Large model-based dynamic control optimization method for heating processes

Abstract

The application discloses a heating process dynamic control optimization method based on a large model and relates to the field of model optimization algorithms. The method comprises the following steps: acquiring multi-spectral real-time monitoring data flow inside a medicine roasting machine; inputting a medicine material stacking discrete element semantic atlas based on the spatial distribution difference of the medicine material particle size into a multi-scale anti-smoothing graph neural network model; performing deep reasoning based on long-tail feature enhancement; decoupling microscopic heat accumulation singularity features; reconstructing a blind area hidden layer thermodynamic state vector based on the microscopic heat accumulation singularity features and in combination with the Brazil nut effect principle in fluid dynamics; calculating a layered solidification risk index; constructing a risk perception type asymmetric objective function based on the blind area hidden layer thermodynamic state vector and the layered solidification risk index; solving nonlinear turbulent disturbance trajectory parameters; and superimposing and compensating a benchmark roasting curve, so that the hidden quality problems of surface underfire and bottom charring caused by physical stratification are fundamentally solved.

Description

Technical Field

[0001] This invention relates to the field of model optimization algorithms, specifically to a dynamic control optimization method for heating processes based on a large model. Background Technology

[0002] In the industrial-scale roasting process of traditional Chinese medicine (TCM) decoction pieces, existing intelligent equipment mostly uses non-contact machine vision or infrared thermal imaging technology to provide feedback on the roasting status. However, for heterogeneous bulk materials composed of root and rhizome slices and debris, the Brazil nut effect in fluid dynamics is easily induced under low-speed roasting conditions, leading to vertical segregation and stratification of particles. Specifically, large-diameter medicinal materials float to the surface of the fluid, while fine debris is pulled down by gravity to the boundary layer at the bottom of the pot. This physical stratification creates an observable trap at the visual perception level, where the uniform color and temperature of the large particles on the surface mask the localized heat accumulation caused by the high contact thermal resistance and poor heat dissipation of the sedimented debris at the bottom. This prevents existing surface visual feedback systems from detecting the early signs of scorching at the bottom, resulting in hidden quality defects such as under-roasting on the surface and carbonization at the bottom.

[0003] From an algorithmic perspective, most mainstream temperature control models are based on convolutional neural networks or graph neural network architectures. Their core mechanism relies on spatial feature aggregation operators (such as max pooling, average pooling, or attention weighting) to extract global state representations. However, the high-temperature features of low-level debris are typical long-tailed data in images or maps, with extremely low pixel proportions and discrete spatial locations. During the layer-by-layer forward propagation of deep networks, these microscopic high-frequency singularity features are easily smoothed by the low-frequency features of the dominant surface large particles, causing local extremum information to vanish during feature map dimensionality reduction. In addition, existing loss function designs mostly focus on optimizing the consistency of global texture or temperature, lacking penalty weights for outliers in extremely small areas. This causes the algorithm to converge to a local minimum point during gradient descent that is optimal in overall statistical indicators but out of control in local physical states, failing to effectively address the challenge of fine-grained temperature control in granular layered scenarios.

[0004] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention

[0005] To address the technical problems mentioned in the background section, this invention is proposed. This invention provides a dynamic control optimization method for heating processes based on a large model.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] The dynamic control optimization method for heating processes based on large models includes the following steps:

[0008] The multispectral real-time monitoring data stream inside the herb roasting machine was acquired, and a discrete element semantic map of herb stacking was constructed based on the spatial distribution differences of herb particle size.

[0009] The stacked discrete element semantic graph of the medicinal materials is input into a multi-scale anti-smoothing graph neural network model to perform deep inference based on long-tail feature enhancement and decouple the microscopic thermal accumulation singularity features.

[0010] Based on the microscopic thermal accumulation singularity characteristics and combined with the Brazil fruit effect principle in fluid dynamics, the thermodynamic state vector of the blind zone hidden layer is reconstructed, and the layer solidification risk index is calculated.

[0011] Based on the thermodynamic state vector of the blind zone hidden layer and the layered solidification risk index, a risk-aware asymmetric objective function is constructed to solve the nonlinear turbulent disturbance trajectory parameters and to superimpose compensation on the benchmark overcooking curve.

[0012] Furthermore, the steps for constructing the stacked discrete element semantic graph of medicinal materials are as follows:

[0013] The visible light image sequence and infrared thermal imaging sequence inside the stir-frying machine are acquired, and physical space reconstruction and multi-dimensional spatiotemporal field calculation based on the inverse ray projection model are performed to generate a heat-texture temporal field data package inside the pot.

[0014] The heat-texture temporal field data package inside the pot is input into the instance segmentation network. Based on the bimodal distribution characteristics of the particle size projection area, the medicinal material targets in the picture are divided into a set of dominant master particles and a set of hidden debris.

[0015] Using the set of dominant master particles as the graph node skeleton and the flow channel saturation index of the set of hidden debris as the edge weight, a discrete element semantic graph of medicinal material stacking is constructed.

[0016] Furthermore, the steps for performing physical space reconstruction and multidimensional spatiotemporal field calculation based on the reverse ray projection model are as follows:

[0017] A reference coordinate system for the pot body is established with the center of the pot bottom as the origin and the normal of the pot bottom plane as the axis. Based on the intrinsic and extrinsic pose data of the visible light camera and the infrared thermal imaging sensor, a reverse ray projection model is constructed. The intersection of the projections of the visible light cone and the infrared cone on the pot bottom plane is calculated using the reverse ray projection model. Specular reflection blind spots and dynamic occlusion blind spots are removed within the mapping area to generate a dynamic effective mask.

[0018] Furthermore, the steps of performing physical space reconstruction and multidimensional spatiotemporal field calculation based on the reverse ray projection model also include:

[0019] The effective mesh cells are locked using the dynamic effectiveness mask. The ambient background radiation temperature and dynamic emissivity spectrum are introduced into the effective mesh cells to calculate the real thermodynamic temperature field. At the same time, dark current subtraction and flat field correction are performed to calculate the normalized texture field.

[0020] Motion compensation interpolation is performed on the asynchronously sampled real thermodynamic temperature field and the normalized texture field to align them to the same physical moment, and then encapsulated into a pot interior heat-texture temporal field data package.

[0021] Furthermore, the calculation steps for the flow channel saturation index are as follows:

[0022] The centroids of the explicit master particle set are connected using a triangulation algorithm to construct a rigid support network for the master particles. The effective pore channel area is obtained by subtracting the area occupied by the master particles from the remaining geometric space in each cell of the rigid support network.

[0023] Furthermore, the step of calculating the flow channel saturation index also includes:

[0024] The hidden debris set is mapped to the rigid support network of the main particles, and the geometric retention area ratio of the debris in the effective pore flow channel area is calculated to generate the flow channel saturation index.

[0025] Furthermore, the steps for performing deep inference based on long-tail feature enhancement are as follows:

[0026] A multi-scale anti-smoothing graph neural network model is constructed, which consists of a dilated convolutional branch, a point cloud attention branch, a feature aggregation layer, and a residual amplification module.

[0027] The stacked discrete element semantic graph of the medicinal materials is input into the multi-scale anti-smoothing graph neural network model. The global smoothing temperature feature matrix is ​​extracted through the dilated convolution branch, and the local blocking weighted gradient vector is locked through the point cloud attention branch.

[0028] In the feature aggregation layer, the global smooth temperature feature matrix is ​​used as the smoothing benchmark, and the residual significance is checked on the local blocking weighted gradient vector. If the residual value exceeds the preset threshold, a cluster of high-temperature pixels with a very small area is obtained.

[0029] Furthermore, the step of performing deep inference based on long-tail feature enhancement also includes:

[0030] The extremely small area high-temperature pixel cluster is input into the residual amplification module, and the thermodynamic feature values ​​are mapped by a nonlinear activation function to output the microscopic thermal accumulation singularity feature.

[0031] Furthermore, the steps for calculating the stratified solidification risk index are as follows:

[0032] The microscopic thermal accumulation singularity features are mapped to the pot body reference coordinate system, and the sedimentation probability of medicinal material particles under the action of gravitational and centrifugal force fields is deduced using a variational autoencoder to generate a layered probability density matrix.

[0033] Based on the layered probability density matrix, the heat retention time is estimated using the thermal diffusion inversion algorithm to generate the thermodynamic state vector of the blind zone hidden layer.

[0034] Based on the thermodynamic state vector of the blind zone, the asynchronous deviation between the temperature rise rate of the blind zone and the temperature rise rate of the surface medicinal material is calculated. Combined with the critical coking temperature of the medicinal material, the layered solidification risk index is output.

[0035] Furthermore, the steps for calculating the nonlinear turbulent disturbance trajectory parameters and superimposing compensation onto the baseline turbulence curve are as follows:

[0036] Construct a risk-aware asymmetric objective function, and determine the risk-dominant state based on the asymmetric objective function;

[0037] Based on the aforementioned risk-dominant state, the nonlinear turbulent disturbance trajectory parameters are calculated, and the preset benchmark overturning curve is superimposed and compensated.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] This invention acquires multispectral real-time monitoring data streams from inside a medicinal herb roasting machine. Based on the spatial distribution differences in the particle size of the medicinal materials, it constructs a discrete element semantic map of stacked medicinal materials. This map is then input into a multi-scale anti-smoothing graph neural network model, performing deep inference based on long-tail feature enhancement to decouple microscopic thermal accumulation singularity features. Based on these features and combined with the Brazil nut effect principle in fluid dynamics, the thermodynamic state vector of the hidden layer in the blind zone is reconstructed, and the layered solidification risk index is calculated. Existing technologies often rely on global pooling operations to extract average features from the surface image, which easily leads to the weak high-temperature signals of the bottom layer debris being submerged by the normal features of the large surface particles, i.e., feature disappearance. This invention innovatively constructs a discrete element semantic map of stacked medicinal materials and combines it with a multi-scale anti-smoothing graph neural network, using long-tail feature enhancement technology to nonlinearly amplify the microscopic thermal accumulation singularities. By calculating the flow channel saturation index, this invention can invert the stratification state of the bottom particles caused by the Brazil nut effect in the visual blind zone, transforming the previously invisible bottom heat retention time into a quantifiable stratification solidification risk index. This allows the system to penetrate the illusion of a moderately heated surface, accurately locate and warn of the early carbonization risk of tiny debris at the bottom, fundamentally solving the hidden quality problem of under-heating of the surface and scorching of the bottom caused by physical stratification.

[0040] This invention constructs a risk-aware asymmetric objective function based on the thermodynamic state vector of the blind zone and the stratification solidification risk index, calculates the nonlinear turbulent disturbance trajectory parameters, and superimposes compensation onto the baseline stir-frying curve. Existing PID or linear feedback control can only adjust the overall rotation speed or temperature, failing to provide the instantaneous destructive force required to break the dense accumulation layer at the bottom. This invention introduces a risk-aware asymmetric objective function. When stratification risk is detected as dominant, it no longer pursues stable stir-frying of the surface layer, but actively outputs a nonlinear turbulent disturbance trajectory containing high-frequency vibration amplitude and transient angular acceleration pulses. This control strategy utilizes high-intensity inertial torque to counteract the static friction of the accumulated medicinal materials, forcibly disrupting the force chain arch structure at the bottom layer, causing the deposited debris to float back to the surface for effective heat exchange due to fluidization. This invention achieves a leap from passively following the surface state to actively reconstructing the underlying physical field, ensuring the uniformity of heat transfer and the inherent safety of the process during the stir-frying of mixed particle size medicinal materials. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not drawn to scale according to the actual size, but are intended to show the main idea of ​​the present invention.

[0042] Figure 1 The flowchart shows the method for dynamic control optimization of the heating process based on a large model.

[0043] Figure 2 This is a schematic diagram of physical space reconstruction and multispectral imaging field of view mapping;

[0044] Figure 3 A schematic diagram illustrating the physical principles of multidimensional field solution and spatiotemporal alignment;

[0045] Figure 4 This is a schematic diagram illustrating the physical construction principle of discrete element semantic graphs. Detailed Implementation

[0046] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are also within the scope of protection of the present invention.

[0047] like Figure 1 As shown, the dynamic control optimization method for the heating process based on a large model includes the following steps:

[0048] Step S100: Obtain the multispectral real-time monitoring data stream inside the herb roasting machine, and construct a discrete element semantic map of herb stacking based on the spatial distribution differences of herb particle size;

[0049] It should be noted that the multispectral real-time monitoring data stream includes synchronously acquired visible light image sequences and infrared thermal imaging sequences.

[0050] The process of constructing the discrete element semantic graph in step S100 specifically includes the following steps:

[0051] Step S101: Obtain the visible light image sequence and infrared thermal imaging sequence inside the stir-frying machine, perform physical space reconstruction and multi-dimensional spatiotemporal field calculation based on the reverse ray projection model, and generate a heat-texture temporal field data package inside the pot;

[0052] Step S101, which involves physical space reconstruction and multidimensional spatiotemporal field calculation based on the reverse ray projection model, specifically includes the following steps:

[0053] like Figure 2 The diagram shown illustrates physical space reconstruction and multispectral imaging field of view mapping.

[0054] Reference Figure 2 Figure ① shows a visible light camera. Figure ② shows an infrared camera. Figure ③ shows the visible light cone projection boundary ray, corresponding to the visible shooting area. The boundary, shown in Figure ④, is the boundary ray of the infrared viewing cone projection, corresponding to the infrared temperature measurement area. The boundary. Additionally, the installation height is shown in the diagram. , Pitch angle , Physical radius of the bottom of the pot The center of the bottom of the pot and common effective measurement area .

[0055] Reference Figure 2 First, take the center of the bottom of the pot as the origin. The normal to the bottom plane of the pot is the axial direction, establishing a unified reference coordinate system for the pot body. Based on the visible light camera shown in Figure ①... Similar to the infrared thermal imaging sensor shown in Figure ② The physical installation parameters are used to obtain the intrinsic and extrinsic pose matrices of both. The extrinsic translation vector is determined by the installation height shown in the diagram. and The rotation matrix is ​​determined by the pitch angle at the center of the field of view shown in the diagram. and Decision. Based on the above parameters, a retroray projection model is constructed. This model is used to establish the pixel coordinates of the two-dimensional image plane. To the three-dimensional pot body reference coordinate system (in this coordinate system, the equation corresponding to the bottom disk plane shown in the figure is...) The mapping relationship is established. Based on this, the polar coordinate parameter domain mesh of the pot body is initialized, and the physical radius is... The bottom area of ​​the pot is divided into Each grid cell enables spatial discretization of the data.

[0056] Step S1011: With the center of the bottom of the pot as the origin and the normal of the bottom plane as the axis, establish a reference coordinate system for the pot body. Based on the intrinsic and extrinsic pose data of the visible light camera and the infrared thermal imaging sensor, construct a reverse ray projection model and initialize the polar coordinate parameter domain mesh of the pot body. The reverse ray projection model is used to establish the mapping relationship between the two-dimensional image plane pixel coordinates and the three-dimensional pot body reference coordinate system.

[0057] like Figure 2 As shown, a unified reference coordinate system for the pot is constructed. The geometric center of the bottom of the stir-frying machine pot is set as the origin of the coordinate system. Taking the direction of the normal to the bottom plane of the pot (i.e., the vertical upward direction) as... The axis is defined by two mutually perpendicular straight lines within the plane of the pot bottom. shaft and Axis. Acquiring visible light from a camera. With infrared camera intrinsic parameter matrix It includes focal length, principal point coordinates, and distortion coefficients, and obtains both relative to the pot body reference coordinate system using the checkerboard calibration method. The extrinsic pose matrix, which is constructed from the physical mounting parameters of the camera: where the translation vector... It is mainly determined by the installation height of the camera, i.e., the installation height of the visible light camera. and the installation height of the infrared camera Rotation matrix The pitch angle determined by the center of the field of view, i.e., the visible light pitch angle. and infrared pitch angle Based on the aforementioned intrinsic and extrinsic parameters, a reverse ray projection model is constructed. This model is an inverse solver based on the pinhole camera model commonly used in computer vision. Its core logic is to use the homography matrix to solve for the intersection point of the ray and the plane, which is a well-known technique in the field of computer vision, and its specific details will not be elaborated further. The input of this model is: any camera ( or Pixel coordinates on the image plane And the preset equation of the bottom plane of the pot (in this embodiment, it is...) ),in Represents the column coordinates of the image matrix. Represents row coordinates; the output is the coordinates of the unique intersection point between the spatial ray corresponding to that pixel and the bottom plane of the pot. , This indicates that the physical point corresponding to the pixel coordinate is on the bottom plane of the pot. Physical distance on the axis This indicates that the physical point corresponding to the pixel coordinate is on the bottom plane of the pot. The physical distance on the axis, the input-output transformation relationship, and the model construction are as follows: using the inverse matrix of the intrinsic parameter matrix. pixel coordinates Convert to a normalized direction vector in the camera coordinate system; use the inverse transformation of the extrinsic matrix. Transform the direction vector to the pot reference coordinate system In this process, a spatial ray equation is formed originating from the optical center of the camera; the equations of this spatial ray equation and the bottom plane equation are then combined. Solving the system of equations yields the unique intersection point coordinates. Since the two cameras are positioned differently, the ray paths corresponding to pixels on their respective image planes are different; therefore, the calculated intersection point coordinates are strongly spatially correlated. This model is used because visible light and infrared images exhibit parallax; only by mapping two-dimensional pixels back to three-dimensional physical space (i.e., the bottom plane of the pot) can pixel-level spatial alignment of cross-modal data be achieved. Based on this, the polar coordinate parameter domain mesh of the pot is initialized, with the physical radius set to... The bottom region of the pot is divided into the following sections along the radial direction: A series of concentric rings, divided along the polar angle direction. Each sector forms Each grid cell is indexed by a unique polar coordinate. Identification; among which: For polar radius index, This indicates which ring layer the grid belongs to; Polar index, This indicates which sector the grid belongs to. This grid is used to subsequently carry the projected texture and temperature data, achieving spatial discretization and unified alignment of the data.

[0058] Reference Figure 2 The field of view projection is calculated using the aforementioned reverse ray projection model. Specifically, it involves extracting data from the visible light camera shown in Figure ①. The vertex pixels of the image sensor are used to calculate their projected coordinates onto the plane of the bottom of the pot, and these coordinates are then connected to form the visible imaging area. (i.e., the left-hand area enclosed by the dotted line in Figure ③); similarly, extract the infrared camera data from Figure ②. The vertex pixel projection forms the infrared temperature measurement area. (That is, the right-hand region enclosed by the dotted line in Figure ④). Perform spatial Boolean operations to calculate... and The geometrically overlapping polygons, the overlapping area being the common effective survey area filled in the shaded area shown in the figure. The system iterates through the initialized polar coordinate grid, determining whether the center of each grid cell lies within the region. Within this area, specular reflection blind zones determined by the incident angle and dynamic occlusion blind zones caused by the stirring spatula are further eliminated. Finally, a dynamic validity mask that identifies the validity of the mesh is generated to ensure that subsequent multi-dimensional field calculations are performed only within the valid physical region.

[0059] Step S1011 establishes a unified physical coordinate benchmark and constructs a reverse ray projection model. The fundamental analytical basis for this is that the inherent parallax problem of non-coaxial binocular perception systems cannot be solved by simple image registration. Due to the different installation positions of the visible light and infrared sensors, the pixel positions of the same medicinal material in the two images exhibit significant nonlinear deviations. Only by establishing a rigorous mathematical mapping from the two-dimensional image plane to the three-dimensional pot bottom plane can visual data from different perspectives be restored to spatial points with real physical measurements. This step is indispensable in the entire method because it is the geometric foundation for all subsequent multimodal data fusion. Without this precise alignment of physical space, subsequent temperature and texture overlay analysis would be like marking coordinates on an incorrect map, inevitably leading to the mismatch of high temperature in the background to the medicinal material in the foreground, resulting in fundamental interpretation errors.

[0060] Step S1012: Calculate the intersection of the projections of the visible light cone and the infrared cone on the bottom plane of the pot using the reverse ray projection model, map the projection intersection onto the polar coordinate parameter domain grid of the pot, and remove the specular reflection blind zone and the dynamic occlusion blind zone within the mapped area to generate a dynamic effectiveness mask.

[0061] To obtain the bottom surface of the viewing cone, specifically by extracting the visible light from the camera... With infrared camera The pixel coordinates of the four vertices of the image sensor, i.e., the four corner points of the image matrix, are used as input into the aforementioned backlight projection model to calculate the projected coordinates of the four vertices on the bottom plane of the pot. The trapezoidal area formed by sequentially connecting these four projected coordinates is the visible imaging area. (correspond Figure 2 The area enclosed by the dashed line on the left side and the infrared temperature measurement area (correspond Figure 2 (The area enclosed by the dashed line on the right). Perform spatial Boolean operations: calculate using the polygon clipping algorithm. and The geometrically overlapping polygons, the overlapping area is the common effective survey area. (correspond Figure 2(Middle shaded area); Traverse all initialized polar coordinate parameter domain meshes of the pot body, and determine whether the center physical coordinates of each mesh cell are located within the region. Within the polygon boundary, if a value is found inside the polygon, it is marked as a valid candidate. Specular reflection blind zone removal: Calculate the incident angle and threshold: For each grid marked as a valid candidate, construct a line-of-sight vector from the center of that grid to the optical center of the camera. Calculate the relationship between this vector and the normal vector of the pot bottom. The angle between them is the angle of incidence of the light ray. The threshold is an empirical value set based on the Fresnel reflection principle. For example, the threshold... ,when When this value is exceeded, the reflectivity of the metal surface increases sharply, causing temperature measurement to fail. Therefore, it is identified as a specular reflection blind zone and removed. Dynamic occlusion blind zone removal: Using background subtraction or optical flow methods, a binary mask image of the moving foreground (i.e., spatula, hand) is extracted from the original two-dimensional image (visible light or infrared image). All pixels with a value of 1 (representing occlusion) in this mask image are substituted back into the reverse ray projection model to calculate the set of projection points of these occluded pixels on the bottom plane of the pot. It is determined which mesh cells fall within the coverage area of ​​these projection point sets, and these meshes are identified as dynamic occlusion blind zones. Only the meshes not covered by the above specular reflection blind zones and dynamic occlusion blind zones are retained to generate a dynamic validity mask that identifies the validity of the meshes.

[0062] Step S1012 generates a dynamic validity mask and eliminates blind spots, based on the Fresnel reflection principle and the geometric projection law of motion occlusion. At specific incident angles, the metal pot exhibits extremely high specular reflectivity, causing the sensor to receive false signals from environmental reflections instead of medicinal material radiation. Simultaneously, the dynamic movement of the stirring spatula periodically obstructs the view. This step is crucial in the method because it acts as a gatekeeper for data quality. Through spatial Boolean operations and optical path geometry analysis, it filters out all potential sources of interference before the data enters the core algorithm, ensuring that the model only processes valid pixels containing real medicinal material information and preventing false triggering of the control system due to environmental noise.

[0063] like Figure 3 The diagram shown illustrates the physical principles of multidimensional field solution and spatiotemporal alignment.

[0064] Reference Figure 3 The diagram shows the visible light texture field calculation process on the left, the true thermodynamic temperature inversion process on the right, and the spatiotemporal alignment and encapsulation process at the bottom. This is the original grayscale matrix (including noise). These are the inherent noise matrix and the brightness gain matrix used for correction, respectively. The normalized texture field is calculated; Total radiance (including reflection). This refers to the background temperature inside the pot. The true thermodynamic temperature field is derived from the inverse; The inter-frame displacement field is estimated based on the optical flow method. To unify the time base, This is the final encapsulated pot interior heat-texture time-series field data package.

[0065] Reference Figure 3 First, a dynamic validity mask is used to lock the effective computation region. Figure 3 In the visible light processing branch on the left, the system acquires the raw grayscale matrix output by the visible light camera. To eliminate inherent sensor noise and uneven illumination in the optical system, a pre-calibrated inherent noise matrix (to eliminate random thermal noise) and a brightness gain matrix (to compensate for luminous flux attenuation at the lens edges) are introduced. Based on these formulas, clean data is provided for subsequent material identification.

[0066] exist Figure 3 In the infrared processing branch on the right, the system acquires the total radiance detected by the infrared sensor. To isolate ambient reflected radiation and correct for emissivity effects, the ambient background radiation temperature acquired by the cold zone probe and a dynamic emissivity coefficient assigned based on texture classification (metal / medicinal material) are introduced. Based on the reverse radiation transfer model formula, a true thermodynamic temperature field is output, eliminating physical errors.

[0067] Step S1013: Lock the effective mesh cells using the dynamic validity mask, introduce the ambient background radiation temperature and dynamic emissivity spectrum into the effective mesh cells to calculate the real thermodynamic temperature field, and simultaneously perform dark current subtraction and flat field correction to calculate the normalized texture field.

[0068] Using dynamic effectiveness mask The effective computation region is locked, and computation is performed only on the mesh cells of the dynamic validity mask. For visible light data, a normalized texture field is computed. Its core lies in eliminating inherent sensor noise and uneven illumination in the optical system. The specific process is as follows: Obtain the original grayscale matrix output by the visible light camera at the current moment. Retrieve the inherent noise matrix that has been calibrated and stored in a completely dark environment. The calibration method for this matrix is ​​as follows: In a completely dark environment, such as after covering the lens cap, continuously collect data. For a single frame of dark-field image, the arithmetic mean of the grayscale value of each pixel is calculated to eliminate fluctuations in random thermal noise and obtain a stable dark current noise floor matrix. One frame, for example, 50 frames; retrieve the brightness gain matrix that has been calibrated and stored beforehand under illumination by a uniform surface light source (such as an integrating sphere). The physical meaning of this matrix is ​​the reciprocal of the light flux attenuation coefficient from the lens center to the edge. The calibration method for this matrix involves pointing the camera at a standard surface light source with uniform brightness, such as the opening of an integrating sphere, acquiring a white field image, normalizing the brightness value of the brightest point in the image center to 1, and calculating the reciprocal of the ratio of the brightness of each pixel in the image to the center brightness. This generates a gain coefficient matrix used to compensate for the light flux attenuation at the lens edge. For each pixel within an effective grid cell, the normalized texture field is calculated using the following formula. :

[0069]

[0070] Indicated in grid cell The normalized texture field values ​​calculated and eliminated from device errors; This represents the original sensor observation grayscale value mapped to this grid cell; and These represent the dark current noise value and the brightness gain coefficient value corresponding to the location of the grid cell, respectively. This formula eliminates dark current noise and lens vignetting, resulting in... The numerical changes are linearly correlated only with the reflectivity characteristics of the medicinal material surface, thus providing pure data for subsequent material identification through texture. For infrared data, the true thermodynamic temperature field is calculated. Its core lies in stripping away environmental reflected radiation and correcting for the influence of emissivity. The specific process is as follows: Obtain the total equivalent radiance detected by the infrared sensor. This refers to the raw signal directly output by the sensor (such as voltage or grayscale value), which, after radiation calibration based on the photoelectric response characteristic curve of the infrared detector, is converted into a physical quantity that is positively correlated with the radiation power of the target surface. Radiation calibration of infrared sensors is a common existing technology, and the calibration mathematical formulas will not be elaborated here. (Obtaining the background temperature inside the pot...) This parameter is acquired in real time by an auxiliary thermocouple or independent infrared probe installed in a non-heated area (such as the cold area above the pot opening) on ​​the inner wall of the stir-frying machine to collect the background temperature. The physical meaning of this value is that it represents the average equivalent temperature of environmental interference heat sources emitted by other high-temperature heat sources inside the pot (such as heating tubes and the hot pot wall) and reflected by the surface of the medicinal materials, in addition to the radiation from the medicinal materials themselves, and entering the infrared sensor. The dynamic emissivity spectrum... A specific dynamic emissivity coefficient is assigned to each grid based on the material of the surface identified from the visible light image at the current moment (such as the metal pot wall, the surface of medicinal materials, or an oil film). Specifically, for each valid grid cell Using the normalized texture field that has already been solved Material classification is performed based on the differences in optical reflection characteristics between metals and organic materials. Metal pot walls exhibit high-brightness specular reflection characteristics under visible light, while medicinal materials exhibit low-brightness diffuse reflection characteristics. A grayscale threshold is preset. For example, 220; if a certain effective grid cell texture value If it is determined to be a highly reflective metallic material, it is assigned a low emissivity, such as Conversely, if the emission rate is not high, it is determined to be a medicinal material and assigned a high emissivity, such as... Based on the Stefan-Boltzmann law-based reverse radiative transfer model, the following formula is used to construct the real thermodynamic temperature field. :

[0071]

[0072] This represents the numerical value of the true thermodynamic temperature field of the object surface at this grid location (unit: Kelvin or degrees Celsius). This represents the total equivalent radiance, including its own radiation and reflected radiation, detected by the infrared sensor at this grid location. This term represents the environmental reflected radiation component, which physically represents the spurious energy in the signal received by the sensor that is actually reflected from the surrounding environment by the medicinal material; this must be subtracted. The formula subtracts the energy reflected from the ambient temperature from the total detected energy. The background radiation component that is induced and reflected on the surface of the object, and the remaining spontaneous emission energy divided by the dynamic emissivity coefficient. Normalization is performed, and the fourth root is used to inversely determine the true physical temperature of the object's surface, thereby eliminating the physical error of different materials having different radiance at the same temperature.

[0073] Step S1013 performs purification and inversion of the multidimensional physical field, based on the radiative transfer equation and the physical nature of photoelectric imaging. The original sensor readings are mixed with inherent dark current noise from the device, edge attenuation from the optical system, and thermal radiation interference from the environmental background, and cannot directly represent the objective properties of the medicinal material. This step serves as a bridge connecting the perceived signal and physical reality. By subtracting background noise and background reflection components, and restoring the total equivalent radiance to the true thermodynamic temperature, the numerical changes in the data are linearly related only to the material properties and thermal state of the medicinal material's surface. This provides pure and physically accurate underlying data support for subsequent accurate differentiation between metals and medicinal materials, as well as the capture of microscopic hotspots.

[0074] Reference Figure 3The bottom-level workflow addresses the potential sampling asynchrony issue between visible light and infrared sensors by acquiring a hardware trigger signal as a unified time reference. Optical flow analysis is then used to analyze the normalized texture field of two consecutive frames to calculate the inter-frame displacement field of the medicinal material particles (e.g., ...). Figure 3 (As shown by the green arrow in the middle), after feature alignment is completed using this displacement field, the encapsulated data packet is used as the input tensor of the subsequent instance segmentation network.

[0075] Step S1014: Based on the hardware time base and the inter-frame displacement estimated by the optical flow method, perform motion compensation interpolation on the asynchronously sampled real thermodynamic temperature field and the normalized texture field to align them to the same physical time and encapsulate them into a pot interior heat-texture temporal field data packet.

[0076] Obtain the hardware trigger signal from the main control system of the herb roasting machine as a unified time reference. Considering visible light cameras With infrared camera Potential sampling asynchrony or transmission delay can be analyzed using optical flow to examine the normalized texture fields of two consecutive frames. The pixel motion vectors between the particles are used to calculate the inter-frame displacement field of the medicinal material particles in polar coordinate grid. ; based on the aforementioned time base For the target time, utilize the inter-frame displacement field Real thermodynamic temperature fields acquired at asynchronous times With normalized texture field Motion compensation interpolation is performed to correct the two data streams to their spatial positions at the same physical moment, eliminating thermal-texture spatial misalignment caused by the rapid tumbling of the medicinal materials; the aligned true thermodynamic temperature field is then used. With normalized texture field According to polar coordinate index Channel merging is performed, and the data is encapsulated to generate a standard in-pot heat-texture temporal field data package. The data in this data packet has already undergone mask-based background removal in step S1013 and is directly used as the input tensor of the instance segmentation network in the subsequent step S102.

[0077] Step S1014 performs spatiotemporal alignment and encapsulation of asynchronous sampling, based on the Lagrange perspective and optical flow estimation theory in fluid mechanics. In the dynamic scene of high-speed tumbling of medicinal materials, even a microsecond-level sampling time difference can cause significant spatial misalignment of high-speed moving particles, resulting in a mismatch between temperature data and texture features. This step, by calculating pixel-level motion vectors and performing motion compensation interpolation, smooths out the deviation caused by asynchronous hardware sampling frequencies in the time dimension, ensuring that the encapsulated temporal field data package accurately describes the transient state of the medicinal materials at the same physical moment, preventing data feature mismatch and causal inversion caused by time slippage.

[0078] Step S102: Input the heat-texture temporal field data package inside the pot into the instance segmentation network. Based on the bimodal distribution characteristics of the particle size projection area, divide the medicinal material target in the picture into a set of dominant principal particles and a set of hidden debris.

[0079] Unpacking the internal heating of the pot—texture temporal field data package Extract the normalized texture field that has been spatiotemporally aligned and has had background interference removed. With the real thermodynamic temperature field Construct a two-channel input tensor The dual-channel input tensor The input is fed into the Mask R-CNN instance segmentation network. This model is chosen because medicinal herbs exhibit significant physical stacking and edge adhesion during the roasting process. Traditional semantic segmentation or threshold segmentation cannot distinguish between these adhered, independent individuals. Mask R-CNN, however, possesses instance-level recognition capabilities, accurately separating physically contacting but logically independent medicinal herb particles into different objects through a Region of Interest (ROI) alignment mechanism. The network preferably uses ResNet-101 combined with a feature pyramid network as the backbone to extract multi-scale deep features. A region generation network generates candidate anchor boxes, and the RoI Align layer achieves feature alignment, connecting parallel classification / regression branches and mask prediction branches. The Mask R-CNN network mainly consists of three cascaded parts: feature extraction backbone, region proposal generation, and functional branch prediction. It uses ResNet-101 combined with a feature pyramid network as the backbone to extract multi-scale deep semantic features from the input dual-channel tensor. A region proposal network generates candidate anchor boxes covering the target medicinal material, and the RoIAlign layer precisely maps features of different sized regions of interest to fixed-size feature vectors to eliminate spatial quantization errors. The network ends with parallel fully connected classification / regression branches and a fully convolutional mask branch, outputting the medicinal material category confidence, bounding box coordinates, and pixel-level binary masks, respectively. During the training phase, the model constructs a multi-task loss function that includes classification error, bounding box regression error, and mask segmentation error. End-to-end supervised training is performed using a stochastic gradient descent (SGD) optimizer. Initial learning rate (e.g., 0.002), momentum parameter (e.g., 0.9), and weight decay coefficient are set. The loss is calculated by forward propagating labeled thermal-texture images of medicinal materials into the input network, and the network weight parameters are iteratively updated using a backpropagation algorithm until the total loss value of the model converges on the overlapping medicinal material segmentation task. This establishes a nonlinear mapping relationship from the multidimensional feature space to the geometric boundaries of independent entities. The correlation between input and output lies in the fact that the network, through supervised learning, establishes a nonlinear mapping relationship from the joint feature space of texture edges and temperature gradients to the geometric boundaries of independent instances; that is, it uses texture abrupt changes in the visible light channel to locate physical edges and uses temperature consistency in the infrared channel to confirm the same material properties, thereby accurately segmenting independent medicinal material individuals in overlapping scenarios. The input of this model is the aforementioned dual-channel input tensor. The output is a set of pixel-level binary masks for all independent medicinal material targets within the current frame. ,in For the target index; the meaning of the pixel-level binary mask set is: for the first identified pixel in the image. For each individual medicinal herb target, a matrix with the same size as the original image is generated. In this matrix, only pixels belonging to that specific medicinal herb target are marked with a 1, while all other pixels are marked with a 0. This achieves exclusive digital labeling for each individual medicinal herb. The remaining pixels include the background and other medicinal herbs. The output set is then traversed. Calculate the number of pixels with a value of 1 in the mask for each target instance, and use this number as the grain size projection area. A particle size distribution histogram was constructed; based on the physical law of the bimodal distribution of large and small particles during the crushing process of medicinal materials, an adaptive threshold algorithm was used to calculate the critical area threshold that distinguishes between main particles and debris. Adaptive thresholding algorithms, such as Otsu's method, will apply to all conditions that meet the following criteria. target instance Classified as a set of dominant principal particles To satisfy all The target instances are classified as latent debris sets. This allows for refined sorting of targets based on multidimensional field data.

[0080] Step S103: Using the set of dominant master particles as the graph node skeleton and the flow channel saturation index of the set of hidden debris as the edge weight, construct a discrete element semantic graph of medicinal material stacking.

[0081] Initialize a non-Euclidean topological data structure, specifically defined as an undirected weighted attribute graph, which is a discrete element semantic graph of medicinal herb stacks. The graph consists of a vertex set Edge set and attribute weight set It consists of three parts. Constructing the vertex set. This set is the set of dominant principal particles output in step S102. The mapping representation in graph data structures is constructed by traversing the set. For each principal particle instance, extract its unique centroid physical coordinates. Generate a corresponding graph node for each master particle, and aggregate all generated nodes to form a vertex set. This establishes the skeleton nodes of the graph. Edge sets are then constructed. This set directly inherits the main particle rigid support network constructed in step S1031. The topological connectivity is used to instantiate each Delaunay triangle physical connection connecting adjacent principal particle centroids in the network as an undirected edge in the graph, thus forming an edge set. This represents the potential heat conduction paths and physical contact relationships between the main particles. Constructing an attribute weight set... Establish a mapping mechanism from physical flow channel state to graph edge weights, traversing each triangular mesh cell calculated in step S1032. and its corresponding flow channel saturation index Because of each edge Physically, it constitutes the common boundary of one or two triangular mesh cells; therefore, the flow channel saturation index of the triangular mesh cell associated with this edge... Assigned as the fluid resistance attribute of that side to the attribute weight set. If two triangles are associated, the mean or maximum value of the two triangles is taken. The generated discrete element semantic graph of stacked medicinal materials is then used. middle, The spatial location of the medicinal materials was anchored. It depicts the neighborhood topology of medicinal materials, and This numerically represents the resistance to hot air flow caused by debris filling on the neighborhood path, providing a standardized tensor input containing complete physical semantics for subsequent decoupling of microscopic thermal accumulation features using a graph neural network model.

[0082] The process of calculating the flow channel saturation index in step S103 specifically includes the following steps:

[0083] like Figure 4 The diagram shown illustrates the physical construction principle of discrete element semantic graphs.

[0084] Reference Figure 4 The diagram shows three spherical objects labeled as main particle A, main particle B, and main particle C (i.e., nodes). The triangle formed by the dashed lines represents a rigid connection (edge), and the orange-filled areas in the corners represent the area occupied by the solid phase. The blue-filled area in the middle of the diagram represents the effective pore flow channel area. The gray circular area in the middle of the diagram and the spots inside it represent the geometric retention area of ​​debris formed by the accumulation of hidden debris. ).

[0085] Reference Figure 4 The following description will be based on a triangular mesh cell consisting of three adjacent master particles. Figure 4 The image shows a rigid support network unit for principal particles constructed by connecting the centroids of adjacent principal particles using the Delaunay triangulation algorithm. The physical coordinates of the centroids of principal particles A, B, and C are extracted and denoted as follows: , and (i.e., the three vertices shown in the diagram). Using these centroid coordinates, a non-overlapping triangular mesh topology is constructed, with each triangular mesh cell physically simulating an independent heat exchange microchamber. Next, the total geometric area of ​​this triangular mesh cell is calculated. And deduct the projected area actually occupied by the main particle entity at the three vertex positions (i.e., the area occupied by the solid phase shown in the figure). Subtracting the two ( The effective pore flow area of ​​the blue region shown in the figure is obtained. This quantifies the size of the free cross-section (theoretical maximum permeability) through which hot air can directly pass through the microchamber gap.

[0086] The hidden debris set is mapped into the network, and it is determined whether the debris point cloud falls within the boundary of the triangular mesh cell. The physical areas of all debris pixels falling inside the cell are accumulated to form the gray debris geometric retention area shown in the figure. Finally, using the formula Calculate the flow channel saturation index. This ratio intuitively reflects the degree to which the effective pore flow channels (blue area) shown in the figure are blocked by debris (gray area), providing an edge weight parameter that accurately reflects the risk of physical blockage for subsequent semantic map construction.

[0087] Step S1031: Use the triangulation algorithm to connect the centroids of the explicit main particle set to construct the main particle rigid support network, and subtract the area occupied by the main particles from the remaining geometric space in each unit of the main particle rigid support network to obtain the effective pore channel area.

[0088] Traversing the set of dominant principal particles Here, the principal particle instance refers to a medicinal herb slice that is determined to have a particle size greater than the critical area threshold and an independent closed boundary in the mask set output by the instance segmentation network, such as a complete Astragalus membranaceus slice or Atractylodes macrocephala slice. The centroid physical coordinates of each principal particle instance in the set are extracted. The specific method for obtaining the coordinates is to calculate the first-order geometric moment of the binarized mask matrix of the particle to obtain its centroid position in the pixel coordinate system, and then combine it with the mesh mapping relationship in step S1011 to convert it into coordinates in physical space, thereby forming a discrete set of centroid points of the main particle. ,in Indicates the first The physical coordinates of the centroids of each principal particle instance are determined. The Delaunay triangulation algorithm is used to analyze the set of centroid points. Perform topology connections to construct a rigid support network for master granules. To clarify the specific construction steps, this process follows the law of an empty circumscribed circle, and the specific execution logic is as follows: In the point set... Search for any three adjacent centroids in the middle. As vertices of the candidate triangle, where Let each of the three adjacent principal particles selected as vertices of the candidate triangle represent their centroid coordinates; construct a triangle through these three vertices. The unique circumcircle, the set of detection points Check if all other remaining centroids are located inside the circumcircle; if the circumcircle does not contain any other centroids, then these three centroids constitute a valid topological connection, and three edges are generated to connect them. And establish an effective triangular mesh unit. The above verification is repeated for all potential combinations within the point set, ultimately forming a triangular mesh topology that is non-overlapping, seamlessly covering, and maximizes the minimum interior angle of all triangles. This network physically simulates the force chain structure of stacked medicinal materials, with each triangular mesh unit... This represents an independent heat exchange microchamber enclosed by three adjacent master particles. Spatial subtraction is performed on each triangular mesh element to extract the effective pore flow channel; specifically, this involves calculating the value of that triangular mesh element. Geometric total area This refers to the mathematical area of ​​the triangle, and the sum of the projected areas actually occupied by the main particle entities at the three vertices within the unit's range, denoted as . The basis for subtracting the two is that, according to the principle of volume conservation in porous media, any infinitesimal space consists of two parts: a solid-occupied region and a gas-flow region. Therefore, the total geometric area... Subtract the area occupied by the solid phase The effective pore flow channel area can then be obtained. This quantifies the free cross-sectional size through which hot airflow can directly pass through the micro-cavity gap without considering internal heat conduction within the particles, thus establishing the geometric denominator for subsequent calculations of debris blockage.

[0089] Step S1032: Map the hidden debris set to the main particle rigid support network, calculate the geometric retention area ratio of the debris in the effective pore flow channel area, and generate the flow channel saturation index;

[0090] Execute hidden debris set The precise coordinate mapping to physical space is specifically based on the reverse ray projection model constructed in step S1011 and the pot reference coordinate system. traversal Each fragment pixel in the image is represented by its image plane coordinates. The inverse kinematics method calculates the physical center coordinates of the object on the bottom plane of the pot, denoted as . ,in, and Corresponding to the definitions in step S1011 shaft and The physical distance value along the axis is used to transform discrete raster pixels into a spatial point cloud with real physical measurements, thus realizing the integration of debris data with the rigid support network of the main particles. Spatial alignment within the same vector coordinate system. Performing spatial inclusion determination and area statistics based on geometric topology, utilizing point-polygon inclusion determination algorithms from computer graphics, such as ray casting, to map the physical center coordinates of the previously mapped debris. Each triangular mesh element in the rigid support network of the main particles The vector boundaries are compared; if the coordinates of a certain debris point are determined... Located in a certain triangular mesh cell Within the geometric boundary, the debris point is classified as retained material within that flow channel. Calculate the geometric retention area. For each triangular mesh cell The sum of the physical areas of all the debris pixels that are determined to fall within it, i.e., the area of ​​a single pixel. The number of pixels falling into this unit, this accumulated value is the geometric retention area. Physically, this represents the cross-sectional area actually occupied by the debris entity within the effective clearance of the microchamber. Using the formula... Calculate the flow channel saturation index for each triangular mesh cell. ; where the denominator The theoretical maximum permeability cross section calculated in step S1031, after deducting the main particle entity, is the molecular... This ratio represents the current actual blocked cross-section. It intuitively reflects the degree of blockage in the flow channel and is a normalized dimensionless physical quantity (range of values). to ),when Approaching When the flow channel contains only a small amount of suspended dust, it is in a dilute phase suspension state of gas-solid two-phase flow, resulting in high heat exchange efficiency; when Approaching When the effective pores are almost filled with debris and are in a dense phase blockage state, they are prone to forming local thermal accumulation singularities because hot air cannot remove the heat. This provides edge weight parameters that accurately reflect the risk of physical blockage for the subsequent construction of semantic graphs.

[0091] Steps S1031 to S1032 construct a discrete element semantic graph and calculate the flow channel saturation. The analysis is based on the volume conservation principle and force chain topology in porous media fluid dynamics. Traditional visual methods struggle to penetrate the porous structure within particle stacks, while the rigid network constructed through Delaunay triangulation can simulate the physical contact relationships between medicinal materials. By calculating the geometric filling ratio of debris in the grid flow channels, the microscopic particle stratification phenomenon is quantified into a specific flow channel saturation index. This step reveals the root cause of hot air channel blockage from a geometrical-physical perspective, transforming the difficult-to-quantify stacking state into edge weight parameters that graph neural networks can understand. This is a crucial step in achieving the leap from macroscopic appearance to microscopic mechanism.

[0092] The dynamic control optimization method for heating processes based on large models also includes the following steps:

[0093] Step S200: Input the stacked discrete element semantic map of the medicinal materials into the multi-scale anti-smoothing graph neural network model, perform deep inference based on long-tail feature enhancement, and decouple the microscopic thermal accumulation singularity features;

[0094] The multi-scale anti-smoothing graph neural network model consists of a dilated convolution branch, a point cloud attention branch, a feature aggregation layer, and a residual amplification module. The model employs a dual-stream parallel feature extraction—nonlinear aggregation—residual amplification composite network topology. Specifically, the model's input is split between the dilated convolution branch and the point cloud attention branch, forming a parallel feature extraction front-end used to capture wide-area background and microscopic anomalies, respectively. The outputs of the dilated convolution branch and the point cloud attention branch are connected to the input interface of the feature aggregation layer, which enables projection alignment and differential fusion of multi-source heterogeneous features. The output of the feature aggregation layer is connected to the residual amplification module, which performs nonlinear gain reconstruction on the selected long-tail features, ultimately outputting microscopic thermally accumulated singularity features.

[0095] The process of decoupling the microscopic thermal accumulation singularity features in step S200 specifically includes the following steps:

[0096] Step S201: Stack the discrete semantic graph of the medicinal materials, extract the global smooth temperature feature matrix through the dilated convolution branch, and lock the local blocking weighted gradient vector through the point cloud attention branch;

[0097] Discrete element semantic graph of medicinal material stacking Feature mounting is performed to construct the original thermodynamic feature matrix of the graph nodes. Specifically, it involves traversing the graph. vertex set Using the centroid physical coordinates of each principal particle node Indexed return to internal heating—texture temporal field data package Extracting the true thermodynamic temperature field The temperature scalar value at the corresponding coordinate position is assigned to the corresponding graph node, thus forming a graph with dimension . The node input matrix, where This represents the total number of medicinal herb particle nodes in the graph; each row of this matrix directly stores the current Celsius temperature value of the corresponding medicinal herb particle. This complete semantic graph carrying temperature features... The input is split into a dilated convolution branch and a point cloud attention branch. For the dilated convolution branch, a graph dilated convolutional network is used, whose specific structure includes multiple stacked graph convolutional layers, and a dilation rate parameter is introduced. (For example This model exponentially expands the receptive field of nodes without increasing computational cost. The reason for using this model is that it requires utilizing the edge set of the graph. By obtaining the background temperature reference value for each node within a generalized spatial range, dilated convolution can efficiently achieve long-distance numerical aggregation across nodes. The input to this branch is the original thermodynamic feature matrix of the graph. With edge set The output is a globally smoothed temperature feature matrix. The globally smoothed temperature feature matrix is ​​a matrix with the same dimension as the input. Floating-point matrix, where The dimension of the feature channel output by the dilated convolution; the th element in the global smoothing temperature feature matrix. The value in the row represents the number of rows. Nodes and their The distance-weighted arithmetic mean of the temperature values ​​of all nodes in the topological neighborhood of order 1, where This represents the unique sequence index of the graph node being traversed within the vertex set, with values ​​ranging from... arrive ,in This indicates the range of hops in the graph covered by the convolution operation. The value is no longer the original temperature of that point, but rather a value reflecting the average heat level of the region after spatial low-pass filtering, serving as the subtraction benchmark for subsequent difference calculations. The correlation between input and output lies in the fact that the model performs a numerical smoothing operation through weighted summation of the convolution kernel on the graph topology, transforming discrete single-point temperatures into regional average temperatures. For the point cloud attention branch, anomalies are locked using graph weights. The model type used is a graph attention network, whose specific structure includes a multi-head attention mechanism layer. The reason for choosing this model is the need to utilize the edge weight attributes of the graph. This branch calculates the microscopic temperature difference between nodes and their neighbors, and nonlinearly amplifies this temperature difference based on the blockage status of the flow channels. The input to this branch is the feature matrix of the graph. and the flow channel saturation index as edge weight The output is a locally blocked weighted gradient vector. The locally blocked weighted gradient vector is a vector with dimension 1. , where The vector represents the feature dimension of the attention network output, where the first element is the first element. The value of the nth element represents the nth... Each node and its components The temperature difference between first-order neighbor nodes is defined as the saturation index of the edge weighted flow channel. The product mapping value; that is, this value intuitively reflects the comprehensive score of the degree of blockage multiplied by the magnitude of the temperature difference. The correlation between input and output lies in the fact that the network, through the calculation of the attention coefficient, incorporates the physical flow channel blockage parameters in the graph. The feature gain coefficients, converted into mathematical values, provide a directional numerical amplification of the original temperature gradient. The graph dilated convolutional network and graph attention network, along with their internal convolution and attention calculation processes, are all existing technologies commonly used in the field of artificial intelligence, and will not be elaborated upon here.

[0098] Step S202: In the feature aggregation layer, using the global smooth temperature feature matrix as the smoothing benchmark, perform residual significance verification on the local blocking weighted gradient vector. If the residual value exceeds a preset threshold, a cluster of high-temperature pixels with a very small area is obtained.

[0099] Constructing a feature aggregation layer, the specific type of this model preferably adopts... A convolutional neural network (CNN), in its specific structure, consists of an input interface, channel projection units, and nonlinear activation units. Essentially, it is a fully connected layer that shares weights across the feature channel dimension. The reason for using this model is that the feature vectors output by the dilated convolution branch and the point cloud attention branch typically have different channel dimensions and inconsistent physical units, making direct subtraction mathematically meaningless; while Convolution can map heterogeneous input features to a feature space of the same dimension by learning a trainable weight matrix. This achieves semantic alignment and dimensional unification of features. The topology of convolutional neural networks and their internal matrix multiplication operations are common techniques in the field of deep learning, and will not be elaborated upon here. The input to this model is a globally smoothed temperature feature matrix. Weighted gradient vector with local blocking The output is the dimension-aligned projected features, denoted as follows: and This is represented as the globally smoothed temperature feature matrix after dimension alignment and the locally blocked weighted gradient vector. The correlation between input and output lies in the fact that the model, through linear projection, preserves the spatial topological information of the graph nodes and only performs linear recombination on the feature channels, thus providing algebraically compatible input for subsequent residual calculation. Residual significance verification is performed using the formula... Calculate each graph node Thermodynamic residual modulus ,in, and They represent the first Each node in a unified feature space The local gradient projection vector and the global smooth projection vector in the model; This represents the L2 norm operation. The physical meaning of this residual modulus is the net anomaly intensity value caused only by channel blockage and localized heat accumulation after eliminating the influence of overall ambient temperature rise (i.e., smoothing the baseline). A preset residual threshold based on statistics is set. The calculation formula is: ,in This represents the arithmetic mean of the thermodynamic residual magnitudes of all graph nodes within the current frame. This represents the standard deviation of the thermodynamic residual magnitudes of all graph nodes within the current frame. This threshold is set according to the three-standard-deviation criterion of the normal distribution, used to identify statistically significant outliers. The entire graph is traversed, and all nodes that meet this criterion are selected. The nodes and their corresponding feature vectors are extracted to form a cluster of extremely small high-temperature pixels. The extremely small area high temperature pixel cluster is a set of discrete nodes in the image whose residual values ​​are significantly higher than the background noise. These nodes often present a discontinuous point or small patch distribution in physical space, corresponding to medicinal material fragments or edges, rather than continuous main areas, thus accurately anchoring the scorching risk points in the long tail distribution at the data level.

[0100] Step S203: Input the extremely small area high temperature pixel cluster into the residual amplification module, use a nonlinear activation function to perform gain mapping on the thermodynamic feature values, and output the microscopic thermal accumulation singularity feature.

[0101] The extremely small area high temperature pixel cluster The residual amplification module, input to the model, has a core parameterized nonlinear transformation layer. The specific computation process utilizes formulas... Map the input features, where, This is a learnable gain weight matrix used to stretch the amplitude of the anomalous signal within the feature space. It is the bias vector; For the selected nonlinear activation function, ReLU or ELU is preferred, as it preserves gradient information in the negative interval and introduces nonlinear excitation. The microscopic thermal accumulation singularity features are output through this operation. Numerically, it is no longer a simple temperature value, but a high-dimensional tensor that has undergone signal amplification and feature reconstruction. It maps the originally faint early scorch signs into a feature expression with high response intensity. The originally faint early scorch signs are transient high temperatures that only appear on a few pixels, ensuring that the subsequent step S300 can maintain high sensitivity to capture tiny thermal accumulation singularities when reconstructing the hidden layer state in the blind zone.

[0102] Steps S201 to S203 perform deep inference based on long-tail feature enhancement, which is based on the long-tail distribution theory in statistics and the feature decoupling idea in deep learning. In conventional convolution operations, extremely small areas of high-temperature anomalies are often averaged out and disappear by the surrounding large areas of normal background features. However, this step extracts the global smooth background and local micro-gradients separately through a parallel branching structure, and uses residual verification technology to specifically screen out those discrete signals that significantly deviate from the background trend and amplify them nonlinearly. This step solves the technical defect of traditional algorithms that ignore early local thermal accumulation due to the smoothing effect, ensuring that the model can capture those tiny scorch signs hidden under the overall uniform appearance as keenly as a microscope.

[0103] The dynamic control optimization method for heating processes based on large models also includes the following steps:

[0104] Step S300: Based on the microscopic thermal accumulation singularity characteristics and combined with the Brazil fruit effect principle in fluid dynamics, reconstruct the thermodynamic state vector of the blind zone hidden layer and calculate the layered solidification risk index.

[0105] Step S300, which involves reconstructing the thermodynamic state vector of the blind zone hidden layer and calculating the stratification solidification risk index, specifically includes the following steps:

[0106] Step S301: Map the microscopic thermal accumulation singularity features to the pot reference coordinate system, use the variational autoencoder to deduce the sedimentation probability of the medicinal material particles under the action of gravity field and centrifugal force field, and generate a layered probability density matrix.

[0107] The mapping to the pot body reference coordinate system is to map the microscopic thermal accumulation singularity features. Each feature node in the equation is associated with its defined centroid physical coordinates. By performing index association, the feature data is anchored to the pot reference coordinate system constructed in step S1011. In the middle, and in the vertical direction (i.e., the Z-axis direction), the space is divided into A discrete virtual depth layer, for example A variational autoencoder is constructed and run. This model belongs to the generative probabilistic graphical model and consists of an encoder network, a reparameterized sampling layer, and a decoder network. Microscopic thermal accumulation singularity features are then utilized. The observation data is input into the encoder network. The encoder network specifically employs a multilayer perceptron or a one-dimensional convolutional neural network structure to map a high-dimensional feature space to a low-dimensional latent probability distribution space. The output of the encoder network contains two vectors: a latent mean vector and a latent mean vector. With the latent log-variance vector These two parameters together define the distribution parameters of the latent space. Their physical meaning is that, for each medicinal particle, the model no longer predicts its definite position, but rather the Gaussian distribution characteristics of its motion state in the physical field, i.e., the expected value and uncertainty range of its position. Latent variables are generated based on these distribution parameters. To ensure the differentiability of the model during training, a reparameterization technique is employed, utilizing the formula... Perform calculations, where From the standard normal distribution The auxiliary noise vector randomly sampled in the middle, where I represents the identity matrix, is the latent variable obtained through this step. In mathematics, this represents a latent state sample obtained by randomly sampling from a predicted probability distribution; it carries the most likely implicit physical properties of the medicinal material particles at the current moment. The latent variables... The vectors are concatenated with the external physical field parameters and then fed into the decoder network as input. The decoder network also employs a multilayer perceptron structure, with its output layer connected to a Softmax activation function, outputting a hierarchical probability density matrix. The external physical field parameters include the gravitational constant. The centrifugal force field parameters determined by the rotation speed of the herb-frying machine , For example, the preset standard gravitational acceleration vector. Based on the physical entity corresponding to the microscopic thermal accumulation singularity characteristics, namely the dominant master particle, the vertex set constructed in step S103 is recalled retrospectively. Extract the unique centroid physical coordinates of each principal particle instance. , the coordinates In the reference coordinate system of the pot body The three-dimensional spatial components are defined as follows: ,in, This indicates that the center of mass of the principal particle is at the bottom plane of the pot. Physical coordinates on the axis Indicates that it is in Physical coordinates on the axis Indicates its vertical direction The physical height value on the axis. Using the coordinate components defined above, calculate the centroid of the main particle relative to the rotation center axis of the pot body (i.e., The horizontal projection distance of the axis is denoted as the radial distance. The calculation formula is: This distance physically represents the radius lever arm of the particle's rotational motion. The rotational speed feedback value of the stir-frying machine motor is read in real time and converted into angular velocity. Using the formula Calculate the scalar of the centrifugal acceleration currently experienced by the particle. To accurately describe the directionality of force (radial outward) to a neural network, a three-dimensional vector is constructed using coordinate components. This vector, acting as a dynamic constraint input to the decoder, corrects the direction of sedimentation probability estimation. It precisely describes the radially outward physical thrust experienced by each particle at its current spatial position due to rotation, providing a dynamic constraint for the decoder to reconstruct the sedimentation probability. The Softmax activation function exponentially normalizes the original values ​​output by the decoder, compressing the values ​​to... arrive Within the range, and ensure that each particle node is within The sum of the probability values ​​on each virtual depth layer is strictly equal to This generates a valid hierarchical probability density matrix. The hierarchical probability density matrix It is a dimension A two-dimensional matrix, where The number of nodes in the graph, and the elements in the matrix. Indicates the first The medicinal material node particles fell into the first The probability value of the virtual depth of the layer. The reason for adopting this model is that the "Brazil fruit effect" (i.e., large particles float and small debris sinks) of medicinal particles during the stir-frying process has a high degree of randomness and nonlinearity. Traditional deterministic physics engines have too much computation and are difficult to run in real time. However, VAE can learn from a large amount of historical data and infer the probability distribution of particle positions from observed thermal features, effectively handling the uncertainty problem. The correlation between input and output lies in the fact that the model establishes a probability mapping mechanism, which maps the microscopic thermal accumulation singularity features of the input (i.e., the observed appearance) to the sedimentation probability distribution of particles in the vertical direction (i.e., the implicit physical position). That is, the closer the feature value is to the charred feature of the particle, the higher the probability that the model infers that it is located at the bottom layer. The variational autoencoder and its internal neural network hierarchical structure, loss function construction and backpropagation calculation process are all existing technologies commonly used in the field of artificial intelligence, and will not be elaborated here.

[0108] Step S302: For the high-density bottom layer region in the stratified probability density matrix, the heat retention time in the current stir-frying cycle is estimated using the heat diffusion inversion algorithm, and the heat retention time is concatenated with the stratified probability density matrix to generate the thermodynamic state vector of the blind zone hidden layer.

[0109] Define the high-density bottom layer region; analyze the stratified probability density matrix. The cumulative probability is summed at the depth level, and the cumulative settlement probability is set to a preset settlement threshold. The set of nodes is defined as the high-density bottom layer region. Subsequent calculations are strictly limited to this region, and the depth dimension is the number of virtual layers. Direction, preset settlement threshold The selection is based on the statistical quantiles of the Brazil nut effect, typically set at 0.8 or 0.9, representing the extremely high probability sedimentation interval with a cumulative probability exceeding 80%. This step, at the data level, identifies those medicinal material particles that are highly likely to deposit at the bottom of the pot and cannot be directly observed visually. The microscopic thermal accumulation intensity value is then obtained. Obtain the extremely small area high temperature pixel clusters selected in step S202. The node indexes are used to directly backtrack and query the original thermodynamic feature matrix constructed in step S201. This matrix stores the true infrared physical temperature after emissivity correction, and the maximum value of the original temperature corresponding to all nodes in the cluster is selected as the [value]. For the high-density bottom layer region in the aforementioned stratified probability density matrix, a heat diffusion inversion algorithm is applied to estimate the heat retention time within the current stir-frying cycle. Specifically, an estimation model is constructed based on the characteristic diffusion time and the dimensionless temperature rise ratio, using the formula... To make an estimate, among which Heat retention time, which physically represents the estimated time during which the particle remains stagnant in the blind zone at the bottom of the pot and fails to be brought back to the surface by stirring. The high-density bottom layer region In the reference coordinate system of the pot body Geometric centroid depth on the axis; The equivalent thermal diffusivity of the medicinal material is obtained by consulting the physical property database of medicinal materials. The physical property database of medicinal materials is constructed in the system memory during the equipment initialization stage by calibrating the thermal properties of specific medicinal material varieties (such as rhizomes and leaves) through standard thermophysical experiments, and then storing the thermal conductivity and specific heat capacity data in the system memory. The background radiation temperature inside the pot is obtained directly from the background temperature variable obtained in step S1013 above. The first term in the above formula represents the measured strength value. This is the characteristic diffusion time, which characterizes the theoretical baseline time required for heat to penetrate to that depth; the second term This is a dimensionless temperature rise ratio, used to correct the baseline duration based on the measured temperature rise. The physical basis for the calculation is that simply knowing the particles are at the bottom and the temperature is high is insufficient to determine scorching; the duration of time must be introduced. The above formula uses the principle of heat diffusion to convert the abnormal temperature difference into a time equivalent, quantifying the duration for which the hot spot remains at the bottom and accumulates heat, thus preventing normal particles that have just fallen to the bottom and have not yet accumulated heat from being misjudged as risk points. Finally, the calculated... and Concatenate to generate the thermodynamic state vector of the hidden layer in the blind zone. This vector realizes the complete physical state reconstruction from cause to effect in terms of spatial probability and time dwell dimension.

[0110] Step S303: Based on the thermodynamic state vector of the blind zone hidden layer, calculate the asynchronous deviation value between the temperature rise rate of the blind zone and the temperature rise rate of the surface medicinal material, and weight-couple the deviation value with the critical coking temperature of the medicinal material to output the layered solidification risk index.

[0111] Calculate the temperature rise rate of the surface medicinal materials Globally smoothed temperature feature matrix The dimension of this matrix is Extract the first feature channel vector of the matrix, i.e., the vector with dimension . The column vector stores the base temperature values ​​after graph-dilated convolution smoothing. The arithmetic mean of all elements in this column vector is calculated to obtain the current average surface temperature. ; Retrieve the average surface temperature of the previous sampling frame stored in system memory, denoted as Combined with the physical sampling time interval between two frames Using the difference formula The first step is to obtain the overall temperature rise rate, representing the temperature rise rate of the surface medicinal material, under normal stir-frying conditions. The second step is to calculate the temperature rise rate in the blind zone. Extracting the thermodynamic state vector of the hidden layer in the blind zone Compared with the measured strength value Based on the principle of energy accumulation, using the formula Calculate the equivalent extrapolated temperature rise rate within the blind zone, i.e., the blind zone temperature rise rate. Its physical essence is the average historical rate, that is, the average rate of temperature rise of the bottom medicinal material during the retention period. Calculate the stratification solidification risk index. Specifically, it refers to the asynchronous deviation value of the calculation rate. In order to uniformly map indicators from different physical dimensions to Probability intervals, for different asynchronous deviation values and temperature approximation term Perform maximum-minimum normalization, where The preset critical temperature for coking medicinal materials is determined based on the current processing specifications for the variety of medicinal materials being processed, such as 200 degrees Celsius or 250 degrees Celsius. The normalized rate deviation is then calculated. Approximation to normalized temperature The exponent is output using a weighted coupling formula. ;in To pre-set the weighting coefficients, these two coefficients are usually predetermined using the analytic hierarchy process (AHP), for example... =0.6, =0.4; when the index When the value approaches 1, it indicates that the bottom not only has severe heat accumulation but also uncontrolled temperature rise far exceeding that of the surface. A solidified layer has formed at the bottom of the pot, requiring immediate triggering of strong disturbance control logic.

[0112] Steps S301 to S303 reconstruct the hidden state of the blind zone and assess the risk of stratification. The analysis is based on a generative probabilistic graphical model and the principle of thermal diffusion in thermodynamics. Since visual sensors cannot directly observe the herbs obscured at the bottom of the pot, this step utilizes a variational autoencoder to learn the sedimentation probability distribution of particles in gravity and centrifugal force fields, and combines this with thermal retention time to estimate the cumulative thermal risk at the bottom layer. This step achieves logical deduction from visible surface data to the hidden state at the bottom layer, quantifies the risk index of stratification and solidification, and enables the system to see through surface illusions and perceive the actual risk of scorching at the bottom, solving the problem of blind zone overheating caused by the inability to see the bottom layer in traditional control systems.

[0113] The dynamic control optimization method for heating processes based on large models also includes the following steps:

[0114] Step S400: Based on the thermodynamic state vector of the blind zone hidden layer and the layered solidification risk index, construct a risk-aware asymmetric objective function, solve the nonlinear turbulent disturbance trajectory parameters, and superimpose compensation on the benchmark overturning curve.

[0115] The process of generating nonlinear turbulent disturbance trajectory parameters in step S400 specifically includes the following steps:

[0116] Step S401: Construct a risk-aware asymmetric objective function, use normalization coefficients to construct a total objective function, and determine the risk-dominant state based on the total objective function;

[0117] A risk-aware asymmetric objective function is constructed to quantify the game relationship between surface consistency and underlying risk. Specifically, the surface consistency loss is calculated. : The global smoothed temperature feature matrix output by backtracking step S201 Calculate the variance of the temperature values ​​of all nodes in the first channel of the matrix, denoted as . This value characterizes the temperature difference dispersion of the surface medicinal material. Calculate the hidden layer risk penalty term. From the thermodynamic state vector of the hidden layer in the blind zone In this process, the channel component that stores the heat residence time value is specifically extracted and defined as the residence time feature vector. Calculate the L2 norm of the vector and combine it with the stratified solidification risk index. Using the formula The risk term is calculated, where γ is a preset sensitivity coefficient. This coefficient is used to adjust the nonlinearity of the model's response to underlying risks. Its value is typically set between 3.0 and 8.0, with the specific value determined based on the stringency of the actual process's zero tolerance for scorching. For example, a larger value, such as 5.0, can be used for valuable medicinal materials to increase sensitivity. To address the issue of inconsistencies in dimensions and differences in orders of magnitude that cannot be directly compared, a normalized scaling factor is used, pre-calculated during the system initialization phase. and Calculate the scaling factors for the surface consistency loss and hidden risk penalty terms based on the average values ​​in historical operating data. and , making Construct a dimensionless overall objective function The criterion for determining whether a risk factor dominates has been revised to compare the dimensionless components after weighting. When the current situation is determined to be risk-dominated, the optimization direction will be forced to shift from maintaining surface uniformity to eliminating bottom risk.

[0118] Step S402: Based on minimizing the gradient direction of the total objective function, calculate the nonlinear turbulent disturbance trajectory parameters and perform superposition compensation on the preset benchmark turbulence curve;

[0119] In situations where risk factors dominate, to ensure that output commands comply with the physical limitations of the mechanical actuator, a pre-calibrated mapping table is used to obtain specific parameters: 1) Amplitude parameters of high-frequency chatter commands. : Calculate the gradient magnitude of the risk penalty term Using this as an index key, a preset vibration amplitude calibration table is retrieved. This table, calibrated through discrete element simulation experiments, establishes a nonlinear correspondence between gradient intensity and optimal vibration amplitude, for example: gradient interval. Corresponding amplitude , interval Corresponding amplitude This outputs the specific physical amplitude. ;2) Transient angular acceleration pulse value First, the real-time rate of change of the stratified solidification risk index is calculated using the difference formula. The calculation formula is as follows: ;in, Indicates the current sampling time The real-time calculated and outputted current frame layered solidification risk index reflects the latest risk status of the medicinal materials. This represents the time interval between two adjacent sampling times. Indicates the previous sampling time The historical frame hierarchical risk index, stored in the system register, serves as a benchmark to determine whether the risk is in a stable or deteriorating state. Multiply by turbulence susceptibility factor The transient angular acceleration pulse value is obtained. The calculation formula is: Turbulence susceptibility coefficient This is a preset system gain constant, which is obtained by constructing a medicinal herb stacking model with different particle size ratios during the discrete element simulation calibration stage before algorithm deployment. This simulates the stratification state caused by the Brazil berry effect. Angular acceleration pulses of different intensities are applied in the simulation to find the minimum effective acceleration threshold that can break the stratified structure without causing physical breakage of the medicinal herb. This threshold is then compared with the risk escalation rate. The ratio of is This coefficient defines the sensitivity of the control system to the worsening risk of charring, ensuring that the motor can output sufficient inertial torque to counteract the static friction of the herb pile. The above calculations... and The output is nonlinear turbulent disturbance trajectory parameters, which are used as dynamic compensation values ​​and mathematically superimposed on the reference stirring curve to drive the actuator to generate a composite motion trajectory of macro-stirring + micro-turbulence. The reference stirring curve is the standard uniform or variable speed stirring trajectory preset in the process package.

[0120] Steps S401 to S402 calculate nonlinear turbulence parameters and implement dynamic control, based on the asymmetric payoff strategy in game theory and the blockage phase transition theory of particulate matter. When stratification risk is detected, traditional linear regulation is ineffective; it is necessary to construct an asymmetric objective function containing a penalty term for the underlying risk to disrupt the system equilibrium and calculate turbulence parameters including high-frequency flutter and transient pulses. This step ensures that the control system can output strong disturbance commands sufficient to overcome the static friction between particles and break the slack structure at critical moments, fundamentally eliminating the particle stratification and accumulation state that leads to scorching from a physical perspective, and achieving the final closed loop of process safety and quality.

[0121] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0122] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0123] The foregoing description is illustrative of the invention and should not be construed as limiting it. Although several exemplary embodiments of the invention have been described, those skilled in the art will readily understand that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the invention. Therefore, all such modifications are intended to be included within the defined scope of the invention. It should be understood that the foregoing description is illustrative of the invention and should not be construed as limiting it to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the appended scope of protection. The invention is defined by its scope of protection and its equivalents.

Claims

1. A dynamic control optimization method for heating processes based on a large model, characterized in that, Includes the following steps: The multispectral real-time monitoring data stream inside the herb roasting machine was acquired, and a discrete element semantic map of herb stacking was constructed based on the spatial distribution differences of herb particle size. The stacked discrete element semantic graph of the medicinal materials is input into a multi-scale anti-smoothing graph neural network model to perform deep inference based on long-tail feature enhancement and decouple the microscopic thermal accumulation singularity features. Based on the microscopic thermal accumulation singularity characteristics and combined with the Brazil fruit effect principle in fluid dynamics, the thermodynamic state vector of the blind zone hidden layer is reconstructed, and the layer solidification risk index is calculated. Based on the thermodynamic state vector of the blind zone hidden layer and the layered solidification risk index, a risk-aware asymmetric objective function is constructed to solve the nonlinear turbulent disturbance trajectory parameters and to superimpose compensation on the benchmark overcooking curve.

2. The dynamic control optimization method for heating process based on a large model according to claim 1, characterized in that, The steps for constructing the stacked discrete element semantic graph of medicinal materials are as follows: The visible light image sequence and infrared thermal imaging sequence inside the stir-frying machine are acquired, and physical space reconstruction and multi-dimensional spatiotemporal field calculation based on the inverse ray projection model are performed to generate a heat-texture temporal field data package inside the pot. The heat-texture temporal field data package inside the pot is input into the instance segmentation network. Based on the bimodal distribution characteristics of the particle size projection area, the medicinal material targets in the picture are divided into a set of dominant master particles and a set of hidden debris. Using the set of dominant master particles as the graph node skeleton and the flow channel saturation index of the set of hidden debris as the edge weight, a discrete element semantic graph of medicinal material stacking is constructed.

3. The dynamic control optimization method for heating process based on a large model according to claim 2, characterized in that, The steps for performing physical space reconstruction and multidimensional spatiotemporal field calculation based on the reverse ray projection model are as follows: A reference coordinate system for the pot body is established with the center of the pot bottom as the origin and the normal of the pot bottom plane as the axis. Based on the intrinsic and extrinsic pose data of the visible light camera and the infrared thermal imaging sensor, a reverse ray projection model is constructed. The intersection of the projections of the visible light cone and the infrared cone on the pot bottom plane is calculated using the reverse ray projection model. Specular reflection blind spots and dynamic occlusion blind spots are removed within the mapping area to generate a dynamic effective mask.

4. The dynamic control optimization method for heating process based on a large model according to claim 3, characterized in that, The steps for performing physical space reconstruction and multidimensional spatiotemporal field calculation based on the reverse ray projection model also include: The effective mesh cells are locked using the dynamic effectiveness mask. The ambient background radiation temperature and dynamic emissivity spectrum are introduced into the effective mesh cells to calculate the real thermodynamic temperature field. At the same time, dark current subtraction and flat field correction are performed to calculate the normalized texture field. Motion compensation interpolation is performed on the asynchronously sampled real thermodynamic temperature field and the normalized texture field to align them to the same physical moment, and then encapsulated into a pot interior heat-texture temporal field data package.

5. The dynamic control optimization method for heating process based on a large model according to claim 2, characterized in that, The calculation steps for the flow channel saturation index are as follows: The centroids of the explicit master particle set are connected using a triangulation algorithm to construct a rigid support network for the master particles. The effective pore channel area is obtained by subtracting the area occupied by the master particles from the remaining geometric space in each cell of the rigid support network.

6. The dynamic control optimization method for heating process based on a large model according to claim 5, characterized in that, The step of calculating the flow channel saturation index further includes: The hidden debris set is mapped to the rigid support network of the main particles, and the geometric retention area ratio of the debris in the effective pore flow channel area is calculated to generate the flow channel saturation index.

7. The dynamic control optimization method for heating process based on a large model according to claim 1, characterized in that, The steps for calculating the stratified solidification risk index are as follows: The microscopic thermal accumulation singularity features are mapped to the pot body reference coordinate system, and the sedimentation probability of medicinal material particles is inferred using a variational autoencoder to generate a layered probability density matrix. Based on the layered probability density matrix, the heat retention time is estimated using the thermal diffusion inversion algorithm to generate the thermodynamic state vector of the blind zone hidden layer. Based on the thermodynamic state vector of the blind zone, the asynchronous deviation between the temperature rise rate of the blind zone and the temperature rise rate of the surface medicinal material is calculated. Combined with the critical coking temperature of the medicinal material, the layered solidification risk index is output.

8. The dynamic control optimization method for heating process based on a large model according to claim 1, characterized in that, The steps for performing deep inference based on long-tail feature enhancement are as follows: A multi-scale anti-smoothing graph neural network model is constructed, which consists of a dilated convolutional branch, a point cloud attention branch, a feature aggregation layer, and a residual amplification module. The stacked discrete element semantic graph of the medicinal materials is input into the multi-scale anti-smoothing graph neural network model. The global smoothing temperature feature matrix is ​​extracted through the dilated convolution branch, and the local blocking weighted gradient vector is locked through the point cloud attention branch. In the feature aggregation layer, the global smooth temperature feature matrix is ​​used as the smoothing benchmark, and the residual significance is checked on the local blocking weighted gradient vector. If the residual value exceeds the preset threshold, a cluster of high-temperature pixels with a very small area is obtained.

9. The dynamic control optimization method for heating process based on a large model according to claim 8, characterized in that, The steps for performing deep inference based on long-tail feature enhancement also include: The extremely small area high-temperature pixel cluster is input into the residual amplification module, and the thermodynamic feature values ​​are mapped by a nonlinear activation function to output the microscopic thermal accumulation singularity feature.

10. The dynamic control optimization method for heating process based on a large model according to claim 1, characterized in that, The steps for calculating the nonlinear turbulent disturbance trajectory parameters and superimposing compensation onto the baseline turbulence curve are as follows: Construct a risk-aware asymmetric objective function, and determine the risk-dominant state based on the asymmetric objective function; Based on the aforementioned risk-dominant state, the nonlinear turbulent disturbance trajectory parameters are calculated, and the preset benchmark overturning curve is superimposed and compensated.

Images