural product marketing data analysis, generate market dynamic warning information, and output the sales strategy suggestions and the market dynamic warning information;

[0040] The structure of the first neural network model includes an input layer, four hidden layers, and an output layer. The number of neurons in the input layer is the same as the dimension of the marketing data features. The number of neurons in the four hidden layers is 1024, 512, 256, and 128 respectively. The number of neurons in the output layer is 64. The hidden layer uses the ReLU activation function, and the output layer uses the Sigmoid activation function;

[0041] The structure of the second neural network model includes three convolutional layers and two fully connected layers. The kernel sizes of the three convolutional layers are 5×5, 3×3, and 3×3 respectively, the numbers of kernels of the three convolutional layers are 32, 64, and 128 respectively, the numbers of neurons of the two fully connected layers are 256 and 128 respectively. A max pooling layer is used after the three convolutional layers, and the pooling kernel size of the max pooling layer is 2×2;

[0042] The improved radial basis kernel function takes into account the contribution weight coefficient based on the sales contribution matrix and the Euclidean distance between the first eigenvector and the second eigenvector, enabling non-linear mapping and feature enhancement of data during the process of mapping the first eigenvector and the second eigenvector to the high-dimensional feature space. The weight coefficient calculated by introducing the sales contribution matrix in the improved radial basis kernel function is used to adjust the influence degree between different features, and the weight coefficient is calculated based on the self-contribution value and the interaction contribution value;

[0043] The agricultural product market competitiveness evaluation model constructs a multi-variable non-linear regression model based on weighted support vector regression. The loss function of the agricultural product market competitiveness evaluation model includes three terms: the first term is the second norm of the weight vector, which is used to control the model complexity; the second term is the slack variable term with a penalty factor, which is used to balance the error tolerance of the model; the third term is the relative entropy term with a balancing factor, which is used to measure the difference between the predicted distribution and the true distribution. The agricultural product market competitiveness evaluation model uses the sequential minimal optimization algorithm to solve the optimal parameters, and the agricultural product market competitiveness evaluation model outputs the market competitiveness score of the agricultural product;

[0044] The agricultural product sales cycle prediction model constructs a time series prediction model based on an improved long short-term memory network. The agricultural product sales cycle prediction model adopts a bidirectional long short-term memory network structure. The agricultural product sales cycle prediction model considers both historical and future time series information at the same time. The input of the agricultural product sales cycle prediction model includes the main frequency component extracted from the frequency domain feature matrix, the historical sales data sequence, and the agricultural product natural growth cycle information. The agricultural product sales cycle prediction model designs an attention mechanism in the network structure to capture periodic features at different time scales. The agricultural product sales cycle prediction model introduces a seasonal decomposition module to separate the long-term trend, seasonal fluctuation, and random fluctuation components of agricultural product sales. The loss function of the agricultural product sales cycle prediction model uses the mean square error with an L1 regularization term, and the parameters are updated through the Adam optimizer. The prediction output of the agricultural product sales cycle prediction model includes the predicted sales volume values for multiple future time windows and the confidence estimation of the prediction interval;

[0045] The agricultural product market competitiveness evaluation model constructs a comprehensive evaluation model based on a graph neural network and a multi-task learning framework. The agricultural product market competitiveness evaluation model represents the competition relationship between different agricultural products as a weighted graph structure. In the graph structure, nodes represent agricultural products, and the edge weights in the graph structure are jointly determined by the sales contribution matrix and the feature distance matrix. The agricultural product market competitiveness evaluation model uses a graph convolutional network for feature extraction. The agricultural product market competitiveness evaluation model designs multiple sub-tasks, including market share prediction, price trend prediction, and product life cycle stage judgment. The sub-tasks share the underlying feature representation but have independent task-specific layers. The loss function of the agricultural product market competitiveness evaluation model adopts a multi-task joint learning strategy, balances the training difficulty of different tasks through dynamic weights, and introduces adversarial training to improve the robustness of the model. The output of the agricultural product market competitiveness evaluation model includes a comprehensive competitiveness score, sub-item scores for each dimension, and decision-making suggestions.

[0046] The specific implementation manners of the above steps are described in detail below. The specific implementation manner of step S01 is to record the agricultural product marketing data in real time through a data acquisition system deployed at the agricultural product sales terminals. This system adopts a distributed data acquisition architecture, arranges data acquisition nodes at each sales outlet, and collects information related to agricultural product sales through a sensor network and radio frequency identification technology. During specific acquisition, first, establish a basic information database for agricultural products, assign a unique commodity identification code to each agricultural product, and attach it to the agricultural product packaging in the form of a barcode or a QR code. When selling, read the commodity identification code through a scanning device and associate it with the sales quantity, sales amount, and sales time of the agricultural product, and at the same time record the sales outlet information as the origin information. In terms of collecting display position information, divide the sales area into several display units, and each display unit is equipped with a position sensor for detecting the placement position of the agricultural product. The system associates the display unit number with the commodity identification code to form agricultural product display position data. The function of this step is to establish a complete agricultural product marketing data acquisition system to provide a data basis for subsequent analysis. The data acquisition frequency is dynamically adjusted according to the sales scale, and generally, a sampling frequency of once every 5 minutes is adopted to ensure the real-time and integrity of the data. For the control of data quality, the system sets an outlier detection mechanism, which automatically marks and manually reviews when the collected data exceeds the preset threshold range. The outlier threshold for the sales amount is set to 3 times the historical average value, and the outlier threshold for the sales quantity is set to 5 times the historical average value.

[0047] The specific implementation of step S02 is to construct a dual neural network model architecture for feature extraction tasks at different levels. The first neural network model adopts a fully connected structure to extract the basic features of agricultural product marketing data. The weight initialization of this model uses the Laplace distribution with a mean of 0 and a scale parameter set to 0.01. This initialization method helps the model maintain stability in the initial stage of training. The second neural network model adopts a convolutional structure to specifically process features with spatial correlation, such as the display position information of agricultural products. The weight initialization of this model uses the Gaussian distribution with a mean of 0 and a standard deviation set to 0.02, which is beneficial for hierarchical feature extraction. The two models are trained using different learning strategies. The first neural network model uses the batch stochastic gradient descent algorithm with an initial learning rate set to 0.001, and the learning rate decays to 0.95 times the original after every 50 training epochs. The second neural network model uses the Adam optimizer with an initial learning rate set to 0.0002, a decay rate of 0.9, and a momentum parameter of 0.99. During the model training process, an early stopping strategy is adopted to prevent overfitting. Training stops when the loss function on the validation set does not decrease for 5 consecutive epochs. The purpose of this step is to construct a deep learning model architecture that can effectively extract agricultural product marketing features.

[0048] The specific implementation of step S03 is to use the first neural network model to extract features from agricultural product marketing data. First, the input data is standardized using the min-max normalization method to scale the features of each dimension to the interval from zero to one. Then the standardized data is input into the first neural network model, which extracts sales trend features, price fluctuation features, and regional distribution features through multiple non-linear transformations. During the feature extraction process, the batch normalization technique is adopted to improve the stability of model training. Each layer of features is non-linearly activated after being normalized. For the sales trend features, the model focuses on the time series change pattern of sales volume and extracts local trend information through the sliding window method with the window size set to 7 days. The price fluctuation features are extracted using the difference method to calculate the price change rate between adjacent time points. The regional distribution features are obtained through cluster analysis of the sales outlet location information using an improved density clustering algorithm with the clustering radius parameter set to the mean of the nearest neighbor distances of the sales outlets. The role of this step is to convert the original marketing data into feature vectors with clear semantics for subsequent analysis.

[0049] The specific implementation of step S04 is to perform a frequency-domain transformation on the sales trend characteristics. The fast Fourier transform algorithm is used to convert the time-domain signal into a frequency-domain representation. First, the sales trend characteristics are preprocessed, and the missing values are processed by linear interpolation to ensure the continuity of the data. Then, a suitable time window length is selected, usually taken as 30 days, and the data is segmented. For each segment of data, the mean value is first subtracted to eliminate the DC component, and then a Hanning window is added to reduce spectral leakage. Next, the fast Fourier transform is performed to obtain the spectrum in complex form. The spectrum is converted into an amplitude spectrum and a phase spectrum to form a frequency-domain feature matrix. The rows of this matrix represent different time windows, and the columns represent different frequency components. The power spectral density analysis method is used to calculate the energy proportion of each frequency component, and the frequency components with an energy proportion exceeding 1% are considered significant periodic components. The purpose of this step is to identify the periodic patterns of agricultural product sales and provide a basis for sales forecasting.

[0050] The specific implementation of step S05 is to construct an agricultural product sales cycle prediction model based on the frequency-domain feature matrix. First, the main frequency components are extracted from the frequency-domain feature matrix. The singular value decomposition method is used to decompose the frequency-domain feature matrix, and the eigenvectors corresponding to the singular values with an energy proportion exceeding 85% are selected as the main frequency components. Then, combined with the natural growth cycle information of agricultural products, a seasonal adjustment model is established. This model takes into account the growth cycle of agricultural products, climate factors, and market supply and demand relationships, and integrates these influencing factors through a weighted combination method. When predicting, a sliding prediction strategy is adopted, and the size of the prediction window is dynamically adjusted according to the characteristics of different agricultural products, usually set to 1 / 4 to 1 / 3 of the natural growth cycle. The reliability evaluation of the prediction results uses the confidence interval method, and the confidence level is set to 95%. The role of this step is to achieve an accurate prediction of the agricultural product sales cycle.

[0051] The specific implementation of step S06 is to calculate the display contribution value of agricultural products. First, a display position scoring system is established, and each display unit is scored according to factors such as the passenger flow, line of sight accessibility, and purchase convenience of the display area. The scoring uses a 100-point system. Then, the sales conversion rate of each display position is statistically calculated. The calculation method is the ratio of the actual sales volume at this position to the product view volume, and the view volume is obtained through a passenger flow sensor. Based on the display position score and the sales conversion rate, the weighted geometric mean method is used to calculate the display contribution value. The weight coefficient is determined by the grey relational analysis method, considering the correlation degree of each factor with the sales performance. The calculation result of the display contribution value is normalized to the interval from zero to one for easy comparison between different agricultural products. The purpose of this step is to quantify the influence degree of the display position on sales.

[0052] The specific implementation of step S07 is to construct a sales contribution matrix. First, calculate the self - contribution value of each agricultural product. Using the multiple regression analysis method, take the sales volume as the dependent variable, and factors such as product price, quality grade, brand awareness, etc. as independent variables to establish a regression model. Then, analyze the sales correlation between different agricultural products. Using the association rule mining algorithm, set the minimum support to 0.1 and the minimum confidence to 0.3 to mine the purchase association rules between agricultural products. Calculate the interaction contribution value based on the mined association rules, using the mutual information as the measurement index. Finally, combine the self - contribution value and the interaction contribution value into a sales contribution matrix. This matrix is a symmetric matrix, where the diagonal elements are the self - contribution values and the non - diagonal elements are the interaction contribution values. The role of this step is to construct a quantitative evaluation system for the sales ability of agricultural products.

[0053] The specific implementation of step S08 is to use the second neural network model for deep feature extraction. Combine the price fluctuation feature and the display contribution value into an input feature vector, and perform feature extraction through the convolutional layer. The first - layer convolution focuses on extracting local feature patterns, while the second - layer and third - layer convolutions gradually extract higher - level abstract features. During the feature extraction process, use the residual connection structure to prevent the problem of gradient disappearance. The extracted features are integrated through the fully - connected layer to obtain a second feature vector containing the price elasticity coefficient and the market saturation index. The price elasticity coefficient reflects the degree of influence of price changes on sales volume, and the market saturation index characterizes the utilization degree of the market capacity. The purpose of this step is to obtain the market response characteristics of agricultural products.

[0054] The specific implementation of step S09 is to establish a kernel function mapping model. Use an improved radial basis kernel function, which introduces a weight adjustment term based on the sales contribution matrix on the basis of the standard radial basis kernel function. The parameter selection of the kernel function uses the grid search method to optimize the parameters on the validation set. The bandwidth parameter of the kernel function is adaptively adjusted according to the statistical characteristics of the input features, generally set to 0.1 to 0.5 times the standard deviation of the features. Map the first feature vector and the second feature vector to a high - dimensional feature space through the kernel function to achieve non - linear feature transformation. The role of this step is to enhance the expression ability of features and capture the non - linear relationship between features.

[0055] The specific implementation of step S10 is to calculate the feature distance matrix in the high - dimensional feature space. Use the Mahalanobis distance to measure the distance between different agricultural product feature vectors. This distance measurement method takes into account the correlation between features. Before calculating the distance, standardize the features to eliminate the influence of the dimension. Then calculate the covariance matrix for the calculation of the Mahalanobis distance. Organize the calculated distance values into a feature distance matrix, which reflects the market similarity between agricultural products. The purpose of this step is to quantify the market association degree between different agricultural products.

[0056] The specific implementation of step S11 is to construct an evaluation model for the market competitiveness of agricultural products. This model adopts a weighted support vector regression framework, and the loss function includes a mean square error term and a relative entropy term. The mean square error term is used to ensure the accuracy of prediction, while the relative entropy term is used to constrain the difference between the predicted distribution and the true distribution. The model training uses the sequential minimal optimization algorithm, which improves the solution efficiency by decomposing a large optimization problem into a series of small optimization problems. During the training process, the cross-validation method is used to select the optimal hyperparameters, including the regularization parameter and the kernel function parameter. The role of this step is to establish an evaluation system for the market competitiveness of agricultural products.

[0057] The specific implementation of step S12 is to generate sales strategy suggestions based on the prediction model and the evaluation model. First, use the sales cycle prediction model to predict future market demand, and the prediction cycle is generally set to 30 days. Then, according to the evaluation results of market competitiveness, analyze the competitive advantages and disadvantages of each agricultural product. Combining the analysis results of the display position, use the integer programming method to optimize the display layout of agricultural products, and the objective function is to maximize the overall sales revenue. For price adjustment suggestions, use game theory methods to analyze the optimal pricing strategy in a competitive environment. The purpose of this step is to provide operational marketing decision support.

[0058] The specific implementation of step S13 is to optimize the first neural network model using the neural network distillation method. Take the second neural network model as the teacher model, and transfer the feature representation ability it has learned to the first neural network model through knowledge distillation. The distillation process uses soft labels with a temperature parameter of 3, which balances the contributions of the original labels and the soft labels. During the training process, adopt a dynamic weight adjustment strategy to dynamically adjust the weight ratio of hard labels and soft labels according to the performance of the validation set. The role of this step is to improve the feature extraction ability of the first neural network model.

[0059] The specific implementation of step S14 is to deploy the optimized first neural network model for real-time data analysis. The system adopts a stream processing architecture to receive and process marketing data in real time. For the detection of abnormal fluctuations, set multi-level warning thresholds. A 30% change in sales volume triggers a first-level warning, a 50% change triggers a second-level warning, and an 80% change triggers a third-level warning. Optionally, the system regularly generates market analysis reports, including sales trend analysis, competitive situation analysis, etc.

[0060] The output vector representation of the first neural network model is as follows:

[0061] h (l) =f(W (l) h (l-1) +b (l) ):

[0062] In the formula, h (l)is the output vector of the l-th layer; W (l) is the weight matrix of the l-th layer; b (l) is the bias vector of the l-th layer; f is the activation function; when l = 1, h (0) is the input marketing data vector; parameter initialization adopts the Laplace distribution, and the probability density function is: where μ = 0 is the location parameter and b = 0.01 is the scale parameter.

[0063] Parameter acquisition method: W (l) is obtained through random initialization and then training, b (l) is initialized as a zero vector and then obtained through training; the training data includes historical sales data, and the parameters are updated by minimizing the prediction error; f uses the ReLU function in the hidden layer: f(x) = max(0, x), and the sigmoid function is used in the output layer:

[0064] The first neural network model initializes the weights using the Laplace distribution instead of the commonly used Gaussian distribution because the Laplace distribution has heavier tails, which helps to increase the sparsity and robustness of the model and can better handle outliers at the same time; the use of the ReLU activation function can alleviate the vanishing gradient problem and speed up the training speed.

[0065] The Fourier transform of the sales trend feature is expressed as follows:

[0066]

[0067] In the formula, x(n) is the time-domain sales data sequence; X(k) is the frequency-domain feature sequence; N is the sequence length, and its value is the number of days of sales data; k is the frequency index, and its value range is from 0 to N - 1; j is the imaginary unit; to reduce spectral leakage, the input sequence is windowed: x w (n) = x(n)w(n), where w(n) is the Hanning window function:

[0068] Parameter acquisition method: x(n) is the daily sales volume data, which is directly obtained through the sales system; N is generally taken as 30 days or 90 days, determined according to the analysis requirements; the Fourier transform result includes the amplitude spectrum |X(k)| and the phase spectrum ∠X(k). The amplitude spectrum is used to analyze the periodic intensity, and the phase spectrum is used to analyze the phase relationship;

[0069] The Fourier transform uses the complex exponential form to completely retain the amplitude and phase information of the data (analogous to a signal), which is conducive to comprehensively analyzing the periodic characteristics of sales; adding the Hanning window function is to reduce spectral leakage and improve the accuracy of spectral analysis; by analyzing the combination of the amplitude spectrum and the phase spectrum, the time characteristics of the sales pattern can be better understood.

[0070] The calculation of the sales contribution degree matrix is shown as follows:

[0071]

[0072] In the formula, C ij is the sales contribution degree matrix; s ii is the self - contribution value of the i - th agricultural product; s ij (i≠j) is the interaction contribution value between agricultural products i and j; n is the total number of agricultural products; the formula for calculating the self - contribution value is: s ii =α1v i +α2p i +α3q i , where v i is the sales volume, p i is the profit rate, q i is the quality grade, and α1, α2, α3 are weight coefficients obtained through multiple regression; the formula for calculating the interaction contribution value is: s ij =β1MI ij +β2SC ij , where MI ij is the mutual information quantity, SC ij is the sales - related coefficient, and β1, β2 are weight coefficients obtained by fitting experimental data.

[0073] Method for obtaining parameters: v i is obtained through the sales system; p i is obtained through the financial system; q i is obtained through the quality evaluation system, and the range is from 1 to 10; MI ij is obtained by calculating the mutual information of the sales volume sequences of two agricultural products: SC ij is obtained by calculating the Pearson correlation coefficient.

[0074] The design of the sales contribution degree matrix comprehensively considers the single - product performance and the correlation between products, and improves the accuracy of evaluation through the weighted combination of multiple influencing factors; the introduction of mutual information can capture non - linear correlation, while the Pearson correlation coefficient reflects the degree of linear correlation. The combination of the two can more comprehensively describe the correlation relationship between products.

[0075] The improved radial basis kernel function is shown as follows:

[0076]

[0077] In the formula, x i , x j is the feature vector; ω ijis the weight coefficient based on the sales contribution matrix; σ is the kernel function bandwidth parameter; λ is the regularization parameter; R ij is the feature correlation term, and its calculation formula is: where Cov represents covariance and Var represents variance; ω ij The calculation formula of is: where s ij is an element in the sales contribution matrix.

[0078] Parameter acquisition method: σ is determined by cross-validation, generally taking 0.1 to 0.5 times the standard deviation of the features; the value range of λ is 0.01 to 0.1; Cov and Var are obtained through statistical calculation of historical data; the correlation threshold is set to 0.3, and the correlation terms below this threshold are set to zero.

[0079] The improved radial basis kernel function makes the feature mapping better reflect the market relationship between products by introducing a weight modulation term and a correlation term; ω ij The design of takes into account the normalization of the sales contribution, avoiding the influence of different product scale differences; R ij The introduction of the term can capture the statistical correlation between features.

[0080] The loss function of the agricultural product market competitiveness evaluation model is expressed as follows:

[0081]

[0082] In the formula, m is the number of samples; y i is the true competitiveness score; is the predicted score; w is the model weight vector; ξ i is the slack variable; KL(p||q) is the relative entropy; γ1, γ2, γ3 are balance parameters; the relative entropy calculation formula is: where p i is the true distribution, and q i is the predicted distribution.

[0083] Parameter acquisition method: y i is obtained through market research and expert scoring, with a range of 0 to 100; γ1, γ2, γ3 are obtained through grid search optimization, and the search ranges are: γ1 ∈ [0.001, 0.1], γ2 ∈ [0.01, 1], γ3 ∈ [0.1, 2]; ξ i is obtained through optimization solving and satisfies the non-negative constraint.

[0084] The design of the loss function of the competitiveness evaluation model consists of four parts: the mean square error term is used to ensure prediction accuracy, the L2 regularization term is used to control the model complexity and prevent overfitting, the slack variable term is used to improve the model's tolerance to outliers, and the relative entropy term is used to constrain the difference between the predicted distribution and the true distribution; the terms are balanced by weight parameters.

[0085] The loss function of neural network distillation is expressed as follows:

[0086]

[0087] In the formula, z t is the output logits of the teacher model; z₀ is the output logits of the student model; T is the temperature parameter, with a value range of 3; α is the balance parameter, with a value of 0.5; L CE is the cross-entropy loss:

[0088] Parameter acquisition method: z t , z s are obtained through the forward propagation of their respective networks; the T parameter is determined through experimental verification, with a range between 2 and 5; α is determined based on the model performance on the validation set.

[0089] The design of the knowledge distillation loss function draws on the model compression theory. The smoothness of the soft labels is controlled by the temperature parameter T. A larger T value will generate a softer probability distribution, which helps with knowledge transfer; the balance parameter α adjusts the importance of the soft labels and the hard labels; the cross-entropy loss ensures the basic prediction ability of the student model.

[0090] Specifically, the principle of the present invention is: The technical principle of the present invention is based on deep learning and market competition theory. First, the design of the dual neural network structure is based on the hierarchical idea of feature extraction. The first neural network model is responsible for extracting temporal features. Its fully connected structure is suitable for processing high-dimensional input data, and through multiple non-linear transformations, it can learn the internal laws of the data; the second neural network model adopts a convolutional structure and is specifically designed to process features with spatial correlation. This structural design fully considers the spatial local correlation of the display positions of agricultural products.

[0091] In terms of feature mapping, the improved radial basis kernel function introduces weight modulation based on the sales contribution degree. This design enables the feature mapping process to be adaptively adjusted according to the market influence of different agricultural products. At the same time, by introducing a correlation term, this kernel function can better capture the non-linear relationship between features. The market competitiveness evaluation model adopts a graph neural network structure. This design takes into account that the agricultural product market is a complex relationship network, where there is both competition and collaboration among products, and the graph structure can naturally express this complex relationship.

[0092] The introduction of the knowledge distillation mechanism not only improves the efficiency of the model but also enhances its generalization ability. By transferring knowledge from the teacher model to the student model, model compression is achieved while maintaining prediction accuracy. The design of the multi-task learning framework is based on the principle of task relevance, and by sharing parameters and feature representations, the learning efficiency and generalization ability of the model are improved.

[0093] A specific Embodiment 1 of the present invention is provided below, and the specific implementation manners of each step in this Embodiment 1 are described in detail as follows.

[0094] The specific implementation manner of step S01 is to collect marketing data of agricultural product sales terminals in real time through a distributed data collection architecture. The collection system adopts a three-layer architecture design, namely a data collection layer, a data processing layer, and a data storage layer. In the data collection layer, data collection nodes are deployed for each sales outlet, and radio frequency identification technology is used to uniquely identify agricultural products. Each agricultural product identifier is expressed as: ID p = E k (t||s||c), where E k is an encryption function, t is a timestamp, s is the outlet number, and c is the commodity category code. In the data processing layer, the collected raw data is cleaned and standardized. The outlier determination in data cleaning uses Mahalanobis distance: where x is the data vector to be detected, μ is the historical data mean vector, ∑ is the covariance matrix, and when D M (x) is greater than the threshold 3.5, it is determined as an outlier. Data standardization uses the min-max standardization method: where x norm is the standardized data, x min and x max are the minimum and maximum values of the data respectively. In the data storage layer, a distributed database is used to store the processed data. The data storage structure design includes a basic information table, a sales record table, and a location information table, and the tables are associated through the commodity identification code. The collection of agricultural product display location information is realized by arranging location sensors in the sales area. The accuracy evaluation of location information uses root mean square error: where p i is the actual location coordinate, is the measured location coordinate, n is the number of measurements, and the location accuracy requirement is that RMSE is less than 0.5 meters. The sampling frequency of sales data is dynamically adjusted according to the sales scale, and the adjustment strategy is: where f s is the sampling frequency, f base is the base frequency of 5 minutes, v is the average sales volume in the current period, v0 is the reference sales volume, f maxThe maximum sampling frequency is 1 minute. A multi-level early warning mechanism is adopted for data quality control. The abnormal threshold for sales volume is 5 times the historical average, the abnormal threshold for sales amount is 3 times the historical average, and the abnormal threshold for commodity inventory is 0.5 times the safety inventory. This step provides a high-quality data basis for subsequent analysis by establishing a complete data collection system.

[0095] The specific implementation of step S02 is to construct a dual neural network model architecture, and design a fully connected neural network and a convolutional neural network respectively. The first neural network model adopts a fully connected structure, and the calculation formula for the output vector of each layer is: h (l) =f(W (l) h (l-1) +b (l) ), where h (l) is the output vector of the l-th layer, W (l) is the weight matrix, b (l) is the bias vector, and f is the activation function. The weight initialization adopts the Laplace distribution: The parameters are set as μ = 0, b = 0.01. The second neural network model adopts a convolutional structure, and the calculation formula for the convolutional layer is: where C (l) is the convolutional output of the l-th layer, is the convolutional kernel. is the input feature map, K l is the number of convolutional kernels, and * is the convolution operator. The weight initialization adopts the Gaussian distribution: The parameters are set as μ = 0, σ = 0.02. Different optimization strategies are adopted for model training. The first neural network model uses batch stochastic gradient descent, and the learning rate update formula is: where η t is the learning rate at the t-th step, η0 = 0.001 is the initial learning rate, and γ = 0.95 is the decay rate. The second neural network model adopts the Adam optimizer, and the parameter update formula is: where θ t is the parameter, is the first moment estimate, is the second moment estimate, η = 0.0002 is the learning rate, and ∈ = 10 -8 is the numerical stability constant. To prevent overfitting, an early stopping strategy is adopted. The training stops if the loss function on the validation set does not decrease for 5 consecutive epochs.

[0096] The specific implementation of step S03 is to extract feature vectors using the first neural network model. The input data is first normalized using the batch normalization method: where μ B and are the mean and variance of the small - batch data, γ and β are learnable scaling and translation parameters, ∈ = 10 - 5 is a numerical stability constant. The sales trend feature extraction uses a sliding window method with a window size of 7 days, and the feature calculation formula is: where f t is the trend feature at time t, w i is the time weight coefficient, and x t-i is the historical sales data. The price fluctuation feature is calculated using the difference method: where Δp t is the price change rate, and p t is the current price. The regional distribution feature is obtained through an improved density clustering algorithm, and the self - adaptive calculation formula for the clustering radius is: where r is the clustering radius, d ij is the distance between sales outlets, and α = 1.5 is the adjustment coefficient.

[0097] The specific implementation of step S04 is to perform frequency - domain analysis on the sales trend features. First, the data is pre - processed, and linear interpolation is used to process missing values: where x i is the point to be interpolated, and x a and x b are known data points. Then, the fast Fourier transform is performed: To reduce spectral leakage, windowing is performed on the input data: x w (n)=x(n)w(n), where w(n) is the Hanning window function: The power spectral density analysis uses the periodogram method: When P(k) / P total > 0.01, the corresponding frequency component is considered significant.

[0098] The specific implementation of step S05 is to construct an agricultural product sales cycle prediction model. First, the main frequency component is extracted from the frequency - domain feature matrix using singular value decomposition: M = U∑V r , and the eigenvectors corresponding to the singular values with an energy ratio exceeding 85% are selected. The model uses a bidirectional long - short - term memory network structure, and the forward - propagation calculation formula is: The back - propagation calculation formula is: The output layer uses an attention mechanism: where the attention weight α ti is calculated through the softmax function:

[0099] The specific implementation of step S06 is to calculate the display contribution value of agricultural products. The display position scoring system uses a multi - factor weighted scoring model: where S pos is the position score, and wi is the weight coefficient, f i is the scoring factor, including passenger flow, line of sight accessibility, purchase convenience, etc. The scoring uses a 100-point system. The sales conversion rate calculation formula is: In the formula, N sale is the actual sales volume, N view is the product view volume, and the view volume is obtained through a passenger flow sensor. The sensor data processing uses the Kalman filtering algorithm: In the formula is the state estimated value, K t is the Kalman gain, z t is the observed value, and H is the observation matrix. The final calculation of the display contribution value uses the weighted geometric mean: In the formula, CV is the display contribution value, f i is the evaluation index, w i is the weight coefficient, and the weight coefficient is determined through grey relational analysis.

[0100] The specific implementation of step S07 is to construct a sales contribution matrix. The self - contribution value is calculated using a multiple regression model: s ii = α1v i + α2p i + α3q i , in the formula, v i is the sales volume, p i is the profit rate, q i is the quality grade, and the parameters are estimated by the least - squares method: α=(X T X) -1 X T y. The interactive contribution value calculation is based on association rule mining, and the Apriori algorithm is used for rule mining. The support formula is: support(A→B)=P(A∪B), the confidence formula is: confidence(A→B)=P(B|A), the minimum support is set to 0.1, and the minimum confidence is set to 0.3. The mutual information amount formula between products is: In the formula, p(x, y) is the joint probability distribution, and p(x) and p(y) are the marginal probability distributions. The final sales contribution matrix is expressed as:

[0101] The specific implementation of step S08 is to use the second neural network model for deep feature extraction. The input feature vector is processed through three convolutional layers. After each convolution, batch normalization and activation function processing are performed. The calculation formula of the l - th convolutional layer is: h l = ReLU(BN(Conv(h l-1))) where Conv represents the convolution operation, BN represents batch normalization, and ReLU represents the activation function. To prevent the vanishing gradient, a residual connection structure is adopted: h l = h l-1 + F(h l-1 ), where F is the residual mapping function. The formula for the price elasticity coefficient is: where Q is the demand quantity and P is the price. The market saturation index adopts the form of a logistic function: where β is the growth rate parameter and x0 is the inflection point parameter.

[0102] The specific implementation of step S09 is to establish a kernel function mapping model. The improved radial basis kernel function expression is: where ω ij is the weight coefficient based on the sales contribution degree matrix: R ij is the feature correlation term: The kernel function parameter optimization adopts the cross-validation method, and the validation index is the root mean square error.

[0103] The specific implementation of step S10 is to calculate the feature distance matrix. The Mahalanobis distance metric is adopted: where ∑ is the feature covariance matrix. Feature standardization adopts the z-score method: where μ is the mean and σ is the standard deviation. The feature distance matrix is used to construct the agricultural product market association network, and the network edge weight is determined by the distance value.

[0104] The specific implementation of step S11 is to construct an agricultural product market competitiveness evaluation model. This model adopts a weighted support vector regression framework, and the loss function is: where the first term is the prediction error term, the second term is the regularization term, the third term is the slack variable term, and the fourth term is the distribution matching term. The formula for relative entropy is: The model solution adopts the sequential minimal optimization algorithm, and two samples that violate the KKT conditions most severely are selected for optimization in each iteration: where E i is the prediction error, K 12 is the kernel function value, and λ is the regularization parameter. Parameter optimization adopts the grid search method, and the parameter search range is: γ1 ∈ [0.001, 0.1], γ2 ∈ [0.01, 1], γ3 ∈ [0.1, 2].

[0105] The specific implementation of step S12 is to generate sales strategy suggestions. First, use the sales cycle prediction model to predict the market demand in the next 30 days. The formula for the prediction confidence interval is: where y t is the predicted value, zα / 2 is the quantile of the standard normal distribution at a confidence level of 95%, is the predicted variance. The display layout optimization adopts an integer programming model: The constraint conditions are: where x ij is a 0-1 decision variable, and pi j is the location revenue. The price adjustment suggestion is based on the game theory method, and the Nash equilibrium is used to solve the optimal pricing strategy: where R i is the revenue function, is the equilibrium price of other products.

[0106] The specific implementation of step S13 is to optimize the model using the neural network distillation method. The distillation loss function is: where z t is the output of the teacher model, z s is the output of the student model, T is the temperature parameter, and α is the balance parameter. The knowledge transfer process adopts a hierarchical distillation strategy, and the intermediate layer feature matching adopts attention transfer: where A t and A s are the attention maps of the teacher model and the student model respectively, and ||·|| F is the Frobenius norm.

[0107] The specific implementation of step S14 is to deploy the optimized model for real-time analysis. The system adopts a stream processing architecture, and the data stream processing adopts a sliding window method: w t =[x t-k+1 , …, x t , and the window size k is dynamically adjusted according to the data characteristics. The anomaly detection adopts a multi-level early warning mechanism, and the early warning threshold calculation formula is: threshold i =μ + i·σ, where μ is the historical mean, σ is the standard deviation, i is the early warning level, i = 3 for the first-level early warning, i = 4 for the second-level early warning, and i = 5 for the third-level early warning. The market dynamic early warning information includes price anomaly early warning, inventory early warning, and sales volume early warning. The confidence score of the early warning information adopts the fuzzy comprehensive evaluation method: where R is the correlation matrix, W is the weight vector, is the fuzzy composition operator. The system performance evaluation adopts two indicators: real-time response time and early warning accuracy rate. The real-time response time requirement is less than 100 milliseconds, and the early warning accuracy rate requirement is greater than 90%. The generation cycle of the market analysis report is once a day, and the report content includes sales trend analysis, competition situation analysis, market risk assessment, and strategy suggestions, and the analysis text is automatically generated using natural language generation technology.

[0108] To better understand and implement the present invention, an embodiment 2 of a specific application scenario of the present invention is provided below:

[0109] In January 2024, a certain unit launched a project on marketing data analysis of agricultural products based on deep learning. The project selected the largest agricultural product wholesale market in the provincial capital city as a pilot, and collected and analyzed the marketing data of 300 common fresh agricultural products such as vegetables and fruits for 3 months.

[0110] First, the project team deployed a data collection system at 50 sales outlets in this wholesale market. Each outlet was equipped with devices such as barcode scanners, weight sensors, and passenger flow detectors. The system assigned a 16-bit commodity identification code to each agricultural product, where the first 4 bits represent the category, the middle 8 bits represent the specific variety, and the last 4 bits represent the origin number. The sales area was divided into 300 standard display units, each with an area of 4 square meters, and equipped with infrared position sensors and temperature and humidity sensors. The system collected data every 5 minutes, and the recorded content included information such as sales quantity (accurate to 0.1 kg), sales amount (accurate to 0.01 yuan), passenger flow, and display position number. The data anomaly detection threshold was set as follows: the sales amount does not exceed 3 times the historical average, and the sales quantity does not exceed 5 times the historical average.

[0111] In terms of the construction of the neural network model, the first neural network model adopted a 4-layer hidden layer structure, with the number of neurons being 1024, 512, 256, and 128 respectively, and 64 neurons in the output layer. The weight initialization adopted the Laplace distribution with the location parameter μ = 0 and the scale parameter b = 0.01. The model training adopted the batch stochastic gradient descent algorithm, with an initial learning rate of 0.001, and the learning rate decayed to 0.95 times the original every 50 epochs. The second neural network model adopted a structure of 3 convolutional layers and 2 fully connected layers. The convolutional kernel sizes were 5×5, 3×3, and 3×3 respectively, the number of convolutional kernels was 32, 64, and 128, and a 2×2 max pooling layer was adopted. The weight initialization adopted the Gaussian distribution with a mean of 0 and a standard deviation of 0.02. The training adopted Ada m optimizer, with an initial learning rate of 0.0002 and a decay rate of 0.9.

[0112] For the extraction of sales trend features, the system used a 7-day sliding window to calculate the following indicators for each agricultural product: daily average sales volume, sales volume standard deviation, growth rate, and volatility coefficient. The price fluctuation features include daily price change rate, price elasticity coefficient, and price fluctuation range. The regional distribution features were obtained through an improved density clustering algorithm, and the clustering radius was set to 0.3 times the mean of the nearest neighbor distances. The following is a partial feature data table of a certain vegetable:

[0113] Table 1 Sales Trend Feature Data (7-day Window)

[0114] Date Average daily sales volume (kg) Sales standard deviation Growth rate (%) Fluctuation coefficient January 1st 523.5 78.6 0 0.15 January 2nd 548.7 82.3 4.8 0.15 January 3rd 592.3 88.8 7.9 0.15 January 4th 567.8 85.2 -4.1 0.15 January 5th 601.2 90.2 5.9 0.15

[0115] Figure 2 It shows the sales trend characteristics of a certain kind of vegetable in 30 days. The blue solid line represents the changing trend of the daily average sales volume, and the light blue area represents the range of the standard deviation of the sales volume. It can be clearly observed from the figure the sales fluctuation law in a 7-day cycle, which is consistent with the weekly shopping habits of consumers. The abscissa in the figure represents the date (days), and the ordinate represents the sales volume (kg / day). It can be clearly seen that the sales volume fluctuates between 500 - 700 kg / day. In terms of frequency domain analysis, a 30-day basic time window is adopted to perform Fourier transform on the sales trend characteristics. Before the transform, a Hanning window function is used for preprocessing, and the window function expression is: where N = 30. The Fourier transform adopts the following expression: By analyzing the amplitude spectrum, it is found that this agricultural product has an obvious 7-day periodicity, and the energy proportion of the frequency component corresponding to the 7-day cycle reaches 37.8%. Figure 3 It shows the frequency domain analysis results of the sales data. The abscissa represents the frequency (days -1 ), and the ordinate represents the energy proportion of each frequency component.

[0116] The construction of the sales contribution degree matrix adopts a multi-dimensional index evaluation system. The formula for the self - contribution value is: s ii = 0.4v i + 0.3p i + 0.3q i , where v i is the normalized value of the sales volume, p i is the profit rate (%), and q i is the quality grade (1 - 10 points). The formula for the interactive contribution value is: s ij = 0.6MI ij + 0.4SC ij , where MI ij is the mutual information quantity, and SC ij is the sales correlation coefficient. The following are the sales contribution degree data of 5 main kinds of vegetables:

[0117] Table 2 Vegetable Sales Contribution Degree Matrix

[0118] Number 0001 0002 0003 0004 0005 0001 0.82 0.45 0.38 0.21 0.15 0002 0.45 0.78 0.42 0.25 0.18 0003 0.38 0.42 0.75 0.35 0.22 0004 0.21 0.25 0.35 0.71 0.41 0005 0.15 0.18 0.22 0.41 0.68

[0119] Figure 4It is a heatmap showing the contribution values of display positions. The darker the color, the higher the contribution value. It can be observed from the figure that the self - contribution values on the diagonal are generally between 0.7 and 0.9, while the interaction contribution values between adjacent positions fluctuate between 0.3 and 0.6. The horizontal and vertical coordinates of the heatmap both represent the display position numbers, clearly showing the contribution relationship between different display positions. The improved radial basis kernel function adopts the following expression: where σ takes 0.3 times the characteristic standard deviation, and ω ij is the normalized weight based on the sales contribution matrix, and R ij is the feature correlation term. Through this kernel function, the feature vector is mapped to a 1024 - dimensional feature space.

[0120] The loss function of the market competitiveness evaluation model is set as: The model training uses the sequential minimal optimization algorithm, with the number of iterations set to 1000 times and the convergence threshold to 0.001. After training, the mean square error of the model on the validation set is 0.0235, and the relative error is 3.8%. Figure 5 The radar chart is used to show the competitiveness evaluation results of 5 main vegetables in four dimensions: market share, price competitiveness, quality score, and display effect. Each vegetable variety is represented by different colors and patterns, and the score range is from 0 to 1. It can be seen from the figure that Vegetable A has advantages in market share and quality score, while Vegetable C is outstanding in price competitiveness.

[0121] During the knowledge distillation process, the loss function adopts: where the temperature parameter T = 3. Through knowledge distillation, the first neural network model has been significantly improved in feature extraction performance, and the feature extraction accuracy on the validation set has increased from 85.6% to 91.3%.

[0122] During the actual operation of the system, a three - level early warning mechanism is set: a 30% change in sales volume triggers a first - level early warning, a 50% change triggers a second - level early warning, and an 80% change triggers a third - level early warning. For early warning information, the system automatically generates an analysis report and pushes it to the relevant person in charge. The following is an example of early warning data on a certain day:

[0123] Table 3 Market Early Warning Information Statistics

[0124] Early warning level Quantity of agricultural products Average price change (%) Main influencing factors Level 1 15 12.5 Weather change Level 2 8 28.3 Supply shortage Level 3 3 45.7 Quality problem

[0125] Through a three-month trial operation, the system has demonstrated remarkable practical effects. The accuracy rate of agricultural product sales prediction has reached 92.5%, the accuracy rate of price fluctuation early warning has reached 88.7%, and the adoption rate of display location optimization suggestions has reached 85.3%. The application of the system has increased the overall turnover of the pilot market by 15.2%, improved the inventory turnover rate by 23.5%, and reduced the product loss rate by 8.7%.

[0126] Traditional data analysis for agricultural product marketing mainly relies on simple statistical methods and empirical judgments, and has the following problems: 1. Data collection is not comprehensive and real-time enough. It often only focuses on sales volume and amount, ignoring important information such as customer flow and display location. 2. The analysis methods are too simple, mainly using basic statistical methods such as moving average and linear regression, and unable to effectively mine the deep patterns in the data. 3. The prediction model has low accuracy and short prediction period, and is difficult to cope with the rapid changes in the market. 4. There is a lack of a systematic competitiveness evaluation system, and the analysis of the correlation between agricultural products is insufficient. 5. The decision-making suggestions lack scientific basis and mainly rely on empirical judgments.

[0127] In contrast, the present invention has the following technical advantages: 1. A complete data collection system is constructed to achieve data acquisition throughout the entire sales process, with a collection frequency of once every 5 minutes, ensuring the real-time nature and integrity of the data. 2. A dual neural network model architecture is adopted, and features are extracted through deep learning methods, capable of discovering complex data patterns. 3. Fourier analysis is innovatively combined with deep learning to achieve accurate periodic prediction. 4. An improved kernel function based on sales contribution degree is proposed, enhancing the feature expression ability. 5. A competitiveness evaluation model with a multi-task learning framework is designed to provide a comprehensive market analysis. 6. The model performance is optimized through the knowledge distillation method, improving the practicality of the system. These innovative points enable the present system to have significant technical advantages in the field of agricultural product marketing data analysis and provide strong support for the refined management of the agricultural product market.

[0128] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Table 4 below.

[0129] Table 4 Variable Explanation Table

[0130]

[0131]

[0132] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. A method for summarizing and analyzing agricultural product marketing data, characterized in that: The method comprises the following steps: collecting marketing data of agricultural product sales terminals, establishing a first neural network model and a second neural network model, wherein the first neural network model is a fully connected neural network for extracting basic features of agricultural product marketing data, and the second neural network model is a convolutional neural network for processing features with spatial correlation, the weight distribution of the first neural network model is a Laplace distribution, and the weight distribution of the second neural network model is a Gaussian distribution, inputting the marketing data into the first neural network model to obtain a first feature vector, the first feature vector includes sales trend features, price fluctuation features, and regional distribution features, establishing a sales contribution matrix, the sales contribution matrix includes self-contribution values ​​and interactive contribution values, the self-contribution values ​​represent the sales ability of a single agricultural product, the interactive contribution values ​​represent the sales correlation between different agricultural products, and the sales trend features and price fluctuation features in the first feature vector are combined into a sales contribution matrix. The second eigenvector is obtained by inputting the second neural network model, wherein the second eigenvector includes a price elasticity coefficient and a market saturation index; a kernel function mapping model is established, wherein the kernel function mapping model adopts an improved radial basis kernel function, wherein the improved radial basis kernel function considers a contribution weight coefficient based on the sales contribution matrix and the Euclidean distance between the first eigenvector and the second eigenvector; the first eigenvector and the second eigenvector are mapped to a high-dimensional feature space; an agricultural product market competitiveness evaluation model is established based on the sales contribution matrix; the agricultural product market competitiveness evaluation model is optimized by a loss function, wherein the loss function includes a mean square error term and a relative entropy term; sales strategy recommendations are generated by using the agricultural product market competitiveness evaluation model; a neural network distillation method is used to optimize and train the first neural network model; and the neural network distillation method uses the second neural network model as a teacher model.

2. The agricultural product marketing data summary and analysis method according to claim 1, characterized in that: The step of collecting marketing data of agricultural product sales terminals specifically collects the sales quantity, sales amount, sales time, category number, origin information, and display location information of agricultural products, and assigns a unique product identification code to each agricultural product through a distributed data collection architecture, which is attached to the agricultural product packaging in the form of a barcode or a QR code.

3. The agricultural product marketing data summary and analysis method according to claim 1, characterized in that: The steps for constructing the first neural network model and the second neural network model are as follows: the first neural network model includes an input layer, four hidden layers and an output layer, the number of neurons in the four hidden layers are 1024, 512, 256, and 128 respectively, the number of neurons in the output layer is 64, the hidden layer adopts a ReLU activation function, and the output layer adopts a Sigmoid activation function; the second neural network model includes three convolutional layers and two fully connected layers, the convolution kernel sizes of the three convolutional layers are 5 times 5, 3 times 3, and 3 times 3 respectively, the number of convolution kernels of the three convolutional layers are 32, 64, and 128 respectively, the number of neurons in the two fully connected layers are 256 and 128 respectively, and a maximum pooling layer is used after the three convolutional layers, and the pooling kernel size of the maximum pooling layer is 2 times 2.

4. The agricultural product marketing data summary and analysis method according to claim 1, characterized in that: The first feature vector includes sales trend features, price fluctuation features, and regional distribution features. The sales trend features in the first feature vector are subjected to Fourier transform to obtain a frequency domain feature matrix, which is used to characterize the cyclical law of agricultural product sales.

5. The agricultural product marketing data summary and analysis method according to claim 4 is characterized in that: The main frequency component is extracted according to the frequency domain feature matrix, and a sales cycle prediction model for agricultural products is established. The sales cycle prediction model for agricultural products is based on the natural growth cycle and the market supply and demand relationship. The sales cycle prediction model for agricultural products adopts a bidirectional long short-term memory network structure and considers historical and future time series information at the same time.

6. The agricultural product marketing data summary and analysis method according to claim 1, characterized in that: The sales contribution matrix includes self-contribution values ​​and interactive contribution values, wherein the self-contribution values ​​represent the sales capability of a single agricultural product, and the interactive contribution values ​​represent the sales correlation between different agricultural products.

7. The agricultural product marketing data summary and analysis method according to claim 1, characterized in that: The kernel function mapping model adopts an improved radial basis kernel function, and the improved radial basis kernel function takes into account the contribution weight coefficient based on the sales contribution matrix and the Euclidean distance between the first eigenvector and the second eigenvector.

8. The agricultural product marketing data aggregation and analysis method according to claim 1, characterized in that: The agricultural product market competitiveness evaluation model builds a comprehensive evaluation model based on a graph neural network and a multi-task learning framework. The agricultural product market competitiveness evaluation model represents the competitive relationship between different agricultural products as a weighted graph structure, and the edge weights in the graph structure are jointly determined by the sales contribution matrix and the feature distance matrix.

9. The agricultural product marketing data summary and analysis method according to claim 1, characterized in that: The loss function of the agricultural product market competitiveness evaluation model includes three items: the first item is the second norm of the weight vector, which is used to control the complexity of the model; the second item is the slack variable item with a penalty factor, which is used to balance the error tolerance of the model; the third item is the relative entropy item with a balancing factor, which is used to measure the difference between the predicted distribution and the true distribution.

10. The agricultural product marketing data summary and analysis method according to claim 1, characterized in that: The trained first neural network model is used for real-time agricultural product marketing data analysis to generate market dynamics early warning information, and the sales strategy recommendations and the market dynamics early warning information are output.

Images