Tobacco leaf curing process optimizing method and system

Abstract

The invention discloses a tobacco leaf curing process optimization method and system, and the method comprises the steps: building a dynamic process diagram with multi-dimensional nodes according to the historical data of a tobacco leaf curing process, and enabling the dynamic process diagram to represent the state and the processing process of tobacco leaves at any processing moment; the weight of the processing technology at the moment is calculated according to the tobacco leaf shape modal data and the processing technology at any moment in the real-time dynamic technology diagram, and the processing technology at the moment is adjusted in real time; integrating the plurality of modal data of the real-time tobacco leaf shape and establishing a training model, wherein the training model is used for outputting optimal process parameters; searching an optimal process path globally according to the optimal process parameters output by the training model, inhibiting parameter fluctuation, and regulating and controlling the real-time process parameters through fuzzy logic control; and performing feature extraction on the to-be-baked tobacco leaves to match the optimal process parameters and the optimal process path for baking. The process is intelligently and dynamically optimized, the quality is accurately controlled, the consumption is reduced, and browning is efficiently and stably inhibited.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of tobacco leaf curing, and particularly relates to a tobacco leaf curing process optimization method and system. BACKGROUND

[0002] Tobacco leaf curing is a key link in tobacco production, and traditionally mainly adopts methods such as earth house curing, three-stage curing and horizontal intensive curing house. Among them, the three-stage curing process includes the yellowing period, the color fixing period and the dry muscle period, and each stage has strict requirements for temperature and humidity. For example, the temperature in the yellowing period needs to gradually rise from 35 degrees to 38 degrees, and the relative humidity needs to be kept at 70%; in the color fixing period, the temperature needs to be further raised and the humidity needs to be controlled to fix the color of the tobacco leaves; in the dry muscle period, higher temperature is needed to completely dry the tobacco leaves.

[0003] However, the traditional process has many problems. Tobacco farmers often judge the state of tobacco leaves and adjust process parameters according to experience, lacking precise control means, resulting in unstable tobacco leaf quality. For example, the curing loss rate in Zhen'an tobacco area in 2022 was as high as 9.42%, of which more than 80% of the loss was caused by human factors. In addition, the temperature control precision of traditional curing houses is low, and the labor intensity is large, which needs frequent manual operation, further affecting the curing quality.

[0004] In recent years, with the development of Internet of Things, big data and artificial intelligence technologies, tobacco leaf curing has begun to develop towards intelligence. For example, the intelligent curing system based on Internet of Things can monitor the temperature and humidity in the curing house in real time, and realize automatic management through remote control. In addition, some areas have introduced new energy curing houses, which use new energy such as biomass and electricity, and are equipped with automatic intelligent control sensing technology to realize accurate control of temperature and humidity. The application of these technologies has significantly improved the curing efficiency and tobacco leaf quality, and reduced labor intensity and cost.

[0005] However, the existing intelligent curing technology still has deficiencies. For example, some systems only realize basic remote monitoring and automatic control, lacking deep optimization capability for the curing process. In addition, due to the great differences in tobacco varieties, ecological environment and production technology in different tobacco areas, more precise and intelligent curing processes and equipment need to be developed according to specific conditions; some tobacco leaf curing optimization methods rely on expert experience data to make static models to find the optimal curing method for tobacco leaves, but such methods have subjective bias and poor processing capability for special cases, and the unpredictability of the results will also increase as the curing time of tobacco leaves is relatively long. Therefore, there is an urgent need for a method and system that can combine historical data and real-time information to intelligently optimize the tobacco leaf curing process, so as to further improve the stability and consistency of the quality of tobacco leaf curing. SUMMARY

[0006] The purpose of this invention is to provide a method and system for optimizing tobacco curing processes, which can solve the technical problems of high subjectivity, poor stability, and difficulty in handling individual cases in existing static model optimization methods.

[0007] To achieve the above objectives, a specific embodiment of the present invention provides a method for optimizing tobacco curing process. The method includes: constructing a dynamic process diagram with multi-dimensional nodes based on historical data of tobacco curing process, wherein the dynamic process diagram represents the state of tobacco leaves and processing technology at any processing time.

[0008] Based on the tobacco leaf modal data and processing technology at any time in the real-time dynamic process diagram, the weight of the processing technology at that time is calculated, and the processing technology at that time is adjusted in real time.

[0009] The real-time tobacco leaf-like modal data are integrated and a training model is established. The training model is used to output the optimal process parameters.

[0010] The optimal process path is searched globally based on the optimal process parameters output by the trained model, while suppressing parameter fluctuations and controlling the real-time process parameters through fuzzy logic control.

[0011] Feature extraction is performed on the tobacco leaves to be cured to match the optimal process parameters and the optimal process path for curing.

[0012] In one or more embodiments of the present invention, the step of constructing the dynamic process diagram includes:

[0013] Construct a data structure including tobacco leaf temperature T, humidity H, moisture content W, and color features C∈R. 12 The process node network, wherein:

[0014] Node definition: v i =(T i H i W i C i ); Definition of an edge: e j =(Δt) j ω j (t)), Δt j For process operation time, ω j (t) represents the dynamic weight; node splitting is introduced to address unexpected situations during the baking process, and the node splitting conditions include:

[0015] When ΔW > 5% or ΔC > 15% within a preset time period, a child node set v is generated. sub ={(T±kΔT,H±kΔH,W,C)}, where ΔT and ΔH are temperature adjustment and humidity adjustment, and k=1 or 2.

[0016] In one or more embodiments of the present invention, when the node corresponding to the tobacco leaf data splits, the weight of its corresponding process will change, and the edge weights will be dynamically adjusted based on spatiotemporal coupling, including: ω j (t)=ω j0 ·e -λt +η·(αΔW+βΔC), where λ is the historical weight decay coefficient to ensure that the influence of old data continues to decrease every hour; the real-time sensitivity factor η∈[0.3, 0.8], which is generated through node splitting and increases to 0.8 during mutation; α and β are the weight coefficients of moisture content and color change.

[0017] In one or more embodiments of the present invention, the real-time tobacco leaf-like multimodal data includes two-dimensional temperature and humidity data, one-dimensional moisture content data, 12-dimensional tobacco leaf image data, 6-dimensional near-infrared spectrophotometry data, and 3-dimensional biochemical index data. Texture features are obtained by highlighting ROI automatic segmentation and median filtering for noise reduction, and then normalized. A 23-dimensional vector X = [T, H, W, C] is input to the training model. 1-12 S 1-6 B 1-3 The output target is the predicted adjustment process parameter Y = [ΔT]. opt ΔH opt The training strategy for the real-time training model uses time series for cross-validation, emphasizes the priority of temperature and humidity control by establishing a custom function, and generates adjustment suggestions for process parameters in real-time inference over a certain period of time.

[0018] In one or more embodiments of the present invention, the step of globally searching for the optimal process path based on the optimal process parameters output by the trained model, while suppressing parameter fluctuations, includes:

[0019] Setting path cost functions includes:

[0020]

[0021] Dynamic weight ω j The real-time calculated values ​​from the preceding steps reflect the historical reliability and current sensitivity of the path.

[0022] Path entropy constraint Entropy(P):

[0023] θ i The angle of parameter change is used to quantify the disorder of the path, to suppress sudden temperature changes and drastic humidity fluctuations, and to reduce the number of overshoots;

[0024] A genetic algorithm is set up, wherein 80% of the initial population of the genetic algorithm comes from the prediction result Y = [ΔT].opt ΔH opt ], 20% are randomly generated; its fitness function is: Fitness = 0.7Q - 0.2E - 0.1T;

[0025] Where Q is the quality score, adjusted based on color difference and moisture content uniformity; E is the system energy consumption, with a target value of ≤2.8kWh / kg; T is the time deviation rate, allowing ±10% fluctuation; optimization termination condition: Pareto front hypervolume change rate less than 1% or preset number of iterations to ensure convergence efficiency.

[0026] In one or more embodiments of the present invention, the overfuzzy logic control for regulating real-time process parameters includes:

[0027] Based on the physiological characteristics of tobacco leaves, typical operating conditions for tobacco curing are set. The optimal process parameters are processed by the membership function input according to the typical operating conditions, and the output membership function is used to match the optimal process parameters with the control accuracy of the curing equipment. An inference is performed once within a preset time period, and the output optimal process parameters are converted into equipment control commands.

[0028] In one or more embodiments of the present invention, the step of extracting features from the tobacco leaves to be cured to match the optimal process parameters and the optimal process path for curing includes the following steps:

[0029] Feature extraction is performed on the tobacco leaves to be cured. The features of the tobacco leaves to be cured include: F = [W, C, S, B], where W is the moisture content, C is the color feature, S is the near-infrared spectral data, and B is the biochemical index.

[0030] Based on the above features, the dynamic process diagram is expanded, and then the process is matched in the training model to generate a dynamic baking scheme, and the baking of the tobacco leaves to be tested begins.

[0031] In one or more embodiments of the present invention, the process matching includes:

[0032]

[0033] Where, γ=[ω w ω c ω s ω b The weights are assigned to W, C, S, and B, emphasizing the dominant role of moisture content and color. Several historical schemes are selected, and those with a similarity Sim value higher than a predetermined value are used as candidate benchmarks.

[0034] In one or more embodiments of the present invention, generating a dynamic baking scheme includes using a transfer learning formula:

[0035]

[0036] Among them, Sim m For the current tobacco leaf characteristics F new With historical craftsmanship F m The weighted cosine similarity (values ​​from 0 to 1) can be calculated using the following formula:

[0037]

[0038] Among them, F new =[W new C new S new B new ] represents the characteristic fingerprint of the current process, F m =[W m C m S m B m [This is] a characteristic fingerprint of historical craftsmanship;

[0039] P m For the parameter set of similar processes in the preset group in the historical process scheme library, ΔP RL The update rule for the incremental adjustments generated by Q-learning is as follows:

[0040] Q(s, a)←Q(s, a)+0.2[r+0.9maxQ(s′, a′)-Q(s, a)];

[0041] Reward function h = 0.7Q new -0.2E+0.1(1-CVD), Q new The quality score for the new process is given, and CVD represents the degree of constraint violation.

[0042] In another aspect of the present invention, a tobacco curing process optimization system is provided, the system comprising:

[0043] The data preprocessing module is used to construct a dynamic process diagram with multi-dimensional nodes based on historical data of tobacco curing process. The dynamic process diagram represents the state of tobacco leaves and processing technology at any processing time.

[0044] The weight calculation module is used to calculate the weight of the processing technology at any given time based on the tobacco leaf modal data and processing technology in the real-time dynamic process diagram, and to make real-time adjustments to the processing technology at that time.

[0045] The model building module is used to integrate multiple modal data of the real-time tobacco leaf and build a training model, which is used to output the optimal process parameters.

[0046] The parameter adjustment module is used to search for the optimal process path globally based on the optimal process parameters output by the training model, while suppressing parameter fluctuations and controlling the real-time process parameters through fuzzy logic control.

[0047] The process generation module is used to extract features from the tobacco leaves to be cured in order to match the optimal process parameters and the optimal process path for curing.

[0048] In another aspect of the invention, an electronic device is provided, comprising: at least one processor; and a memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform a method for optimizing a tobacco curing process.

[0049] In another aspect of the invention, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of a method for optimizing tobacco curing processes.

[0050] Compared with existing technologies, the tobacco curing process optimization method of this invention achieves precise control of key tobacco indicators through multimodal data fusion and intelligent modeling. It utilizes color features (12-dimensional parameters such as RGB and Lab space) to monitor the browning process in real time, and combines near-infrared spectroscopy to retrieve starch and moisture content, constructing a dynamic process network. When browning risk (such as a sudden change in Lab value b*) or abnormal moisture content is detected, a temperature and humidity adjustment scheme is quickly generated based on a hybrid optimization algorithm to suppress quality deterioration. Fuzzy logic control transforms expert experience into 49 closed-loop adjustment rules, ensuring execution accuracy (e.g., temperature error ±0.2℃). Compared with traditional methods relying on fixed templates, this scheme significantly improves the uniformity of tobacco color, moisture content stability, and finished product grade rate, especially showing outstanding results in high-value tobacco varieties (such as Yunnan KRK26).

[0051] This invention employs a global-local collaborative hybrid optimization strategy to achieve a multi-objective balance between energy consumption and time while prioritizing quality. By improving the A* algorithm to screen low-fluctuation process paths and combining it with a genetic algorithm for fine-tuning temperature and humidity parameters, energy consumption is reduced by approximately 20% compared to traditional methods. Simultaneously, an adaptive process generation mechanism based on transfer learning and reinforcement learning can quickly reuse historical high-quality solutions (similarity > 0.7) and dynamically explore incremental optimization space through Q-learning, shortening the new process generation time by more than 60%. This intelligent decision-making model significantly reduces the cost of manual trial and error, making it particularly suitable for large-scale production scenarios with multiple varieties and batches.

[0052] To address unforeseen conditions during the baking process (such as sudden increases in moisture content or equipment malfunctions), this invention constructs a graded response mechanism. Through node splitting technology in the dynamic process diagram, it covers extreme operating conditions of ±4℃ / ±10%RH; the real-time sensitivity factor η automatically increases to 0.8 upon detecting a sudden change, accelerating the algorithm's response speed. The collaborative mechanism of fuzzy control and optimization algorithms allows seamless switching between the global optimal solution and local fine-tuning, avoiding batch losses caused by parameter over-adjustment in traditional methods. Furthermore, the feature-driven knowledge accumulation system continuously expands the process rule base, enabling the system to adapt across production areas and varieties, ensuring stable production in complex agricultural environments. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart of a method for optimizing tobacco curing process in one embodiment of the present invention;

[0055] Figure 2 This is a flowchart illustrating the matching process for tobacco leaf curing in one embodiment of the present invention;

[0056] Figure 3 This is a block diagram of a dynamic evaluation system for tobacco leaf quality according to one embodiment of the present invention;

[0057] Figure 4 This is a hardware structure diagram of a computing device according to an embodiment of the present invention. Detailed Implementation

[0058] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0059] As described in the background section, existing technologies cannot adapt to the technical problems of tobacco variety mutations, long process production time, process matching relying on human experience, and insufficient utilization of deep features such as spectral and biochemical properties by traditional methods.

[0060] To address the aforementioned technical issues, such as Figure 1As shown, this invention introduces a method for optimizing tobacco curing processes, including the following steps:

[0061] S101. Construct a dynamic process diagram with multi-dimensional nodes based on historical data of tobacco curing process, wherein the dynamic process diagram represents the state of tobacco leaves and processing technology at any processing time;

[0062] S102. Calculate the weight of the processing technology at any given moment based on the tobacco leaf modal data and processing technology in the real-time dynamic process diagram, and make real-time adjustments to the processing technology at that moment.

[0063] S103. Integrate the multiple modal data of the real-time tobacco leaf shape and establish a training model, the training model being used to output the optimal process parameters;

[0064] S104. Establish an optimization model based on the optimal process parameters output by the training model and globally search for the optimal process path. Use fuzzy logic control to regulate the real-time process parameters.

[0065] S105. Extract features from the tobacco leaves to be cured to match the optimal process parameters and the optimal process path for curing.

[0066] In a specific real-time process, the steps for constructing a dynamic process diagram include:

[0067] Construct a data structure including tobacco leaf temperature T, humidity H, moisture content W, and color features C∈R. 12 The process node network, wherein:

[0068] Node definition: v i =(T i H i W i C i );

[0069] Definition of an edge: e j =(Δt) j ω j (t)), Δt j For process operation time, ω j (t) represents the dynamic weights;

[0070] To address unforeseen circumstances during the baking process, node splitting is introduced, and the node splitting conditions include:

[0071] When ΔW > 5% or ΔC > 15% within a preset time period, a child node set v is generated. sub ={(T±kΔT,H±kΔH,W,C)}, where ΔT=2℃,ΔH=5%RH,k=1,2.

[0072] In this embodiment, by constructing a system based on tobacco leaf temperature T, humidity H, moisture content W, and color feature C∈R... 12 This is a process network of nodes that dynamically reflects changes in the state of tobacco leaves. Through a node splitting mechanism (e.g., triggering child node generation when ΔW>5%), it addresses the problem of delayed response to sudden states in traditional directed graphs, laying the foundation for adjustments in response to subsequent emergencies. For example, when a sudden change in a certain type of tobacco leaf is detected:

[0073] Mutation detection: Mid-yellowing stage ΔW = +6.5% (target ΔW ≤ 5%), ΔC = 18% (color difference exceeds threshold).

[0074] Child node generation: Original parameters: T = 40℃, H = 88%RH Generated nodes: {36℃ / 90%RH, 40℃ / 80%RH, 34℃ / 95%RH, 42℃ / 76%RH}

[0075] Optimization results: The T=40℃ / H=80%RH path was selected to allow the moisture content to return to the target curve within 1.5 hours, avoiding a 12% increase in the smoke rate. Through graded adjustments (k=1 for fine-tuning and k=2 for strong intervention), the precision of process control was improved while ensuring safety, reducing the number of over-adjustments by 41% compared to traditional methods.

[0076] In a further embodiment, when the node corresponding to the tobacco leaf data splits, the weight of its corresponding process will change. The edge weights are dynamically adjusted based on spatiotemporal coupling, including: ω j (t)=ωj0·e-λt+η·(αΔW+βΔC), where λ is the historical weight decay coefficient to ensure that the influence of old data continues to decrease every hour; η∈[0.3,0.8], the real-time sensitivity factor η is generated through node splitting and increases to 0.8 during mutation; α and β are the weight coefficients of moisture content and color change. In this embodiment, α=0.3 and β=0.4, which can be adjusted according to actual needs.

[0077] The node status is input into the weight calculation formula in real time, driving dynamic weight updates. Example: A node v i Moisture content W i The increase from 82% to 85% triggered a recalculation of all values ​​in v. i The edge weights starting from the origin are used to guide the topology optimization in step 1: high-weight edges (ω) j The child nodes corresponding to (t)>0.7) are preferentially retained, and their k value range is increased in subsequent splits (e.g., allowing k=2 or greater). Low-weight edges (ω) j Child nodes triggered by (t)<0.3) will be merged or deleted to reduce computational redundancy.

[0078] Mutation detection: When the mid-yellowing stage (ΔW = +6.5%) occurs (target ΔW ≤ 5%), the weight of the corresponding edge in the process changes, and the real-time sensitivity factor η is increased to 0.8. In the weight formula, αΔW = 0.6 × 6.5 = 3.9. Calculate the weight of each child node edge: Edge (38℃ → 36℃, 90% RH): ω j =0.4×e-0.3×1+0.8×3.9=3.2; Side (38℃→40℃, 80%RH): ω j =0.6×e-0.3×1+0.8×3.9=3.4; This allows subsequent steps to select the path with the highest weight. In practice, the moisture content is reduced to 82% within 2 hours to avoid the generation of soot. Through a topology-weight dynamic coupling mechanism, closed-loop optimization of the tobacco curing process is achieved. Nodes are split into topology optimizations that provide real-time data sources and whose weight calculations feed back into the optimization, forming an adaptive process network. Compared with traditional static methods, this two-way interactive mechanism has significant advantages in terms of anomaly response speed, process stability, and control accuracy.

[0079] In a further embodiment, the real-time multimodal data of tobacco leaves includes two dimensions of temperature and humidity data, one dimension of moisture content data, 12 dimensions of tobacco leaf image data, six dimensions of near-infrared spectrophotometry data, and three dimensions of biochemical index data, as shown in Table 1:

[0080] Table 1. Multiple Modal Data and Their Meanings

[0081]

[0082]

[0083] Based on the above multimodal parameter settings, the parameter design of the LightGBM model is optimized based on the tobacco curing characteristics. Its hyperparameters and settings are shown in Table 2:

[0084] Table 2. Parameter Design of LightGBM Model

[0085]

[0086]

[0087] The final model output target is the predicted adjusted process parameter Y = [ΔT] opt ΔH opt The training strategy for the real-time training model uses time series for cross-validation. A custom function is established to emphasize the priority of temperature and humidity control, and real-time inference is performed over a specified period to generate adjustment suggestions for process parameters, where Y = [ΔT]. opt ΔH optThe domain of adjustment can be adjusted according to actual needs, matching ΔT and ΔH in step 1, to guide node splitting.

[0088] In a further embodiment, the normalization formula is: T norm =15T-35,H norm =100H, the baking start temperature threshold is 35℃ (the lower limit of tolerance for fresh tobacco leaves), half of the temperature range (35℃~50℃) is 15, ensuring the normalized value range [-1, 1]; loss function weight: L=0.7·MSE(ΔT)+0.3·MSE(ΔH), where 0.7 / 0.3 means that temperature control has a higher priority than humidity (temperature error accounts for 70% of the quality impact). For the selection of tobacco leaves in this embodiment, it can be adjusted as needed. The mean square error MSE is used to strengthen the penalty for large deviations (such as when |ΔT|>2℃, the loss increases with the square of k).

[0089] In a further embodiment, the optimal process path is searched globally based on the optimal process parameters output by the trained model, while suppressing parameter fluctuations, including:

[0090] Setting the path cost function using the A* algorithm includes:

[0091]

[0092] Dynamic weight ω j The real-time calculated values ​​from the preceding steps reflect the historical reliability and current sensitivity of the path; path entropy constraint Entropy(P): θ i The angle of parameter change (such as the angle between the temperature adjustment direction and the target curve) is used to quantify the disorder of the path; it suppresses sudden temperature changes and drastic humidity fluctuations, reducing the number of over-adjustments by 42%.

[0093] A genetic algorithm is set up, wherein 80% of the initial population of the genetic algorithm comes from the prediction result Y = [ΔT]. opt ΔH opt ], 20% randomly generated; Fitness function: Fitness = 0.7Q - 0.2E - 0.1T, where Q is the quality score: based on color difference (ΔE < 5) and moisture content uniformity (σW < 2%); E is the system energy consumption: target value ≤ 2.8 kWh / kg; T is the time deviation rate: allow ±10% fluctuation; Termination condition: Pareto front hypervolume change rate < 1% or more than 50 iterations to ensure convergence efficiency.

[0094] This step is related to the previous steps in the following way: node and edge data: the process network node (v) generated in step 1. i ) and edge (e jThe search space of the A* algorithm is defined as follows: When a mutation (such as ΔW>5%) is detected in step S101 and a new node is generated, the A* algorithm recalculates the path entropy and prioritizes the path with low fluctuation.

[0095] Weights are updated in real time: ω calculated in step S102 j (t) directly affects the path cost of the A* algorithm; high-weight edges (ω) j >0.7) is selected first; Sensitivity linkage: If the real-time sensitivity factor η in step S102 increases to 0.8 (when mutation occurs), the quality weight Q in the fitness function of the genetic algorithm is simultaneously strengthened.

[0096] Initial population generation: ΔT output from step S103 opt and ΔH opt As a baseline value for the initial population of 80% in the genetic algorithm, it accelerates convergence; Feature data support: Color feature C∈R 12 (Step S103 Input) is used to calculate the color difference item in the quality score Q.

[0097] For example, if a mutation is detected and the Lab value b*>40 (severe browning risk); Step S104 response: The A* algorithm selects a cooling path (T: 42℃→38℃), and the path entropy constraint suppresses the sudden temperature drop; The genetic algorithm converges after 28 generations of iteration, and the fitness is increased by 18% to reduce the Lab value.

[0098] Step S104 organically combines dynamic process diagrams, real-time weights, and multimodal prediction (step S103) using a hybrid optimization algorithm to achieve full-process optimization of tobacco curing. Its core value lies in: global-local equilibrium: the A* algorithm covers a wide search range, while the genetic algorithm provides fine-tuning; fluctuation suppression: path entropy constraints reduce parameter mutations, decreasing overtuning by 87%; and collaborative efficiency: the initial population is generated based on prediction results, improving convergence speed by 37.5%. This step is the technical hub of the entire method, ensuring optimal performance across multiple objectives: quality, energy consumption, and time.

[0099] In a further embodiment, the control of real-time process parameters through fuzzy logic control includes: setting typical working conditions for tobacco curing based on the physiological characteristics of tobacco leaves; processing the optimal process parameters by inputting a membership function according to the typical working conditions; outputting a membership function to match the optimal process parameters with the control precision of the curing equipment; setting a preset time period to perform an inference once; and converting the output optimal process parameters into equipment control commands.

[0100] The typical operating conditions for tobacco curing can be obtained based on expert database rules or historical curing data. When there is a deviation between the fuzzy rule suggestions in the typical operating conditions and the optimization results, Y = [ΔT] will be executed first. opt ΔH optSimultaneously, the actual execution values ​​ΔT and ΔH of the fuzzy control are fed back to the weight formula in step S102. When the fuzzy control detects an execution error (such as a temperature deviation > 1℃), the η value in step S102 is increased from 0.3 to 0.8, accelerating the weight update. The execution results of the fuzzy control (such as the actual temperature and humidity change curve) serve as training data for the multimodal model in step S103, improving prediction accuracy. Furthermore, the execution results of the fuzzy control executed by the device can be added to the expert rule base to form dynamic knowledge accumulation.

[0101] like Figure 2 As shown, the process of extracting features from the tobacco leaves to be cured in order to match the optimal process parameters and the optimal process path for curing includes the following steps:

[0102] S1051. Extract features from the tobacco leaves to be cured, wherein the features of the tobacco leaves to be cured include: F = [W, C, S, B], where W is the moisture content, C is the color feature, S is the near-infrared spectral data, and B is the biochemical index;

[0103] S1052. Based on the above features, the dynamic process diagram is expanded and then the process is matched in the training model to generate a dynamic baking scheme, and the baking of the tobacco leaves to be tested begins.

[0104] In a further embodiment, process matching includes:

[0105]

[0106] Where, γ=[ω w ω c ω s ω b The weight allocation corresponds to W, C, S, and B, emphasizing the dominant role of moisture content and color. Several historical schemes are selected, and schemes with a similarity Sim value higher than the predetermined value are used as candidate benchmarks.

[0107] Generating dynamic baking schemes involves using transfer learning formulas:

[0108]

[0109] Among them, Sim m For the current tobacco leaf characteristics F new With historical craftsmanship F m The weighted cosine similarity (values ​​from 0 to 1) can be calculated using the following formula:

[0110]

[0111] Among them, F new =[W new C new S new Bnew ], F m =[W m C m S m B m ]: The characteristic fingerprints of historical craftsmanship;

[0112] P m For the parameter set of similar processes in the preset group in the historical process scheme library, ΔP RL The incremental adjustment amount generated by Q-learning is updated according to the following rules:

[0113] Q(s, a)←Q(s, a)+0.2[r+0.9maxQ(s′, a′)-Q(s, a)];

[0114] Reward function h = 0.7Q new -0.2E+0.1(1-CVD), Q new The quality score for the new process is based on color difference and moisture content uniformity, while for CVD it is based on constraint violation, such as the number of temperature exceedances and the extent of humidity deviations.

[0115] Historical process weighting: Historical solutions with a similarity Simm ≥ 0.7 are prioritized for reuse, reducing trial-and-error costs (examples show a 68% reduction in process generation time). Feature matching: Color features (C∈R12) and near-infrared spectroscopy (S∈R6) are weighted at 60% (γ2+γ3=0.4+0.2) to ensure that key quality indicators dominate the migration direction. Q-learning increment: ΔP RL The optimization is achieved through exploration and balance generation, for example: Action space: adjusting temperature and humidity setpoints (±2℃ / ±5%RH), extending the yellowing period (±4h); State space: real-time monitoring values ​​of tobacco leaf characteristics F=[W, C, S, B]. Reward mechanism: The quality weight of 70% is consistent with the fitness function (0.7Q) in step S4, ensuring a unified global optimization objective. CVD penalty: If the new process violates equipment safety thresholds (e.g., temperature > 70℃ or humidity > 95%RH), a penalty is applied to the 0.1(1-CVD) term in the reward function, forcing the strategy to converge to the feasible region.

[0116] The following describes the technical solution of this invention using a tobacco variety as an example:

[0117] Example 1

[0118] Initial operating conditions: Tobacco variety: Henan Xuchang Yunyan 87; Curing stage: Color fixing period (target parameters: temperature 52℃±2℃, humidity 65%RH±5%); Initial parameters: temperature T=52℃, humidity H=65%RH, moisture content W=72% (target range 70-75%), color characteristics C=[R=110, G=90, B=75, a* =3, b * =18] (Low risk of browning);

[0119] Step S101: Dynamic process diagram construction

[0120] Node state: Current node v i =(52℃, 65%RH, 72%, C);

[0121] No mutations (ΔW = 1.2% < 5%, ΔC = 4% < 15%), no node splitting is triggered.

[0122] Process path selection: along the historically optimal path (node ​​sequence: yellowing period 38℃ → color fixing period 52℃ → drying period 68℃).

[0123] Step S102: Dynamic weight calculation

[0124] Weighting formula: ω j (t)=3.0·e-0.3×60+0.3·(0.6×1.2+0.4×4)=0.55+0.3×3.52=1.61;

[0125] Parameter characteristics: Real-time sensitivity factor η = 0.3 (no mutation, low sensitivity mode); historical weight dominates (accounting for 0.55 / 1.61≈34%).

[0126] Step S103: Multimodal data fusion modeling

[0127] Input features:

[0128] X=[52℃, 65%RH, 72%, C, S, B]∈R 23 ,in:

[0129] Near-infrared spectroscopy (S) showed that the starch content was stable (4.5%).

[0130] Biochemical index B = [starch% = 4.5, chlorophyll a / b = 0.6].

[0131] Model prediction: LightGBM output Y = [ΔT] opt = +0.5℃, ΔH opt =-2%RH).

[0132] Step S104: Hybrid Optimization Algorithm

[0133] A* Algorithm Path Verification: Path cost f(n) = 1.61 × 60 + 0 + 0.3 × 0.8 = 96.9, confirming that the current path is the historical best;

[0134] Genetic algorithm fine-tuning: Initial population: 80% based on predicted value ΔT opt= +0.5℃, 20% random perturbation (±0.2℃ / ±1%RH); converged after 15 iterations (hypervolume change rate = 0.8% < 1%), fitness = 85 → 86 (improvement of 1.2%).

[0135] Fuzzy logic control

[0136] Rule matching: Trigger steady-state rule: "ΔW∈[-1%,+1%] maintain current temperature and humidity", which can be derived from a database of typical operating conditions and provided by expert experience;

[0137] Results: Actual temperature adjustment ΔT = +0.3℃ (target +0.5℃, equipment error ±0.2℃); humidity maintained at 65%RH (target -2%RH, but retained due to equipment accuracy).

[0138] Step S105: Adaptive process feature fingerprint generation and matching: Matching to historical scheme P2 (Sim = 0.92), directly reusing process parameters; Q-learning increment: reward function r = 0.7 × 0.95 - 0.2 × 2.5 + 0.1 × 1 = 0.665 - 0.5 + 0.1 = 0.265; generating increment ΔP RL = +0.1℃ / h (fine-tuning the heating rate).

[0139]

[0140]

[0141] Example 2

[0142] In case of emergencies: Initial operating conditions include

[0143] Tobacco variety: Yunnan KRK26; Curing stage: Yellowing stage (target duration 40h); Initial parameters: Temperature T = 38℃, Humidity H = 85%RH, Moisture content W = 82%, Color characteristics C = [R = 120, G = 95, B = 80, H = 45°, S = 0.6, L = 50, a * =5, b * =25, GLCM contrast ratio=18];

[0144] Mutation event: ΔW = +6.5% and ΔC = 18% (accelerated browning) were detected within 2 hours.

[0145] Step S101: Dynamic process diagram construction

[0146] Node splitting condition triggered: ΔW = 6.5% > 5%, generating a set of child nodes:

[0147] v sub={(36℃, 80%RH), (40℃, 90%RH), (34℃, 75%RH), (42℃, 95%RH)}

[0148] Parameters based on: ΔT = 2℃, ΔH = 5%RH, k = 1, 2.

[0149] Step S102: Dynamic weight calculation

[0150] Weight update: Initial weight ω j0 =3.0; Real-time sensitivity factor η = 0.8 (because |ΔW| > 3%); Calculate edge weight: ω j (t)=3.0·e-0.3×120+0.8·(0.6×6.5+0.4×18)=0.15+0.8×10.2=8.31;

[0151] Step S103: Multimodal data fusion modeling

[0152] Input characteristics: X = [38℃, 85%RH, 82%, C, S, B] ∈ R23; Near-infrared spectrum S = [0.32@950nm, 0.45@1450nm, ...] (starch content inversion value = 6.8%); Biochemical index B = [starch% = 6.8, chlorophyll a / b = 0.7, cell wall thickness = 14μm].

[0153] Model prediction: LightGBM output Y = [ΔT] opt = -1.8℃, ΔH opt = +3%RH).

[0154] Step S104: Hybrid Optimization Algorithm

[0155] A* Algorithm Path Selection: Path cost f(n) = 18.7; Selected path T = 36℃, H = 80%RH.

[0156] Genetic Algorithm Optimization:

[0157] Initial population: 80% based on ΔT opt = -1.8℃, 20% randomly generated;

[0158] After 35 iterations, the algorithm converged, and the fitness improved from 60 to 74 (+23%).

[0159] Fuzzy logic control

[0160] Rule Trigger: Matching rule: ΔW = +5%, ΔT = -1.5℃; Actual Execution: ΔT = -1.8℃, ΔH = +3%RH. Control Effect: Temperature dropped to 36.2℃ (error ±0.2℃) within 2 hours, and humidity was adjusted to 88%RH.

[0161] Step 5: Adaptive Process Generation

[0162] Feature fingerprint matching: The similarity to the historical Yunnan KRK26 scheme was calculated to Sim = 0.76;

[0163] Q-learning incremental generation: reward function r=0.7×0.89-0.2×2.6+0.1×(1-0)=0.203; generation increment ΔPRL=+0.4℃ / h.

[0164] Final process scheme: yellowing period extended to 44h, stage temperature control curve: 36℃→40℃ (heating rate 0.4℃ / h).

[0165] Effect comparison:

[0166]

[0167] Dynamic process diagram expansion: Newly added sub-nodes (36℃, 80%RH) are stored in the historical process library;

[0168] Model iteration optimization: Actual execution data is fed back to step 3, and the LightGBM model error rate is reduced from 5% to 3%;

[0169] Rule base update: The heating rate strategy generated by Q-learning has been transformed into a new fuzzy rule: "IF starch% > 6% THEN heating rate ≤ 0.5℃ / h".

[0170] Conclusion: This embodiment fully demonstrates the coordinated execution of the six steps, and quantitative data verifies the method's effectiveness in improving quality (browning). 40% reduction in energy consumption 19%) and efficiency optimization (time) Significant effects in 68% of cases.

[0171] like Figure 3 As shown, a system for optimizing tobacco curing processes according to a specific embodiment of the present invention is described. The system includes:

[0172] The data preprocessing module 201 is used to construct a dynamic process diagram with multi-dimensional nodes based on historical data of tobacco curing process. The dynamic process diagram represents the state of tobacco leaves and processing technology at any processing time.

[0173] The weight calculation module 202 is used to calculate the weight of the processing technology at any time based on the tobacco leaf modal data and processing technology at any time in the real-time dynamic process diagram, and make real-time adjustments to the processing technology at that time.

[0174] Model building module 203 is used to integrate multiple modal data of the real-time tobacco leaf and build a training model, which is used to output the optimal process parameters.

[0175] The parameter adjustment module 204 and the parameter adjustment module 20 are used to search for the optimal process path globally based on the optimal process parameters output by the training model, while suppressing parameter fluctuations and controlling the real-time process parameters through fuzzy logic control.

[0176] The process generation module 205 is used to extract features from the tobacco leaves to be cured in order to match the optimal process parameters and the optimal process path for curing.

[0177] Figure 4 The hardware structure diagram of a computing device 30 for a method of optimizing tobacco curing process is shown according to an embodiment of this specification. The computing device 30 may include at least one processor 301, a memory 302 (e.g., non-volatile memory), a main memory 303, and a communication interface 304, and the at least one processor 301, memory 302, main memory 303, and communication interface 304 are connected together via a bus 305. At least one processor 301 executes at least one computer-readable instruction stored or encoded in the memory 302.

[0178] It should be understood that the computer-executable instructions stored in memory 302, when executed, cause at least one processor 301 to perform the above-described combinations in the various embodiments of this specification. Figures 2-4 The description includes various operations and functions.

[0179] In the embodiments of this specification, the computing device 30 may include, but is not limited to: personal computer, server computer, workstation, desktop computer, laptop computer, notebook computer, mobile computing device, smartphone, tablet computer, cellular phone, personal digital assistant (PDA), handheld system, messaging device, wearable computing device, consumer electronic device, etc.

[0180] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0181] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0182] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0183] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0184] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0185] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for optimizing tobacco leaf curing process, characterized in that, The method includes: A dynamic process diagram with multi-dimensional nodes is constructed based on historical data of tobacco curing process. The dynamic process diagram represents the state of tobacco leaves and processing technology at any processing time. Based on the tobacco leaf modal data and processing technology at any time in the real-time dynamic process diagram, the weight of the processing technology at that time is calculated, and the processing technology at that time is adjusted in real time. The real-time tobacco leaf-like modal data are integrated and a training model is established. The training model is used to output the optimal process parameters. The optimal process path is searched globally based on the optimal process parameters output by the trained model, while suppressing parameter fluctuations and controlling the real-time process parameters through fuzzy logic control. Feature extraction is performed on the tobacco leaves to be cured to match the optimal process parameters and the optimal process path for curing.

2. The method for optimizing tobacco curing process according to claim 1, characterized in that, The steps for constructing the dynamic process diagram include: Construct a data structure including tobacco leaf temperature T, humidity H, moisture content W, and color features C∈R 12 The process node network, wherein: Node definition: v i =(T i H i W i C i ); Definition of an edge: e j =(Δt) j ω j (t)), Δt j For process operation time, ω j (t) represents the dynamic weight; node splitting is introduced to address unexpected situations during the baking process, and the node splitting conditions include: When ΔW > 5% or ΔC > 15% within a preset time period, a child node set v is generated. sub ={(T±kΔT,H±kΔH,W,C)}, where ΔT and ΔH are temperature adjustment and humidity adjustment, and k=1 or 2.

3. The method for optimizing tobacco leaf curing process according to claim 2, characterized in that, When the node corresponding to the tobacco leaf data splits, the weight of its corresponding process will change. The edge weights are dynamically adjusted based on spatiotemporal coupling, including: ω j (t)=ω j0 ·e -λt +η·(αΔW+βΔC), where λ is the historical weight decay coefficient to ensure that the influence of old data continues to decrease every hour; the real-time sensitivity factor η∈[0.3, 0.8], which is generated through node splitting and increases to 0.8 during mutation; α and β are the weight coefficients of moisture content and color change.

4. The method for optimizing tobacco curing process according to claim 1, characterized in that, The real-time tobacco leaf-like multimodal data includes two dimensions of temperature and humidity data, one dimension of moisture content data, 12 dimensions of tobacco leaf image data, six dimensions of near-infrared spectrophotometry data, and three dimensions of biochemical index data. Texture features are obtained and normalized by using ROI automatic segmentation and median filtering for noise reduction. The 23-dimensional vector X = [T, H, W, C] is then input into the training model. 1-12 S 1-6 B 1-3 The output target is the predicted adjustment process parameter Y = [ΔT]. opt ΔH opt The training strategy for the real-time training model uses time series for cross-validation, emphasizes the priority of temperature and humidity control by establishing a custom function, and generates adjustment suggestions for process parameters in real-time inference over a certain period of time.

5. The method for optimizing tobacco leaf curing process according to claim 4, characterized in that, The step of globally searching for the optimal process path based on the optimal process parameters output by the trained model, while simultaneously suppressing parameter fluctuations, includes: Setting path cost functions includes: Dynamic weight ω j The real-time calculated values ​​from the preceding steps reflect the historical reliability and current sensitivity of the path. Path entropy constraint Entropy(P): θ i The angle of parameter change is used to quantify the disorder of the path, to suppress sudden temperature changes and drastic humidity fluctuations, and to reduce the number of overshoots; A genetic algorithm is set up, wherein 80% of the initial population of the genetic algorithm comes from the prediction result Y = [ΔT]. opt ΔH opt ], 20% are randomly generated; its fitness function is: Fitness = 0.7Q - 0.2E - 0.1T; Where Q is the quality score, adjusted based on color difference and moisture content uniformity; E is the system energy consumption, with a target value of ≤2.8kWh / kg; T is the time deviation rate, allowing ±10% fluctuation; optimization termination condition: the hypervolume change rate of the Pareto front is less than the predetermined value or the preset number of iterations, ensuring convergence efficiency.

6. The method for optimizing tobacco curing process according to claim 1, characterized in that, The overfuzzy logic control enables the regulation of real-time process parameters, including: Based on the physiological characteristics of tobacco leaves, typical operating conditions for tobacco curing are set. The optimal process parameters are processed by the membership function input according to the typical operating conditions, and the output membership function is used to match the optimal process parameters with the control accuracy of the curing equipment. An inference is performed once within a preset time period, and the output optimal process parameters are converted into equipment control commands.

7. The method for optimizing tobacco curing process according to claim 6, characterized in that, The process of extracting features from the tobacco leaves to be cured in order to match the optimal process parameters and the optimal process path for curing includes the following steps: Feature extraction is performed on the tobacco leaves to be cured. The features of the tobacco leaves to be cured include: F = [W, C, S, B], where W is the moisture content, C is the color feature, S is the near-infrared spectral data, and B is the biochemical index. Based on the above features, the dynamic process diagram is expanded, and then the process is matched in the training model to generate a dynamic baking scheme, and the baking of the tobacco leaves to be tested begins.

8. The method for optimizing tobacco curing process according to claim 7, characterized in that, The process matching includes: Where, γ=[ω w ω c ω s ω b The weights are assigned to W, C, S, and B, emphasizing the dominant role of moisture content and color. Several historical schemes are selected, and those with a similarity Sim value higher than a predetermined value are used as candidate benchmarks.

9. The method for optimizing tobacco leaf curing process according to claim 8, characterized in that, The generation of the dynamic baking scheme includes the use of a transfer learning formula: Among them, Sim m For the current tobacco leaf characteristics F new With historical craftsmanship F m The weighted cosine similarity (values ​​from 0 to 1) can be calculated using the following formula: Among them, F new =[W new C new S new B new ] represents the characteristic fingerprint of the current process, F m =[W m C m S m B m [This is] a characteristic fingerprint of historical craftsmanship; P m For the parameter set of similar processes in the preset group in the historical process scheme library, ΔP RL The update rule for the incremental adjustments generated by Q-learning is as follows: Q(s, a)←Q(s, a)+0.2[r+0.9maxQ(s′, a′)-Q(s, a)]; Reward function h = 0.7Q new -0.2E+0.1(1-CVD), Q new The quality score for the new process is given, and CVD represents the degree of constraint violation.

10. A tobacco leaf curing process optimization system, characterized in that, The system includes: The data preprocessing module is used to construct a dynamic process diagram with multi-dimensional nodes based on historical data of tobacco curing process. The dynamic process diagram represents the state of tobacco leaves and processing technology at any processing time. The weight calculation module is used to calculate the weight of the processing technology at any given time based on the tobacco leaf modal data and processing technology in the real-time dynamic process diagram, and to make real-time adjustments to the processing technology at that time. The model building module is used to integrate multiple modal data of the real-time tobacco leaf and build a training model, which is used to output the optimal process parameters. The parameter adjustment module is used to search for the optimal process path globally based on the optimal process parameters output by the training model, while suppressing parameter fluctuations and controlling the real-time process parameters through fuzzy logic control. The process generation module is used to extract features from the tobacco leaves to be cured in order to match the optimal process parameters and the optimal process path for curing.

Images