A konjac pasting uniformity recognition control method based on GNN

Abstract

The application relates to a GNN-based konjac pasting uniformity recognition control method, and belongs to the technical field of data recognition, which comprises the following steps: step S1: data acquisition and graph structure construction: constructing parameter data of a pasting state into graph structure data; step S2: graph neural network model processing: inputting the constructed graph structure data into a pre-trained graph neural network model, and finally outputting a spatial distribution heat map; step S3: uniformity recognition and decision: judging whether the current pasting state reaches a preset uniformity standard according to the spatial distribution heat map of the pasting uniformity; step S4: feedback control execution: generating a control instruction according to the uniformity recognition and positioning result, so as to promote pasting uniformization; and step S5: iterative optimization: forming a closed-loop feedback control system, so as to complete precise control of konjac pasting; the application has the beneficial effects that the control recognition precision of the pasting uniformity is ensured, and control lag is prevented.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data recognition, and particularly relates to a konjac gelatinization uniformity recognition control method based on GNN. BACKGROUND

[0002] Konjac is a kind of food and medicine homologous material rich in glucomannan. Gelatinization is a core process in the processing of konjac. After konjac powder is mixed with water, a gelatinous structure is formed under heating conditions. The gelatinization uniformity directly determines the taste, texture, nutrient retention rate and shelf life of the final product. If the gelatinization is uneven, there will be hard cores that are not gelatinized or soft and rotten phenomena that are over-gelatinized, which will seriously affect the market competitiveness.

[0003] The existing konjac gelatinization control method mainly relies on two technical paths: one is experience-based control, which observes the color and viscosity changes of the gelatinization system by artificial observation or relies on the operator's sense of touch. This method is highly subjective and has large errors, and is difficult to adapt to industrial continuous production. The second is traditional sensor control, which uses temperature sensors, viscometers and other devices to collect single-point or local data, and adjusts the heating power, stirring rate and other parameters based on the preset threshold.

[0004] However, the konjac gelatinization system is a heterogeneous complex fluid, and the temperature, humidity and material concentration have significant differences in spatial distribution. Single-point data cannot reflect the overall uniformity, and traditional control algorithms (such as PID) can only achieve parameter feedback adjustment and cannot establish a nonlinear correlation model between multiple factors and uniformity, resulting in strong control hysteresis and a uniformity qualification rate of less than 85%. SUMMARY

[0005] The application provides a konjac gelatinization uniformity recognition control method based on GNN, which is used to solve the technical problems of low recognition accuracy and control hysteresis in konjac gelatinization uniformity control. It outputs a spatial distribution heat map reflecting the overall gelatinization uniformity and a spatial distribution heat map identifying the position of the uneven gelatinization area, realizes the recognition of gelatinization uniformity, and ensures the control recognition accuracy of gelatinization uniformity. According to the spatial distribution heat map of gelatinization uniformity, it is judged whether the current gelatinization state reaches the preset uniformity standard. If not, the gelatinization defect area is located combined with the spatial distribution heat map, which ensures the control recognition accuracy of gelatinization uniformity. A closed-loop feedback control system is formed until the gelatinization uniformity of the whole reaction system reaches the preset standard, so as to complete the accurate control of konjac gelatinization and prevent control hysteresis.

[0006] In order to achieve the above purpose, the application realizes the following technical scheme:

[0007] A konjac gelatinization uniformity recognition control method based on GNN, comprising the following steps:

[0008] Step S1: data acquisition and graph structure construction: a multi-modal sensor network is arranged in the konjac gelatinization reaction container, real-time acquisition of parameter data of the gelatinization state, and construction of the parameter data of the gelatinization state into graph structure data; wherein the parameter data of the gelatinization state includes temperature gradient, concentration gradient and velocity gradient, the graph structure data takes the spatial position point as the graph node, the graph node features are composed of multi-modal data of the corresponding position sensors, and the connection between the graph nodes is connected according to the correlation of fluid dynamics;

[0009] Step S2: graph neural network model processing: input the constructed graph structure data into the pre-trained graph neural network model, the graph neural network model aggregates neighborhood information through the message passing mechanism, extracts and fuses the local features and global topological structure features of each graph node, and finally outputs a spatial distribution heat map reflecting the gelatinization uniformity and a spatial distribution heat map identifying the gelatinization uneven area position;

[0010] Step S3: uniformity recognition and decision: according to the spatial distribution heat map of the gelatinization uniformity, it is judged whether the current gelatinization state reaches the preset uniformity standard; if not, the spatial distribution heat map is combined to locate the gelatinization defect area;

[0011] Step S4: feedback control execution: according to the uniformity recognition and positioning result, a control instruction is generated to drive the execution mechanism to perform directional and accurate regulation and control operation on the gelatinization defect area, which includes but is not limited to adjusting the local temperature, the stirring paddle speed, the direction or the feeding rate, so as to promote the gelatinization uniformization;

[0012] Step S5: iterative optimization: repeat steps S1 to S4 to form a closed-loop feedback control system until the gelatinization uniformity of the whole reaction system reaches the preset standard, so as to complete the precise control of konjac gelatinization.

[0013] Optionally, in step S1, the fluid dynamics related edges are used to connect the graph nodes. In the konjac gelatinization process, the flow state of the material directly reflects the gelatinization uniformity, and the connection needs to be established through the fluid parameter correlation:

[0014] Let the node and the fluid velocity vector at time is:

[0015] and ;

[0016] Wherein, represents the velocity vector or velocity component of the material at time and position , , and are the velocity components of the particle at position in the material along the three coordinate axes, and respectively; is the velocity vector or velocity component of the particle at position at time , and are the velocity components of the particle at position in the material along the three coordinate axes, and respectively;

[0017] The component difference of and is calculated by the velocity vectors at different positions, the velocity gradient is calculated, and the uniformity of the shear stress distribution is analyzed, and finally the gelatinization uniformity is related.

[0018] Optionally, the linear correlation degree of the velocity vector is measured :

[0019] ;

[0020] wherein, is used to describe the velocity vector in which the velocity components and change together over time; if , it indicates that the change trends of and are consistent; it indicates that the change trends are opposite; it indicates that there is no linear common change between the two;

[0021] is a standardization factor, is the fluctuation amplitude of itself; is the fluctuation amplitude of itself; after squaring the product, the standardization function is played, so that the final correlation coefficient is constrained between [−1, 1], and the interference of the correlation degree of the quantity dimension and the fluctuation amplitude of itself is eliminated.

[0022] Optionally, the konjac gelatinization is a process of coupling temperature field and concentration field, thermal mass gradient coupling, and the correlation degree of the length is calculated: the temperature gradient and the concentration gradient between the nodes .

[0023] ​​​ ;

[0024] wherein, represents a ternary gradient term of heat-mass coupling, is a ternary or triple gradient, and represent different positions, is time, and is a physical quantity describing the coupling strength of the common transmission process of heat and mass gradients;

[0025] is a temperature gradient, is temperature, and the temperature gradient is the spatial variation rate, is the difference in the distribution of temperature in space, which is the core driving force of heat transfer, and heat is transferred from high temperature to low temperature, which is driven by the temperature gradient;

[0026] is a concentration gradient, is the mass concentration or particle number concentration, and similar to the temperature gradient, the concentration gradient is the core driving force of mass transfer;

[0027] is a velocity gradient, is the motion velocity of fluid or particles, and the velocity gradient reflects the shear or stretching characteristics of the flow field, and in the convection process, the velocity distribution will simultaneously affect heat transfer and mass transfer;

[0028] is the normalized length of the heat, mass, and flow gradients, is an infinitesimal quantity to avoid a zero denominator;

[0029] is the dot product of the temperature gradient, concentration gradient, and velocity gradient, reflecting the degree of synergistic coupling of the three.

[0030] Optionally, in step S2, for the spatial distribution thermodynamic map reflecting the gelatinization uniformity, starting from the node features of the map, a multi-scale spatial interaction model is established, and finally a continuous value thermodynamic map reflecting the gelatinization uniformity is generated :

[0031] ;

[0032] wherein, is a plurality of coordinates The output thermodynamic map of the gelatinization uniformity reflects the gelatinization uniformity at the position , and is limited to the interval [0, 1] through subsequent truncation operations, and the closer the value is to 1, the higher the gelatinization uniformity; the closer the value is to 0, the lower the gelatinization uniformity;

[0033] This represents a multilayer perceptron, whose function is to process the features of the input. Nonlinear transformation and feature extraction are performed to map the input features to the gelatinization uniformity values ​​and capture data related to gelatinization uniformity in the features. Is input to The feature vectors are derived from multi-dimensional perception of the material.

[0034] It is a truncation operation, which is to... The output value is limited to between 0 and 1;

[0035] Through multilayer perceptron Characteristics that reflect the properties of materials The process involves processing to obtain a preliminary value related to the uniformity of gelatinization; then, a Clip operation is used to constrain this value within the [0,1] interval, ultimately yielding each coordinate on the heatmap. A continuous value that reflects the uniformity of gelatinization at that location. When continuous values ​​are distributed in space, a heatmap of gelatinization uniformity is formed. .

[0036] Optionally, in step S2, the spatial distribution heatmap for identifying the locations of unevenly pasted areas is as follows: Binary region identifiers are generated based on uniformity features to highlight non-uniform regions:

[0037] ;

[0038] in, Representing multiple coordinates The output heatmap of gelatinization non-uniformity reflects the location The degree of unevenness in gelatinization; It is multiple coordinates The output heatmap of gelatinization uniformity The gradient, the magnitude of which reflects the drasticness of the field change, and the direction which reflects the trend of the field change, are used to capture the location where inhomogeneity begins to appear. Is input to The feature vectors in the data are derived from multi-dimensional perception of the material. The connection integrates the original data with the gradient information of the current field, allowing the model to more comprehensively identify uneven regions.

[0039] It is another multilayer perceptron, which, as a nonlinear mapping module, is used to model complex relationships between the fused features of the original basis and gradient.

[0040] is a temperature-dependent activation or transformation function, is emphasized in association with the temperature sensitivity of pasting, which is a typical temperature-driven process, and the pasting behavior varies greatly at different temperatures. This function is used to adapt the influence of temperature on the unevenness judgment;

[0041] is a step function, is a decision threshold; when the output of exceeds, it is determined that the uneven region of pasting is uneven; otherwise, it is a uniform region. Its role is to binarize the continuous model output, clearly distinguishing between uneven and uniform regions.

[0042] Optionally, for the boundary of the uneven region, the accuracy is enhanced to improve the accuracy of the heat map, in order to make the boundary of the uneven region clearer, the accuracy of the uneven region is introduced:

[0043] ;

[0044] wherein, is a plurality of coordinates output of the heat map of the unevenness of pasting with accuracy; represents a plurality of coordinates output of the heat map of the unevenness of pasting, reflecting the unevenness of pasting at the position ; is a plurality of coordinates output of the heat map of the unevenness of pasting gradient, reflecting the rate and direction of change in space; is a plurality of coordinates output of the heat map of the uniformity of pasting gradient;

[0045] is the norm of the gradient of , used to measure the strength of the gradient; is a sign function, used to judge the consistency of the direction of two gradients, the value of is greater than 1, then the direction of the two gradients is the same, the value of is less than 1, then the direction of the two gradients is opposite; is a hyperparameter, used to control the influence degree of gradient interaction on the loss;

[0046] is the gradient direction alignment term, the dot product of the two gradients reflects the consistency of the direction:

[0047] is the gradient strength enhancement term, is the strength of the gradient, by adjusting the gradient, the area with strong gradient is amplified, let the edge, the details of the gradient significant position, combined with the norm of the gradient , can be adjusted according to the strength of the gradient of the role of .

[0048] Optionally, in step S3, based on the GCI thermal map gradient information of the paste uniformity and the paste non-uniformity, the contour is driven to shrink to the GCI mutation, so as to locate the paste defect area, and the contour is driven to move to the area with large GCI gradient modulus:

[0049] ;

[0050] wherein, the paste defect area; is the parametric representation of the contour evolving over time, that is, the parameter at each contour point, is the curve parameter of the contour, is time, which describes the evolution process of the contour;

[0051] is the gradient of the GCI, representing the relevant index of the paste uniformity, the modulus reflects the change rate of the GCI, the area with uniform paste, the GCI changes gently, at the defect boundary, the paste changes from uniform to non-uniform, and the GC changes sharply;

[0052] is the feature factor of the point on the contour, and is the signal strength of the GCI at the contour point , or the attribute of the contour itself; is the contribution of the entire contour from the parameter to ;

[0053] When the contour is close to the paste defect boundary, since the GCI is mutated here, will be significantly increased, resulting in a large increase in the value of the integral function .

[0054] Optionally, in step S4, the execution step of the feedback control is:

[0055] Step S4.1: control instruction refinement generation;

[0056] Step S4.2: multi-type actuator collaborative regulation;

[0057] Step S4.3: real-time closed-loop monitoring of the regulation process;

[0058] Step S4.4: dynamic optimization based on the regulation parameters of the gelatinization stage;

[0059] Step S4.5: backup regulation strategy under abnormal working conditions.

[0060] Optionally, in step S5, after performing the regulation operation in step S4, the current gelatinization state data is collected again in step S1, the model analysis in step S2, and the uniformity judgment in step S3 after a preset effective duration, if the judgment result is still not up to standard, the next round of iteration of steps S1-S3 is triggered immediately; if it is up to standard but there is an edge critical region, intermittent monitoring iteration is started.

[0061] If the multi-modal data collected in step S1 mutates or the non-uniform area position thermodynamic map output by step S2 shows that the defect area expands by more than 3%, the preset effective duration is skipped and emergency iteration is directly triggered to avoid deterioration of non-uniformity.

[0062] Advantages of the present application:

[0063] 1. The present application inputs the constructed graph structure data into the pre-trained graph neural network model, the graph neural network model aggregates neighborhood information through the message passing mechanism, extracts and fuses the local features and global topological structure features of each graph node, and finally outputs a spatial distribution thermodynamic map reflecting the entire gelatinization uniformity and a spatial distribution thermodynamic map identifying the gelatinization non-uniform area position, realizing the recognition of gelatinization uniformity and ensuring the control recognition accuracy of gelatinization uniformity; uniformity recognition and decision: according to the spatial distribution thermodynamic map of gelatinization uniformity, it is judged whether the current gelatinization state reaches the preset uniformity standard; if it does not reach the standard, the gelatinization defect area is located combined with the spatial distribution thermodynamic map, ensuring the control recognition accuracy of gelatinization uniformity.

[0064] 2. The present application forms a closed-loop feedback control system until the gelatinization uniformity of the entire reaction system reaches the preset standard, thereby completing the precise control of konjac gelatinization and preventing control lag. BRIEF DESCRIPTION OF DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0066] Figure 1 The workflow diagram of the present application;

[0067] Figure 2A feedback control workflow schematic diagram of the present application. DETAILED DESCRIPTION

[0068] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0069] Embodiment 1;

[0070] As Figure 1 shown, the present embodiment provides a GNN-based konjac gelatinization uniformity recognition control method, comprising the following steps:

[0071] Step S1: data acquisition and graph structure construction: a multi-modal sensor network is arranged in the konjac gelatinization reaction container, and parameter data of the gelatinization state is acquired in real time, and the parameter data of the gelatinization state is constructed into graph structure data; wherein the parameter data of the gelatinization state includes temperature gradient, concentration gradient and velocity gradient, the graph structure data takes the spatial position point as the graph node, the graph node features are composed of multi-modal data of the corresponding position sensors, and the connection between the graph nodes is connected according to the correlation of fluid dynamics;

[0072] Step S2: graph neural network model processing: input the constructed graph structure data into the pre-trained graph neural network model, the graph neural network model aggregates neighborhood information through the message passing mechanism, extracts and fuses the local features and global topological structure features of each graph node, and finally outputs a spatial distribution heat map reflecting the entire gelatinization uniformity and a spatial distribution heat map identifying the position of the gelatinization uneven area;

[0073] Step S3: uniformity recognition and decision: according to the spatial distribution heat map of the gelatinization uniformity, it is judged whether the current gelatinization state reaches the preset uniformity standard; if not, the spatial distribution heat map is combined to locate the gelatinization defect area;

[0074] Step S4: feedback control execution: according to the uniformity recognition and positioning result, a control instruction is generated to drive the actuator to perform directional and accurate regulation and control operation on the gelatinization defect area, and the regulation and control operation includes but is not limited to adjusting the local temperature, the stirring paddle speed, the direction or the feeding rate, so as to promote the gelatinization uniformization;

[0075] Step S5: iterative optimization: steps S1 to S4 are repeatedly executed to form a closed-loop feedback control system until the gelatinization uniformity of the entire reaction system reaches the preset standard, thereby completing the accurate control of konjac gelatinization.

[0076] Embodiment 2;

[0077] Based on example 1, in step S1, the fluid dynamics related edges, i.e. the connections between graph nodes based on the velocity vector correlation, are adopted. In the konjac gelatinization process, the flow state (velocity, shear stress) of the material directly reflects the gelatinization uniformity, and the connection needs to be established through the fluid parameter correlation.

[0078] Let the nodes and be at time and the fluid velocity vectors of the nodes and

[0079] be and respectively.

[0080] where denotes the velocity vector or velocity component of a particle at time and position in the material, , and are the velocity components of the particle at position in the material in the directions of the three coordinate axes , and respectively. is the velocity vector or velocity component of a particle at time and position in the material, , and are the velocity components of the particle at position in the material in the directions of the three coordinate axes , and respectively.

[0081] If the velocity vectors (or components) of different positions (such as: and ) are highly correlated (such as: the velocity magnitude, the direction is close, and the velocity gradient is uniform), it indicates that the material flow is more consistent, the heat and moisture transfer is more uniform, and the gelatinization is more uniform; on the contrary, if the velocity difference is large and the correlation is low, it is easy to appear local flow dead angle, leading to uneven gelatinization.

[0082] The shear stress of the fluid, reflecting the shear effect inside the material, affects the intermolecular interaction and the gelatinization process, is directly related to the velocity gradient, such as: Newtonian fluid satisfies , is the shear stress, is the viscosity, is the velocity gradient. Through the velocity vectors of different positions, such as: and The component difference can be used to calculate the velocity gradient, and then analyze the uniformity of shear stress distribution, ultimately relating it to the uniformity of gelatinization.

[0083] The linear correlation of velocity vectors is used as a measure. :

[0084] ;

[0085] in, Used to describe velocity components and The velocity vector that changes together with time; if ,illustrate and The trend of change is consistent; This indicates that the trends of change are opposite; This explains the common changes in the wireless properties of both.

[0086] For standardization factors, It is a description Its own fluctuation range; It is a description Its own fluctuation range. Taking the square root of the product has a standardizing effect, which constrains the final correlation coefficient value to the range of [−1,1], eliminating the interference between the dimension and its own fluctuation range on the degree of correlation.

[0087] Measuring the linear correlation of velocity vectors It measures the strength and direction of the linear relationship. ,but and They exhibit a perfectly positive linear correlation and show completely identical trends. They are completely negatively linearly correlated, and their trends are completely opposite. There is no obvious linear correlation. In fluid mechanics (e.g., turbulence analysis, velocity field structure studies), the linear correlation of velocity vectors is used to measure the degree of correlation. It is used to correlate velocity components in different directions at the same location (e.g., correlation of fluctuating velocities), analyze the synchronicity of velocities at different locations (e.g., spatial correlation characteristics of velocity fields), and realize the evolution law of flow field.

[0088] Furthermore, konjac gelatinization is a process in which the temperature field (heat transfer) and the concentration field (mass transfer) are coupled. That is, the temperature gradient drives the diffusion of starch molecules, and the concentration gradient affects the heat absorption efficiency. The original edge weights only consider a single physical quantity (velocity, shear stress) and do not quantify the multi-field coupling effect.

[0089] Thermal-mass gradient coupling, multi-length collaborative correlation, computational nodes and The degree of coordination between the temperature gradient and the concentration gradient:

[0090] ;

[0091] where, represents the thermal-mass coupled ternary gradient action term, is ternary or triple gradient, and represent different positions, is time, which is a physical quantity describing the coupling strength of the common transmission process of thermal and mass gradients;

[0092] is the temperature gradient, is the temperature, and the temperature gradient is the spatial variation rate, is the difference in the distribution of temperature in space, which is the core driving force of heat transfer, and heat is transferred from high temperature to low temperature, which is driven by the temperature gradient.

[0093] is the concentration gradient, is the mass concentration or particle number concentration, similar to the temperature gradient, the concentration gradient is the core driving force of mass transfer (substance diffuses from high concentration to low concentration);

[0094] is the velocity gradient, is the motion velocity of fluid or particles, and the velocity gradient reflects the shear or stretching characteristics of the flow field, in the convection process, the velocity distribution will simultaneously affect heat transfer (convective heat transfer) and mass transfer (convective diffusion).

[0095] In the field of thermal physical material transmission, thermal-mass coupling describes the process of mutual influence between heat transfer and mass transfer (or particle transport), and the gradients of heat (temperature) and mass (mass / particles) are not isolated, but interact and jointly determine the transmission law, which is a quantitative coupling effect.

[0096] is the normalized processing of the moduli (overall strength) of the three gradients of heat, mass, and flow, is an infinitesimal quantity to avoid zero denominator;

[0097] is the dot product (or multiple coupling effect) of the temperature gradient, concentration gradient, and velocity gradient, reflecting the degree of synergistic coupling of the three.

[0098] The coupling effect of the three gradients of heat, mass, and flow is the ratio of the overall strength of the three gradients, plus an infinitesimal quantity to correct the environment where heat, mass, and flow exist simultaneously, and the coupling effect of the three is relative to the strength of the sum of the independent effects of each.

[0099] If the numerator (coupling term) is much larger than the denominator (sum of independent terms), it means that the coupling effect is dominant, i.e., the gradients of heat, mass and flow are mutually enhanced, and the transfer process is controlled by coupling; if the numerator is much smaller than the denominator, it means that the coupling effect is very weak, i.e., the transfer of heat, mass and flow occurs independently, and the coupling of the transfer process is ignored.

[0100] The coupling effect of the gradient is divided by the overall strength of the gradient to quantify the degree of mutual influence among heat transfer, mass transfer and fluid flow, thereby realizing the analysis of complex multi-field coupled transfer processes.

[0101] Embodiment 3;

[0102] Based on embodiment 1, in step S2, the message passing mechanism aggregates neighborhood information, which is to let each node change its own state by sending and receiving messages from neighboring nodes, so as to realize the aggregation of local and even global information, i.e., the new feature of a node is the original feature of itself and the mixed message of the features of all its neighbors. The message passing mechanism aggregates neighborhood information, which is to collect the feature information of all neighboring nodes for each node. This process is regarded as each node receiving information from surrounding nodes, fusing the aggregated neighbor information with the original information of the node itself to generate a new and more representative node feature, which contains the aggregated neighborhood information.

[0103] The local feature is generated by multiple rounds of message passing, and the feature of each node gradually contains the information of the surrounding area (such as 3-5 connected sensor ranges), generating feature information that interacts in the local feature space.

[0104] The global topological structure feature is to gather the features of all nodes into a global embedding vector of the entire graph (Graph Embedding), which encodes the macroscopic state of the entire paste.

[0105] For the spatial distribution heat map reflecting the uniformity of the paste, starting from the node features of the graph, through multi-scale spatial interaction modeling, a continuous value heat map reflecting the uniformity of the paste is finally generated :

[0106] ;

[0107] wherein, is a plurality of coordinates The output heat map of the uniformity of the paste reflects the uniformity of the paste at the position , and is limited to the interval [0, 1] through subsequent truncation operation. The closer the value is to 1, the higher the uniformity of the paste; the closer the value is to 0, the lower the uniformity of the paste;

[0108] MLP, a basic neural network structure, is used to perform nonlinear transformation and feature extraction on the input features to map the input features to the values of the paste uniformity, capturing the data related to paste uniformity in the features;

[0109] is the feature vector input to , which is obtained from multi-dimensional perception of the material (e.g., starch food material related to paste) such as temperature, moisture, and microstructure information at different positions after pre-processing, and is the original basis for subsequent calculation of paste uniformity;

[0110] is a clipping operation that limits the value output by to the range of 0 to 1. Because paste uniformity is a relative and quantifiable degree indicator, using continuous values in the interval [0, 1] to represent it not only conforms to the intuitive representation (gradual change from very uneven to very uniform), but also facilitates the visualization of the heat map (different values correspond to different colors, clearly showing the spatial distribution of uniformity);

[0111] Through the multi-layer perception , the features reflecting the characteristics of the material are processed to obtain a preliminary value related to paste uniformity; then the value is constrained in the interval [0, 1] using the Clip operation, and finally a continuous value reflecting the paste uniformity at each coordinate on the heat map is obtained . When the continuous values are distributed in space, a heat map of paste uniformity is formed .

[0112] For the spatial distribution heat map identifying the paste uneven area position , a binary region identification is generated based on the uniformity feature, highlighting the uneven area:

[0113] ;

[0114] wherein represents the heat map of paste unevenness output by multiple coordinates , reflecting the paste unevenness at position ; is the gradient of the heat map of paste uniformity output by multiple coordinates , the size of the gradient reflects the intensity of the field change (paste unevenness), and the direction reflects the trend of the field change (unevenness direction), and the gradient is used to capture the position where unevenness begins to appear; is input to Feature vector in the input of the model, which is from the multi-dimensional perception of the material (e.g. starch food related to gelatinization), such as temperature, moisture and microstructure information at different positions after pre-processing, is the original basis for subsequent calculation of gelatinization unevenness. Connection, to fuse the original basis and the gradient information of the current field, so that the model can more comprehensively judge the uneven area.

[0115] Another multi-layer perception machine, as a non-linear mapping module, is to model the complex relationship of the fused features of the original basis + gradient. Because gelatinization unevenness is a non-linear result of multiple factors (temperature, moisture and material), a neural network is needed to develop implicit rules.

[0116] Temperature-related activation or transformation function, Emphasizes the correlation with temperature sensitivity of gelatinization. Gelatinization is a typical temperature-driven process, and the gelatinization behavior differs greatly at different temperatures. This function is used to adapt the influence of temperature on unevenness judgment (e.g. different unevenness judgment standards in high and low temperature zones).

[0117] Step function (or threshold function), Threshold value; when The output of exceeds , it is determined as a gelatinization uneven area; otherwise, it is a uniform area. Its role is to binarize the continuous model output and clearly distinguish between uneven and uniform areas.

[0118] From the field distribution to the heat map identification of the gelatinization process, the uneven distribution of temperature and moisture fields is the essence of the uneven area. The following methods are used to generate a heat map that identifies uneven areas.

[0119] Capture the source characteristics of unevenness with Capture the changes in the field (potential signs of unevenness), combined with Introduce the original basis to comprehensively describe the possibility of unevenness.

[0120] Modeling complex non-linear relationships is done with Calculate the non-linear relationship between field changes + original data and gelatinization unevenness, because gelatinization is a complex physical process of multiple factor coupling.

[0121] Adapt the temperature effect with Adjust the model output to match the decisive role of temperature on gelatinization unevenness, i.e. under different temperatures, the same field change may correspond to different unevenness degrees.

[0122] Binarize the uneven area: use Step function and threshold The continuous unevenness is converted into 0-1 decision results, which exceed the threshold The position is marked as an uneven area, otherwise as a uniform area.

[0123] Finally, these 0-1 decision results are presented in the form of a heat map, with uneven areas highlighted with high brightness and uniform areas with low brightness, thereby visually identifying the spatial distribution of uneven areas.

[0124] Further, the uneven area boundary is enhanced to improve the accuracy of the heat map, so that the boundary of the uneven area is clearer, and the accuracy of the uneven area is introduced:

[0125] ;

[0126] wherein, is a plurality of coordinates The output heat map of the unevenness of the paste has accuracy; represents a plurality of coordinates The output heat map of the unevenness of the paste reflects the unevenness of the paste at the position ; is a plurality of coordinates The output heat map of the unevenness of the paste The gradient reflects the rate of change and direction in space (i.e. the gradient is large at the edge or contour); is a plurality of coordinates The output heat map of the uniformity of the paste The gradient of the heat map of the uniformity of the paste.

[0127] is the norm of the gradient , which is used to measure the strength of the gradient; is a sign function, which is used to judge the consistency of the direction of the two gradients, If the value of is greater than 1, the directions of the two gradients are the same, If the value of is less than 1, the directions of the two gradients are opposite; is a hyperparameter, which is used to control the influence degree of the gradient interaction on the loss;

[0128] is the gradient direction alignment term, the dot product of the two gradients reflects the consistency of the direction: if the dot product is positive, it means that the gradient of the continuous value heat map of the unevenness of the paste and the gradient of the continuous value heat map of the uniformity of the paste are in the same direction (e.g. both rising at the target edge); if it is negative, it means that the directions are opposite; when the directions of the two gradients are consistent, the ​​enhanced; when the direction is opposite, it is weakened, so that the gradient direction is better aligned with the edge, ensuring that the area represented matches the edge of the target (represented by ), improving the accuracy and consistency of the results.

[0129] the gradient strength enhancement term, is the strength of the gradient itself (the more obvious the edge, the larger the norm). By adjusting the , the area with strong gradient will be amplified, making more focused on the edge, the gradient significant position of the details. Combined with the norm of the gradient , it can dynamically adjust the effect on according to the strength of the gradient itself. The area with strong gradient (usually the area with rich details) will be given greater weight, so that these areas can have a more significant impact in subsequent optimization or calculation, thereby highlighting the key details and making the results more detailed. align the gradients of the continuous value heat map of the blurring non-uniformity and the continuous value heat map of the blurring uniformity

[0130] , thereby enhancing the continuous value heat map of the blurring non-uniformity the ability to depict the perception of information of edges and details, making the results more detailed. Example 4;

[0131] Based on Example 1, in step S3, for judging whether the current blurring state reaches the preset uniformity standard:

[0132] Evaluate the area state: A area (red / white edge area) is over-blurring; the state is that the temperature of these areas is too high or the heating time is too long, and the material has been over-blurring. The change in physical properties is that the material may be carbonized and hardened, affecting subsequent processing (such as pumping and mixing) and the taste of the final product.

[0133] Over-blurring material may adhere to the inner wall of the equipment, making it difficult to clean, and long-term accumulation can affect heat transfer efficiency and increase equipment maintenance costs.

[0134] B area (central blue / green area) is under-blurring; the state is that the temperature of these areas is too low or the heating time is insufficient, and the material has not yet reached the conditions for sufficient blurring. If the starch is not fully blurring, it will affect the final texture, taste (such as too hard, with a raw heart) and water retention of the product.

[0135] B area (central blue / green area) is under-blurring; the state is that the temperature of these areas is too low or the heating time is insufficient, and the material has not yet reached the conditions for sufficient blurring. If the starch is not fully blurring, it will affect the final texture, taste (such as too hard, with a raw heart) and water retention of the product. ​​

[0136] Zone C (the yellow / orange area in the middle) indicates good gelatinization. The condition is that these gelatinized areas are uniform in color, indicating that the temperature and time are properly controlled. Although such areas exist, they are surrounded by over-gelatinized or under-gelatinized areas, and the area may be insufficient, indicating that the overall process parameter settings need to be adjusted.

[0137] Based on the GCI thermal map gradient information of gelatinization uniformity and non-uniformity, the profile is driven to shrink towards the GCI abrupt change (defect boundary) to locate the gelatinization defect region, and the profile is driven to move towards the region where the GCI gradient modulus is large (abrupt change = defect boundary). Therefore, the external potential energy is defined as a negative function of the GCI gradient modulus, that is, the larger the gradient modulus, the lower the potential energy, and the easier it is for the profile to move towards this location.

[0138] ;

[0139] in, This is a region with gelatinization defects; This is a parameterized representation of the contour that evolves over time, i.e., the parameters at each contour point. These are the curve parameters of the contour (e.g., the parameters of the entire contour being traversed). It is time, describing the evolution of the outline;

[0140] The gradient of GCI, Relevant indicators representing the uniformity of gelatinization, gradient modulus It reflects the rate of change of GCI. In regions with uniform gelatinization, the GCI changes gradually (small gradient magnitude). At the defect boundary, gelatinization changes abruptly from uniform to non-uniform, and the GC changes drastically (large gradient magnitude).

[0141] The eigenvalues ​​of points on the contour are denoted by , and the GCI is denoted by at the contour points. The signal strength at the location, or the properties of the contour itself; For the entire contour, from the parameters arrive The contributions are accumulated.

[0142] When the contour approaches the boundary of the gelatinization defect, the GCI abruptly changes at that point. It will increase significantly, causing the integrand to... The value increased significantly.

[0143] The negative sign before the formula makes the contour closer to the defect boundary (the area with a large gradient magnitude). The smaller the value, the smaller the result (because the integral result is large, and the whole is smaller after the negative sign).

[0144] As the region shrinks, the outline will spontaneously shift... The smaller area (i.e. the defect boundary of GCI mutation) shrinks, eventually fitting the boundary of the defect, and realizing the positioning of the paste defect area. The paste defect area is constructed by the mutation characteristics of GCI gradient , and the profile is automatically moved to the defect boundary of the paste uniformity mutation under the driving of the area reduction, so as to accurately position the paste defect area.

[0145] Embodiment 5;

[0146] Based on embodiment 1, as shown in step S4, the feedback control execution step is: Figure 2

[0147] Step S4.1: Control instruction refinement generation; based on the characteristics of the positioned paste defect area (including defect type: temperature deviation / insufficient stirring / uneven feeding; defect degree: deviation value, influence range; defect position: spatial coordinates), combined with the process parameter threshold of the current paste stage (such as: the temperature in the middle of paste needs to be maintained at 85-90℃, the stirring Reynolds number ≥5000), the quantitative control instruction is generated by the following rules:

[0148] If the defect type is temperature deviation: the instruction needs to include the spatial coordinates of the defect area, the target temperature adjustment value (such as: the coordinates of area A (X1, Y1, Z1), the temperature is increased by 3℃ to 88℃, the temperature control rate (such as: increase the temperature by 1℃ per minute to avoid local overheating));

[0149] If the defect type is insufficient stirring: the instruction needs to clearly indicate the stirring paddle number corresponding to the area, the speed adjustment value (such as: the stirring paddle No. 3 is increased from 200r / min to 250r / min), and the stirring direction correction (such as: the stirring paddle No. 2 is changed from clockwise to counterclockwise to enhance the fluid circulation in area B);

[0150] If the defect type is uneven feeding: the instruction needs to specify the feeding port number and the local feeding rate adjustment ratio (such as: the feeding rate of feeding port No. 1 to area C is increased by 15%, and feeding port No. 2 remains unchanged), while the stirring parameters are adjusted in association (such as: the speed of stirring paddle No. 4 is increased synchronously to avoid material accumulation).

[0151] Step S4.2: Multi-type actuator collaborative control; according to the instruction type, the corresponding actuator is activated to realize directional and accurate control + cross-institutional collaboration:

[0152] Temperature control mechanism: for the temperature deviation area, start the distributed infrared heating module (corresponding to the defect area coordinates) or the micro cooling pipe (only when the local temperature is too high), and correct the heating / cooling power in real time through the PID algorithm to ensure that the temperature smoothly approaches the target value;

[0153] ​Stirring control mechanism: based on stirring instructions, drive servo motor to adjust the speed and direction of the target stirring paddle, and activate the paddle torque sensor to monitor the change of stirring resistance (if the resistance increases suddenly, it is judged that the material is caked, and the low-speed reverse-rotation forward rotation cycle is triggered to break the caking operation);

[0154] Feed control mechanism: adjust the instantaneous feed rate of the target feed port through the electromagnetic flow valve, simultaneously open the dispersion nozzle (pore diameter 0.5 mm) below the feed port to avoid feed accumulation, and collect material distribution data in the feed area every 10 seconds (related to S1 sensor network).

[0155] Step S4.3: Real-time closed-loop monitoring of the regulation process; while performing the regulation operation, a "data collection-effect judgment-instruction correction" closed loop is constructed:

[0156] Data collection: call multi-modal sensor network to collect high-frequency data (sampling frequency 5 Hz) of parameters (temperature, viscosity, fluid flow rate) in the defect area and surrounding 10 cm range, and convert them into graph structure data in real time;

[0157] Effect judgment: input real-time graph structure data into S2 pre-trained graph neural network model, output updated gelatinization uniformity heat map within 10 seconds, compare with the heat map before regulation, calculate the uniformity improvement rate of the defect area (such as: the uniformity of area A is improved from 65% to 82%);

[0158] Instruction correction: if the improvement rate is ≥15% (preset threshold), maintain the current regulation instruction; if 5% ≤ improvement rate <15%, fine-tune the regulation parameters (such as: temperature adjustment value increases by 1℃, stirring speed increases by 20r / min); if the improvement rate is <5%, re-analyze the defect reasons (such as: exclude temperature / stirring factors, judge as feed port blockage), generate new instructions;

[0159] Step S4.4: Dynamic optimization of regulation parameters based on the gelatinization stage; optimize the regulation strategy combined with the stage characteristics of the whole gelatinization cycle:

[0160] Early gelatinization (material not fully hydrated, low viscosity): preferentially regulate the feed rate and stirring direction to avoid material settling, and temperature regulation mainly focuses on slow heating (rate ≤0.8℃ / min);

[0161] Mid-gelatinization (material is largely gelatinized, viscosity increases sharply): focus on temperature and stirring speed regulation, temperature fluctuation is controlled within ±0.5℃, and stirring speed is adjusted in real time according to viscosity changes (viscosity increases by 100 mPa・s, stirring speed increases by 15 r / min);

[0162] Late gelatinization stage (gelatinization degree ≥ 90%, tends to be stable): reduce the control intensity, maintain the temperature at the target value ± 0.3℃, and slightly adjust the stirring speed by ≤10 r / min, while reducing the feeding control frequency (once every 30 seconds).

[0163] Step S4.5: backup control strategy under abnormal conditions; for actuator failure or sudden disturbance, a pre-device scheme is provided:

[0164] Actuator failure: if a certain area heating module fails, start the compensation heating of the adjacent 2 heating modules (power increase by 20%, expand the heating range to cover the defect area); if the stirring paddle is stuck, turn off the faulty paddle, increase the speed of the adjacent 2 paddles by 30%, and at the same time, extend the stirring time;

[0165] Sudden disturbance (such as voltage fluctuation, material composition deviation): if the sensor detects a temperature drop of ≥5℃, immediately start the "emergency heating" of all heating modules (power increase by 50%, for 30 seconds), and simultaneously pause the feeding, and after the temperature rises, resume normal control; if the material viscosity is abnormally high (20% higher than the standard value), reduce the stirring speed by 10%, and turn on the cooling pipe for slight cooling (1-2℃), to avoid over gelatinization.

[0166] Example 6;

[0167] Based on example 1, in step S5, after performing the control operation in step S4, a pre-set effective duration (usually 30-60s according to the konjac gelatinization kinetics characteristics) is set, and the current gelatinization state data is collected again, the model analysis is performed, and the uniformity judgment is performed, if the judgment result is still not up to standard (such as: the proportion of uneven area is >5%), the next round of iteration of steps S1-S3 is triggered immediately; if it meets the standard but there is an edge critical area (such as: the local uniformity is close to the threshold), intermittent monitoring iteration is started (repeat steps S1-S3 every 2 minutes, stop if the same standard is met for 3 times).

[0168] Abnormal state trigger: if the multi-modal data collected in step S1 mutates (such as: the temperature fluctuation in a certain area is >±2℃, the stirring torque mutates >10%), or the uneven area position thermal map output by step S2 shows that the defect area expands >3%, the pre-set effective duration is skipped, and the emergency iteration is directly triggered, to avoid the deterioration of the unevenness.

[0169] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A GNN-based identification control method for the uniformity of konjac paste gelatinization, characterized in that, Comprise the following steps: Step S1: data acquisition and graph structure construction: a multimodal sensor network is arranged in the konjac paste reaction container, the parameter data of the paste state is collected in real time, and the parameter data of the paste state is constructed as graph structure data; wherein the parameter data of the paste state includes temperature gradient, concentration gradient and velocity gradient, the graph structure data takes the spatial position point as the graph node, the graph node feature is composed of the multimodal data of the corresponding position sensor, and the connection between the graph nodes is connected according to the correlation of fluid dynamics; In the step S1, the fluid dynamics related edge is used to connect the graph nodes, and in the konjac paste process, the flow state of the material directly reflects the paste uniformity, which needs to be connected through the correlation of fluid parameters: Set node With The fluid velocity vector at the time is: and ; wherein denotes the time instant and the position of the material the velocity vector or velocity component of a particle at the position of the material , and are the velocity components of a particle at the position of the material in the directions of the three coordinate axes , and ; is the velocity vector or velocity component of a particle at the position of the material and the time instant , , and are the velocity components of a particle at the position of the material in the directions of the three coordinate axes , and ; Calculate using velocity vectors at different positions and The component difference is used to calculate the velocity gradient, and then the uniformity of shear stress distribution is analyzed, which is finally correlated with the gelatinization uniformity. Using linear correlation of velocity vectors is: ； wherein for describing the velocity component and velocity vectors that vary together over time; if indicates and varying trends are consistent; indicates varying trends are opposite; indicates a common variation in both non-linearity; is a standardization factor, is a standardization factor, is a standardization factor, is a standardization factor, is a standardization factor, after square root of the product, the standardization function makes the final correlation coefficient value be constrained in [−1, 1], eliminating the interference of dimension and self fluctuation amplitude on the correlation degree. Konjac gelatinization is a process involving the coupling of temperature and concentration fields, thermo-mass gradient coupling, and multi-length collaborative correlation, with computational nodes. and The degree of coordination between the temperature gradient and the concentration gradient: ； wherein, represents a ternary gradient term of heat-mass coupling, is a ternary or triple gradient, and represent different positions, is time, and is a coupling strength physical quantity describing the joint transmission process of heat and mass gradients. is the temperature gradient, is the temperature, and the temperature gradient is the rate of change in space, is the difference in the distribution of temperature in space, which is the core driving force of heat transfer. Heat transfers from high temperature to low temperature, which is driven by the temperature gradient; concentration gradient, concentration gradient, like temperature gradient, concentration gradient is the core driving force of mass transfer; is the velocity gradient, is the velocity of the fluid or particle, the velocity gradient reflects the shear or stretching characteristics of the flow field, and in the convective process, the velocity distribution will affect both heat transfer and mass transfer; is the normalized length of the thermal, mass, and flow gradients, is a very small quantity to avoid division by zero; is the dot product of the temperature gradient, the concentration gradient, and the velocity gradient, and reflects the degree of coupling between the three; Step S2: graph neural network model processing: input the constructed graph structure data into the pre-trained graph neural network model, the graph neural network model aggregates neighborhood information through message passing mechanism, extracts and fuses local features and global topological structure features of each graph node, and finally outputs a spatial distribution heat map reflecting the whole paste uniformity and a spatial distribution heat map identifying the paste uneven area position; Step S3: uniformity recognition and decision: according to the spatial distribution heat map of paste uniformity, it is judged whether the current paste state reaches the preset uniformity standard; if not, the spatial distribution heat map is combined to locate the paste defect area; Step S4: feedback control execution: according to the uniformity recognition and positioning result, a control instruction is generated to drive the execution mechanism to carry out directional and accurate regulation and control operation on the paste defect area, which includes but is not limited to adjusting local temperature, stirring paddle speed, direction or feeding rate, so as to promote paste uniformization; Step S5: iterative optimization: repeat steps S1 to S4 to form a closed loop feedback control system until the paste uniformity of the whole reaction system reaches the preset standard, so as to complete the accurate control of konjac paste.

2. The GNN-based identification and control method for the uniformity of konjac paste gelatinization according to claim 1, characterized in that, In the step S2, for the spatial distribution heat map reflecting the whole gelatinization uniformity, starting from the graph node features, through multi-scale spatial interaction modeling, a continuous value heat map reflecting the gelatinization uniformity is finally generated : ； wherein, a plurality of coordinates a heat map of the outputted roasting uniformity, reflecting the uniformity of roasting at the positions is limited in the interval [0, 1] through subsequent truncation operations, the value closer to 1, the higher the roasting uniformity; the closer to 0, the lower the roasting uniformity; represents a multi-layer perceptron, which is used to perform nonlinear transformation and feature extraction on the input features to map the input features to the dough uniformity values, capturing the data in the features that are relevant to the dough uniformity; is the feature vector input to the multi-dimensional perception of the material; is a truncation operation, which is to the output value is limited to between 0 and 1; By a multilayer perceptron Features reflecting material properties The processing is done to obtain a preliminary value related to the degree of uniformity of the gelatinization; then a Clip operation is used to constrain the value in the interval [0, 1] to obtain a continuous value at each coordinate on the heat map that reflects the degree of uniformity of the gelatinization at that position The distribution of the continuous values in space forms the heat map of the degree of uniformity of the gelatinization .

3. The GNN-based identification and control method for the uniformity of konjac paste gelatinization according to claim 1, characterized in that, In the step S2, for the spatial distribution thermal map identifying the mal-distributed region position, a binary region identification is generated based on the uniformity feature, highlighting the mal-distributed region: , ； wherein, representing a plurality of coordinates a heat map of outputted paste unevenness, reflecting the location where the paste is uneven; is a plurality of coordinates a heat map of outputted paste evenness a gradient, the size of the gradient reflecting the degree of field change, the direction reflecting the trend of field change, the gradient being to capture the location where unevenness begins to appear; is a feature vector inputted into , the feature vector coming from multi-dimensional perception of the material, using to connect, fusing the original basis with the gradient information of the current field, letting the model more comprehensively judge uneven areas; is another multi-layer perceptron, acting as a non-linear mapping module, modeling complex relationships on the original gradient-based fused features; is a temperature dependent activation or transformation function, is a function that emphasizes the association with temperature sensitivity of pasting, which is a typical temperature driven process with large differences in pasting behavior at different temperatures, and is used to fit the influence of temperature on the unevenness determination; It is a step function. It is the threshold for judgment; when The output exceeds When the region is non-uniform, it is identified as a region of non-uniform gelatinization; otherwise, it is identified as a region of uniformity. Its function is to binarize the continuous model output and clearly distinguish between non-uniform and uniform regions.

4. The GNN-based identification and control method for the uniformity of konjac paste gelatinization according to claim 3, characterized in that, For the boundary enhancement of uneven area to improve the accuracy of heat map, in order to make the boundary of uneven area clearer, the accuracy of uneven area is introduced: ； in, For multiple coordinates Output a heatmap with accurate gelatinization inhomogeneity; Representing multiple coordinates The output heatmap of gelatinization non-uniformity reflects the location The degree of unevenness in gelatinization; For multiple coordinates The output heatmap of gelatinization non-uniformity The gradient reflects The rate and direction of change in space; It is multiple coordinates The output heatmap of gelatinization uniformity The gradient; is a gradient is a norm of is a norm of is a sign function is a sign function is a sign function is a hyper-parameter For gradient direction alignment term, dot product of two gradients Reflects the direction consistency: For gradient intensity enhancement term, is The intensity of the gradient itself, by Adjusting, the area with strong gradient will be amplified, let More focused on the edge, the gradient significant location of details, combined with the norm of the gradient According to the intensity of the gradient itself, it can dynamically adjust the effect on .​ 5. The GNN-based identification and control method for the uniformity of konjac paste gelatinization according to claim 1, characterized in that, In the step S3, based on the GCI heat map gradient information of paste uniformity and paste non-uniformity, the contour is driven to shrink to the GCI mutation place to locate the paste defect area, and the contour is driven to move to the area with large GCI gradient module: ； wherein is a pasting defect region; is a parametric representation of the profile over time, i.e. a parameter at each profile point, is a curve parameter of the profile, is time, describing the evolution of the profile; a gradient of the GCI, a correlation index representing the uniformity of the paste, a modulus of the gradient reflects the rate of change of the GCI, the area of uniform paste, the GCI changes gently, the defect boundary, the paste changes from uniform to non-uniform, the GC changes sharply; is the feature factor for the profile point, is the signal strength of the GCI at the profile point or a property of the profile itself; is the contribution of the entire profile, summed up from the parameters to ;​ When the profile is close to the boundary of the mashing defect, since the GCI mutates at this point, it will increase significantly, causing the value of the integral function to rise sharply.

6. The GNN-based identification and control method for the uniformity of konjac paste gelatinization according to claim 1, characterized in that, In the step S4, the execution step of feedback control is: Step S4.1: control instruction refinement generation; Step S4.2: multi-type execution mechanism collaborative regulation; Step S4.3: real-time closed loop monitoring of regulation process; Step S4.4: dynamic optimization of regulation parameters based on paste stage; Step S4.5: backup regulation strategy under abnormal working condition.

7. The GNN-based identification and control method for the uniformity of konjac paste gelatinization according to claim 1, characterized in that, In the step S5, after the regulation operation in step S4, interval preset effective duration, reexecute step S1 to collect current paste state data, step S2 model analysis, step S3 uniformity judgment, if the judgment result is still not up to standard, the next round of iteration of step S1-step S3 is triggered immediately; if it reaches the standard but there is an edge critical area, intermittent monitoring iteration is started; If the multimodal data collected in step S1 has a mutation, or the non-uniformity area position heat map output in step S2 shows that the defect area has expanded by more than 3%, the preset effective duration is skipped, and an emergency iteration is directly triggered to avoid the deterioration of non-uniformity.

Images