Method for establishing heat and humidity transfer and air flow model of uniform-temperature bulk curing barn

Abstract

The invention provides a method for establishing a heat and humidity transfer and air flow model of a uniform-temperature bulk curing barn, and belongs to the technical field of tobacco curing. The method comprises the following steps: firstly, constructing a basic framework of a multi-resolution neural network, and collecting model input parameters through an experimental device; then constructing a heat transfer sub-model, a tobacco leaf moisture migration sub-model and an air flow sub-model by utilizing a heat and mass transfer principle, and integrating to form a multi-physics coupling neural network model; establishing a training data set by utilizing experimental data and computational fluid mechanics simulation data, training a model by applying a deep learning algorithm, and establishing a function relational expression between a neural network weight coefficient and the rotating speed of the circulating fan; and finally, based on the trained model, optimizing circulating fan parameters and position configuration of a return port, realizing uniform distribution of a temperature field and a humidity field in the tobacco loading chamber, outputting an optimal circulating fan parameter configuration scheme, and effectively solving the technical problem of non-uniform distribution of the temperature field, the humidity field and airflow in the curing barn.

Description

Technical Field

[0001] The present invention belongs to the technical field of tobacco leaf curing, and in particular relates to a method for establishing a heat and moisture transfer and air flow model in a uniform temperature intensive curing barn. Background Art

[0002] Flushing is a critical step in tobacco processing. The temperature control performance and temperature uniformity during the technical execution process directly impact the final quality of the tobacco leaves. Traditional flue-curing barn technology relies primarily on subjective human control, employing fixed heating patterns and ventilation designs, combined with traditional empirical judgment to adjust barn parameters. With the advancement of tobacco processing equipment and technology, automatic control systems based on temperature and humidity sensors have emerged, enabling real-time monitoring of the flue-curing barn environment and enabling simple parameter adjustments. In recent years, researchers have begun experimenting with computational fluid dynamics (CFD) methods to simulate the internal flue-curing barn environment in order to optimize barn structure and control strategies.

[0003] However, traditional flue-curing barn control technology has obvious defects: empirical control relies heavily on the subjective judgment of operators, making it difficult to achieve precise regulation; automatic control systems based on simple sensors can only obtain monitoring data from limited locations and cannot fully grasp the complex heat and humidity distribution inside the flue-curing barn; and although computational fluid dynamics simulation is theoretically complete, it has high computational complexity and is difficult to achieve real-time control, and the model construction and solution process lacks accurate assessment of the quality and baking characteristics of fresh tobacco leaves.

[0004] Therefore, the core challenge facing existing technologies is how to effectively establish a model that can accurately describe the temperature, humidity, and airflow distribution within the flue-curing barn, and based on this model, achieve uniformity control of the flue-curing barn environment. Especially under high-density tobacco loading conditions, there is a complex coupling relationship between heat transfer, moisture migration, and air flow within the flue-curing barn. Existing technologies have difficulty effectively capturing this multi-physics coupling characteristic, resulting in uneven distribution of temperature, humidity, and airflow within the flue-curing barn, which affects the stability and consistency of tobacco leaf curing quality. In other words, existing technologies suffer from the technical problem of uneven distribution of temperature, humidity, and airflow within the flue-curing barn. Summary of the Invention

[0005] In view of this, the present invention provides a method for establishing a heat and moisture transfer and air flow model in a uniform temperature dense flue-curing room, which can solve the technical problems of uneven temperature and humidity fields and airflow distribution in the flue-curing room in the prior art.

[0006] The present invention is implemented as follows: The present invention provides a method for establishing a heat and moisture transfer and air flow model in a uniform temperature intensive flue-curing barn, including: constructing a multi-resolution neural network basic architecture; collecting model input parameters such as the initial moisture content of tobacco leaves and tobacco filling density; constructing a heat transfer sub-model; constructing a tobacco leaf moisture migration sub-model; establishing an air flow sub-model; integrating the three sub-models to form a multi-physics field coupled neural network model; establishing a model training data set; and training the multi-physics field coupled neural network model to obtain a heat and moisture transfer and air flow model.

[0007] Among them, the basic architecture of the multi-resolution neural network refers to a neural network architecture based on the heat transfer equation, the mass transfer equation, and the momentum transfer equation, and a mathematical model structure is constructed by dividing the smoke chamber area into multiple control volumes.

[0008] Among them, the multi-resolution neural network refers to a deep learning model with a multi-level structure, which can simultaneously capture physical features of different scales. It includes a high-resolution layer for obtaining local detail information and a low-resolution layer for obtaining global feature information, and realizes the fusion of features of different resolutions through jump connections.

[0009] The heat transfer equation is a mathematical expression that describes the conservation of energy within the system. The input includes temperature gradient, thermal conductivity, convection heat transfer coefficient, radiation coefficient, and heat source intensity. The output is the temperature distribution and heat flux density at each point in the system.

[0010] The mass transfer equation is a mathematical expression that describes the material migration process. The input includes concentration gradient, diffusion coefficient, mass source term, and flow velocity. The output is the concentration distribution and mass transfer rate of each point in the system.

[0011] The momentum transfer equation is a mathematical expression that describes the motion state of a fluid. The input includes pressure gradient, viscosity coefficient, volume force, and boundary conditions, and the output is the velocity distribution and pressure distribution of each point in the system.

[0012] The heat transfer sub-model is constructed using the principles of heat and mass transfer to characterize the effects of temperature gradient, convective heat transfer coefficient, and heat transfer area on the heat transfer process.

[0013] The tobacco leaf moisture migration sub-model includes a moisture diffusion equation, an evaporation-condensation equation, and a phase change latent heat equation, and calculates the rate at which moisture in the tobacco leaf migrates to the surface and evaporates into the air.

[0014] The air flow sub-model is established based on the principles of fluid mechanics, combining the continuity equation, momentum equation, and energy equation to simulate the distribution of air pressure field, velocity field, and temperature field in the smoke chamber.

[0015] It also includes the steps of establishing a functional relationship between the neural network weight coefficient and the circulation fan speed; and optimizing the circulation fan parameters and the return port position configuration to achieve uniform distribution of temperature and humidity fields in the smoke chamber.

[0016] The present invention divides the tobacco loading chamber into multiple control volumes and integrates three sub-models: heat transfer, moisture migration, and air flow. This constructs a complete multi-physics coupling model, achieving an accurate description of the internal environment of the flue-curing barn. This method cleverly combines physical mechanism constraints with deep learning technology, maintaining the correctness of physical laws while improving the computational efficiency of the model. Compared with traditional methods, the present invention overcomes the subjectivity and inaccuracy of empirical control. Through the multi-physics coupling model, it can comprehensively describe the complex internal environment of the flue-curing barn. At the same time, it avoids the defects of limited information acquisition of simple sensor systems and achieves a global prediction of the temperature, humidity, wind speed, and pressure fields in the tobacco loading chamber. In addition, this method is more computationally efficient than pure computational fluid dynamics simulation, can support real-time control decisions, and ensures that the model prediction results conform to physical laws through the physical constraint loss function. By optimizing the circulation fan parameters and the position configuration of the return air outlet, the present invention effectively solves the technical problems of uneven temperature, humidity, and airflow distribution in the flue-curing barn, improving the quality stability and consistency of the tobacco leaf curing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0018] 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.

[0019] like Figure 1 FIG. 1 is a flow chart of a method for establishing a heat and moisture transfer and air flow model for a uniform temperature intensive baking room provided by the present invention. The method includes the following steps:

[0020] S01. Construct a mathematical model structure, divide the smoke chamber area into multiple control volumes, and establish a basic multi-resolution neural network architecture based on the heat transfer equation, mass transfer equation, and momentum transfer equation;

[0021] S02. Using the experimental device, measure and collect the initial moisture content of tobacco leaves, tobacco packing density, ambient temperature, ambient humidity, flue-curing room structural parameters, heating power, and circulating fan operating parameters as model input parameters;

[0022] S03. Using the principles of heat and mass transfer, a sub-model of heat transfer between tobacco leaves and air within a controlled volume was constructed to characterize the effects of temperature gradient, convective heat transfer coefficient, and heat transfer area on the heat transfer process.

[0023] S04. Construct a tobacco leaf moisture migration sub-model, including the moisture diffusion equation, the evaporation-condensation equation, and the phase change latent heat equation, to calculate the rate of moisture migration from the tobacco leaf to the surface and evaporation into the air;

[0024] S05. Based on the principles of fluid mechanics, an air flow sub-model is established. Combining the continuity equation, momentum equation, and energy equation, the distribution of air pressure field, velocity field, and temperature field in the smoke chamber is simulated.

[0025] S06. Integrate the heat transfer sub-model, moisture migration sub-model, and air flow sub-model to build a complete heat and moisture transfer and air flow multi-physics field coupled neural network model;

[0026] S07. Using experimental data and computational fluid dynamics software Fluent simulation data, a model training data set was established, including the distribution of temperature field, humidity field, wind speed field, and pressure field in the smoke chamber under different working conditions;

[0027] S08. Using a deep learning algorithm to train the multi-physics field coupled neural network model to obtain a heat and moisture transfer and air flow model, using a loss function to optimize model parameters, and establishing a functional relationship between the neural network weight coefficient and the circulation fan speed;

[0028] S09. Optionally, it also includes optimizing the air volume, air pressure parameters and return port position configuration of the circulation fan based on the trained model to achieve uniform distribution of temperature and humidity fields in the smoke chamber and output the optimal circulation fan parameter configuration plan.

[0029] The control volume refers to a number of tiny space units into which the smoke chamber space of the flue-curing barn is divided according to a certain grid division method. The physical quantities in each control volume are considered to be uniformly distributed, and matter and energy are exchanged between the control volumes through the interface.

[0030] Among them, the multi-resolution neural network refers to a deep learning model with a multi-level structure, which can simultaneously capture physical features of different scales. It includes a high-resolution layer for obtaining local detail information and a low-resolution layer for obtaining global feature information, and realizes the fusion of features of different resolutions through jump connections.

[0031] The heat transfer equation is a mathematical expression that describes the conservation of energy within the system. The input includes temperature gradient, thermal conductivity, convection heat transfer coefficient, radiation coefficient, and heat source intensity. The output is the temperature distribution and heat flux density at each point in the system.

[0032] The mass transfer equation is a mathematical expression that describes the material migration process. The input includes concentration gradient, diffusion coefficient, mass source term, and flow velocity. The output is the concentration distribution and mass transfer rate of each point in the system.

[0033] The momentum transfer equation is a mathematical expression that describes the motion state of a fluid. The input includes pressure gradient, viscosity coefficient, volume force, and boundary conditions, and the output is the velocity distribution and pressure distribution of each point in the system.

[0034] The temperature gradient refers to the rate of change of temperature in space, which is measured by the thermocouple array in the experimental device and is expressed in Kelvin per meter.

[0035] The diffusion coefficient refers to a measure of the diffusion ability of a substance molecule in a medium, obtained by experimental measurement, and is expressed in square meters per second.

[0036] The mass source term refers to the amount of matter produced or disappeared in the control body per unit volume per unit time, which is calculated by the evaporation and condensation equation and is expressed in kilograms per cubic meter per second.

[0037] The flow velocity refers to the movement rate of a fluid element in space, which is measured by a hot wire anemometer and is expressed in meters per second.

[0038] The pressure gradient refers to the rate of change of pressure in space, which is measured by a micro differential pressure gauge and is expressed in Pascals per meter.

[0039] The viscosity coefficient refers to the ability of a fluid to resist deformation, which is obtained from a fluid physical property database and is expressed in Pascal seconds.

[0040] The volume force refers to the force acting on a unit volume of the fluid, including gravity, which is calculated by multiplying the acceleration due to gravity by the density, and the unit is Newton per cubic meter.

[0041] Among them, the boundary conditions refer to mathematical expressions that describe the values of physical quantities on the model boundary, including the wall temperature of the smoke chamber, the air velocity of the air supply outlet, the air supply outlet temperature, the air supply outlet humidity, and the exhaust outlet pressure, which are measured by the sensors of the experimental device.

[0042] Among them, the functional relationship refers to a mathematical expression that describes the corresponding relationship between the neural network weight coefficient and the circulation fan speed. The input is the circulation fan speed, and the output is the weight coefficient matrix of the key layer in the neural network.

[0043] Among them, the specific structure of the multi-physics field coupled neural network model is a U-shaped network architecture including an encoder and a decoder. The encoder is composed of multiple downsampling convolution modules, each of which includes a convolution layer, a batch normalization layer and an activation function layer for extracting features of different scales; the decoder is composed of multiple upsampling deconvolution modules, each of which includes a deconvolution layer, a batch normalization layer and an activation function layer for restoring features to the original resolution; the middle layer of the model embeds a physical mechanism equation constraint module, including a heat transfer constraint module, a mass transfer constraint module and a momentum transfer constraint module, which guides the learning process of the neural network through physical knowledge; the jump connection structure of the network fuses the feature maps of the corresponding levels of the encoder and decoder, retaining high-frequency detail information; the final output layer of the network is a multi-channel structure, corresponding to the prediction results of the temperature field output channel, the humidity field output channel, the wind speed field output channel and the pressure field output channel respectively; the weight matrix of the third convolution layer of the encoder satisfies a functional relationship with the speed of the circulating fan, and when the speed increases, the value related to the airflow velocity in the weight matrix increases proportionally.

[0044] Among them, the steps of establishing the model training data set specifically include collecting temperature distribution data of a real baking room under different working conditions through experimental measurement, using a thermocouple array to be arranged at key positions in the smoke loading room to measure the temperature; using a humidity sensor array to measure the relative humidity and moisture content at key positions in the smoke loading room; using a hot wire anemometer to measure the wind speed distribution at key positions in the smoke loading room; using a micro differential pressure meter to measure the pressure difference at key positions in the smoke loading room; at the same time, using Fluent software to establish a computational fluid dynamics model of the smoke loading room, and by setting different boundary conditions and initial conditions, simulating the temperature field, humidity field, wind speed field and pressure field distribution in the smoke loading room under different working conditions; dividing the experimental data and simulation data into a training set and a validation set in a ratio of 7 to 3; normalizing the data to eliminate the dimensional differences between different physical quantities; filtering the noise data to improve the data quality; recording the circulation fan speed corresponding to each set of data, and establishing a corresponding relationship between the circulation fan speed and the physical field distribution in the smoke loading room.

[0045] Among them, the steps of model training specifically include initializing the neural network parameters and using the normal distribution initialization method to improve the training efficiency; designing a composite loss function, including a mean square error loss term, a physical constraint loss term and a regularization loss term, wherein the physical constraint loss term is used to ensure that the model prediction results meet the basic laws of the heat transfer equation, the mass transfer equation and the momentum transfer equation; using a batch gradient descent algorithm to update the network parameters; using a learning rate decay strategy, using a larger learning rate to accelerate convergence in the early stage of training, and using a smaller learning rate to fine-tune the parameters in the later stage; implementing an early stopping strategy to prevent model overfitting based on the performance of the validation set; when the loss function value on the validation set no longer decreases or reaches the preset number of iterations, the training process is terminated; using model integration technology, multiple models with different initialization parameters are trained, and their average output is taken as the final prediction result to improve the model generalization ability; for training data with different circulation fan speeds, the changes in the weight coefficients of the key layers in the network are recorded, and the functional relationship between the weight coefficients and the circulation fan speed is obtained by fitting.

[0046] The convective heat transfer coefficient is a parameter that characterizes the intensity of heat transfer between a fluid and a solid surface. It is affected by fluid flow rate, fluid properties, surface shape, and surface roughness, and is expressed in watts per square meter of Kelvin. The phase change latent heat equation is a mathematical expression that describes the heat absorbed or released by a substance during a phase change. Its inputs include the phase change temperature, phase change pressure, and material state parameters, and its output is the heat required for the phase change process. The loss function is a mathematical expression that evaluates the difference between the model's predicted results and the true value. It includes a data-driven loss term and a physical constraint loss term. The data-driven loss term measures the deviation between the predicted value and the measured value, while the physical constraint loss term measures the degree to which the predicted result complies with physical laws. The initial moisture content of tobacco leaves refers to the percentage of moisture contained in the tobacco leaves before entering the flue-curing barn. It is measured by a tobacco moisture sensor and input into the moisture transfer sub-model. The unit is percentage. The loading density refers to the mass of tobacco leaves placed per unit volume in the loading chamber. It is measured by a weighing system and input into the heat transfer sub-model and the moisture transfer sub-model. The unit is kilograms per cubic meter. The ambient temperature refers to the air temperature outside the flue-curing barn, measured by a temperature sensor and used as a model boundary condition in degrees Celsius. The ambient humidity refers to the relative humidity outside the flue-curing barn, measured by a humidity sensor and used as a model boundary condition in percentage. The flue-curing barn structural parameters include the dimensions of the smoke chamber, the layout of the hot air ducts, the location and dimensions of the supply and return air vents, the thickness and thermal conductivity of the insulation material, and are obtained from design drawings and material parameter tables. They are used to establish the geometric boundaries of the air flow sub-model. The heating power refers to the heat provided by the heating system, measured by a power meter, and used as the heat source intensity input for the heat transfer sub-model in kilowatts. The circulating fan operating parameters include fan speed, blade angle, and motor power, measured by the fan control system and sensors, and used as boundary conditions for the air flow sub-model. The fan speed is measured in revolutions per minute, the blade angle is measured in degrees, and the motor power is measured in kilowatts. The temperature field refers to the spatial distribution of temperature at various points within the flue-curing barn. It is calculated jointly by the heat transfer sub-model and the air flow sub-model and is used to assess temperature uniformity and baking quality in the flue-curing barn. The temperature field refers to the spatial distribution of temperature at various points within the flue-curing barn. It is calculated by the heat transfer sub-model and the air flow sub-model. It is used to assess temperature uniformity and baking quality in the flue-curing barn. The temperature field is measured in degrees Celsius. The humidity field refers to the spatial distribution of relative humidity and moisture content at each point in the smoke-loading chamber, which is calculated by the moisture migration sub-model and is used to evaluate the dehumidification effect and baking quality of the baking room, where the unit of relative humidity is percentage and the unit of moisture content is kilograms per kilogram of dry air. The wind speed field refers to the spatial distribution of air flow velocity at each point in the smoke-loading chamber, which is calculated by the air flow sub-model and is used to evaluate the rationality of the airflow organization in the baking room, with the unit being meters per second. The pressure field refers to the spatial distribution of air pressure at each point in the smoke-loading chamber, which is calculated by the air flow sub-model and is used to evaluate the pressure balance in the baking room, with the unit being Pascal. The circulating fan air volume refers to the volume of air delivered by the circulating fan per unit time, which is calculated by the optimization model and is used to guide the actual operating parameter setting of the baking room, with the unit being cubic meters per second.The air pressure parameter refers to the air pressure generated by the circulating fan, calculated by the optimization model and used to guide the actual operating parameter setting of the curing barn. The unit is Pascal. The supply and return air vent location configuration refers to the spatial arrangement of the supply and return air vents in the smoke chamber, calculated by the optimization model and used to guide the structural design improvements of the curing barn. It includes the location coordinates and opening dimensions.

[0047] The specific implementation of the above steps is described in detail below.

[0048] The specific implementation of step S01 is to first divide the smoke chamber space into multiple regular control volumes using a hexahedral meshing method based on the smoke chamber's geometric structure. The size of the control volumes is determined by the chamber's geometric characteristics and is typically a 10-20 cm cube. Based on the control volume principle, mass conservation equations, momentum conservation equations, and energy conservation equations are established for each control volume, forming basic physical constraints. Then, a multi-resolution deep convolutional neural network architecture is designed, consisting of five layers of downsampling and five layers of upsampling. Each downsampling layer uses a convolution operation with a stride of 2, and each upsampling layer uses a transposed convolution. A physical constraint module is embedded in the network's intermediate layers, and the physical mechanism equations are integrated into the network training process using soft constraints. The purpose of this step is to construct a basic deep learning model architecture that can both learn data statistical laws and satisfy physical constraints, laying the foundation for the subsequent accurate simulation of heat and moisture transfer and air flow.

[0049] The specific implementation of step S02 is to collect the input parameters required by the model through a variety of sensors. A capacitive moisture sensor is used to measure the initial moisture content of tobacco leaves, with a measurement range of 10% to 80% and an accuracy of ±1%. A precision electronic scale is used to measure the amount of tobacco and the density of tobacco is calculated in combination with the volume of the tobacco chamber. The density of tobacco is usually controlled at 80 to 120 kilograms per cubic meter. A PT100 platinum resistance temperature sensor is used to measure the ambient temperature, with a measurement range of -20 to 100°C and an accuracy of

[0050] ±0.1℃. A capacitive humidity sensor is used to measure the ambient humidity, with a measurement range of 0 to 100% and an accuracy of ±2%. The structural parameters of the flue-curing barn are obtained from the design drawings and structural material specifications, including the size of the smoke chamber, the layout of the hot air duct, the location and size of the supply and return air outlets, the thickness of the insulation material, and the thermal conductivity coefficient. A power analyzer is used to measure the power of the heating system, with a measurement range of 0 to 50 kilowatts and an accuracy of ±0.5%. A speed sensor, an inclination sensor, and a power meter are used to measure the speed, blade angle, and motor power of the circulating fan, with a speed measurement range of 0 to 3000 revolutions per minute and an accuracy of ±1 revolution per minute. The purpose of this step is to obtain the key parameters that affect heat and moisture transfer and air flow during tobacco leaf curing, and to provide accurate input data and boundary conditions for the model.

[0051] The specific implementation of step S03 is to establish a heat transfer sub-model between tobacco leaves and air based on the basic principles of convective heat transfer. First, the temperature difference between the tobacco leaf surface and the air is calculated, which is used as the driving force for heat transfer. Then, based on the air flow velocity and the geometric characteristics of the tobacco leaf, the convective heat transfer coefficient is calculated using the Dittus-Boelter correlation. For low-velocity areas (0.5 meters per second), the natural convection heat transfer coefficient calculation formula is used. When calculating the effective heat transfer area, the overlap of the tobacco leaves is taken into account and an effective contact coefficient is introduced, typically ranging from 0.6 to 0.8. The temperature difference, convective heat transfer coefficient, and effective heat transfer area are combined to calculate the amount of heat transferred per unit time. For the heat conduction process, Fourier's law of heat conduction is used, taking into account the thermal conductivity of the tobacco leaf itself. The thermal conductivity coefficient is typically 0.2 to 0.4 watts per meter Kelvin. This heat transfer sub-model is discretized using the finite volume method and solved within each time step, which is set to 10 seconds. The purpose of this step is to accurately describe the heat exchange process between tobacco leaves and air, providing a theoretical basis for subsequent temperature field distribution prediction.

[0052] The specific implementation of step S04 is to construct a mathematical model to describe the process of water migration and evaporation into the air in tobacco leaves. First, the water diffusion equation inside tobacco leaves is established, and Fick's second law is used to describe the diffusion process of water in tobacco leaf tissue. The diffusion coefficient is related to temperature and moisture content, and the approximate relationship is an exponential function. The base diffusion coefficient is 10 -10 ~10 -9 Square meters per second. Then, the surface evaporation and condensation equation is established to calculate the evaporation rate based on the moisture concentration difference between the tobacco leaf surface and the air and the mass transfer coefficient. The mass transfer coefficient is usually proportional to the convective heat transfer coefficient, with a proportionality coefficient of 0.9 to 1.1. The phase change latent heat equation is established to calculate the heat absorbed during the evaporation of water. The latent heat value varies with temperature and is approximately 2.3 to 2.4 megajoules per kilogram in the range of 60 to 90°C. These three equations are coupled and solved to obtain the variation pattern of tobacco leaf moisture content with time and space and the evaporation rate to the air. The solution process uses an implicit difference format to ensure numerical stability. The purpose of this step is to accurately simulate the moisture migration phenomenon during the tobacco leaf baking process and lay the foundation for predicting the humidity field distribution in the tobacco loading room.

[0053] The specific implementation of step S05 is to establish a mathematical model based on the Navier-Stokes equations to describe the air flow in the smoke chamber. First, a continuity equation is established to describe the conservation of air mass. Then, a momentum equation is established to describe the dynamic characteristics of air flow, taking into account the effects of pressure gradient, viscosity, and gravity. The air viscosity coefficient is 1.8×10 -5Pascal seconds. An energy equation is established to describe the change in air temperature, taking into account the effects of convection, conduction and internal heat sources. The resistance of tobacco leaves to the airflow is used as the source term in the momentum equation. The resistance coefficient is determined by the loading density and arrangement of the tobacco leaves, and is usually between 1.5 and 2.5. The air model is discretized using the finite volume method, with a grid size of 5 to 10 cm and a time step of 1 second. The boundary conditions include the air velocity, temperature and humidity at the air supply outlet, the return air outlet pressure, and the temperature and heat flow conditions of the wall of the smoke loading chamber. The purpose of this step is to accurately simulate the air flow characteristics in the smoke loading chamber and provide a theoretical basis for predicting the distribution of wind velocity field, pressure field and temperature field.

[0054] The specific implementation of step S06 is to integrate the above three sub-models into a complete multi-physics field coupled neural network model. A loose coupling strategy is adopted, and the heat transfer sub-model, moisture migration sub-model and air flow sub-model are first solved separately in each time step, and then coupling is achieved through information exchange through the interface. In the neural network structure, the underlying features of the encoder are shared, the intermediate layer and the decoder are separated, and three parallel branches are formed corresponding to the three sub-models respectively. In the physical constraint layer, heat conservation constraints, mass conservation constraints and momentum conservation constraints are set, and the loss function is added in the form of penalty terms. A jump connection structure is designed to pass the encoder features directly to the corresponding decoder layer to maintain high-frequency spatial information. Residual connections are used to improve gradient propagation, and two layers of convolution plus one residual connection are used inside each convolution block. An attention mechanism is set to perform weighted fusion of features of different physical fields, and the weight coefficients are obtained through adaptive learning. The purpose of this step is to construct a composite model that conforms to physical laws and has data-driven learning capabilities, which can accurately simulate the complex coupling process of heat and moisture transfer and air flow.

[0055] The specific implementation of step S07 involves establishing a high-quality dataset required for model training. First, a sensor array was placed in the experimental flue-curing barn, including a 32-point thermocouple array (accuracy ±0.5°C), a 16-point humidity sensor array (accuracy ±3%), a 12-point hot wire anemometer (accuracy ±0.1 meters per second), and an 8-point micro-differential pressure gauge (accuracy ±1 Pascal). Experimental measurements were conducted under different operating conditions, with parameters including five tobacco loading densities (80-120 kilograms per cubic meter), four initial moisture contents (65%-80%), three circulating fan speeds (800-1600 revolutions per minute), and three heating powers (15-35 kilowatts), resulting in a total of 60 operating conditions. A three-dimensional model of the tobacco loading barn was constructed using Fluent software, employing a standard k-ε turbulence model. A porous media region was set to simulate the tobacco leaf layer. Boundary conditions were consistent with the experimental conditions, with a grid size of 1-2 million, a solution time step of 1 second, and a total simulation time of 1 hour. The experimental and simulation data were combined and randomly divided into training and validation sets in a ratio of 7 to 3. All physical quantities were normalized to a range of 0 to 1. A moving average filter was used to remove noise from the experimental data with a window width of 5. The purpose of this step was to construct a comprehensive training dataset encompassing various operating conditions, providing a data foundation for subsequent neural network model training.

[0056] The specific implementation method of step S08 is to use a deep learning algorithm to train the multi-physics field coupled neural network model to obtain a heat and moisture transfer and air flow model, and establish a functional association between the neural network parameters and the physical parameters. First, the neural network parameters are initialized using a normal distribution (mean 0, standard deviation 0.01). A composite loss function is designed, including a mean square error loss term (weight 1.0), a physical constraint loss term (weight 0.5), and an L2 regularization loss term (weight 0.0001). The network is trained using a batch gradient descent algorithm with a batch size of 32, an initial learning rate of 0.001, an exponential decay strategy, and a decay rate of 0.9 every 500 steps. During the training process, the model performance is evaluated on the validation set every 50 steps, and early stopping is performed if the performance does not improve after 10 consecutive evaluations. Five model instances with different random initialization parameters are trained, and their average output is taken as the final prediction result.

[0057] The detailed structure of the heat and moisture transfer and air flow model is a multi-scale U-Net architecture and a physical constraint fusion network, which includes an encoding path and a decoding path. The encoding path consists of 5 downsampling blocks, each of which contains two layers of three-dimensional convolutional layers (convolution kernel size is 3×3×3, and padding is 1), a batch normalization layer, and a LeakyReLU activation function (negative slope is 0.2). Downsampling uses maximum pooling with a step size of 2. The decoding path consists of 5 upsampling blocks, each of which contains a three-dimensional deconvolution layer (convolution kernel size is 2×2×2, and step size is 2), two layers of three-dimensional convolutional layers, a batch normalization layer, and a ReLU activation function. The jump connection splices the feature map of the encoding path with the feature map of the corresponding level of the decoding path to retain spatial detail information.

[0058] The model's middle layer embeds a physical constraint module, consisting of three parallel branches: a heat transfer constraint branch, a mass transfer constraint branch, and a momentum transfer constraint branch. The heat transfer constraint branch contains four fully connected layers (with 256, 128, 64, and 32 neurons, respectively). Its inputs are parameters such as temperature gradient, thermal conductivity, and convective heat transfer coefficient, and its output is the heat transfer rate. The mass transfer constraint branch contains three fully connected layers (with 128, 64, and 32 neurons, respectively). Its inputs are parameters such as concentration gradient and diffusion coefficient, and its output is the water transfer rate. The momentum transfer constraint branch contains five fully connected layers (with 512, 256, 128, 64, and 32 neurons, respectively). Its inputs are parameters such as pressure gradient and fluid viscosity coefficient, and its output is the fluid velocity field. The output of the constraint module is fused with the main network features through an attention mechanism, with the attention weight matrix obtained through adaptive learning.

[0059] The model's output layer consists of four prediction heads, one for temperature, one for humidity, one for wind speed, and one for pressure. Each prediction head consists of a 1×1×1 convolutional layer. The model has approximately 12 million parameters, including 5 million for the encoder, 3 million for the physical constraint module, and 4 million for the decoder.

[0060] For the training data of different circulating fan speeds, the weight coefficient changes of the third convolution layer of the encoder in the network are recorded. This layer is mainly responsible for extracting the mesoscale flow characteristics. The functional relationship between the weight coefficient and the circulating fan speed is obtained by least squares fitting, and the goodness of fit R 2 Greater than 0.92. The functional relationship is: the k-th channel value related to the airflow velocity in the weight coefficient matrix W is equal to the reference value W0 multiplied by Where α is the first-order proportionality coefficient (ranging from 0.35 to 0.45), β is the second-order proportionality coefficient (ranging from 0.05 to 0.15), n is the circulating fan speed, and n0 is the reference speed (1200 rpm). When the speed variation exceeds 30%, the nonlinear term β becomes dominant, more accurately describing the turbulent flow characteristics at high speeds. The goal of this step is to train a heat and moisture transfer and air flow model that is both physically consistent and highly accurate. Simultaneously, a quantitative relationship between the neural network parameters and the actual physical control parameters is established, improving the model's physical interpretability and extrapolation capabilities.

[0061] Step S09 is an optional step, and its specific implementation method is to optimize the operating parameters of the circulating fan and the configuration of the return port based on the trained model. First, the optimization target is set as the uniformity index of the temperature field and humidity field, which is defined as minimizing the standard deviation of temperature and relative humidity. The standard deviation threshold of the temperature field is ±1.5°C, and the standard deviation threshold of the humidity field is ±3%. The circulating fan air volume (10-30 cubic meters per second), wind pressure (100-300 Pascals), and the return port position coordinates are used as optimization variables. A Bayesian optimization algorithm is used, with the uniformity index of the temperature field and humidity field as the objective function, a Gaussian process regression model is set as the proxy model, and the acquisition function selects expected improvement. The initial number of sampling points is 10 and the maximum number of iterations is 100. For each set of optimization parameters, the trained model is used to predict the distribution of the temperature field, humidity field, wind speed field, and pressure field in the smoke chamber, and the uniformity index is calculated. The optimization is terminated when the uniformity index improves by no more than 0.1% after 10 consecutive iterations. The final output is the optimal configuration of circulating fan air volume, air pressure parameters, and return port locations, generating a visualization of the physical field distribution within the tobacco loading chamber. This step aims to optimize key operating parameters of the flue-curing barn based on an accurate physical field prediction model, achieving uniform temperature and humidity distribution within the tobacco loading chamber and improving tobacco curing quality.

[0062] Specifically, the core principle of this invention is to construct a multi-physics field coupled neural network model that accurately describes the complex thermal and hygroscopic environment within the curing barn, and to optimize curing barn operating parameters based on this model. This method, based on theoretical foundations derived from heat transfer, mass transfer, and fluid mechanics, accurately describes the physical processes within the curing barn by establishing conservation equations for energy, mass, and momentum within a control volume.

[0063] The key issues that traditional methods struggle to address are the complexity of the multi-physics coupling process and the conflict between model computational efficiency and accuracy. This paper innovatively introduces a multi-resolution neural network architecture, improving computational efficiency while maintaining physical correctness. Specifically, the multi-resolution structure can simultaneously capture physical features at different scales. The high-resolution layer acquires local detail information, while the low-resolution layer acquires global feature information. Skip connections are used to achieve the fusion of features at different resolutions, effectively resolving the problem that traditional single-resolution models struggle to balance local accuracy and global characteristics.

[0064] Another key innovation of this invention is the incorporation of physical mechanism constraints into the neural network training process. By incorporating physical constraint terms into the loss function, the model's predictions are ensured to satisfy the fundamental laws of the heat, mass, and momentum transfer equations, thereby avoiding the risk of violating physical laws inherent in purely data-driven approaches. This physics-informed deep learning approach effectively combines the interpretability of mechanistic models with the efficiency of data-driven models, forming a novel hybrid modeling paradigm.

[0065] Furthermore, the present invention establishes a functional relationship between the neural network weight coefficients and the circulating fan speed, enabling the model to flexibly respond to different operating conditions. This parameterization method not only improves the model's generalization capabilities but also provides a theoretical basis for optimizing operating parameters. By analyzing the influence of circulating fan speed on the network weights, fan parameters can be directly optimized to achieve a uniform distribution of the physical field within the smoke chamber, thereby resolving the technical problem of uneven temperature, humidity, and airflow distribution within the smoke-curing barn.

[0066] 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.

[0067] The specific implementation of step S01 is to first divide the smoke chamber space into multiple regular control volumes based on the smoke chamber's geometric structure using a hexahedral meshing method. The size of the control volume is determined according to the smoke chamber's geometric characteristics and is typically a 10-20 cm cube. Based on the control volume principle, the mass conservation equation, momentum conservation equation, and energy conservation equation are established for each control volume, forming a basic physical constraint relationship. The mass conservation equation is specifically expressed as follows:

[0068]

[0069] Where ρ is the air density in kilograms per cubic meter; t is the time in seconds; is the velocity vector in meters per second; is the divergence operator.

[0070] The momentum conservation equation is specifically expressed as follows:

[0071]

[0072] Where p is the pressure in Pascals; τ is the stress tensor in Pascals; is the gravitational acceleration vector, in meters per second squared; It is the resistance of the porous medium of tobacco leaves to airflow, measured in Newtons per cubic meter.

[0073] The energy conservation equation is specifically expressed as follows:

[0074]

[0075] Where h is the specific enthalpy in joules per kilogram; k is the thermal conductivity in watts per meter Kelvin; T is the temperature in Kelvin; Φ is the viscous dissipation term in watts per cubic meter; S h It is the heat source term, including the heat released or absorbed by tobacco leaves, and its unit is watts per cubic meter.

[0076] Then, a multi-resolution deep convolutional neural network architecture is used to design a U-shaped network structure consisting of 5 layers of downsampling and 5 layers of upsampling. Each downsampling layer uses a convolution operation with a stride of 2, and each upsampling layer uses a transposed convolution. The mathematical expression of the U-shaped network is as follows:

[0077]

[0078] Where, F i is the feature map of the i-th layer encoder, i = 0, 1, 2, 3, 4, F0 is the input data; D is the downsampling function; is the downsampling weight matrix of the i-th layer; G j is the j-th layer decoder feature map, j = 0, 1, 2, 3, 4, G4 = F4; U is the upsampling function; is the j-th layer upsampling weight matrix.

[0079] A physical constraint module is embedded in the middle layer of the network, and the physical mechanism equation is integrated into the network training process using a soft constraint method. The physical constraint term expression is:

[0080]

[0081] Where, L phys is the physical constraint loss; α1, α2, and α3 are weight coefficients, typically ranging from 0.1 to 1.0; ||·||2 represents the L2 norm. The goal of this step is to build a basic deep learning model architecture that can both learn data statistics and satisfy physical constraints, laying the foundation for subsequent accurate simulation of heat and moisture transfer and air flow.

[0082] The specific implementation of step S02 is to collect the input parameters required by the model through a variety of sensors. A capacitive moisture sensor is used to measure the initial moisture content of tobacco leaves, with a measurement range of 10% to 80% and an accuracy of ±1%. A precision electronic scale is used to measure the amount of tobacco and the density of tobacco is calculated in combination with the volume of the tobacco chamber. The density of tobacco is usually controlled at 80 to 120 kilograms per cubic meter. A PT100 platinum resistance temperature sensor is used to measure the ambient temperature, with a measurement range of -20 to 100°C and an accuracy of

[0083] ±0.1℃. A capacitive humidity sensor is used to measure the ambient humidity, with a measurement range of 0 to 100% and an accuracy of ±2%. The structural parameters of the flue-curing barn are obtained from the design drawings and structural material specifications, including the size of the smoke chamber, the layout of the hot air duct, the location and size of the supply and return air outlets, the thickness of the insulation material, and the thermal conductivity coefficient. A power analyzer is used to measure the power of the heating system, with a measurement range of 0 to 50 kilowatts and an accuracy of ±0.5%. A speed sensor, an inclination sensor, and a power meter are used to measure the speed, blade angle, and motor power of the circulating fan, with a speed measurement range of 0 to 3000 revolutions per minute and an accuracy of ±1 revolution per minute. The purpose of this step is to obtain the key parameters that affect heat and moisture transfer and air flow during tobacco leaf curing, and to provide accurate input data and boundary conditions for the model.

[0084] The specific implementation of step S03 is to establish a heat transfer sub-model between tobacco leaves and air based on the basic principle of convective heat transfer. The heat transfer sub-model is based on the principle of conservation of energy and considers three parts: convective heat transfer between tobacco leaves and air, heat conduction within the tobacco leaves, and latent heat of water evaporation. First, the convective heat transfer between the tobacco leaf surface and the air is calculated as follows:

[0085] Q conv =h c A s (T s -T a );

[0086] Where Q conv is the convective heat transfer, in watts; h c is the convective heat transfer coefficient, in watts per square meter Kelvin; A s is the effective surface area of tobacco leaves, in square meters; T s is the surface temperature of tobacco leaves, in Kelvin; T a is the air temperature in Kelvin.

[0087] Convective heat transfer coefficient h c Calculated using the Dittus-Boelter correlation:

[0088]

[0089] Where, ReD is the Reynolds number, dimensionless; P r is the Prandtl number, dimensionless; n is an exponent, which is 0.4 when the airflow heats the tobacco leaves and 0.3 when the airflow cools the tobacco leaves; k a is the thermal conductivity of air, in watts per meter Kelvin; D h The characteristic length is the equivalent diameter of the tobacco leaf, in meters.

[0090] Reynolds number D The calculation formula is:

[0091]

[0092] Where, ρ a is the air density in kilograms per cubic meter; v a is the air velocity in meters per second; μ a is the dynamic viscosity of air, in Pascal seconds.

[0093] The calculation formula of Prandtl number Pr is:

[0094]

[0095] Where c p，a is the specific heat capacity of air at constant pressure, expressed in joules per kilogram Kelvin.

[0096] For low-speed areas (flow rate less than 0.5 meters per second), the natural convection heat transfer coefficient calculation formula is used:

[0097]

[0098] where h c The unit is watts per square meter Kelvin.

[0099] Tobacco leaf effective surface area A s The calculation formula considering the overlapping factors of tobacco leaves is:

[0100] A s =2A L η s η o ;

[0101] Where A L is the projected area of a single tobacco leaf, in square meters; η s is the tobacco leaf surface area correction coefficient, ranging from 1.2 to 1.5; η o is the overlap coefficient, and its value range is 0.6 to 0.8.

[0102] The calculation formula for heat conduction inside tobacco leaves is:

[0103]

[0104] Where Q cond is the heat conduction, unit is watt; k L is the thermal conductivity of tobacco leaves, ranging from 0.2 to 0.4 watts per meter Kelvin; T c is the core temperature of tobacco leaves, in Kelvin; d L It is the effective thickness of tobacco leaves, in meters.

[0105] The heat transfer submodel is discretized using the finite volume method and solved within each time step, which is set to 10 seconds. The purpose of this step is to accurately describe the heat exchange process between the tobacco leaves and the air, providing a theoretical basis for subsequent temperature field distribution prediction.

[0106] The specific implementation of step S04 is to construct a mathematical model to describe the process of water migration and evaporation into the air in tobacco leaves. First, the water diffusion equation inside tobacco leaves is established, and Fick's second law is used to describe the diffusion process of water in tobacco leaf tissue:

[0107]

[0108] Where M is the moisture content of tobacco leaves, expressed in kilograms of water per kilogram of dry matter; D m is the water diffusion coefficient, in square meters per second.

[0109] Diffusion coefficient D m It is related to temperature and moisture content, and the approximate relationship is an exponential function:

[0110]

[0111] Where D0 is the reference diffusion coefficient, which ranges from 10 -10 ~10 -9 Square meters per second; E a is the diffusion activation energy, ranging from 20 to 30 kilojoules per mole; R is the gas constant, taking 8.314 joules per mole Kelvin; β is the water content influence coefficient, ranging from 1.5 to 2.5.

[0112] The calculation formula for the evaporation rate of tobacco leaf surface is:

[0113] J e =k m A s (Y s -Y a );

[0114] Where, J e is the evaporation rate in kilograms per second; k m is the mass transfer coefficient in meters per second; Y sis the mass fraction of water vapor in the air at the tobacco leaf surface; Y a is the mass fraction of water vapor in the mainstream air.

[0115] Mass transfer coefficient k m and the convection heat transfer coefficient h c Proportional to:

[0116]

[0117] Where Le is the Lewis number, which is usually between 0.9 and 1.1.

[0118] The calculation formula for the heat absorbed during water evaporation is:

[0119] Q evap =J e ΔH vap ;

[0120] Where Q evap is the latent heat of vaporization, in watts; ΔH vap It is the latent heat of vaporization of water, which varies with temperature and is approximately 2.3 to 2.4 megajoules per kilogram in the range of 60 to 90°C.

[0121] By coupling and solving the above equations, we can obtain the variation of tobacco leaf moisture content over time and space and the evaporation rate to the air. The solution process uses the implicit difference format:

[0122]

[0123] Where, is the moisture content of the control volume i at the nth time step; Δt is the time step in seconds; V i is the volume of control body i, in cubic meters; N f is the number of faces of control volume i; D m，j is the diffusion coefficient at surface j; A f，j is the area of surface j, in square meters; d ij is the distance from the center of control volume i to surface j, in meters.

[0124] The purpose of this step is to accurately simulate the moisture migration phenomenon during tobacco leaf baking, laying the foundation for predicting the humidity field distribution in the tobacco loading room.

[0125] The specific implementation of step S05 is to establish a mathematical model describing the air flow in the cigarette compartment based on the Navier-Stokes equations. First, the air continuity equation is established:

[0126]

[0127] Where, ρ a is the air density in kilograms per cubic meter; is the air velocity vector in meters per second.

[0128] The air momentum equation is expressed as:

[0129]

[0130] Where p a is the air pressure in Pascal; μ a is the air dynamic viscosity, which is 1.8×10 -5 Pascal second; is the gravitational acceleration vector, in meters per second squared; is the momentum source term, including the resistance of tobacco leaves to airflow, and its unit is Newton per cubic meter.

[0131] The calculation formula of the resistance of the tobacco leaf layer to airflow as a porous medium is:

[0132]

[0133] Where K p is the permeability of porous media, in square meters, with a value range of 10 -7 ~10 -6 square meters; C f is the inertial resistance coefficient, and its value range is 1.5 to 2.5.

[0134] The air energy equation is expressed as:

[0135]

[0136] Where h a is the specific enthalpy of air, in joules per kilogram; k a is the thermal conductivity of air in watts per meter Kelvin; T a is the air temperature in Kelvin; S h It is the energy source item, including the heat released by tobacco leaves and the heat absorbed by water evaporation, and its unit is watts per cubic meter.

[0137] The moisture transfer equation is expressed as:

[0138]

[0139] Where Y a is the mass fraction of water vapor in the air; D v，a is the diffusion coefficient of water vapor in air, in square meters per second; S m It is a mass source term, mainly the water evaporated from tobacco leaves, and its unit is kilograms per cubic meter per second.

[0140] The air flow model is discretized using the finite volume method, with a grid size of 5 to 10 cm and a time step of 1 second. The boundary conditions include the air outlet velocity boundary condition:

[0141]

[0142] T a | inlet =T in ;

[0143] Y a | inlet =Y in ;

[0144] Return air outlet pressure boundary conditions:

[0145] p a | outlet =p out ;

[0146] Boundary conditions of the smoke chamber wall:

[0147]

[0148] Where, is the air outlet wind speed, in meters per second; T in is the air outlet temperature in Kelvin; Y in is the mass fraction of water vapor at the air outlet; p out is the return air pressure in Pascals; is the wall normal unit vector; U is the total wall heat transfer coefficient, in watts per square meter Kelvin; T amb is the ambient temperature in Kelvin.

[0149] The purpose of this step is to accurately simulate the air flow characteristics in the smoke chamber and provide a theoretical basis for predicting the distribution of wind speed field, pressure field and temperature field.

[0150] The specific implementation of step S06 is to integrate the above three sub-models into a complete multi-physics field coupled neural network model. The overall architecture of the multi-physics field coupled neural network model is a U-shaped network structure, with the input being the initial conditions and boundary conditions, and the output being the predicted physical field distribution. The mathematical expression of the model is:

[0151]

[0152] Where, T is the temperature field prediction result; Y is the humidity field prediction result; V is the wind speed field prediction result; P is the pressure field prediction result; is the neural network function; X is the input parameter, including the initial moisture content of tobacco leaves, tobacco filling density, ambient temperature, ambient humidity, flue-curing room structural parameters, heating power, circulating fan operating parameters, etc.; θ is the neural network parameter.

[0153] The neural network structure includes an encoder and a decoder. The encoder expression is:

[0154] h0=X;

[0155]

[0156] Where h i is the feature map of the i-th layer encoder; σ is the activation function, which uses the ReLU function; is the weight matrix of the i-th layer encoder; is the bias vector of the i-th layer encoder; N e is the number of encoder layers, which is 5.

[0157] Each convolutional layer in the encoder is expressed as:

[0158]

[0159] Where h i (x, y, z) is the value of the feature map of the i-th layer at the spatial position (x, y, z); C i-1 is the number of channels in the i-1th layer; k is the radius of the convolution kernel, which is 1; is the convolution kernel weight of the i-th layer encoder; is the bias of the cth channel of the i-th layer encoder.

[0160] The decoder expression is:

[0161]

[0162] Where g i is the feature map of the i-th layer decoder; is the weight matrix of the i-th layer decoder; is the bias vector of the i-th layer decoder; [g i+1 , h i ] represents feature concatenation and implements skip connection.

[0163] The final output layer expression is:

[0164] T=W T g0+b T ;

[0165] Y=W Y g0+b Y ;

[0166] V=W Vg0+b V ;

[0167] P=W P g0+b P ;

[0168] Where W T 、W Y 、W V 、W P are the weight matrices of the output layers of temperature field, humidity field, wind speed field and pressure field respectively; b T 、b Y 、b V 、b P are the corresponding bias vectors respectively.

[0169] The loss function expression of the physical constraint module is:

[0170] L phys =λ1L heat +λ2L mass +λ3L mom ;

[0171] Where, L heat is the heat transfer constraint loss; L mass is the mass transfer constraint loss; L mom is the momentum transfer constraint loss; λ1, λ2, and λ3 are weight coefficients, ranging from 0.1 to 1.0.

[0172] The heat transfer constraint loss expression is:

[0173]

[0174] Where N is the number of sampling points; the subscript i indicates that the physical constraint is evaluated at the i-th sampling point.

[0175] The mass transfer constraint loss expression is:

[0176]

[0177] The momentum transfer constraint loss expression is:

[0178]

[0179] The purpose of this step is to build a composite model that conforms to physical laws and has data-driven learning capabilities, which can accurately simulate the complex coupling process of heat and moisture transfer and air flow.

[0180] The specific implementation of step S07 involves establishing a high-quality dataset required for model training. First, a sensor array was placed in the experimental flue-curing barn, including a 32-point thermocouple array (accuracy ±0.5°C), a 16-point humidity sensor array (accuracy ±3%), a 12-point hot wire anemometer (accuracy ±0.1 meters per second), and an 8-point micro-differential pressure gauge (accuracy ±1 Pascal). Experimental measurements were conducted under different operating conditions, with parameters including five tobacco loading densities (80-120 kilograms per cubic meter), four initial moisture contents (65%-80%), three circulating fan speeds (800-1600 revolutions per minute), and three heating powers (15-35 kilowatts), resulting in a total of 60 operating conditions. A three-dimensional model of the tobacco loading barn was constructed using Fluent software, employing a standard k-ε turbulence model. A porous media region was set to simulate the tobacco leaf layer. Boundary conditions were consistent with the experimental conditions, with a grid size of 1-2 million, a solution time step of 1 second, and a total simulation time of 1 hour. The experimental and simulation data were combined and randomly divided into training and validation sets in a ratio of 7 to 3. All physical quantities were normalized to a range of 0 to 1. A moving average filter was used to remove noise from the experimental data with a window width of 5. The purpose of this step was to construct a comprehensive training dataset encompassing various operating conditions, providing a data foundation for subsequent neural network model training.

[0181] The specific implementation method of step S08 is to use a deep learning algorithm to train the multi-physics field coupled neural network model. First, the neural network parameters are initialized using a normal distribution (mean 0, standard deviation 0.01). A composite loss function is designed, including a mean square error loss term (weight 1.0), a physical constraint loss term (weight 0.5), and an L2 regularization loss term (weight 0.0001). The network is trained using a batch gradient descent algorithm with a batch size of 32, an initial learning rate of 0.001, and an exponential decay strategy with a decay rate of 0.9 every 500 steps. During the training process, the model performance is evaluated on the validation set every 50 steps, and early stopping is performed if the performance does not improve after 10 consecutive evaluations. Five model instances with different random initialization parameters are trained, and their average output is taken as the final prediction result.

[0182] For the training data of different circulating fan speeds, the changes in the weight coefficient of the third convolutional layer of the encoder in the network are recorded, and the functional relationship between the weight coefficient and the circulating fan speed is obtained by least squares fitting:

[0183]

[0184] Where W k (n) is the weight coefficient of the kth channel when the speed is n; W k(n0) is the weight coefficient of the kth channel at the reference speed n0; n is the circulation fan speed, in revolutions per minute; n0 is the reference speed, which is 1200 revolutions per minute; α is the first-order proportional coefficient, which is 0.35-0.45; β is the second-order proportional coefficient, which is 0.05-0.15.

[0185] The purpose of this step is to train a heat and moisture transfer and air flow model that conforms to physical laws and has high prediction accuracy, while establishing a quantitative relationship between the neural network parameters and the actual physical control parameters to improve the physical interpretability and extrapolation ability of the model.

[0186] The specific implementation of step S09 is to optimize the operating parameters of the circulating fan and the configuration of the return port based on the trained model. First, the optimization target is set as the uniformity index of the temperature field and the humidity field. The temperature field uniformity index is defined as the temperature standard deviation, and the humidity field uniformity index is defined as the relative humidity standard deviation. The comprehensive uniformity index is:

[0187] J=w T J T +w Y J Y ;

[0188] Where, J is the comprehensive uniformity index; w T is the weight of the temperature field uniformity index, which is 0.6; w Y is the humidity field uniformity index weight, which is 0.4; J T is the standard deviation of the temperature field; J Y is the standard deviation of the humidity field.

[0189] The optimization variables include the circulation fan air volume (10 to 30 cubic meters per second), air pressure (100 to 300 Pascals), and the return port coordinates. A Bayesian optimization algorithm was used, with the temperature and humidity field uniformity indicators as the objective function. A Gaussian process regression model was used as the surrogate model. The expected improvement was selected as the acquisition function, with an initial sampling point of 10 and a maximum number of iterations of 100.

[0190] For each set of optimized parameters, the trained model is used to predict the distribution of the temperature, humidity, wind speed, and pressure fields within the tobacco loading chamber, and a uniformity index is calculated. Optimization is terminated when the uniformity index improves by no more than 0.1% after 10 consecutive iterations. The optimal configuration of the circulating fan air volume, air pressure parameters, and return port location is ultimately output, generating a visualization of the physical field distribution within the tobacco loading chamber. This step aims to optimize key operating parameters of the flue-curing barn based on an accurate physical field prediction model, achieving uniform temperature and humidity distribution within the tobacco loading chamber and improving tobacco curing quality.

[0191] To better understand and implement the present invention, Example 2 of a specific application scenario is provided below. A tobacco research team conducted practical research using a method for modeling heat and moisture transfer and air flow in a uniform, densely packed flue-curing barn. The research team selected a flue-curing barn with internal dimensions of 6.2m × 4.5m × 3.8m. The barn was equipped with a circulating fan system and a heating system for curing tobacco leaves. Uneven temperature and humidity distribution within the barn resulted in varying quality of the cured tobacco leaves. The research team decided to apply the present invention to develop a heat and moisture transfer and air flow model to optimize the barn's parameter configuration.

[0192] First, the research team divided the tobacco loading room into 15cm×15cm×15cm control volume units, forming a total of approximately 52,000 grid points. Based on the heat transfer equation, mass transfer equation, and momentum transfer equation, a multi-resolution neural network architecture was constructed, using a U-shaped network structure with 5 layers of downsampling and 5 layers of upsampling. The research team measured the initial moisture content of the tobacco leaves to be 78.5% and the tobacco loading density to be 105kg / m 3 , the ambient temperature is 25.2℃, the ambient humidity is 68%, the heating power is 28kW, the circulating fan speed is 1200r / min, and the blade angle is 35°.

[0193] The research team established a heat transfer sub-model based on the finite volume method and calculated the convective heat transfer coefficient between tobacco leaves and air. The convective heat transfer coefficients at different wind speeds are shown in Table 1:

[0194] Table 1 Convective heat transfer coefficient at different wind speeds

[0195] Wind speed (m / s) <![CDATA[Convective heat transfer coefficient (W / (m 2 ·K))]]> 0.3 4.8 0.6 7.3 0.9 9.2 1.2 11.5 1.5 13.6 1.8 15.4 2.1 17.2

[0196] The water migration sub-model uses the diffusion equation based on Fick's second law and calculates the water diffusion coefficient inside tobacco leaves to be 2.85×10 -10 m 2 / s, and the evaporation mass transfer coefficient is 4.2×10 -3 m / s. The air flow submodel uses the standard k-ε turbulence model, and the porous medium permeability is 3.76×10 -7 m 2 , the inertial drag coefficient is 1.85.

[0197] The research team used an experimental setup to deploy a 24-point thermocouple array, a 12-point humidity sensor array, and a 10-point hot-wire anemometer within the baking room. Using Fluent software, they performed CFD simulations under various operating conditions. A total of 45 sets of experimental data and 180 sets of simulation data were collected, 70% of which were used for training and 30% for validation. Table 2 shows some of the experimental measurement data.

[0198] Table 2 Physical field data of some measurement points in the baking room

[0199] Measuring point position (x, y, z) m Temperature (℃) Relative humidity (%) Wind speed (m / s) Static pressure (Pa) (1.0,1.0,1.0) 62.8 75.3 0.82 -8.5 (1.0,3.0,1.0) 59.3 78.6 0.56 -6.3 (3.0,1.0,1.0) 68.4 69.8 1.35 -12.7 (3.0,3.0,1.0) 64.2 72.5 0.95 -9.8 (5.0,1.0,1.0) 61.7 76.8 0.78 -7.9 (5.0,3.0,1.0) 58.6 79.2 0.50 -5.8 (1.0,1.0,2.5) 65.9 71.2 1.08 -10.2 (3.0,1.0,2.5) 70.3 67.5 1.42 -13.5 (5.0,1.0,2.5) 63.8 74.3 0.86 -8.7

[0200] The multi-physics field coupled neural network model was trained using the Adam optimizer, with an initial learning rate of 0.001, a batch size of 32, and 50,000 training iterations. The root mean square error on the validation set reached 0.035, meeting the accuracy requirements. By recording the changes in network weights under different fan speeds, the parameters α = 0.41 and β = 0.12 in the relationship between the weight coefficient and the circulating fan speed were fitted, and the goodness of fit r 2 =0.94.

[0201] Based on the trained model, the research team optimized the circulation fan parameters and the supply and return air vent configuration. The temperature and humidity uniformity before and after optimization are shown in Table 3:

[0202] Table 3 Comparison of temperature and humidity uniformity before and after optimization

[0203] Evaluation indicators Before optimization After optimization Improvement rate (%) Temperature field standard deviation (℃) 4.86 1.32 72.8 Humidity field standard deviation (%) 6.35 2.18 65.7 Wind speed unevenness coefficient 0.62 0.28 54.8 Energy consumption (kW·h) 58.3 47.6 18.4

[0204] The final optimized circulating fan air volume is 24.5m 3 / s, wind pressure was 185Pa, the optimal position of the air outlet was adjusted to 0.4m downward from the original position, and the optimal position of the return air outlet was adjusted to 0.3m upward from the original position. The research team applied the optimized configuration parameters to the actual flue-curing barn and cured five batches of tobacco leaves. The tobacco leaf grading results are shown in Table 4:

[0205] Table 4 Comparison of tobacco leaf rating results before and after optimization

[0206]

[0207]

[0208] The traditional method of optimizing the temperature and humidity distribution in the flue-curing barn mainly relies on the staff's experience to adjust the circulating fan speed and damper opening, and gradually finds the optimal parameters through multiple experiments, which is time-consuming and has limited effect. Another traditional method is to use a large number of CFD simulation calculations to compare the internal flow field distribution of the flue-curing barn under different parameters, but each calculation takes 10 to 20 hours, making it difficult to perform global optimization. The present invention integrates physical mechanisms and deep learning methods to establish a heat and moisture transfer and air flow model that conforms to physical laws and has learning capabilities. On the one hand, it can accurately predict the temperature and humidity distribution in the flue-curing barn. On the other hand, the model has a fast prediction speed, and it only takes 0.2 seconds to evaluate a set of parameters, making global optimization possible. Compared with traditional methods, the present invention reduces the standard deviation of the flue-curing barn temperature field by 72.8%, the standard deviation of the humidity field by 65.7%, and the proportion of high-quality tobacco leaves by 23.2 percentage points. At the same time, energy consumption is reduced by 18.4%, significantly improving the tobacco leaf baking quality and energy utilization efficiency.

[0209] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 5, 6 and 7 below.

[0210] Table 5 Variable Explanation Table (Part 1)

[0211]

[0212] Table 6 Variable Explanation Table (Part 2)

[0213]

[0214] Table 7 Variable Explanation Table (Part 3)

[0215]

[0216]

[0217] 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. A method for establishing a heat and moisture transfer and air flow model for a uniform temperature dense baking room, characterized in that: include: Construct the basic architecture of a multi-resolution neural network; collect model input parameters such as initial moisture content of tobacco leaves and tobacco filling density; and construct a heat transfer sub-model; Construct a tobacco leaf moisture migration sub-model; establish an air flow sub-model; integrate the three sub-models to form a multi-physics field coupled neural network model; establish a model training data set; train the multi-physics field coupled neural network model to obtain a heat and moisture transfer and air flow model.

2. The method for establishing a heat and moisture transfer and air flow model for a uniform temperature intensive baking barn according to claim 1, characterized in that: The multi-resolution neural network basic architecture refers to a neural network architecture based on the heat transfer equation, mass transfer equation, and momentum transfer equation, and a mathematical model structure is constructed by dividing the smoke chamber area into multiple control volumes.

3. The method for establishing a heat and moisture transfer and air flow model for a uniform temperature intensive baking barn according to claim 2, characterized in that: The multi-resolution neural network refers to a deep learning model with a multi-level structure, which is used to simultaneously capture physical features of different scales. It includes a high-resolution layer for obtaining local detail information and a low-resolution layer for obtaining global feature information, and realizes the fusion of features of different resolutions through jump connections.

4. The method for establishing a heat and moisture transfer and air flow model for a uniform temperature dense baking barn according to claim 3, characterized in that: The heat transfer equation is a mathematical expression that describes the conservation of energy within a system. The input includes temperature gradient, thermal conductivity, convection heat transfer coefficient, radiation coefficient, and heat source intensity. The output is the temperature distribution and heat flux density at each point in the system.

5. The method for establishing a heat and moisture transfer and air flow model for a uniform temperature intensive baking barn according to claim 4, characterized in that: The mass transfer equation is a mathematical expression that describes the material migration process. The input includes concentration gradient, diffusion coefficient, mass source term, and flow velocity. The output is the concentration distribution and mass transfer rate of each point in the system.

6. The method for establishing a heat and moisture transfer and air flow model for a uniform temperature intensive baking barn according to claim 5, characterized in that: The momentum transfer equation is a mathematical expression that describes the motion state of a fluid. The input includes pressure gradient, viscosity coefficient, volume force, and boundary conditions, and the output is the velocity distribution and pressure distribution of each point in the system.

7. The method for establishing a heat and moisture transfer and air flow model for a uniform temperature intensive baking barn according to claim 6, characterized in that: The heat transfer sub-model is constructed using the principles of heat and mass transfer, and characterizes the effects of temperature gradient, convective heat transfer coefficient, and heat transfer area on the heat transfer process.

8. The method for establishing a heat and moisture transfer and air flow model for a uniform temperature intensive baking barn according to claim 7, characterized in that: The tobacco leaf moisture migration sub-model includes a moisture diffusion equation, an evaporation-condensation equation, and a phase change latent heat equation, and calculates the rate at which moisture in the tobacco leaf migrates to the surface and evaporates into the air.

9. The method for establishing a heat and moisture transfer and air flow model for a uniform temperature intensive baking barn according to claim 8, characterized in that: The air flow sub-model is established based on the principles of fluid mechanics, and combines the continuity equation, momentum equation, and energy equation to simulate the distribution of air pressure field, velocity field, and temperature field in the smoke chamber.

10. The method for establishing a heat and moisture transfer and air flow model for a uniform temperature intensive baking barn according to claim 1, characterized in that: It also includes the steps of establishing a functional relationship between the neural network weight coefficient and the circulation fan speed; and optimizing the circulation fan parameters and the return port position configuration to achieve uniform distribution of temperature and humidity fields in the smoke loading room.

Images