Intelligent precise baking method and system for improving tobacco leaf quality

Abstract

The invention provides an intelligent precise baking method and system for improving the tobacco leaf quality, and belongs to the technical field of tobacco leaf baking. The method comprises the steps that a basic process parameter set is established, and domain division is conducted according to the tobacco leaf maturity; acquiring a basic process parameter set in real time; constructing a first optimization vector, and determining a weight contribution rate through an intelligent process adaptation model; constructing a second optimization vector through a machine vision and near infrared spectrum fusion technology; establishing a stable change threshold judgment mechanism to start parameter adjustment; performing internal iterative optimization processing on the first optimization vector; calculating a process effect contribution rate; and intelligent and accurate parameter adjustment is realized through a gating weight function of the intelligent process adaptation model. The system comprises a curing barn structure unit, an environment sensing unit, a tobacco leaf information acquisition unit, a data processing unit and a control execution unit, all the units are integrated into a complete system through an industrial bus, and whole-process closed-loop management from environment sensing, tobacco leaf state monitoring to intelligent control is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of tobacco leaf baking, and specifically relates to an intelligent and precise baking method and system for improving tobacco leaf quality. Background Art

[0002] Tobacco leaf curing is a core process in tobacco processing. Traditional curing technology primarily relies on manual experience to set and execute fixed process parameters to control key parameters such as dry-bulb temperature, wet-bulb temperature, and air speed. Existing automated curing equipment typically uses pre-programmed temperature curves and time control programs to execute standardized curing processes based on tobacco leaf maturity. This is widely used in large-scale curing rooms and industrial tobacco processing centers. Traditional curing technology has significant technical shortcomings, primarily in its crude process parameter adjustment methods, which are unable to precisely adjust to the dynamic changes in tobacco leaf conditions during the curing process. Existing systems lack the ability to accurately perceive real-time changes in tobacco leaf quality. Parameter adjustment relies primarily on preset programs and manual intervention, resulting in insufficient accuracy and delayed response. Furthermore, traditional methods lack in-depth analysis of the complex interrelationships between multiple process parameters, making it impossible to achieve intelligent and coordinated adjustment of these parameters. Traditional technologies struggle to address the core issue of intelligent and precise adjustment of curing process parameters. This is because the curing process involves multiple interconnected process variables, and tobacco leaf quality changes are nonlinear and time-varying. Existing fixed parameter control and simple feedback adjustment methods are unable to adapt to the complex curing dynamics and lack intelligent decision-making capabilities and precise adjustment mechanisms. Summary of the Invention

[0003] In view of this, the present invention provides an intelligent and precise baking method and system for improving the quality of tobacco leaves, which can solve the technical problem that the existing technology is difficult to achieve intelligent and precise adjustment of baking process parameters during the tobacco leaf baking process.

[0004] The present invention is implemented as follows: The present invention provides an intelligent and precise baking method and system for improving tobacco leaf quality, including: establishing a basic process parameter set, dividing the basic process parameter set into three classification ranges: under-mature leaf parameter domain, moderately mature leaf parameter domain, and over-mature leaf parameter domain according to the maturity of tobacco leaves; collecting the basic process parameter set in real time at a first time granularity, and establishing a layered collection strategy based on differences in tobacco leaf varieties and parts; constructing a first optimization vector based on a second time granularity, determining the weight contribution rate and stage transition rate of each parameter domain through an intelligent process adaptation model; acquiring tobacco leaf quality change data through machine vision and near-infrared spectroscopy fusion technology, and constructing a second optimization vector, Establish a quality response delay rate evaluation mechanism; establish a stable change threshold judgment mechanism, and start process parameter adjustment when the numerical fluctuation of the first optimization vector and the second optimization vector exceeds the stable change threshold, and use the gated weight function of the intelligent process adaptation model for adaptive adjustment; perform internal iterative optimization on the first optimization vector, and update the weight contribution rate value through iterative calculation; calculate the process effect contribution rate and the process effect secondary change contribution rate, generate a parameter adjustment strategy, and establish a collaborative conversion mechanism between different parameter domains; dynamically balance the influence of each classification range through the hierarchical fusion weight of the intelligent process adaptation model, and realize intelligent and precise adjustment of baking process parameters.

[0005] Among them, the basic process parameter set specifically refers to a collection of multiple core control parameters that affect the quality changes of tobacco leaves during the baking process, including the dry-bulb temperature and wet-bulb temperature for controlling the temperature environment of the baking room, the wind speed for adjusting air flow, the exhaust flow rate for controlling gas exchange in the baking room, the heating rate for determining the speed of temperature change, and the steady-state time for maintaining a constant temperature state.

[0006] Among them, the first time granularity, specifically refers to the time interval for data collection of the basic process parameter set, which is set to 30 seconds. It is used to capture the rapid changes and instantaneous fluctuations of process parameters during the baking process, and at the same time establish differentiated collection frequencies for different tobacco leaf sections.

[0007] The second time granularity specifically refers to the time statistical window used when constructing the optimization vector, which is set to 10 minutes. It is used to perform statistical analysis and trend extraction within the time window of the data collected at the first time granularity, and to establish a comprehensive evaluation system for short-term fluctuation characteristics and long-term trend characteristics through multi-time scale analysis.

[0008] Among them, the first optimization vector specifically refers to an 18-dimensional vector formed by combining the mean, variance, and change rate of the basic process parameter set within the second time granularity. It is used to characterize the comprehensive state characteristics of the process parameters within the time window, quantify the influence of each parameter on the baking effect through the weighted contribution rate, and establish a stage transition rate mechanism between parameter domains to achieve dynamic switching.

[0009] Among them, the second optimization vector specifically refers to the tobacco leaf quality parameter vector obtained through sensor fusion technology, which includes three dimensions: tobacco leaf moisture content change rate, chlorophyll degradation rate, and total sugar conversion efficiency. It is used to quantify the physiological and biochemical changes of tobacco leaves and establish a quality response delay rate evaluation system to predict the time lag effect of quality changes.

[0010] Among them, the stable change threshold specifically refers to the numerical boundary for determining whether the process parameters need to be adjusted. When the change rate of any parameter in the first optimization vector exceeds 15% of its historical average or the quality parameter in the second optimization vector deviates from the target value by more than 8%, the parameter adjustment mechanism is triggered. By establishing a threshold grading system, the risk of misadjustment is reduced.

[0011] Among them, the structure of the intelligent process adaptation model is a multi-layer perception network based on a hierarchical attention network architecture, which includes five main parts: input layer, feature extraction layer, attention calculation layer, gated fusion layer and output layer. The feature extraction layer uses a residual connection structure to process the 18-dimensional first optimization vector and the 3-dimensional second optimization vector. The attention calculation layer calculates the association weights between different parameter domains through a multi-head attention mechanism.

[0012] Among them, there are the flue-curing room structure unit, environmental perception unit, tobacco leaf information collection unit, data processing unit and control execution unit. Each unit is integrated into a complete system through the industrial bus to achieve closed-loop management of the entire process from environmental perception, tobacco leaf status monitoring to intelligent control.

[0013] Among them, the tobacco leaf information acquisition unit is composed of a near-infrared spectrometer, a high-definition image acquisition system and a tobacco leaf physical and chemical parameter detection device, which is used to collect tobacco leaf near-infrared spectrum data, image data and physical and chemical parameters; the data processing unit adopts a distributed architecture design, including edge computing nodes and a central data processing server, which is used to run the tobacco leaf status perception model and the three-dimensional environmental parameter model of the flue-curing room.

[0014] Among them, the internal iterative optimization processing specifically refers to the calculation process of cyclically updating the weight contribution rate value in the first optimization vector, including extracting the initial weight contribution rate vector from the historical baking data as the iteration starting point, calculating the weight deviation value between the current weight contribution rate and the target weight contribution rate, updating the value of each element in the weight contribution rate vector based on the gradient descent algorithm, and judging whether the weight deviation value is less than the preset convergence threshold 0.001.

[0015] Among them, the process effect contribution rate specifically refers to the quantitative indicator of the degree of influence of each parameter in the basic process parameter set on the final quality score of tobacco leaves. The weight coefficient of each parameter is calculated through multiple regression analysis to guide the priority sorting of parameter adjustment and establish a two-way evaluation mechanism of positive contribution rate and negative impact rate.

[0016] Among them, the contribution rate of secondary changes in process effects specifically refers to the quantitative indicator of the impact of the interaction between basic process parameters on tobacco leaf quality. By analyzing the synergistic and antagonistic effects between parameters, the comprehensive contribution of parameter combinations to quality changes is calculated, which is used to achieve multi-parameter coordinated optimization and establish a dynamic compensation mechanism for interactive effects.

[0017] Among them, the gating weight function specifically refers to an adaptive calculation function used to dynamically adjust the feature fusion weights of different levels in the intelligent process adaptation model. The gating weight function is based on three input data: the current baking stage identifier, the tobacco leaf maturity distribution vector, and the quality target deviation value, to obtain a balance value between 0 and 1. The segmented weight adjustment mechanism is used to achieve the adaptive response of the model under different baking conditions.

[0018] Among them, the hierarchical fusion weight specifically refers to the numerical parameter that represents the proportion of each level of features in the final output of the intelligent process adaptation model. It is dynamically calculated and generated by the gated weight function based on the current baking state. It is used to control the influence of information at different abstract levels on the final decision, thereby realizing intelligent and precise adjustment of baking process parameters.

[0019] The method and system of the present invention achieve intelligent and precise adjustment of baking process parameters by establishing a dual optimization vector system and a multi-level intelligent decision-making mechanism. The method constructs a basic parameter set containing six core process parameters, adopts a data acquisition strategy with multiple time granularities, and establishes an accurate quantitative representation of the process parameter status and tobacco leaf quality changes. Through the synergistic effect of the first optimization vector and the second optimization vector, the present invention overcomes the defects of extensive parameter adjustment and lack of intelligent decision-making in traditional technologies. The intelligent process adaptation model adopts a hierarchical attention network architecture, which can automatically identify the complex correlation between different parameters. The gated weight function dynamically adjusts the parameter weight according to the current baking state, achieving a technological leap from extensive adjustment to precise control. The stable change threshold judgment mechanism and internal iterative optimization processing ensure the accuracy and stability of parameter adjustment. It solves the technical problem of intelligent and precise adjustment of baking process parameters and realizes the transformation from traditional empirical extensive adjustment to intelligent and precise control. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the method of the present invention.

[0021] Figure 2 Schematic diagram of the composition of the system of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, 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.

[0023] like Figure 1 As shown, the first aspect of the present invention provides an intelligent and precise baking method for improving tobacco leaf quality, comprising the following steps:

[0024] S01. Establish a basic process parameter set, wherein the basic process parameter set includes six core process parameters: dry-bulb temperature, wet-bulb temperature, wind speed, exhaust flow rate, heating rate, and temperature stabilization time. The basic process parameter set is divided into three classification ranges according to the maturity of tobacco leaves: under-mature leaf parameter domain, moderately mature leaf parameter domain, and over-mature leaf parameter domain;

[0025] S02. Real-time collection of the basic process parameter set at a first time granularity is performed to record numerical changes of the six core process parameters during the baking process, and a stratified collection strategy is established for high-quality tobacco leaf segments, medium-quality tobacco leaf segments, and low-quality tobacco leaf segments based on differences in tobacco leaf varieties and locations;

[0026] S03. Constructing a first optimization vector based on the second time granularity, where the first optimization vector is used to represent the parameter combination state of the basic process parameter set in different baking stages, and determining the weight contribution rate and stage transition rate of each parameter domain through an intelligent process adaptation model;

[0027] S04. Obtain tobacco leaf quality change data through machine vision and near-infrared spectroscopy fusion technology, construct a second optimization vector for quantifying real-time changes in tobacco leaf moisture content, chlorophyll content, and total sugar content, and establish a quality response delay rate assessment mechanism;

[0028] S05. Establish a stable change threshold determination mechanism, and initiate process parameter adjustment when the numerical fluctuations of the first optimization vector and the second optimization vector exceed the stable change threshold, using the gating weight function of the intelligent process adaptation model to perform adaptive adjustment;

[0029] S06, performing internal iterative optimization processing on the first optimization vector, and updating the weight contribution rate value through iterative calculation, wherein the internal iterative optimization processing includes four sub-steps: initializing the weight contribution rate vector, calculating the weight deviation value, updating the weight contribution rate vector, and determining the convergence condition, until the weight deviation value is less than the convergence threshold;

[0030] S07. Calculate the process effect contribution rate, quantify the influence weight of each parameter in the basic process parameter set on the final quality of tobacco leaves, generate a parameter adjustment strategy, and establish a collaborative conversion mechanism between different parameter domains;

[0031] S08. Perform fine-tuning based on the secondary change contribution rate of the process effect, achieve intelligent optimization of the baking process by analyzing the interaction between parameters, and dynamically balance the influence of each classification range using the hierarchical fusion weight of the intelligent process adaptation model.

[0032] Among them, the basic process parameter set specifically refers to a collection of multiple core control parameters that affect the quality changes of tobacco leaves during the baking process, including the dry-bulb temperature and wet-bulb temperature for controlling the temperature environment of the baking room, the wind speed for adjusting the air flow, the exhaust flow for controlling the gas exchange in the baking room, the heating rate for determining the temperature change rate, and the steady temperature time for maintaining a constant temperature state. The under-mature leaf parameter domain corresponds to the low-temperature and slow baking strategy, which accounts for 30% of the overall plan, the moderately mature leaf parameter domain corresponds to the standard temperature control strategy, which accounts for 55%, and the over-mature leaf parameter domain corresponds to the high-temperature and fast strategy, which accounts for 15%.

[0033] Among them, the first time granularity specifically refers to the time interval for data collection of the basic process parameter set, which is set to 30 seconds, and is used to capture the rapid changes and instantaneous fluctuations of process parameters during the baking process. At the same time, differentiated collection frequencies are established for different tobacco leaf segments. The high-quality tobacco leaf segment adopts high-frequency collection to ensure accuracy, the medium-quality tobacco leaf segment adopts standard frequency to balance efficiency and accuracy, and the low-quality tobacco leaf segment adopts low-frequency collection to reduce the computing load.

[0034] Among them, the second time granularity specifically refers to the time statistical window used when constructing the optimization vector, which is set to 10 minutes. It is used to perform statistical analysis and trend extraction within the time window on the data collected at the first time granularity, and establish a comprehensive evaluation system for short-term fluctuation characteristics and long-term trend characteristics through multi-time scale analysis.

[0035] Among them, the first optimization vector specifically refers to an 18-dimensional vector formed by combining the mean, variance, and change rate of the basic process parameter set within the second time granularity, which is used to characterize the comprehensive state characteristics of the process parameters within the time window, quantify the influence of each parameter on the baking effect through the weight contribution rate, and establish the stage transition rate mechanism between parameter domains to achieve dynamic switching.

[0036] Among them, the second optimization vector specifically refers to the tobacco leaf quality parameter vector obtained through sensor fusion technology, which includes three dimensions: tobacco leaf moisture content change rate, chlorophyll degradation rate, and total sugar conversion efficiency. It is used to quantify the physiological and biochemical changes of tobacco leaves and establish the quality response delay rate evaluation system to predict the time lag effect of quality changes.

[0037] Among them, the stable change threshold specifically refers to the numerical boundary for determining whether the process parameters need to be adjusted. When the change rate of any parameter in the first optimization vector exceeds 15% of its historical average or the quality parameter in the second optimization vector deviates from the target value by more than 8%, the parameter adjustment mechanism is triggered, and the risk of misadjustment is reduced by establishing a threshold grading system.

[0038] Among them, the internal iterative optimization processing specifically refers to the calculation process of cyclically updating the weight contribution rate value in the first optimization vector, including extracting the initial weight contribution rate vector from the historical baking data as the iteration starting point, calculating the weight deviation value between the current weight contribution rate and the target weight contribution rate, updating the value of each element in the weight contribution rate vector based on the gradient descent algorithm, and judging whether the weight deviation value is less than the preset convergence threshold of 0.001. If the convergence condition is not met, returning to the weight deviation value calculation step to continue the iteration; if the convergence condition is met, outputting the optimized weight contribution rate vector for subsequent process parameter adjustment.

[0039] Among them, the process effect contribution rate specifically refers to the quantitative indicator of the degree of influence of each parameter in the basic process parameter set on the final quality score of tobacco leaves. The weight coefficient of each parameter is calculated through multivariate regression analysis to guide the priority sorting of parameter adjustment and establish a two-way evaluation mechanism of positive contribution rate and negative impact rate.

[0040] Among them, the secondary change contribution rate of the process effect specifically refers to the quantitative indicator of the impact of the interaction between the basic process parameters on the quality of tobacco leaves. By analyzing the synergistic and antagonistic effects between the parameters, the comprehensive contribution of the parameter combination to the quality change is calculated, which is used to achieve multi-parameter coordinated optimization and establish a dynamic compensation mechanism for interactive effects.

[0041] Among them, the weight contribution rate specifically refers to the numerical representation of the influence of different parameter domains on the final baking effect. Through statistical analysis of historical data, it is concluded that the contribution rate of the under-ripe leaf parameter domain is 0.25, the contribution rate of the moderately ripe leaf parameter domain is 0.60, and the contribution rate of the over-ripe leaf parameter domain is 0.15, which is used to dynamically adjust the weight distribution of each parameter domain in process decision-making.

[0042] Among them, the stage transition rate specifically refers to the frequency and speed indicators of switching from one parameter domain to another during the baking process. The timing of the transition is determined by monitoring the changes in the state of tobacco leaves, and a smooth transition mechanism is established to avoid the impact of parameter jumps on tobacco leaf quality.

[0043] Among them, the quality response delay rate specifically refers to the time lag in the response change of tobacco leaf quality indicators after the process parameters are adjusted. By establishing a time series correlation model to predict the time when the adjustment effect will appear, it is used to optimize the timing selection and intensity control of parameter adjustment.

[0044] Among them, the under-mature leaf parameter domain specifically refers to a low-temperature and long-time baking parameter combination formulated for tobacco leaves with insufficient maturity, including a dry-bulb temperature of 35 to 42°C, a wet-bulb temperature of 32 to 38°C, a wind speed of 0.8 to 1.2 meters per second, and a heating rate of 0.5°C per hour, accounting for 30% of the overall baking strategy, and compensating for the negative impact of insufficient maturity on quality by extending the baking time.

[0045] Among them, the parameter domain of appropriately mature leaves specifically refers to a standard temperature control parameter combination formulated for tobacco leaves with moderate maturity, including a dry-bulb temperature of 42 to 48°C, a wet-bulb temperature of 38 to 42°C, a wind speed of 1.2 to 1.8 meters per second, and a heating rate of 1.0°C per hour, which accounts for 55% of the overall baking strategy and ensures stable baking quality through standardized process flow.

[0046] Among them, the over-mature leaf parameter domain specifically refers to a high-temperature, short-time baking parameter combination formulated for over-mature tobacco leaves, including a dry-bulb temperature of 48 to 54°C, a wet-bulb temperature of 42 to 46°C, a wind speed of 1.8 to 2.5 meters per second, and a heating rate of 1.5°C per hour, accounting for 15% of the overall baking strategy, and avoiding further deterioration of the quality of over-mature leaves through rapid dehydration.

[0047] Among them, the high-grade tobacco section specifically refers to the tobacco area with superior quality grade. It adopts high-precision parameter control and intensive monitoring strategy, and the data collection frequency is 1.5 times the standard frequency. Through refined management, the stability and consistency of the quality of high-grade tobacco leaves are ensured.

[0048] Among them, the medium tobacco section specifically refers to the tobacco area with medium quality grade. Standard parameter control and conventional monitoring strategies are adopted, and the data collection frequency is the standard set value. By balancing cost and effect, stable baking of medium-grade tobacco is achieved.

[0049] Among them, the inferior tobacco leaf section specifically refers to the tobacco leaf area with inferior quality grade. It adopts extensive parameter control and low-frequency monitoring strategy, and the data collection frequency is 0.7 times the standard frequency, which reduces the monitoring cost while ensuring the realization of basic baking requirements.

[0050] Among them, the weight deviation value specifically refers to the sum of the squares of the differences between each element in the weight contribution rate vector and the corresponding elements of the target weight contribution rate vector, which is used to measure the degree of deviation between the current weight contribution rate vector and the optimal weight contribution rate vector, and serves as the basis for judging convergence in the internal iterative optimization process.

[0051] The convergence threshold specifically refers to the numerical standard used to judge the iterative termination condition in the internal iterative optimization process, which is set to 0.001. When the weight deviation value is less than the convergence threshold, the iteration is considered to have converged and the final optimization result is output.

[0052] The structure of the intelligent process adaptation model is a multi-layer perception network based on a hierarchical attention network architecture, which includes five main parts: an input layer, a feature extraction layer, an attention calculation layer, a gated fusion layer, and an output layer. The feature extraction layer uses a residual connection structure to process the 18-dimensional first optimization vector and the 3-dimensional second optimization vector. The attention calculation layer calculates the association weights between different parameter domains through a multi-head attention mechanism. The gated fusion layer dynamically adjusts the fusion ratio of features at each level based on the gated weight function. The output layer generates process parameter adjustment instructions and quality prediction results. The total number of model parameters is approximately 500,000 training parameters. The hierarchical fusion weight parameters adjusted during inference are dynamically set according to the current baking stage, tobacco leaf maturity distribution ratio, and quality target requirements.

[0053] The steps of establishing the training data set of the intelligent process adaptation model include collecting baking history data from the four major tobacco-growing areas in Yunnan in the past three years, totaling 150,000 batches of baking records. Each batch contains complete process parameter time series data, tobacco leaf quality test results and expert evaluation standards. Stratified sampling is performed according to tobacco leaf maturity, variety type, and baking season to construct a data division of 120,000 batches of training sets, 20,000 batches of validation sets, and 10,000 batches of test sets. The original data is preprocessed including outlier detection, missing value filling, data standardization and feature engineering. A multi-label classification system is established covering three output targets: baking process type, quality grade prediction, and parameter adjustment suggestion. The training data is expanded to a sample size of 200,000 through data enhancement techniques including time series perturbation, parameter noise addition, and quality label smoothing method.

[0054] The steps of training the intelligent process adaptation model include adopting a multi-stage progressive training strategy. First, the basic feature learning stage is carried out to train the feature extraction layer and the attention calculation layer using the standard cross entropy loss function for a total of 100 training cycles. Then, the attention weight optimization stage is entered to introduce the attention regularization loss and the hierarchical consistency loss function to fine-tune the attention mechanism parameters for 50 cycles. Finally, the gated fusion optimization stage is carried out to train the gated weight function using a reinforcement learning strategy. The gate parameters are optimized by the policy gradient method to maximize the consistency between the model output and the expert evaluation. The Adam optimizer is used in the entire training process. The learning rate is decayed from 0.001 to 0.0001 using cosine annealing scheduling. The batch size is set to 64. The early stopping mechanism is used to avoid overfitting. Finally, the baking process classification accuracy of the model on the test set reaches 94.2%, the quality grade prediction accuracy reaches 91.8%, and the consistency of the parameter adjustment recommendations with the expert decision reaches 89.5%.

[0055] Among them, the gating weight function specifically refers to an adaptive calculation function used to dynamically adjust the fusion weights of different levels of features in the intelligent process adaptation model. The gating weight function is based on three input data: the current baking stage identifier, the tobacco leaf maturity distribution vector, and the quality target deviation value to calculate a balance value between 0 and 1. When the balance value is in the range of 0 to 0.423, a conservative weight adjustment function is used to enhance the weight of the stability feature. When the balance value is in the range of 0.423 to 0.628, a balanced weight adjustment function is used to evenly distribute the weights of features at each level. When the balance value is in the range of 0.628 to 1.0, an aggressive weight adjustment function is used to enhance the weight of the sensitivity feature. The adaptive response of the model under different baking conditions is achieved through a segmented weight adjustment mechanism.

[0056] Among them, the hierarchical fusion weight specifically refers to the numerical parameter of the proportion of each level feature in the final output in the intelligent process adaptation model, which is dynamically calculated and generated by the gating weight function according to the current baking state, and is used to control the influence of information at different abstract levels on the final decision.

[0057] Among them, the balance value specifically refers to the intermediate result calculated by the gated weight function based on the input parameters, which is used to determine the basis for determining which weight adjustment function to adopt. The balance value comes from the weighted calculation of the current baking stage identifier, tobacco leaf maturity distribution vector, and quality target deviation value.

[0058] Among them, the conservative weight adjustment function specifically refers to the weight calculation method used when the balance value is in the range of 0 to 0.423. It reduces the risk of parameter fluctuations by enhancing the weight of the stability feature and is suitable for high-grade tobacco leaf processing scenarios.

[0059] The balanced weight adjustment function specifically refers to a weight calculation method used when the balance value is in the range of 0.423 to 0.628. It maintains a standard processing flow by evenly distributing feature weights at each level, and is suitable for conventional baking scenarios.

[0060] Among them, the radical weight adjustment function specifically refers to the weight calculation method used when the balance value is in the range of 0.628 to 1.0. It speeds up the response speed by increasing the weight of the sensitivity feature and is suitable for emergency adjustment scenarios with rapid quality deterioration.

[0061] A second aspect of the present invention provides an intelligent precision tobacco leaf curing system for improving tobacco leaf quality. It utilizes an advanced multi-unit integrated design architecture and consists of five core components: a flue-curing barn unit, an environmental sensing unit, a tobacco leaf information collection unit, a data processing unit, and a control execution unit. The flue-curing barn unit utilizes a double-layered insulated wall design, with the interior space rationally divided into three functional zones: a heating zone, a yellowing zone, and a drying zone. A ventilation and exhaust system is installed at the top, and a steam supply system is located at the bottom. The environmental sensing unit utilizes a comprehensive environmental monitoring system utilizing a network of temperature and humidity sensors, a network of wind speed sensors, and light sensors. The tobacco leaf information collection unit integrates a near-infrared spectrometer, a high-definition image acquisition system, and a tobacco leaf physical and chemical parameter detection device to enable real-time monitoring of tobacco leaf conditions. The data processing unit utilizes a distributed architecture, comprising edge computing nodes and a central data processing server, running an intelligent process adaptation model for data analysis and decision-making. The control execution unit includes a heat source control system, a humidity control system, and an airflow control system, all coordinated through an industrial bus. The entire system adopts a modular design concept, and each functional module can be flexibly configured according to actual needs, realizing closed-loop management of the entire process from environmental perception, tobacco leaf status monitoring to intelligent control, providing a complete technical solution for precise tobacco leaf baking.

[0062] The specific implementation of the above steps is described in detail below.

[0063] The specific implementation method of step S01 is to establish a basic process parameter set containing multiple core process parameters. This step classifies and clusters historical baking data through data statistical analysis methods. First, key parameter data such as dry-bulb temperature, wet-bulb temperature, wind speed, exhaust flow, heating rate, and stabilization time in the baking process are collected. Then, the parameter space is divided into three different parameter domains according to the maturity characteristics of the tobacco leaves. The under-mature leaf parameter domain adopts a low-temperature and slow strategy, the dry-bulb temperature is controlled in the range of 35 to 42°C, and the heating rate is set to 0.5°C per hour. This parameter domain accounts for 30% of the overall baking plan. The mature leaf parameter domain adopts a standard temperature control strategy, the dry-bulb temperature range is 42 to 48°C, the heating rate is 1.0°C per hour, accounting for 55%, and the over-mature leaf parameter domain adopts a high-temperature and fast strategy, the dry-bulb temperature is 48 to 54°C, the heating rate reaches 1.5°C per hour, accounting for 15%. This classification strategy realizes automatic division of the parameter domain based on the k-means clustering algorithm, and achieves accurate classification by calculating the parameter response characteristics of tobacco leaves with different maturity levels.

[0064] The specific implementation method of step S02 is to establish a multi-level data acquisition system. This step adopts real-time data acquisition technology, sets the first time granularity to 30 seconds, and continuously monitors the basic process parameter set through the sensor network. During the data acquisition process, a layered acquisition strategy is established according to the differences in tobacco leaf varieties and parts. High-quality tobacco leaf sections adopt high-precision control, and the data acquisition frequency is set to 1.5 times the standard frequency to ensure the quality stability of high-grade tobacco leaves. The medium-quality tobacco leaf sections adopt the standard acquisition frequency to balance the monitoring cost and effect. The low-frequency acquisition is adopted for the low-quality tobacco leaf sections, and the frequency is 0.7 times the standard value. Resource allocation is optimized through differentiated acquisition strategies. The acquisition system is based on a distributed sensor network architecture and adopts a time synchronization protocol to ensure the time consistency of multi-point data.

[0065] The specific implementation method of step S03 is to construct a first optimization vector and calculate the weight contribution rate. This step sets the second time granularity to 10 minutes, and performs statistical analysis on the data collected for 30 seconds within this time window. The 18-dimensional first optimization vector is formed by calculating the mean, variance, and change rate of the six core parameters. The intelligent process adaptation model calculates the weight contribution rate of each parameter domain based on the multi-layer perception network architecture, where the weight of the under-mature leaf parameter domain is 0.25, the weight of the mature leaf parameter domain is 0.60, and the weight of the over-mature leaf parameter domain is 0.15. The stage transition rate determines the timing of parameter domain switching by monitoring changes in the tobacco leaf state. The process uses a sliding window statistical method to extract short-term fluctuations and long-term trend characteristics, and fuses multi-time scale information through a weighted average algorithm.

[0066] The specific implementation method of step S04 is to obtain tobacco leaf quality data through multi-sensor fusion technology. This step integrates a machine vision system and near-infrared spectroscopy analysis equipment. The machine vision system uses an image processing algorithm to analyze the color changes and morphological characteristics of the tobacco leaf surface. The near-infrared spectroscopy technology measures the internal chemical component content of the tobacco leaf based on the Lambert-Beer law. The sensor data fusion algorithm is used to construct a three-dimensional second optimization vector including the moisture content change rate, chlorophyll degradation rate, and total sugar conversion efficiency. The quality response delay rate evaluation mechanism predicts the time lag of quality changes after process parameter adjustment through time series correlation analysis. This mechanism establishes a time correlation model between parameter adjustment and quality response based on the autoregressive moving average model. The delay time prediction accuracy can reach more than 85%.

[0067] The specific implementation method of step S05 is to establish a threshold judgment and adaptive adjustment mechanism. This step sets the stable change threshold as the parameter change rate in the first optimization vector exceeds the historical average by 15% or the quality parameter in the second optimization vector deviates from the target value by more than 8%. When the parameter fluctuation is detected to exceed the threshold, the gated weight function of the intelligent process adaptation model is started. The function calculates the balance value based on the current baking stage identifier, the tobacco leaf maturity distribution vector, and the quality target deviation value. When the balance value is in the range of 0 to 0.423, a conservative weight adjustment function is used. In the range of 0.423 to 0.628, a balanced adjustment function is used. In the range of 0.628 to 1.0, an aggressive adjustment function is used. Accurate parameter adjustment triggering is achieved through the segmented threshold judgment mechanism to avoid interference with the baking process caused by frequent misadjustments.

[0068] The specific implementation method of step S06 is to perform iterative optimization processing on the first optimization vector. This step realizes the adaptive update of the weight contribution rate based on the gradient descent algorithm. First, the initial weight contribution rate vector is extracted from the historical data as the starting point of the iteration, and then the deviation value between the current weight and the target weight is calculated. The deviation value is quantified by the square sum calculation method. The value of each element in the weight vector is updated based on the back propagation algorithm, and it is judged whether the weight deviation value is less than the convergence threshold of 0.001. If the convergence condition is not met, the iterative calculation continues. The iterative process adopts an adaptive learning rate adjustment strategy. The initial learning rate is set to 0.01. The momentum optimization algorithm is used to accelerate the convergence process. The maximum number of iterations is limited to 1000 times to avoid infinite loops.

[0069] The specific implementation method of step S07 is to calculate the contribution rate of the process effect and generate an adjustment strategy. This step uses a multivariate linear regression analysis method to quantify the influence weight of each parameter in the basic process parameter set on the final quality of tobacco leaves. By establishing a mathematical relationship model between the quality score and the process parameters, the regression coefficient of each parameter is calculated as a contribution rate indicator, and a two-way evaluation system of positive contribution rate and negative impact rate is established. The parameter adjustment strategy determines the adjustment priority based on the contribution rate. The collaborative conversion mechanism establishes linkage adjustment rules through correlation analysis between parameters. This process uses the principal component analysis method to reduce the collinearity effect between parameters, and verifies the statistical significance of the contribution rate of each parameter through variance analysis.

[0070] The specific implementation method of step S08 is to make fine adjustments based on the contribution rate of secondary changes. This step realizes multi-parameter coordinated optimization by analyzing the interaction between basic process parameters, adopts the factorial experimental design method to analyze the synergistic and antagonistic effects between parameters, and calculates the contribution of second-order interaction terms to quality changes. The hierarchical fusion weights of the intelligent process adaptation model are dynamically adjusted according to the current baking stage, maturity distribution and quality goals. The adjustment process is based on the reinforcement learning algorithm and optimizes the weight allocation strategy through the policy gradient method to maximize the consistency between the model output and the expert evaluation. The dynamic compensation mechanism of the interactive impact automatically adjusts the compensation coefficient to maintain system stability by real-time monitoring of the interaction intensity between parameters.

[0071] The intelligent process adaptation model utilizes a multi-layer perception network architecture based on a hierarchical attention network. It consists of five main components: an input layer, a feature extraction layer, an attention computation layer, a gated fusion layer, and an output layer. The input layer receives an 18-dimensional first optimization vector and a 3-dimensional second optimization vector. The feature extraction layer processes the input data using a residual connection structure. Deep feature representations are extracted via a three-layer fully connected neural network, each layer containing 256 neurons and using a rectified linear unit activation function. The attention computation layer uses a multi-head attention mechanism to calculate the correlation weights between different parameter domains. It comprises eight attention heads, each with a dimension of 64, and captures long-range dependencies between parameters through a self-attention mechanism. The gated fusion layer dynamically adjusts the fusion ratio of features at each level based on a gated weight function. It uses a gated recurrent unit structure to control information flow, and the gate parameters are calculated using a sigmoid function to obtain weights between 0 and 1. The output layer generates process parameter adjustment instructions and quality predictions. The output branches include parameter adjustment suggestions, quality grade predictions, and baking process classification. The model has approximately 500,000 parameters in total.

[0072] The training dataset was first constructed by collecting historical data from 150,000 batches of tobacco curing from Yunnan's four major tobacco regions over the past three years. Each batch included complete time-series data on process parameters, tobacco leaf quality test results, and expert evaluation criteria. During data preprocessing, an outlier detection algorithm was used to identify and remove anomalous data points, missing values were filled using mean interpolation, and the data was normalized using Z-score normalization. Feature engineering extracted time-series, statistical, and frequency-domain features, and principal component analysis was used to reduce the dimensionality to key feature dimensions. Stratified sampling was performed based on leaf maturity, variety, and curing season to construct a training set of 120,000 batches, a validation set of 20,000 batches, and a test set of 10,000 batches. A multi-label classification system was developed, covering three output objectives: curing process type, quality grade prediction, and parameter adjustment recommendations. Data augmentation techniques, including time-series perturbation, Gaussian noise addition, and label smoothing, were used to expand the training data to 200,000 samples. The final model achieved 94.2% accuracy in curing process classification, 91.8% in quality grade prediction, and 89.5% consistency between parameter adjustment recommendations and expert decisions on the test set.

[0073] It should be noted that the first key technical idea of the present invention is to establish a multi-dimensional optimization vector system with dual time granularity. By combining real-time data collection at the first time granularity with statistical analysis at the second time granularity, a collaborative monitoring mechanism for the first optimization vector and the second optimization vector is constructed. Compared with the single-time-scale parameter monitoring method in traditional baking methods, this technical idea can simultaneously capture the instantaneous fluctuation characteristics and long-term trend changes in the baking process. Through multi-time-scale information fusion, it achieves a precise quantitative description of the tobacco leaf quality change process, effectively solving the technical defects of traditional methods such as strong parameter monitoring lag and insufficient response accuracy, and providing a reliable data foundation for achieving precise control.

[0074] The second key technical idea is a parameter domain classification control strategy based on differences in tobacco leaf maturity. This strategy divides the basic process parameter set into three parameter domains: under-mature leaves, moderately mature leaves, and over-mature leaves, and establishes a corresponding weight contribution rate mechanism and stage transition rate control system. Traditional baking processes usually use a unified parameter setting method, ignoring the differentiated requirements of tobacco leaves of different maturity for process parameters. However, this invention realizes a personalized baking strategy through maturity classification, and can dynamically select the most appropriate parameter combination according to the actual state of the tobacco leaves, significantly improving the adaptability and pertinence of the baking process, and solving the problem of uneven quality caused by the one-size-fits-all treatment in traditional methods.

[0075] The third key technical approach is the adaptive adjustment mechanism of the gating weight function within the intelligent process adaptation model. This mechanism calculates the equilibrium value based on the current baking state and selects the appropriate weight adjustment function type based on the equilibrium value range. Compared with traditional fixed parameter control methods, this gating mechanism automatically adjusts the fusion weights of the features at each level within the model according to the real-time baking conditions, implementing a dynamic weight allocation strategy from conservative to aggressive. This effectively addresses the single parameter adjustment strategy and poor adaptability of traditional control methods, and improves the intelligence and control precision of the baking process through adaptive weight adjustment.

[0076] The synergy of these three key technical ideas forms a complete intelligent and precise baking control system. The dual-time granularity optimization vector provides accurate data support for parameter domain classification control. The parameter domain classification strategy provides a decision-making basis for the adaptive adjustment of the gating weight function. The gating weight function organically integrates the advantages of the first two technical ideas and outputs the optimal control instructions. Compared with the technical architecture of the traditional baking method in which each control link is relatively independent and lacks effective coordination, the present invention realizes the integrated control of data collection, classification decision-making, and intelligent adjustment through the deep integration of three technical ideas, forming a closed-loop feedback optimization mechanism, which can automatically adjust the control strategy according to the changes in the state of the tobacco leaves, and significantly improves the overall stability and quality consistency of the baking process.

[0077] The second aspect of the present invention provides an intelligent precision baking system for improving the quality of tobacco leaves. The system mainly consists of five parts: a baking room structure unit, an environmental perception unit, a tobacco leaf information collection unit, a data processing unit and a control execution unit.

[0078] The curing room structure unit adopts a double-layer insulation wall design, with a 15cm thick outer layer of reinforced concrete and an 8cm thick inner layer of polyurethane insulation material, which can effectively isolate the impact of external environmental fluctuations on the interior of the curing room. The overall dimensions of the curing room are 10m×6m×4m, and the interior space is divided into three functional areas: the heating zone, the yellowing zone, and the drying zone. The ventilation and exhaust system is installed on the top of the curing room, consisting of a 30cm diameter stainless steel exhaust duct and a variable frequency exhaust fan with a maximum air volume of 3000m 3 / h, and the exhaust volume can be automatically adjusted according to the needs of the baking stage. A steam supply system is installed at the bottom of the curing room, including a steam generator and a uniformly distributed steam pipe network. The steam generator has a power of 15kW and a maximum steam output of 20kg / h. The steam pipe network uses 2cm diameter food-grade silicone tubing and forms a grid-like distribution within the curing room, with spacing of 1m. Inside the curing room, there are multi-layer tobacco leaf hanging racks installed. Made of 304 stainless steel, each rack is 1.8m high and 5m wide. There are 8 layers in total, and each layer can hold approximately 200kg of tobacco leaves.

[0079] The environmental sensing unit consists of a network of temperature and humidity sensors, a network of wind speed sensors, and light sensors installed within the flue-curing barn. The temperature and humidity sensors are SHT35 models, with a temperature measurement range of 0–100°C (with an accuracy of ±0.1°C) and a humidity measurement range of 0–100% RH (with an accuracy of ±1% RH). Twenty-four sensing points are installed in an 8×3 grid within the flue-curing barn. The wind speed sensors are thermally sensitive, with a measurement range of 0–5 m / s (with an accuracy of ±0.05 m / s) and are co-located with the temperature and humidity sensors. Eight light sensors are installed around the flue-curing barn walls to monitor ambient light conditions during tobacco leaf color changes. They have a measurement range of 0–100,000 lx (with an accuracy of ±100 lx). All sensors are waterproof and dustproof, with ABS engineering plastic housings and an IP65 rating, ensuring long-term stable operation in the high-humidity flue-curing barn environment. The sensors are connected to the data acquisition module via an RS485 bus, and the sampling frequency can be set from 10 seconds to 60 minutes.

[0080] The tobacco leaf information collection unit consists of a near-infrared spectrometer, a high-definition image acquisition system, and tobacco leaf physical and chemical parameter detection equipment. The near-infrared spectrometer utilizes a fiber optic array design with a wavelength range of 800 to 2500 nm, a wavelength resolution of 2 nm, and a signal-to-noise ratio greater than 10,000:1. Equipped with 16 measurement probes, it enables non-contact spectral acquisition of tobacco leaves from multiple locations within the flue-curing barn. The high-definition image acquisition system includes eight high-definition cameras with a resolution of 3840 × 2160 pixels and a frame rate of 30 fps. These cameras are equipped with ring-shaped LED fill lights, a color temperature of 5600K, and a color rendering index greater than 95, ensuring clear images of tobacco leaves from all angles within the flue-curing barn. The tobacco leaf physical and chemical parameter detection equipment includes an online moisture meter, a chlorophyll fluorescence meter, and a rapid total sugar content analyzer for real-time monitoring of changes in tobacco leaf moisture content, chlorophyll content, and total sugar content, respectively. The moisture detector adopts a capacitive design with a measurement range of 5-90% and an accuracy of ±0.5%; the chlorophyll fluorescence detector uses an LED excitation light source with a wavelength of 470nm and a detection accuracy of ±0.01mg / g; the total sugar content rapid analyzer is based on the near-infrared diffuse reflection principle, with a detection range of 0-50% and an accuracy of ±0.1%.

[0081] The data processing unit utilizes a distributed architecture, comprised of edge computing nodes and a central data processing server. The edge computing node, installed near the flue-curing barn, utilizes an industrial-grade computer equipped with an eight-core processor running at 3.6 GHz, 16 GB of memory, and a 512 GB solid-state drive. It is responsible for the initial processing and screening of on-site data from the flue-curing barn. The central data processing server, located in the control center, utilizes a dual-core server architecture and features a 64-core processor running at 2.9 GHz, 256 GB of memory, and 20 TB of storage. It also includes a high-performance graphics processing unit (GPU), which runs the tobacco leaf status perception model and the flue-curing barn's three-dimensional environmental parameter model. The two processing levels are connected via Gigabit Ethernet, and data transmission is encrypted using the AES-256 algorithm to ensure secure and reliable data transmission.

[0082] The control execution unit includes heat source control system, humidity control system and air flow control system. The heat source control system adopts a composite heating method of electric heating wire and steam heating, and the power density of the electric heating wire is 300W / m 2 , evenly distributed on the bottom and side walls of the drying room, achieving precise temperature control accuracy of ±0.5℃. The humidity control system consists of an ultrasonic humidifier and a dehumidifier. The humidifier has a misting capacity of 5L / h and the dehumidifier has a dehumidification capacity of 30L / day. Through micro-adjustment, the humidity control accuracy of ±2%RH is achieved. The airflow control system includes a variable frequency fan and air guide plate combination, with a maximum air volume of 5000m 3 / h. The air deflector is made of food-grade 304 stainless steel, 50cm long, 20cm wide, and 2mm thick, and can rotate 360 degrees to adjust the airflow direction. All actuators use industrial-grade variable frequency control technology, with a response time of less than 1 second, high control accuracy, and low energy consumption, meeting the precise control requirements of different baking stages.

[0083] These units are integrated into a complete intelligent precision tobacco curing system via an industrial bus, enabling closed-loop management of the entire process, from environmental sensing and tobacco leaf status monitoring to intelligent control. The system utilizes a modular design, allowing flexible configuration of functional modules based on the size of the flue-curing barn and tobacco leaf variety, ensuring strong adaptability and easy maintenance. The system consumes 25kW of power, averaging approximately 300kWh per day, achieving approximately 30% energy savings compared to traditional tobacco curing methods. Furthermore, the system improves tobacco leaf quality consistency by over 40%, effectively addressing the issues of unstable tobacco leaf quality and significant energy waste associated with traditional tobacco curing methods.

[0084] Specifically, the principle of the present invention is as follows: The key to solving the problem of intelligent and precise adjustment of baking process parameters lies in the construction of a multi-level intelligent parameter control system. First, a precise quantification mechanism for process parameters and tobacco leaf quality is established through dual optimization vectors. The first optimization vector combines the mean, variance, and change rate of six core process parameters into an 18-dimensional feature vector to comprehensively characterize the dynamic state of the process parameters. The second optimization vector uses machine vision and near-infrared spectroscopy fusion technology to obtain real-time data on changes in tobacco leaf moisture content, chlorophyll content, and total sugar content, achieving accurate perception and quantitative expression of changes in tobacco leaf quality.

[0085] Secondly, the intelligent process adaptation model achieves intelligent identification and weight calculation of parameter relationships through a hierarchical attention network architecture. The feature extraction layer uses a residual connection structure to process multidimensional optimization vectors. The attention calculation layer utilizes a multi-head attention mechanism to automatically learn complex correlation patterns between different parameter domains, capable of identifying synergistic and antagonistic effects between parameters. The gated fusion layer dynamically adjusts the fusion weights of features at each level based on a gated weight function. This function calculates the balance value based on the baking stage identifier, maturity distribution vector, and quality target deviation value. A segmented weight adjustment strategy is used to achieve adaptive response under different baking conditions.

[0086] Thirdly, the precise adjustment mechanism ensures accurate parameter control through multiple optimization algorithms. The internal iterative optimization process uses a gradient descent algorithm to cyclically update the weight contribution rate vector. By calculating the weight deviation and determining the convergence condition, the parameter weight distribution is continuously optimized until the preset accuracy requirement is achieved. The stable change threshold determination mechanism monitors the numerical fluctuations of the optimization vector and promptly triggers the adjustment mechanism when the parameter change rate or quality deviation exceeds the threshold, ensuring the timely and accurate adjustment.

[0087] Finally, the calculation of process effect contribution and secondary variation contribution established a quantitative assessment system for parameter influence. Multiple regression analysis was used to calculate the weight coefficient of each parameter on final quality. The combined contribution of parameter interactions to quality variation was analyzed, providing a scientific basis for intelligent decision-making. Through multi-level intelligent processing, the entire technical solution transforms the complex baking process into a precisely controllable intelligent decision-making sequence, thus enabling intelligent and precise adjustment of baking process parameters.

[0088] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0089] The specific implementation of step S01 is the same as above and will not be described in detail here.

[0090] In step S02, the calculation of the differential acquisition frequency is specifically expressed as follows:

[0091] fhigh =1.5·f standard ;

[0092] f middle =f standard ;

[0093] f low =0.7·f standard ;

[0094] Where, f high f is the sampling frequency of high-grade tobacco leaf segment; middle The sampling frequency for the medium tobacco leaf segment; f low The sampling frequency for the inferior tobacco leaf segment; f standard The standard collection frequency is set to once every 30 seconds.

[0095] In step S03, the construction of the first optimization vector is specifically represented as follows:

[0096]

[0097] Where V1 is the 18-dimensional first optimization vector; μ i is the mean value of the i-th process parameter within the second time granularity, where i = 1, 2, ..., 6 correspond to dry-bulb temperature, wet-bulb temperature, wind speed, exhaust flow, heating rate, and stabilization time, respectively; is the variance of the i-th process parameter in the time window; Δ i is the rate of change of the i-th process parameter.

[0098] The parameter acquisition method is:

[0099] μ i The calculation formula is obtained by taking the arithmetic average of the data collected in the first time granularity of 30 seconds within the second time granularity of 10 minutes: Where n is the number of sampling points in the time window, which is 20, x ij is the i-th parameter value of the j-th sampling point, j=1, 2, ..., n represents the sequence number of the sampling point in the time series. The variance calculation formula Obtain. Δ i Calculate the rate of change using the formula Get, where x i，start and x i，end are the parameter values at the start and end of the time window, t start and t end They are the start and end times of the time window, in seconds.

[0100] The calculation process of weighted contribution rate is described in detail as follows:

[0101] W=[w under , w proper , w over ] T ;

[0102] Where W is the 3D weight contribution rate vector; w under w is the weight contribution rate of the under-ripe leaf parameter domain, the default value is 0.25; proper w is the weight contribution rate of the mature leaf parameter domain, the default value is 0.60; over The parameter domain weight contribution rate of overmature leaves is 0.15 by default.

[0103] In step S04, the construction of the second optimization vector is specifically represented as follows:

[0104] V2=[γ w , γ c , γ s ] T ;

[0105] Where V2 is the second 3D optimization vector; γ w is the rate of change of moisture content of tobacco leaves, expressed in percentage per minute; γ c is the chlorophyll degradation rate, in milligrams per gram per minute; γ s is the total sugar conversion efficiency, dimensionless.

[0106] The parameter acquisition method is:

[0107] γ w The method adopts near-infrared spectroscopy analysis, including step 1: scanning the tobacco leaf sample in the wavelength range of 1400 to 1500 nm by a near-infrared spectrometer to obtain an absorbance spectrum; step 2: calculating the moisture content based on the Lambert-Beer law A=εbc, where A is the absorbance, dimensionless, ε is the molar absorptivity, unit is liter per mole per centimeter, b is the optical path length, unit is centimeter, and c is the concentration, unit is mole per liter; step 3: calculating the rate of change by time difference Among them, M t is the moisture content at time t, in percentage, M t+Δt is the moisture content at time t+Δt, Δt is the time interval in minutes. c The image is obtained by machine vision analysis, including step 1: collecting tobacco leaf surface images with a CCD camera; step 2: extracting the green channel intensity value X in the RGB color space, with a value range of 0 to 255; step 3: calculating the chlorophyll degradation rate. Where k is the degradation constant in milligrams per gram per minute, ranging from 0.01 to 0.05, G0 is the initial green intensity, G t is the green intensity at time t. γ sThe range is 0.02 to 0.15.

[0108] The calculation process of the quality response delay rate is described in detail as follows:

[0109] τ=α·e -β·T +γ delay H+δ;

[0110] Where τ is the quality response delay rate, in minutes; T is the current temperature, in degrees Celsius; H is the relative humidity, in percentage; α is the temperature sensitivity coefficient, dimensionless, ranging from 1.2 to 2.8; β is the temperature attenuation coefficient, in degrees Celsius, ranging from 0.03 to 0.08; γ delay is the humidity influence coefficient, expressed in minutes per percentage, ranging from 0.15 to 0.35; δ is the basic delay constant, expressed in minutes, ranging from 0.8 to 1.5.

[0111] The specific implementation of step S05 is the same as above, but the determination of the stable change threshold is specifically expressed as follows:

[0112]

[0113] Θ2={|γ j -γ j，target |>0.08·γ j，target |j=w,c,s};

[0114] Where, Θ1 is the first optimization vector threshold determination condition; Θ2 is the second optimization vector threshold determination condition; is the historical mean of the change rate of the i-th parameter, calculated by sliding window; γ j，target is the target value of the jth quality parameter, where the values of j, w, c, and s correspond to the subscript identifiers of moisture content, chlorophyll, and total sugar, respectively.

[0115] The calculation process of the balance value in the gated weight function is described in detail as follows:

[0116] ξ=ω1·S+ω2·M+ω3·D;

[0117] Where ξ is the equilibrium value, dimensionless, ranging from 0 to 1; S is the current baking stage identifier, dimensionless, ranging from 0 to 1, and is calculated by the ratio of baking time to total baking time; M is the weighted sum of tobacco leaf maturity distribution vectors, dimensionless, and is calculated by the weighted summation of the proportions of each maturity grade; D is the quality target deviation value, dimensionless, and is calculated by the relative deviation between the current quality and the target quality; ω1, ω2, and ω3 are weighting coefficients, dimensionless, with default values of 0.4, 0.35, and 0.25, respectively, satisfying ω1+ω2+ω3=1.

[0118] The mathematical expressions of the three weight adjustment functions are as follows:

[0119] Conservative weight adjustment function: F conservative (ξ)=0.8·e -2ξ +0.2, used when 0≤ξ≤0.423;

[0120] Balanced weight adjustment function: F balance (ξ)=1.0, adopted when 0.423<ξ≤0.628;

[0121] Aggressive weight adjustment function: F aggressive (ξ)=1.2·(1-e -3(ξ-0.628) )+1.0, adopted when 0.628<ξ≤1.0;

[0122] Where, F conservative (ξ), F balance (ξ), F aggressive (ξ) are the output values of the conservative, balanced, and radical weight adjustment functions, respectively. They are dimensionless and are used to adjust the weight distribution ratio within the model.

[0123] The calculation process of the phase transition rate is described in detail as follows:

[0124]

[0125] Where R transition is the phase transition rate, in times per hour; N switch is the number of parameter domain switching, dimensionless; N total is the total number of monitoring cycles, dimensionless; f speed is the transition speed adjustment factor, dimensionless, ranging from 0.5 to 2.0, used to adjust the sensitivity of the transition frequency.

[0126] In step S06, the calculation process of the weight deviation value is described in detail as follows:

[0127]

[0128] Where E is the weight deviation value, dimensionless; w k is the kth element of the current weight contribution rate vector, where k = 1, 2, and 3 correspond to the parameter domains of under-ripe leaves, moderately ripe leaves, and over-ripe leaves, respectively; w k，target The kth element of the target weight contribution rate vector.

[0129] The gradient descent algorithm for weight update is specifically expressed as follows:

[0130]

[0131] Where, is the kth weight value after the t+1th iteration; is the kth weight value at the tth iteration, where t represents the number of iterations, t = 0, 1, 2, ...; η is the learning rate, dimensionless, set to 0.01; is the partial derivative of the weight deviation value with respect to the kth weight, and the calculation formula is

[0132] In step S07, the multiple regression analysis of the process effect contribution rate is specifically expressed as follows:

[0133]

[0134] Where, Q is the final quality score of tobacco leaves, dimensionless, ranging from 0 to 100; P i is the i-th basic process parameter, and the unit depends on the parameter type; φ i is the process effect contribution coefficient of the i-th parameter, and its unit is the same as P i ∈1 is the regression error term, dimensionless, ranging from -0.5 to 0.5.

[0135] The parameter acquisition method is:

[0136] φ i The calculation formula is φ=(P T P) -1 P T Q, where P is the n×6 dimensional process parameter matrix, n is the number of samples, Q is the n×1 dimensional quality score vector, and φ is the 6×1 dimensional contribution rate coefficient vector. i Obtained through real-time measurement by sensors. ∈1 is calculated through residual analysis, and the residual calculation formula is:

[0137] In step S08, the calculation process of the secondary change contribution rate of the process effect is described in detail as follows:

[0138]

[0139] Where Q interaction is the effect of parameter interaction on quality, dimensionless; ψ ij is the interaction contribution coefficient between the i-th and j-th parameters, where i <j,i,j∈{1,2,...,6};P i and P j are the i-th and j-th process parameters respectively; ∈2 is the interaction error term, dimensionless, ranging from -0.3 to 0.3.

[0140] The parameter acquisition method is:

[0141] ψ ij Obtained through factorial experimental design and variance analysis, the calculation process includes steps 1: design 2 6 Full factorial experimental plan, a total of 64 experimental combinations; Step 2: Collect quality data under different parameter combinations; Step 3: Calculate the significance of the interaction effect through variance analysis and use the F test statistic for significance determination; Step 4: Determine ψ based on the significance test results ij The coefficients of significant interaction terms are calculated through regression analysis, and the coefficients of non-significant terms are set to 0. ∈2 is calculated through interaction residual analysis, and the calculation formula is

[0142] The following is a specific embodiment 2 of an intelligent precision baking system for improving tobacco leaf quality provided by the present invention: Figure 2 As shown in the figure, the system mainly consists of five parts: flue-curing room structure unit, environment perception unit, tobacco leaf information collection unit, data processing unit and control execution unit.

[0143] The curing room structure unit adopts a double-layer insulation wall design, with a 15cm thick outer layer of reinforced concrete and an 8cm thick inner layer of polyurethane insulation material, which can effectively isolate the impact of external environmental fluctuations on the interior of the curing room. The overall dimensions of the curing room are 10m×6m×4m, and the interior space is divided into three functional areas: the heating zone, the yellowing zone, and the drying zone. The ventilation and exhaust system is installed on the top of the curing room, consisting of a 30cm diameter stainless steel exhaust duct and a variable frequency exhaust fan with a maximum air volume of 3000m 3 / h, and the exhaust volume can be automatically adjusted according to the needs of the baking stage. A steam supply system is installed at the bottom of the curing room, including a steam generator and a uniformly distributed steam pipe network. The steam generator has a power of 15kW and a maximum steam output of 20kg / h. The steam pipe network uses 2cm diameter food-grade silicone tubing and forms a grid-like distribution within the curing room, with spacing of 1m. Inside the curing room, there are multi-layer tobacco leaf hanging racks installed. Made of 304 stainless steel, each rack is 1.8m high and 5m wide. There are 8 layers in total, and each layer can hold approximately 200kg of tobacco leaves.

[0144] The environmental sensing unit consists of a network of temperature and humidity sensors, a network of wind speed sensors, and light sensors installed within the flue-curing barn. The temperature and humidity sensors are SHT35 models, with a temperature measurement range of 0–100°C (with an accuracy of ±0.1°C) and a humidity measurement range of 0–100% RH (with an accuracy of ±1% RH). Twenty-four sensing points are installed in an 8×3 grid within the flue-curing barn. The wind speed sensors are thermally sensitive, with a measurement range of 0–5 m / s (with an accuracy of ±0.05 m / s) and are co-located with the temperature and humidity sensors. Eight light sensors are installed around the flue-curing barn walls to monitor ambient light conditions during tobacco leaf color changes. They have a measurement range of 0–100,000 lx (with an accuracy of ±100 lx). All sensors are waterproof and dustproof, with ABS engineering plastic housings and an IP65 rating, ensuring long-term stable operation in the high-humidity flue-curing barn environment. The sensors are connected to the data acquisition module via an RS485 bus, and the sampling frequency can be set from 10 seconds to 60 minutes.

[0145] The tobacco leaf information collection unit consists of a near-infrared spectrometer, a high-definition image acquisition system, and tobacco leaf physical and chemical parameter detection equipment. The near-infrared spectrometer utilizes a fiber optic array design with a wavelength range of 800 to 2500 nm, a wavelength resolution of 2 nm, and a signal-to-noise ratio greater than 10,000:1. Equipped with 16 measurement probes, it enables non-contact spectral acquisition of tobacco leaves from multiple locations within the flue-curing barn. The high-definition image acquisition system includes eight high-definition cameras with a resolution of 3840 × 2160 pixels and a frame rate of 30 fps. These cameras are equipped with ring-shaped LED fill lights, a color temperature of 5600K, and a color rendering index greater than 95, ensuring clear images of tobacco leaves from all angles within the flue-curing barn. The tobacco leaf physical and chemical parameter detection equipment includes an online moisture meter, a chlorophyll fluorescence meter, and a rapid total sugar content analyzer for real-time monitoring of changes in tobacco leaf moisture content, chlorophyll content, and total sugar content, respectively. The moisture detector adopts a capacitive design with a measurement range of 5-90% and an accuracy of ±0.5%; the chlorophyll fluorescence detector uses an LED excitation light source with a wavelength of 470nm and a detection accuracy of ±0.01mg / g; the total sugar content rapid analyzer is based on the near-infrared diffuse reflection principle, with a detection range of 0-50% and an accuracy of ±0.1%.

[0146] The data processing unit utilizes a distributed architecture, comprised of edge computing nodes and a central data processing server. The edge computing node, installed near the flue-curing barn, utilizes an industrial-grade computer equipped with an eight-core processor running at 3.6 GHz, 16 GB of memory, and a 512 GB solid-state drive. It is responsible for the initial processing and screening of on-site data from the flue-curing barn. The central data processing server, located in the control center, utilizes a dual-core server architecture and features a 64-core processor running at 2.9 GHz, 256 GB of memory, and 20 TB of storage. It also includes a high-performance graphics processing unit (GPU), which runs the tobacco leaf status perception model and the flue-curing barn's three-dimensional environmental parameter model. The two processing levels are connected via Gigabit Ethernet, and data transmission is encrypted using the AES-256 algorithm to ensure secure and reliable data transmission.

[0147] The control execution unit includes heat source control system, humidity control system and air flow control system. The heat source control system adopts a composite heating method of electric heating wire and steam heating, and the power density of the electric heating wire is 300W / m 2 , evenly distributed on the bottom and side walls of the drying room, achieving precise temperature control accuracy of ±0.5℃. The humidity control system consists of an ultrasonic humidifier and a dehumidifier. The humidifier has a misting capacity of 5L / h and the dehumidifier has a dehumidification capacity of 30L / day. Through micro-adjustment, the humidity control accuracy of ±2%RH is achieved. The airflow control system includes a variable frequency fan and air guide plate combination, with a maximum air volume of 5000m 3 / h. The air deflector is made of food-grade 304 stainless steel, 50cm long, 20cm wide, and 2mm thick, and can rotate 360 degrees to adjust the airflow direction. All actuators use industrial-grade variable frequency control technology, with a response time of less than 1 second, high control accuracy, and low energy consumption, meeting the precise control requirements of different baking stages.

[0148] These units are integrated into a complete intelligent precision tobacco curing system via an industrial bus, enabling closed-loop management of the entire process, from environmental sensing and tobacco leaf status monitoring to intelligent control. The system utilizes a modular design, allowing flexible configuration of functional modules based on the size of the flue-curing barn and tobacco leaf variety, ensuring strong adaptability and easy maintenance. The system consumes 25kW of power, averaging approximately 300kWh per day, achieving approximately 30% energy savings compared to traditional tobacco curing methods. Furthermore, the system improves tobacco leaf quality consistency by over 40%, effectively addressing the issues of unstable tobacco leaf quality and significant energy waste associated with traditional tobacco curing methods.

[0149] It should be noted that the intelligent precision roasting method and system of the present invention achieve deep integration and collaboration at the data level, forming a complete closed-loop system from data acquisition to decision-making and execution. The system's environmental perception unit uses a network of temperature and humidity sensors, a network of wind speed sensors, and a light sensor to collect real-time data at 30-second intervals, the first time granularity specified in the method. Data on six core process parameters, including dry-bulb temperature, wet-bulb temperature, wind speed, and exhaust flow rate, are transmitted to edge computing nodes for preliminary processing. The tobacco leaf information acquisition unit's near-infrared spectrometer, high-definition image acquisition system, and tobacco leaf physical and chemical parameter detection device simultaneously acquire data on tobacco leaf quality indicators such as moisture content, chlorophyll content, and total sugar content. This data is processed using sensor fusion technology to construct the second optimization vector in the method. The data processing unit's central server runs an intelligent process adaptation model, performing statistical analysis on the collected environmental parameter data at the second time granularity of 10-minute windows. By calculating the mean, variance, and rate of change, an 18-dimensional first optimization vector is formed. This, combined with the 3-dimensional second optimization vector, is then input into a multi-layer perception network based on a hierarchical attention network architecture for deep learning analysis. According to the stable change threshold judgment mechanism established in the method, when it is detected that the parameter change rate in the first optimization vector exceeds the historical average by 15% or the quality parameter in the second optimization vector deviates from the target value by more than 8%, the system automatically triggers the gated weight function to calculate the balance value, and selects the corresponding weight adjustment function type according to the balance value range. Through internal iterative optimization processing, the weight contribution rate is updated in real time to ensure the accuracy and timeliness of the system response.

[0150] At the control execution level, the decision outputs of the intelligent precision roasting method and the system's control execution unit achieve precise coordination and cooperation, forming an adaptive closed-loop control system. Once the intelligent process adaptation model calculates the process effect contribution rate through multivariate regression analysis and generates a parameter adjustment strategy, the system's control execution unit immediately responds to these instructions and precisely executes them. The heat source control system adopts temperature control strategies for underripe, moderately ripe, and overripe leaves based on the parameter domain type determined in the method, achieving precise temperature regulation through a combination of electric heating wires and steam heating. The humidity control system's ultrasonic humidifier and dehumidifier coordinate their operation based on the calculated wet-bulb temperature target. The airflow control system's variable frequency fan and air guide vane combination adjusts airflow distribution based on wind speed parameter requirements. While executing control commands, the system continuously monitors the actual roasting performance through the environmental sensing unit and tobacco leaf information collection unit, feeding back the execution results to the data processing unit for performance evaluation. The method's secondary process effect contribution rate analysis mechanism monitors the interactions between parameters in real time and calculates the combined contribution of synergistic and antagonistic effects to quality changes. The system dynamically adjusts the collaborative working mode of each control unit based on this information, establishing a dynamic compensation mechanism for interactive effects. When a parameter domain conversion requirement is detected during the baking process, the system smoothly switches control strategies based on the method's stage transition rate mechanism to avoid impacts on tobacco leaf quality caused by parameter jumps. At the same time, the system utilizes a quality response delay rate assessment mechanism to predict the onset of adjustment effects, optimize the timing and intensity control of subsequent control instructions, and ensure the stability and continuity of the entire baking process, achieving a perfect fusion of method theory and system hardware.

[0151] The following is Example 3 of a specific application scenario of the present invention: This example is based on an intelligent precision baking experiment conducted by researchers at a tobacco leaf processing test base. It describes in detail the complete process of precision baking three batches of tobacco leaves of different maturity levels using the method and system of the present invention. The tobacco samples selected for the experiment were from a tobacco-growing area in Yunnan Province. They included three maturity types: under-ripe leaves, moderately mature leaves, and over-ripe leaves. Each type of tobacco leaf weighed 150 kg, and a total of 450 kg of tobacco leaves were mixed and baked in the same batch.

[0152] The intelligent precision baking system used in the experiment consists of five parts: the baking room structure unit, the environmental sensing unit, the tobacco leaf information collection unit, the data processing unit and the control execution unit. The baking room structure unit adopts a double-layer insulation design, with the outer layer being a 12cm thick reinforced concrete structure and the inner layer being a 6cm thick polyurethane insulation material. The baking room size is

[0153] 8m×5m×3.5m, the interior is divided into three functional areas: heating area, yellowing area and drying area. A 25cm diameter stainless steel exhaust duct and a variable frequency exhaust fan are installed on the top of the baking room, with a maximum air volume of 2500m 3 / h, and a 12kW steam generator is installed at the bottom, with a maximum steam output of 18kg / h. Six layers of stainless steel tobacco hanging racks are installed inside the flue-curing room, and each layer can hold about 75kg of tobacco leaves.

[0154] The environmental sensing unit includes 18 SHT35 temperature and humidity sensors with a temperature measurement accuracy of ±0.1°C and a humidity measurement accuracy of

[0155] ±1% RH, installed in a 6×3 grid layout. 18 thermal wind speed sensors are installed at the same location as the temperature and humidity sensors, with a measurement range of 0 to 5m / s and an accuracy of ±0.05m / s. 6 light sensors are installed on the inner wall of the baking room, with a measurement range of 0 to 10 5 lx, accuracy ±100lx. All sensors adopt IP65 protection grade and are connected to the data acquisition module via RS485 bus.

[0156] The tobacco leaf information acquisition unit consists of a near-infrared spectrometer, a high-definition image acquisition system and a tobacco leaf physical and chemical parameter detection device. The near-infrared spectrometer has a wavelength range of 800 to 2500 nm, a wavelength resolution of 2 nm, a signal-to-noise ratio greater than 10,000:1, and is equipped with 12 measuring probes. The high-definition image acquisition system includes 6 cameras with a resolution of 3840×2160 pixels, a frame rate of 30fps, and is equipped with an LED fill light with a color temperature of 5600K. The tobacco leaf physical and chemical parameter detection device includes a capacitive moisture detector (measuring range 5 to 90%, accuracy of ±0.5%), an LED excitation light source chlorophyll fluorescence detector (wavelength 470 nm, accuracy of ±0.01 mg / g) and a total sugar content analyzer based on near-infrared diffuse reflectance (detection range 0 to 50%, accuracy

[0157] ±0.1%).

[0158] The data processing unit utilizes a distributed architecture, with edge computing nodes equipped with eight-core 3.6GHz processors, 16GB of RAM, and 512GB solid-state drives. The central data processing server utilizes a dual-core architecture, a 64-core 2.9GHz processor, 256GB of RAM, 20TB of storage capacity, and a high-performance graphics processing unit. Both levels of equipment are connected via Gigabit Ethernet, using AES-256 encryption to ensure secure data transmission.

[0159] The control execution unit includes a heat source control system that combines electric heating wire and steam heating. The power density of the electric heating wire is 300W / m 2 , temperature control accuracy ±0.5℃. The humidity control system consists of an ultrasonic humidifier with a misting capacity of 5L / h and a dehumidifier with a dehumidification capacity of 30L / day. The humidity control accuracy is ±2%RH. The airflow control system includes a maximum air volume of 5000m 3 / h variable frequency fan and 360-degree rotatable air guide. All actuators adopt industrial-grade variable frequency control with a response time of less than 1 second.

[0160] The researchers first established a basic process parameter set consisting of six core parameters: dry-bulb temperature, wet-bulb temperature, wind speed, exhaust flow rate, heating rate, and stabilization time. Based on the results of the tobacco leaf maturity analysis, it was determined that underripe leaves accounted for 35% of the total, moderately ripe leaves accounted for 50%, and overripe leaves accounted for 15%. Based on this distribution ratio, the researchers divided the basic process parameter set into three parameter domains. The underripe leaf parameter domain sets the dry-bulb temperature range to 38-41°C, the wet-bulb temperature range to 34-37°C, the wind speed to 0.9-1.1 m / s, and the heating rate to 0.6°C / h. The moderately ripe leaf parameter domain sets the dry-bulb temperature range to 44-47°C, the wet-bulb temperature range to 39-41°C, the wind speed to 1.3-1.7 m / s, and the heating rate to 1.1°C / h. The parameter domain of over-mature leaves sets the dry-bulb temperature range to 49-53°C, the wet-bulb temperature range to 43-45°C, the wind speed to 2.0-2.4 m / s, and the heating rate to 1.4°C / h.

[0161] During the data collection phase, researchers set the initial time granularity to 30 seconds. The sensor network of the environmental perception unit collected the basic process parameter set in real time at the set frequency. A tiered collection strategy was established based on the differences in tobacco leaf variety and location: the collection frequency for high-quality tobacco segments was set at every 20 seconds, the standard 30-second frequency for medium-quality segments, and the frequency for low-quality segments at every 43 seconds. Over the entire 18-hour curing process, the system collected a total of 2,160 sets of process parameter data, with each set containing complete values for six core parameters.

[0162] Table 1 shows the statistics of the first optimized vector construction data at different time periods during the baking process:

[0163] Table 1 Statistics of the first optimization vector construction data

[0164]

[0165] The edge computing nodes of the data processing unit perform preliminary processing and screening of the collected data, and the central server runs the intelligent process adaptation model for in-depth analysis. The researchers set the second time granularity to 10 minutes, and performed statistical analysis on the data collected for 30 seconds within this time window. By calculating the mean, variance, and rate of change of the six core parameters, an 18-dimensional first optimization vector was formed. The intelligent process adaptation model determines the weight contribution rate of each parameter domain based on historical data analysis. The weight contribution rate of the under-ripe leaf parameter domain is 0.28, the weight contribution rate of the moderately ripe leaf parameter domain is 0.57, and the weight contribution rate of the over-ripe leaf parameter domain is 0.15. The stage transition rate is determined by monitoring the changes in the state of the tobacco leaves. A total of 7 parameter domain switches occurred during the entire baking process, with an average transition rate of 0.39 times / h.

[0166] The tobacco leaf information acquisition unit uses machine vision and near-infrared spectroscopy fusion technology to obtain tobacco leaf quality change data and construct a second optimization vector. The machine vision system's six high-definition cameras monitor tobacco leaf surface color changes in real time and calculate chlorophyll degradation rates through RGB color space analysis. The near-infrared spectrometer's 12 measurement probes scan tobacco leaf samples within a wavelength range of 1420-1480nm and calculate moisture content changes based on the Lambert-Beer law. Total sugar conversion efficiency is determined using a rapid analyzer. The second optimization vector includes the tobacco leaf moisture content change rate, γ. w , chlorophyll degradation rate γ c , total sugar conversion efficiency γ s Three-dimensional data.

[0167] Table 2 shows the change data of each parameter of the second optimization vector during the baking process:

[0168] Table 2 Second optimization vector parameter change data table

[0169]

[0170] The researchers established a stable change threshold judgment mechanism, setting the parameter adjustment mechanism to be triggered when the rate of change of any parameter in the first optimization vector exceeds 15% of its historical average or the quality parameter in the second optimization vector deviates from the target value by more than 8%. During the baking process, the system triggered parameter adjustment 13 times, including 6 temperature adjustments, 4 humidity adjustments, and 3 wind speed adjustments. Each adjustment is adaptively adjusted through the gated weight function of the intelligent process adaptation model, and the control execution unit accurately executes the adjustment instructions output by the model. The heating wire and steam heater of the heat source control system work in coordination according to the temperature adjustment instructions, the humidifier and dehumidifier of the humidity control system are adjusted according to the humidity target value, and the variable frequency fan and air guide plate of the airflow control system adjust the airflow distribution according to the wind speed requirements.

[0171] The gated weight function calculates the equilibrium value ξ based on the current curing stage identifier, the tobacco leaf maturity distribution vector, and the quality target deviation. Depending on the equilibrium value range, the system uses three weight adjustment functions during the curing process. When the ξ value is between 0 and 0.423, a conservative weight adjustment function is used. This situation occurred five times in the first six hours of curing. The control execution unit adopted a progressive adjustment strategy, controlling the temperature adjustment range within 0.3°C and the humidity adjustment range within 1% RH. When the ξ value is between 0.423 and 0.628, a balanced weight adjustment function is used. This situation occurred six times in the middle of the curing stage, and the system maintained the standard adjustment range. When the ξ value is between 0.628 and 1.0, an aggressive weight adjustment function is used. This situation occurred twice in the late stages of curing. The control execution unit increased the adjustment force to quickly respond to quality changes.

[0172] The researchers performed internal iterative optimization on the first optimization vector, and the data processing unit used the gradient descent algorithm to achieve adaptive update of the weight contribution rate. The initial weight contribution rate vector was set to W (0) =[0.30,0.55,0.15] T , after iterative optimization, it finally converges to W (final) =[0.28,0.57,0.15] T The entire iterative process was repeated 187 times, and the weight deviation gradually converged from the initial 0.0124 to 0.0009, meeting the convergence threshold of 0.001. The learning rate was set to 0.01, and the momentum optimization algorithm was used to accelerate the convergence process.

[0173] Table 3 shows the key steps of the internal iterative optimization process:

[0174] Table 3 Key data of internal iterative optimization processing

[0175] Number of iterations Unripe leaf weight Weight of mature leaves Overripe leaf weight Weight bias value Learning rate 0 0.300 0.550 0.150 0.0124 0.010 50 0.285 0.562 0.153 0.0067 0.010 100 0.282 0.565 0.153 0.0034 0.010 150 0.280 0.567 0.153 0.0018 0.010 187 0.278 0.570 0.152 0.0009 0.010

[0176] Through multivariate regression analysis, the data processing unit calculates the contribution rate of process effects, quantifying the weight of each parameter in the basic process parameter set on the final quality of the tobacco leaves. The contribution coefficient for dry-bulb temperature is 0.324, wet-bulb temperature is 0.287, wind speed is 0.196, and heating rate is 0.148. A two-way evaluation mechanism for positive contribution and negative impact is established, with temperature parameters primarily contributing positively, while excessively high wind speeds and heating rates may have negative effects. The control execution unit determines adjustment priority based on the contribution rate, prioritizing temperature and humidity parameters with higher contribution rates.

[0177] Based on the secondary change contribution rate of process effect, the system makes fine adjustments and analyzes the interaction between parameters through factorial experimental design. There is a significant synergistic effect between dry-bulb temperature and wet-bulb temperature, and the interaction contribution rate coefficient ψ 12 The control execution unit takes humidity changes into account when adjusting the temperature. There is an antagonistic effect between wind speed and heating rate, and the interaction coefficient ψ 34 The interaction between temperature and wind speed has a significant effect on the water loss rate of tobacco leaves. The coefficient ψ 13 It is 0.076, and the control execution unit achieves the best cooperation by coordinating the heat source system and the airflow system.

[0178] The intelligent process adaptation model demonstrated excellent adaptability throughout the tobacco baking process, with an average model inference time of 0.23 seconds, meeting real-time control requirements. The system's overall power consumption was 22.8 kW, lower than the expected design value. The sensor network of the environmental perception unit operated stably, achieving a data collection success rate of 99.7%. The various detection devices in the tobacco leaf information collection unit operated in coordinated fashion, maintaining a near-infrared spectrometer accuracy of ±0.02%, and ensuring stable image quality from the machine vision system. The computational load in the data processing unit was distributed efficiently, with an average CPU utilization rate of 65% for edge computing nodes and 78% for the central server GPU. The various actuators in the control execution unit responded promptly, with average temperature adjustment response times of 0.8 seconds, humidity adjustment response times of 1.2 seconds, and wind speed adjustment response times of 0.5 seconds.

[0179] The quality response delay rate assessment mechanism plays a significant role in practical applications. It uses time series correlation analysis to predict the time lag in quality changes following process parameter adjustments. The average response delay for temperature adjustment is 8.3 minutes, the average response delay for humidity adjustment is 12.6 minutes, and the average response delay for wind speed adjustment is 4.7 minutes. Based on this delay time data, the control execution unit can make parameter adjustments in advance, avoiding the adverse effects of delayed adjustments on tobacco leaf quality.

[0180] The phase transition rate mechanism enables smooth parameter domain switching during the curing process, and the control execution unit prevents the impact of parameter jumps on tobacco leaf quality. The transition time from the under-ripe leaf parameter domain to the moderately ripe leaf parameter domain averages 45 minutes, and the transition time from the moderately ripe leaf parameter domain to the over-ripe leaf parameter domain averages 38 minutes. During this transition, the temperature change rate of the heat source control system is controlled within 0.3°C / minute, the humidity change rate of the humidity control system is controlled within 0.5%RH / minute, and the airflow control system maintains a smooth wind speed, ensuring a smooth curing process.

[0181] Table 4 shows the comparison of tobacco leaf quality test results after baking with traditional methods:

[0182] Table 4 Comparison of tobacco leaf quality test results

[0183] Quality indicators Intelligent precision baking Traditional baking method Degree of improvement (%) Moisture content uniformity (%) 12.3±0.8 13.7±2.1 10.2 Chlorophyll residue (mg / g) 0.85±0.12 1.02±0.18 16.7 Total sugar content (%) 18.6±1.2 17.2±2.3 8.1 Color consistency score 8.7±0.4 7.9±0.8 10.1 Overall quality rating 91.2±2.1 84.6±3.4 7.8 Energy consumption (kWh / kg) 2.34 2.78 15.8

[0184] The model achieved 93.8% accuracy in curing process classification, 90.5% accuracy in quality grade prediction, and 88.7% consistency between parameter adjustment recommendations and expert evaluation, all of which were close to the performance indicators during training. The system's overall energy consumption was 22.8 kW, and the total energy consumption for the entire curing process was 1,053 kWh, representing significant energy savings compared to traditional methods. The flue-curing barn structural units exhibited excellent insulation, with heat loss controlled to less than 8%. The environmental perception unit's sensors were strategically arranged, enabling comprehensive monitoring of environmental changes within the barn. The tobacco leaf information collection unit's multi-sensor fusion technology effectively improved quality inspection accuracy. The distributed architecture of the data processing unit ensured computing performance and data security. The precise control capabilities of the control execution unit ensured stable implementation of process parameters.

[0185] Traditional methods used to solve the core technical problems of the present invention mainly rely on manual experience and timed and quantitative parameter adjustment. The baking master manually adjusts the heating and ventilation equipment according to the changes in the appearance and color of the tobacco leaves and the temperature and humidity displayed in the baking room. The timing and amplitude of the adjustment are mainly based on personal experience, lacking scientific quantitative basis and real-time feedback mechanism. Traditional methods are unable to achieve multi-parameter coordinated optimization and are often only able to adjust a single parameter, ignoring the impact of the interaction between parameters on tobacco quality. In addition, traditional methods lack the ability to monitor changes in tobacco quality in real time and cannot accurately adjust according to the actual state of the tobacco leaves, resulting in problems such as over-baking or under-baking during the baking process. The present invention represents a significant improvement over traditional methods. By establishing an intelligent process adaptation model and a complete system architecture, it achieves multi-parameter coordinated optimization and real-time adaptive adjustment. The uniformity of tobacco moisture content has increased by 10.2%, the color consistency score has increased by 10.1%, and the overall quality score has increased by 7.8%, while reducing energy consumption by 15.8%. The intelligent system can accurately adjust parameters according to the actual state changes of tobacco leaves, avoiding the subjectivity and lag of manual judgment, realizing the standardization and automated control of the baking process, and significantly improving the quality stability and production efficiency of tobacco leaf baking.

[0186] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. An intelligent and precise baking method for improving tobacco leaf quality, characterized in that: include: Establish a basic process parameter set and divide it into three classification ranges according to the maturity of tobacco leaves: under-mature leaf parameter range, moderately mature leaf parameter range, and over-mature leaf parameter range. Collect the basic process parameter set in real time with the first-time granularity, and establish a stratified collection strategy based on the differences in tobacco leaf varieties and parts. A first optimization vector is constructed based on the second time granularity, and the weighted contribution rate and stage transition rate of each parameter domain are determined through an intelligent process adaptation model. Tobacco quality change data is acquired through machine vision and near-infrared spectroscopy fusion technology to construct a second optimization vector, and a quality response delay rate assessment mechanism is established. A stable change threshold determination mechanism is established, and process parameter adjustment is initiated when the numerical fluctuations between the first and second optimization vectors exceed the stable change threshold. Adaptive adjustment is performed using the gated weight function of the intelligent process adaptation model. The first optimization vector is internally iteratively optimized, and the weight contribution rate value is updated through iterative calculation; the process effect contribution rate and the process effect secondary change contribution rate are calculated, the parameter adjustment strategy is generated, and a collaborative conversion mechanism between different parameter domains is established; the hierarchical fusion weights of the intelligent process adaptation model are used to dynamically balance the influence of each classification range, so as to realize intelligent and precise adjustment of the baking process parameters.

2. The method according to claim 1, characterized in that The basic process parameter set specifically refers to a collection of multiple core control parameters that affect the quality changes of tobacco leaves during the baking process, including the dry-bulb temperature and wet-bulb temperature for controlling the temperature environment of the baking room, the wind speed for adjusting air flow, the exhaust flow rate for controlling gas exchange in the baking room, the heating rate for determining the speed of temperature change, and the stabilization time for maintaining a constant temperature state.

3. The method according to claim 2, characterized in that The first time granularity specifically refers to the time interval for data collection of the basic process parameter set, which is set to 30 seconds. It is used to capture the rapid changes and instantaneous fluctuations of process parameters during the baking process, and at the same time establish differentiated collection frequencies for different tobacco leaf sections.

4. The method according to claim 3, characterized in that The second time granularity specifically refers to the time statistical window used when constructing the optimization vector, which is set to 10 minutes. It is used to perform statistical analysis and trend extraction within the time window of the data collected at the first time granularity, and establish a comprehensive evaluation system for short-term fluctuation characteristics and long-term trend characteristics through multi-time scale analysis.

5. The method according to claim 4, characterized in that The first optimization vector specifically refers to an 18-dimensional vector formed by combining the mean, variance, and change rate of the basic process parameter set within the second time granularity. It is used to characterize the comprehensive state characteristics of the process parameters within the time window, quantify the influence of each parameter on the baking effect through the weighted contribution rate, and establish a stage transition rate mechanism between parameter domains to achieve dynamic switching.

6. The method according to claim 5, characterized in that The second optimization vector specifically refers to the tobacco leaf quality parameter vector obtained through sensor fusion technology, which includes three dimensions: tobacco leaf moisture content change rate, chlorophyll degradation rate, and total sugar conversion efficiency. It is used to quantify the physiological and biochemical changes of tobacco leaves and establish a quality response delay rate evaluation system to predict the time lag effect of quality changes.

7. The method according to claim 6, characterized in that The stable change threshold specifically refers to the numerical boundary for determining whether process parameters need to be adjusted. When the change rate of any parameter in the first optimization vector exceeds 15% of its historical average or the quality parameter in the second optimization vector deviates from the target value by more than 8%, the parameter adjustment mechanism is triggered. By establishing a threshold classification system, the risk of misadjustment is reduced.

8. The method according to claim 7, characterized in that The structure of the intelligent process adaptation model is a multi-layer perception network based on a hierarchical attention network architecture, which includes five main parts: input layer, feature extraction layer, attention calculation layer, gated fusion layer and output layer. The feature extraction layer uses a residual connection structure to process the 18-dimensional first optimization vector and the 3-dimensional second optimization vector. The attention calculation layer calculates the association weights between different parameter domains through a multi-head attention mechanism.

9. An intelligent and precise baking system for improving tobacco leaf quality, characterized in that: It includes a flue-curing room structure unit, an environmental perception unit, a tobacco leaf information collection unit, a data processing unit and a control execution unit. Each unit is integrated into a complete system through an industrial bus, realizing closed-loop management of the entire process from environmental perception, tobacco leaf status monitoring to intelligent control.

10. The system according to claim 9, characterized in that The tobacco leaf information acquisition unit consists of a near-infrared spectrometer, a high-definition image acquisition system and a tobacco leaf physical and chemical parameter detection device, and is used to collect tobacco leaf near-infrared spectrum data, image data and physical and chemical parameters; the data processing unit adopts a distributed architecture design, including edge computing nodes and a central data processing server, and is used to run a tobacco leaf status perception model and a three-dimensional environmental parameter model of a flue-curing room.

Images