Process temperature regulation system and method based on multi-parameter coupling

Abstract

The application relates to the technical field of temperature regulation, and discloses a process temperature regulation system and method based on multi-parameter coupling. The system divides a thermal unit of a process space, constructs a thermal impedance topology graph, and provides bottom-layer data support conforming to the essence of a physical process for subsequent distortion identification and accurate regulation. A space-time graph neural network is used to predict the evolution trend of the thermal impedance topology graph, so that the distortion development trend of a future period is deduced in advance. A generative adversarial network is used to separate the contribution of a disturbance source to the topology distortion, and a collaborative weight matrix is constructed by combining an attention mechanism to match the adjustment efficiency of an actuator, so that disturbance source quantification and multi-actuator accurate collaborative regulation are realized, and the collaboration and adaptability of multi-device linkage temperature control are greatly improved. The collaborative instructions of the actuators are calculated based on the collaborative weight matrix as a constraint, and dynamic reverse hedging compensation is realized based on the change rate and change direction of the topology difference, so that a closed-loop temperature control logic is constructed.

Description

Technical Field

[0001] This application relates to the field of temperature control technology, specifically to a process temperature control system and method based on multi-parameter coupling. Background Technology

[0002] Industrial process temperature control is a core control process in industrial production fields such as precision machining, polymer processing and thermoforming. The temperature uniformity and heat exchange stability within the process space determine the material processing uniformity, product mechanical properties and batch production stability. Temperature runaway can easily lead to process defects such as local overheating and scorching of materials, insufficient plasticization, and uneven processing.

[0003] Currently, the industry has seen multiple generations of technological iterations in process temperature control. Traditional control methods use independent parameters such as single-point temperature, medium flow rate, and equipment power as control objects, and achieve single-parameter closed-loop control based on classic PID algorithms, which can meet the basic temperature control requirements under normal stable operating conditions.

[0004] With the increasing sophistication of industrial processes, the complexity of material formulations, and the high precision and standardization of production, the limitations of single-parameter control have become increasingly apparent. Existing technologies are gradually shifting towards multi-parameter coupled control modes, which achieve multi-variable linkage control by associating and collecting various process parameters such as temperature and medium flow rate. This improves the basic control accuracy of process temperature to a certain extent and has become the mainstream technical solution for industrial temperature control at present.

[0005] However, existing methods all rely on discrete process parameters for control decisions, focusing only on the numerical fluctuations of the parameters themselves and ignoring the differentiated heat transfer characteristics of multi-scale heat transfer regions in the process space and the physical laws of bidirectional thermal impedance asymmetry between units. At the same time, existing methods rely solely on historical parameter value changes to predict operating conditions, which cannot adapt to the complex operating conditions of industrial processes with strong nonlinearity, heat transfer lag, and uneven spatial heat transfer distribution, making it difficult to achieve refined, forward-looking, and coordinated temperature control.

[0006] In view of the inherent defects of the prior art, the technical problem to be solved in this application is how to achieve full-range, refined and coordinated control of process temperature through multi-parameter coupling in complex environments. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the purpose of this application is to provide a process temperature control system and method based on multi-parameter coupling, thereby improving process stability and batch consistency of products.

[0008] To achieve the above objectives, this application adopts the following technical solution: In the first aspect, this application provides a process temperature control method based on multi-parameter coupling, including: acquiring sensing data of the current process environment, and generating a thermal impedance topology map with thermal units in the process space as nodes, heat transfer relationships between thermal units as edges, and sensing data as node attributes; The thermal impedance topology map is input into the spatiotemporal graph neural network to predict the evolution trend of the thermal impedance topology map, and the predicted topology map is obtained. It is then compared with the standard thermal equilibrium topology map to extract topological distortion features. Based on the characteristics of topological distortion, the contribution of the disturbance source to the topological distortion is separated by generative adversarial network, and the regulatory effectiveness of the actuator in suppressing topological distortion is determined by attention mechanism, generating a cooperative weight matrix. With the collaborative weight matrix as a constraint and the topological difference between the predicted topology map and the standard thermal equilibrium topology map as the objective, the collaborative instructions of the actuator are calculated and executed. When the rate of change of the topological difference is detected to be greater than the change threshold, a reverse compensation instruction is generated according to the direction of change to execute the hedging control.

[0009] Further, the generation of the thermal impedance topology map includes: Acquire the current process environment sensing data and perform preprocessing; The heat load gradient in the process space is calculated based on the preprocessed sensing data. The heat load gradient is then compared with the gradient threshold group and divided into multiple thermal units. Using thermal units as nodes, connection edges are constructed based on the heat transfer relationship between thermal units, and node attributes are determined based on preprocessed sensing data. Perform connectivity checks on all nodes and connecting edges. After adding virtual heat transfer edges to isolated nodes identified by the checks, generate a thermal impedance topology diagram.

[0010] Furthermore, the heat load gradient is the difference in heat change per unit distance within the process space, formed by temperature, medium flow rate, and equipment output power; thermal units of different scales are divided according to the gradient threshold group, and each thermal unit is filled with only the same heat exchange medium. The heat transfer relationship is divided into two heat transfer directions: heat inflow and heat outflow. The difficulty of heat transfer in the two heat transfer directions corresponds to different thermal resistance values. The connectivity verification checks the physical heat transfer correlation of the thermal units to identify isolated nodes, matches the heat transfer intensity of isolated nodes with the heat transfer level of historical steady-state conditions, and adds virtual heat transfer edges to isolated nodes.

[0011] Furthermore, the extraction of topological distortion features includes: After separately separating the node attributes and the thermal impedance of the connecting edges in the historical thermal impedance topology, they are spliced ​​together in chronological order to form a time series diagram sample sequence. By reading the parameter variation patterns in the time series sample sequence through the spatiotemporal graph neural network, the trend of node attributes and thermal impedance of connecting edges within a preset time period is deduced along the time axis to obtain the predicted topology graph. Establish a correspondence between the nodes and connecting edges in the predicted topology and the nodes and connecting edges corresponding to the spatial coordinates of the same thermal unit in the standard thermal equilibrium topology; Calculate the parameter difference between the corresponding node attributes and the connecting edges according to the comparison relationship, select the nodes and connecting edges with parameter differences greater than the preset threshold as distortion carriers, and summarize them to obtain the topological distortion features.

[0012] Furthermore, the parameter difference includes the offset of node attributes and the increase in thermal resistance of the connecting edge. Nodes with an offset greater than a first preset threshold represent regions of disordered energy diffusion, which are regions where the temperature and medium flow rate in the thermal unit deviate from the steady-state range, causing local heat transfer imbalance. Connecting edges with an increase in thermal resistance greater than a second preset threshold represent regions of high thermal resistance barriers, which are heat transfer interfaces where the thermal resistance difference between adjacent thermal units exceeds the steady-state range.

[0013] Furthermore, the generation of the collaborative weight matrix includes: The topological distortion features are input into a generative adversarial network, and the proportion of distortion coverage corresponding to each type of disturbance source is statistically analyzed to separate the contribution of each disturbance source to the topological distortion. The contribution of the disturbance source, the spatial coordinates of the thermal unit corresponding to the topological distortion, and the weighted distortion amplitude are input into the attention mechanism. Based on the spatial coverage relationship, the distortion region of the actuator is determined, and the actual suppression capability of the actuator against the topological distortion is determined one by one to obtain the regulation efficiency of the actuator. The basic weight is obtained by multiplying the contribution of the same thermal unit and actuator pairing with the regulation efficiency, and then generating a collaborative weight matrix with thermal units as rows and actuators as columns.

[0014] Furthermore, the number of thermal units that undergo topological distortion corresponding to the disturbance source under the two types of distortion carriers, namely the energy disorder diffusion region and the high thermal resistance barrier region, is counted separately to determine the proportion of the distortion coverage area and use it as the contribution of the corresponding disturbance source. The attention mechanism prioritizes matching the spatial coverage relationship between the actuator and the corresponding thermal unit of the topological distortion. The adjustment efficiency of the same actuator for the configuration of fine-scale thermal units is greater than that for coarse-scale thermal units.

[0015] Furthermore, the generation of reverse compensation instructions includes: By comparing the predicted topology map with the standard thermal balance topology map, the topology differences corresponding to the thermal units are obtained. The topology differences of each thermal unit are then split according to the weight ratio of the actuators in the collaborative weight matrix to obtain the individual actuator's sub-item control requirements for the thermal unit. The individual control requirements of all thermal units matched by the same actuator are summarized, the coordinated instructions of the actuator are obtained and executed; The adjusted topology difference is obtained and the rate of change of the topology difference is calculated. When the rate of change is greater than the change threshold, a reverse compensation instruction is generated according to the direction of change of the topology difference. This instruction is then superimposed on the current cooperative instruction and output to perform hedging control.

[0016] Furthermore, when generating the reverse compensation command, the cooperative weight that matches the thermal unit that generates the topological difference is retrieved. The larger the cooperative weight, the greater the control force of the corresponding reverse compensation command. Reverse compensation is initiated only when the rate of change of topological differences exceeds the change threshold for multiple consecutive sampling periods.

[0017] Secondly, this application provides a process temperature control system based on multi-parameter coupling, including a topology module, a prediction module, a generation module and a control module; The topology module is used to acquire the sensing data of the current process environment, and generates a thermal impedance topology map with thermal units in the process space as nodes, heat transfer relationships between thermal units as edges, and sensing data as node attributes. The prediction module is used to input the thermal impedance topology map into the spatiotemporal graph neural network, predict the evolution trend of the thermal impedance topology map, obtain the predicted topology map, and compare it with the standard thermal equilibrium topology map to extract topological distortion features. The generation module, based on topological distortion features, separates the contribution of perturbation sources to topological distortion through generative adversarial networks, and determines the regulatory effectiveness of the actuator in suppressing topological distortion through an attention mechanism, thereby generating a collaborative weight matrix. The control module is used to calculate and execute the actuator's cooperative instructions with the cooperative weight matrix as a constraint and the topological difference between the predicted topology map and the standard thermal equilibrium topology map as the objective. When the rate of change of the topological difference is detected to be greater than the change threshold, a reverse compensation instruction is generated according to the direction of change to execute the hedging control.

[0018] Compared with the prior art, the beneficial effects achieved by this application are as follows: By dividing the process space into thermal units and constructing a thermal impedance topology diagram, we can accurately depict the actual heat transfer state of the process links, providing underlying data support that fits the physical essence of the process for subsequent distortion identification and precise control. We also use a spatiotemporal graph neural network to predict the evolution trend of the thermal impedance topology diagram, enabling forward-looking prediction of topological distortion and anticipating the distortion development trend in future periods.

[0019] By generating adversarial networks to separate the contribution of disturbance sources to topology distortion, and combining attention mechanisms to match actuator regulation efficiency and construct a collaborative weight matrix, disturbance source quantification and precise collaborative control of multiple actuators are achieved, significantly improving the coordination and adaptability of multi-device linkage temperature control. The collaborative weight matrix is ​​used as a constraint to solve the collaborative instructions of the actuators, and dynamic reverse hedging compensation is achieved based on the rate and direction of change of topology differences, constructing a closed-loop temperature control logic, effectively suppressing temperature regulation oscillation and overshoot problems. Attached Figure Description

[0020] Figure 1 The flowchart shows a process temperature control method based on multi-parameter coupling. Figure 2 A logic flowchart for generating a thermal impedance topology diagram; Figure 3 This is a schematic diagram of a process temperature control system based on multi-parameter coupling. Detailed Implementation

[0021] The technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of this application, rather than limitations thereof.

[0022] Example 1 like Figure 1 As shown, this embodiment provides a process temperature control method based on multi-parameter coupling, including: S1. Obtain the sensing data of the current process environment, and generate a thermal impedance topology diagram with thermal units in the process space as nodes, heat transfer relationships between thermal units as edges, and sensing data as node attributes.

[0023] Specifically, such as Figure 2 As shown, generating the thermal impedance topology map includes: Acquire the current process environment sensing data and perform preprocessing; The heat load gradient in the process space is calculated based on the preprocessed sensing data. The heat load gradient is then compared with the gradient threshold group and divided into multiple thermal units. Using thermal units as nodes, connection edges are constructed based on the heat transfer relationship between thermal units, and node attributes are determined based on preprocessed sensing data. Perform connectivity checks on all nodes and connecting edges. After adding virtual heat transfer edges to isolated nodes identified by the checks, generate a thermal impedance topology diagram.

[0024] The heat load gradient is the difference in heat change per unit distance within the process space, formed by temperature, medium flow rate and equipment output power. The heat units are divided into different scales according to the gradient threshold group, and each heat unit is filled with only the same heat exchange medium. Heat transfer is divided into two directions: heat inflow and heat outflow. The difficulty of heat transfer in the two directions corresponds to different thermal resistance values. Connectivity verification checks the physical heat transfer correlation of thermal units to identify isolated nodes, matches the heat transfer intensity of isolated nodes with the heat transfer level of historical steady-state conditions, and adds virtual heat transfer edges to isolated nodes.

[0025] In continuous mixing operations, the composite sensing elements deployed on the inner wall of the closed chamber of the rubber-plastic blending internal mixer, the rotor surface, and the material filling area will generate two types of objective interference data. One type is the instantaneous jump sampling value caused by the friction of the rotor's high-speed rotation, and the other type is the short-term interference data caused by the airflow disturbance in the chamber due to the opening of the silo door during the feeding and discharging processes, which deviates from the steady-state range. Neither type of data can characterize the steady-state mixing heat transfer distribution law of the internal mixer. If it is directly used for heat load gradient calculation, it will cause the calculation of spatial heat change trend to be distorted, resulting in a systematic shift in the boundary of the thermal unit space division and the failure of the underlying data benchmark for topology modeling.

[0026] The sensing data refers to three types of physical parameters acquired synchronously: material / cavity wall temperature at spatial points in the mixing chamber, agitation flow rate of the rubber and plastic media conveyor / rotor, and output power of the heating and cooling jacket of the mixing mill. Preprocessing involves standardizing the raw continuous sampling data by screening out abnormal samples, aligning the sampling time axis, and unifying the dimensions of physical quantities. All sensing points are arranged in a uniform grid pattern in the three-dimensional space of the mixing chamber, with the spatial distance between adjacent sensing points fixed at 0.4 meters. All sensing elements trigger sampling actions synchronously, with a uniform sampling time interval fixed at 1.5 seconds, and the duration of a single complete sampling cycle is set to 40 seconds.

[0027] Specifically, for each piece of raw sampling data, the three parameters of temperature, medium flow rate, and jacket power are compared with the long-term steady-state operating range of the rubber and plastic compound. If any parameter exceeds the preset upper or lower limit range of the steady state, it is judged as abnormal data and removed. For example, the steady-state temperature range is 120℃ to 180℃, the steady-state medium flow rate range is 0.15m / s to 1.0m / s, and the steady-state jacket output power range is 8kW to 50kW. All sampling samples that exceed this range are not retained. After the abnormal data is removed, the units of all retained valid samples are uniformly converted. All data are linearly arranged and stored in the order of sampling time to form time-aligned sensing data.

[0028] This step automatically removes interfering data by using a fixed interval threshold, eliminating calculation deviations caused by material feeding and discharging disturbances and rotor sensor vibrations from the data source. The final output standardized dataset fully matches the physical characteristics of stable rubber and plastic mixing conditions, ensuring the stability of the subsequent gradient calculation process throughout.

[0029] The intensity of heat exchange varies continuously and objectively across different spatial regions within the internal mixing chamber of a rubber-plastic blending mill. The shear heat generation and parameter variation are high in the material-filled areas where the rotor is in contact with the material, while the heat exchange is extremely low in the empty areas at the corners and edges of the chamber. If a uniform fixed size is used to divide the spatial units, two problems will arise: the unit size in the hot material areas will be too large to capture subtle shear heat exchange changes, and the unit size in the inert corner areas will be too small, resulting in redundant calculations. Calculating the spatial heat change gradient based solely on a single temperature parameter cannot reflect the heat difference caused by the coupling effect of the rubber-plastic medium flow rate and the jacket heating power, and the gradient calculation results cannot reproduce the true mixing heat exchange state.

[0030] The heat load gradient refers to the difference in heat change formed by superimposing and converting three parameters—temperature, medium flow rate, and equipment output power—over a unit straight distance in the mixing chamber process space. This differs from the spatial gradient index calculated solely based on a single temperature difference. The gradient threshold group includes two fixed numerical benchmarks: a high gradient judgment threshold and a low gradient judgment threshold. Based on statistical analysis of long-term steady-state operation data of the mixing mill, the mean and standard deviation of the heat load gradient in the rotor material contact area are 15.2 and 2.1, respectively. The mean minus one standard deviation is taken as the high threshold. The mean and standard deviation of the gradient in the corner area of ​​the chamber are 2.8 and 0.9, respectively. The mean plus one standard deviation is taken as the low threshold. Those skilled in the art can calibrate and adjust the threshold according to the heat exchange characteristics of the specific equipment model and material formulation. A heat unit refers to an independently divided closed heat exchange space area inside the mixing chamber. The process mandates that only a single type of heat exchange medium is allowed to be filled in any heat unit.

[0031] The heat load gradient is G = (W) T ×ΔT+W v ×Δv+W P ×ΔP) / L, where G is the corresponding spatial location heat load gradient; W T The temperature coupling weight is set to 0.5, where ΔT is the temperature difference between adjacent sensing points, and W... v The coupling weight for the medium flow velocity is 0.3, where Δv is the difference in flow velocity between adjacent sensing points in the rubber-plastic medium. P The jacket power coupling weight is set to 0.2, ΔP is the difference in jacket output power between adjacent sensing points, and L is the linear spatial distance between adjacent sensing points. Among them, temperature is the most direct indicator of heat exchange state in the internal mixing process and contributes the most to the heat change gradient, so it is given the highest weight. The medium flow velocity affects the heat exchange efficiency through forced convection and contributes the second most. The jacket power is the input of the external heat source and contributes the least. The sum of the weights of the three is 1. Those skilled in the art can make adaptive adjustments to the weights according to the heat exchange characteristics of the specific internal mixing process.

[0032] For example, the high gradient judgment threshold is set to 13, and the low gradient judgment threshold is set to 3.5. The heat load gradient calculated at each point in the entire space of the mixing chamber is divided into thermal unit scales according to fixed rules. The rotor material contact area with a gradient greater than 13 is divided into a fine-scale thermal unit with a side length of 0.4 meters. The mixing area in the middle of the chamber with a gradient between 3.5 and 13 is divided into a medium-scale thermal unit with a side length of 0.8 meters. The empty areas at the corners of the chamber with a gradient less than 3.5 are divided into coarse-scale thermal units with a side length of 1.6 meters. The medium division constraint is executed simultaneously. The interface between the molten rubber material, the molten plastic material and the air in the cavity is used as the boundary for thermal unit division. If the media on both sides of the interface are different, they are separated into separate thermal units. Cross-media merging is not allowed.

[0033] This step fully integrates three types of heat exchange-related physical parameters: shear temperature, rubber-plastic medium flow velocity, and jacket heating power. It accurately reproduces the overall heat change trend of the entire mixing chamber. Two-level gradient thresholds are automatically matched with corresponding spatial scale thermal units, achieving the objective effect of refined spatial modeling of material hot spots and simplified modeling of inert corner areas. It eliminates parameter coupling errors caused by mixed modeling of different rubber-plastic heat exchange media, and the spatial boundary of the thermal unit conforms to the medium interface and the distribution law of shear heat transfer intensity.

[0034] The heat exchange process between adjacent internal mixing heat units exhibits an objective physical characteristic of asymmetric bidirectional heat transfer resistance in both inflow and outflow directions. There is a fixed difference between the heat transfer resistance from the high-temperature rubber melting unit to the low-temperature plastic unit and the reverse heat transfer resistance. If undirected unified connecting edges are used for modeling, the core feature of bidirectional differentiated heat transfer resistance will be lost, and the topology cannot reproduce the real heat transfer behavior of rubber-plastic mixing. At the same time, the steady-state mixing heat transfer state of each heat unit needs to be retained in the topology as an inherent identifier of the node to participate in subsequent time-series deduction calculations. Relying solely on spatial location cannot fully characterize the basic state of unit shear heat transfer.

[0035] The heat transfer relationship is divided into two heat transfer directions: heat inflow and heat outflow. The difficulty of heat transfer in each direction is matched with different thermal resistance values. Node attributes refer to the pre-processed sensing data bound to a single thermal unit, characterizing the basic state of steady-state rubber-plastic compound heat exchange in that unit. The formula for unidirectional thermal resistance is Z=ΔT. stable / (k×S), where Z is the thermal resistance corresponding to the heat transfer direction; ΔT stableThe steady-state mixing temperature difference between two adjacent heat units in the heat transfer direction; k is the thermal conductivity of the dominant heat transfer medium in the heat transfer path, with a fixed value of 0.12 for rubber molten medium, a fixed value of 0.18 for plastic molten medium, and a fixed value of 0.03 for air medium; S is the mutual contact heat transfer area between the two heat units; for two spatially adjacent heat units, the thermal conductivity of the forward heat transfer medium and the thermal conductivity of the reverse heat transfer medium are substituted respectively to calculate the thermal resistance values ​​corresponding to the two unidirectional connection edges. The two sets of thermal resistance values ​​are not averaged or unified, but are stored independently in the corresponding connection edge parameters.

[0036] For example, fine-scale rubber thermal units and medium-scale plastic thermal units are arranged adjacent to each other. The calculated thermal resistance of heat flowing from the rubber unit into the plastic unit is 0.92, and the calculated thermal resistance of heat flowing from the plastic unit back into the rubber unit is 2.75. These two sets of values ​​are bound to two unidirectional connection edges without merging. The node attribute assignment rule is to package and store the preprocessed sensing data of all sensing points within the corresponding thermal unit's spatial range and bind it to the topology node. The node attributes are updated synchronously with the mixing and refining process. The node always retains the latest set of three types of physical parameter datasets from the sampling period. This process enforces the constraint that only adjacent thermal units with rubber-plastic material contact surfaces and fluid convection channels are allowed to build physical heat exchange connection edges. Adjacent units without any heat exchange paths do not generate connection edges.

[0037] This step involves independently constructing connecting edges in each direction and matching differentiated thermal resistance values ​​to restore the true physical law of bidirectional asymmetry in heat transfer between adjacent units in rubber and plastic mixing. The thermal resistance value of each connecting edge accurately corresponds to the heat transfer resistance in a single direction. The nodes are bound to a complete set of mixing sensing data as inherent attributes, so that the topology structure simultaneously carries the spatial geometric information of the mixing chamber and the global shear heat transfer parameter information. The topology parameters fully conform to the steady-state continuous mixing operation conditions of the rubber and plastic internal mixer.

[0038] The closed corners at both ends of the rubber and plastic internal mixer cavity and the isolation area without material filling are surrounded by no adjacent units with material contact surfaces and fluid convection channels. There is no heat transfer path that can build physical heat exchange connection edges. After the first three process modeling, isolated nodes without any connection edges will be formed in the topology graph. Isolated nodes cannot form temporal correlation relationships with other topology units. Directly inputting them into the temporal inference calculation of the spatiotemporal graph neural network will result in incomplete temporal sample sequences. The local unit shear heat transfer evolution trend cannot be calculated normally, and the complete topology evolution process is directly interrupted.

[0039] Connectivity verification refers to checking each internal mixing chamber thermal unit individually to see if there are adjacent units with physical heat transfer relationships. Units without material contact or fluid convection heat transfer paths are determined to be isolated nodes. Historical steady-state operating conditions refer to a long-term continuous working period in which the internal mixer continuously and stably mixes materials, the external ambient temperature and humidity do not fluctuate, the jacket heating power is constant, and the conveying speed of rubber and plastic materials remains at a fixed value. The system continuously stores the sensing dataset under this operating condition. Virtual heat transfer edges are auxiliary topological connection structures used only to complete the topological connectivity structure. They do not represent the existence of real physical heat exchange between units. The heat transfer intensity value corresponding to the virtual heat transfer edge is generated by matching the long-term average heat transfer intensity of the isolated unit under historical steady-state operating conditions.

[0040] The connectivity verification rule involves traversing each cell to search all adjacent spatial thermal cells, checking for the existence of rubber-plastic material contact surfaces or fluid convection channels between cells. Cells without any type of heat transfer path are marked as isolated cells. After completing the full traversal, all isolated cells are summarized. The formula for the steady-state average heat transfer intensity of an isolated cell is: In the formula The long-term steady-state average heat transfer intensity of the isolated unit; N is the total number of steady-state sampling periods; Let be the instantaneous heat transfer intensity of the isolated unit during the i-th sampling period.

[0041] For example, the coarse-scale isolated thermal unit at the left corner of the internal mixer retrieves 12 hours of steady-state mixing sampling data, with a total number of sampling periods N=28800. The instantaneous heat transfer intensity is extracted period by period to complete the summation and average calculation. The calculated average comprehensive heat transfer intensity is 1.24. Based on this value, the corresponding virtual heat transfer edge parameters are matched.

[0042] The process forces virtual heat transfer edges to participate only in topology connectivity completion. In the subsequent distortion feature extraction calculation, filtering rules are set so that the parameters corresponding to the virtual heat transfer edges do not participate in the distortion difference calculation and are only used to ensure the complete connectivity of the topology time series deduction process. After completing the virtual heat transfer edge completion of all isolated units, all nodes of the entire mixing cavity topology diagram have at least one connection edge, generating a thermal impedance topology diagram.

[0043] This step thoroughly screens all isolated units in the mixing chamber without physical heat exchange pathways using a full-domain, unit-by-unit verification rule. It matches the heat exchange intensity of virtual heat transfer edges based on the unit's long-term steady-state mixing conditions, preventing the introduction of false heat exchange parameters that deviate from the equipment's normal rubber and plastic mixing operation. Virtual heat transfer edges are independently marked to distinguish them from physical heat exchange connection edges, achieving isolation and filtering in subsequent calculations. This avoids interfering with the extraction accuracy of real rubber and plastic shear heat transfer distortion data, ultimately resulting in a fully connected topology. The thermal impedance topology serves as the input data source for the spatiotemporal graph neural network, participating in the splicing of time-series sample sequences across multiple sampling periods, providing a pre-marking basis for filtering invalid virtual edge parameters in the distortion extraction stage.

[0044] S2. Input the thermal impedance topology map into the spatiotemporal graph neural network to predict the evolution trend of the thermal impedance topology map, obtain the predicted topology map, and compare it with the standard thermal equilibrium topology map to extract the topological distortion features.

[0045] Specifically, extracting topological distortion features includes: After separately separating the node attributes and the thermal impedance of the connecting edges in the historical thermal impedance topology, they are spliced ​​together in chronological order to form a time series diagram sample sequence. By reading the parameter variation patterns in the time series sample sequence through the spatiotemporal graph neural network, the trend of node attributes and thermal impedance of connecting edges within a preset time period is deduced along the time axis to obtain the predicted topology graph. Establish a correspondence between the nodes and connecting edges in the predicted topology and the nodes and connecting edges corresponding to the spatial coordinates of the same thermal unit in the standard thermal equilibrium topology; Calculate the parameter difference between the corresponding node attributes and the connecting edges according to the comparison relationship, select the nodes and connecting edges with parameter differences greater than the preset threshold as distortion carriers, and summarize them to obtain the topological distortion features.

[0046] Among them, the parameter difference includes the offset of node attributes and the increase of thermal impedance of the connection edge. Nodes with an offset greater than the first preset threshold represent regions of disordered energy diffusion, which are regions where the temperature and medium flow rate in the thermal unit deviate from the steady state range, causing local heat transfer imbalance. Connection edges with an increase of thermal impedance greater than the second preset threshold represent regions of high thermal resistance barriers, which are heat transfer interfaces where the thermal impedance difference between adjacent thermal units exceeds the steady state range. The distortion weight of the thermal element corresponding to the distortion carrier is greater than that of the thermal element at the coarse scale. The distortion weight is multiplied by the parameter difference of the corresponding distortion carrier to obtain the weighted distortion amplitude. The sum of these values ​​yields the topological distortion feature.

[0047] The mixing process in a rubber-plastic compound internal mixer exhibits continuous temporal evolution characteristics. Shear heat generation, rubber-plastic medium flow, and chamber wall heat transfer states undergo lag-like changes over the mixing time. A single-moment thermal impedance topology diagram can only represent the instantaneous static heat transfer state and cannot reflect the dynamic changes of parameters over time. Inputting this into a spatiotemporal graph neural network will prevent the network from learning the temporal coupling law of internal mixing heat transfer. At the same time, the thermal impedance topology includes virtual heat transfer edges and corresponding parameters used for topology completion. Virtual heat transfer edges do not have real physical heat transfer behavior, and their parameters have no temporal evolution reference value. If they are not isolated and included in sample splicing, they will introduce invalid interference data, distort the heat transfer change law learned by the network, and cause distortion of the topology evolution inference results.

[0048] The historical thermal impedance topology map is a collection of thermal impedance topology maps continuously generated and stored over multiple consecutive sampling cycles of the internal mixer; node attributes refer to the pre-processed sensing data of temperature, medium flow rate, and jacket equipment power bound to a single thermal unit; the thermal impedance of the connection edge refers to the heat transfer resistance value corresponding to the unidirectional physical heat exchange path between adjacent thermal units; and the time series sample sequence refers to the input sample formed by linearly stacking and combining the effective topology parameters after splitting multiple sets of different sampling times according to the sampling time from earliest to latest.

[0049] The sampling interval for generating a thermal impedance topology map by the internal mixer is fixed at 1.5 seconds, and the topology map set generated by 30 consecutive sampling cycles is continuously stored as historical topology data. During the splitting operation, only the thermal impedance parameters of the real physical heat exchange connection edge and the node attribute data corresponding to all thermal units are extracted. The virtual heat transfer edge parameters of isolated units are stored separately and are not included in the timing splicing process.

[0050] For example, 30 historical topology maps are generated in 30 consecutive sampling cycles. In each topology map, the virtual heat transfer edge parameters of the empty coarse-scale isolated units at the four corners of the mixing chamber are all isolated and removed. Only the bidirectional physical heat transfer edge of the rotor material area and the node data of all scale thermal units in the entire domain are retained. They are stacked and spliced ​​in sequence according to the sampling time to finally form a time series map sample sequence with a time series length of 30.

[0051] This step isolates the physically meaningless parameters corresponding to the virtual heat transfer edges through parameter classification and splitting operations, eliminating the interference of topology completion auxiliary structures on the learning of temporal evolution laws from the source of the samples; it retains the correlation characteristics of temperature, medium flow rate and bidirectional thermal impedance changing continuously over time during the rubber-plastic compounding process, and the generated time series samples have three standardized features: unified data format, continuous time dimension, and only containing real heat transfer parameters, providing a spatiotemporal graph neural network with a data source without redundant interference and with complete temporal correlation.

[0052] The shear heat generation inside the rubber and plastic blending internal mixer exhibits a significant time lag. The thermal impedance topology collected at the current sampling moment can only characterize the instantaneous heat transfer state. Material hotspots and high thermal resistance heat transfer barriers across units will gradually develop as the mixing process continues. If control is implemented only based on the current instantaneous topology, the control action will have a significant lag and will not be able to suppress the continuous aggravation of heat transfer imbalance in advance. The spatiotemporal graph neural network simultaneously mines the spatial heat transfer coupling relationship between the thermal units in the mixing chamber and the temporal change law of heat transfer parameters over time. Based on historical time series samples, it completes the extrapolation of the full-domain heat transfer parameters within a fixed future time period and predicts the trend of distortion development in advance.

[0053] The spatiotemporal graph neural network refers to a graph structure network that synchronously learns the spatial coupling relationship of thermal units and the temporal variation law of heat transfer parameters; the preset duration refers to the fixed extrapolation duration for the equipment to predict the heat transfer distortion of the mixing process in advance; the predicted topology graph is a topology structure formed by reorganizing the node attributes and physical heat transfer edge thermal impedance parameters obtained from the extrapolation of the future time according to the original spatial arrangement rules of the thermal units in the mixing chamber, and the thermal unit division scale and three-dimensional spatial coordinates inside the topology are completely consistent with the generated topology.

[0054] The spatiotemporal graph neural network adopts a combined architecture of graph convolutional network and gated temporal convolution. The graph convolutional layer is responsible for aggregating the node attributes of adjacent thermal units along the connection edge direction and extracting spatial heat transfer coupling features. The gated temporal convolutional layer is responsible for sequential modeling of spatial features of continuous sampling periods along the time axis and capturing the temporal lag variation law of heat transfer parameters. The time series graph sample sequences collected under historical steady-state and non-steady-state intensive mixing conditions are used as the training set. The steady-state samples are used to learn the baseline of normal heat transfer evolution, and the non-steady-state samples are used to learn the distortion development trend. The mean square error of the parameters between the predicted topology and the actual future topology is used as the training loss function. After offline training, the network parameters are solidified for online inference.

[0055] Specifically, the time-series sample sequence is input into the spatiotemporal graph neural network. The spatiotemporal graph neural network synchronously and in parallel analyzes two types of change patterns. The first type is the time-series fluctuation pattern of the temperature, medium flow rate, and power parameters of a single thermal unit over time. The second type is the spatial coupling pattern of the bidirectional thermal impedance between adjacent thermal units as the operating conditions change. After completing the pattern learning, the network deduces the attribute values ​​of all nodes and the thermal impedance values ​​of each real physical heat transfer edge corresponding to a fixed preset time period in the forward time direction. Based on all the deduced parameters, the network is reorganized according to the original arrangement rules of the mixing chamber units to generate a predicted topology. No thermal units of any scale are added, deleted, or displaced within the topology.

[0056] For example, the internal mixer's preset simulation time is fixed at 12 seconds. After the input time series sample sequence with a time series length of 30 completes the pattern learning, the simulation obtains the estimated values ​​of temperature, medium flow rate and jacket power corresponding to each fine-scale, medium-scale and coarse-scale thermal unit in the entire internal mixing chamber after 12 seconds. At the same time, it outputs the estimated values ​​of thermal impedance corresponding to all bidirectional physical heat exchange edges. Following the original three-dimensional spatial unit arrangement rules of the internal mixing chamber, it reorganizes to obtain a predicted topology map in which the structure, unit coordinates and unit scale all match the real-time topology.

[0057] This step utilizes network synchronization to mine the dual objective laws of spatial coupling and temporal lag changes in heat exchange during mixing, enabling early prediction of local hot spots and high thermal resistance heat transfer barriers across units. The predicted topology map and the real-time generated topology have a unified spatial unit division benchmark, eliminating the problem of misalignment in subsequent parameter comparisons caused by inconsistent topology structures, and alleviating the defects of local overheating imbalance in rubber and plastics caused by the lag in control actions from the source.

[0058] The three-dimensional space of the rubber and plastic mixing chamber is composed of a large number of fine-scale material units, medium-scale mixing units and coarse-scale corner units. The number of internal units in the topology is large and the spatial arrangement structure is complex. If there is no unified matching judgment benchmark, the pairing of two internal units in the topology and the bidirectional heat exchange edge will be disordered, resulting in the complete misalignment of the matching objects for subsequent node and thermal resistance parameter difference calculations, and the overall distortion of the distortion carrier screening process.

[0059] The standard thermal balance topology refers to the baseline topology generated offline and stored according to the aforementioned unified rules under the steady-state condition of a rubber and plastic internal mixer without long-term external disturbance and constant mixing parameters. The spatial coordinates of the thermal unit refer to the three-dimensional spatial positioning value assigned to each unit in the thermal unit division stage, which serves as an identifier to distinguish different thermal units. The three-dimensional spatial coordinate values ​​corresponding to each thermal unit in the predicted topology and the standard thermal balance topology are extracted respectively. The equality of the two sets of coordinate values ​​is used as the matching judgment condition, and the thermal units with the same coordinates in the two topologies are judged as mutual comparison units. The node attributes bound inside the unit and the bidirectional physical heat exchange connection edges between units establish a comparison and pairing relationship. The virtual heat transfer edges do not participate in the coordinate matching comparison process at all.

[0060] The standard thermal balance topology diagram is generated before the internal mixer is put into formal production or during the periodic calibration stage. This is done by ensuring that the equipment is in a long-term stable operating condition (no material feeding or discharging disturbances, constant jacket heating power, constant conveying speed of rubber and plastic materials, and constant temperature and humidity in the workshop environment). Following the complete process of acquiring sensing data, calculating thermal load gradients, dividing thermal units, constructing connecting edges, and verifying connectivity, multiple sets of thermal impedance topology diagrams are continuously generated. The statistical average of node attributes and connecting edge thermal impedance is then taken to eliminate steady-state micro-fluctuations and solidify the data for storage.

[0061] For example, the three-dimensional coordinates of a fine-scale thermal unit in the core material region of the mixing chamber rotor are set as (1.2, 0.8, 0.5). The thermal unit corresponding to this coordinate is retrieved in the predicted topology map and the standard steady-state topology map respectively. The two sets of units are matched and paired. The two bidirectional physical heat exchange connection edges between the units are paired. The coordinates of the empty coarse-scale units at both ends of the cavity are not extracted and do not participate in any matching operation.

[0062] This step establishes precise unit and heat exchange connection edge mapping relationships based on spatial coordinates, eliminating the technical risk of misalignment in multi-scale complex dense topology pairing; the virtual heat transfer edge is isolated throughout the process and does not participate in the matching, avoiding the generation of invalid control pairings by auxiliary topology structures that have no real heat exchange significance, reducing the data processing volume of subsequent difference calculations, and eliminating the problem of misidentification and omission of distorted carriers caused by pairing misalignment from the matching logic level.

[0063] Under steady-state standard operating conditions, the temperature, medium flow rate, and equipment power parameters of each thermal unit node, as well as the thermal resistance of the bidirectional heat exchange side, all have a fixed steady-state fluctuation range. Small parameter deviations are within the allowable fluctuation range of normal mixing processes and have no value for control intervention. Only when the parameter difference exceeds the preset threshold does it indicate that there is a significant heat exchange imbalance in the corresponding heat exchange area with process risks. By using the threshold boundary, a layered screening is completed to filter out small normal fluctuations that have no control significance, and only abnormal heat exchange units and heat exchange interfaces that require subsequent decomposition of disturbances and allocation of control weights are retained.

[0064] The parameter differences are divided into two categories: node attribute offset and connection edge thermal impedance increase. The preset threshold is a quantitative judgment benchmark to distinguish between normal process fluctuations and abnormal heat transfer imbalances, and is divided into node judgment threshold and connection edge judgment threshold. The node attribute offset is composed of the sum of the absolute values ​​of the offsets of the three types of sensing parameters, and the connection edge thermal impedance increase is the direct difference between the predicted thermal impedance and the steady-state standard thermal impedance. node =|T pre -T std |+|v pre -v std |+|P pre -P std |,D Z =Z pre -Z std In the formula D node T represents the node attribute offset of a single set of control nodes. pre v pre P pre The parameters are, in order, the temperature, rubber-plastic medium flow rate, and jacket output power corresponding to the predicted topology nodes, T. std v std P std The following are the steady-state reference parameters for the corresponding coordinate nodes within the standard thermal equilibrium topology, D. Z Z represents the increase in thermal resistance of a single set of control connection sides. pre To predict the thermal resistance of the topological heat transfer edge, Z std This represents the steady-state thermal impedance of the heat transfer edge corresponding to the standard topology.

[0065] Specifically, for each set of nodes that have completed coordinate matching and bidirectional physical heat exchange connection edges, the corresponding difference is calculated. The node attribute offset is compared with the node's corresponding preset threshold, and the thermal impedance increase is compared with the connection edge's corresponding preset threshold. Only nodes and connection edges with difference values ​​greater than the corresponding threshold are marked as distortion carriers. Other units and heat exchange edges with small fluctuations are directly discarded and not included in the subsequent feature processing.

[0066] For example, the node attribute offset is matched with a first preset threshold of 10. This threshold is determined based on the statistical upper limit of the normal fluctuation range of the node's sensing parameters under steady-state conditions of the internal mixer. That is, the sum of the long-term mean and standard deviation of the absolute values ​​of the offsets of the three parameters of temperature, flow rate, and power under steady-state conditions is about 8 to 10. The upper limit is taken as the judgment benchmark. The connection edge thermal resistance increase is matched with a second preset threshold of 0.4. This threshold is determined based on the statistical analysis of the normal fluctuation range of the thermal resistance of the bidirectional heat exchange edge under steady-state conditions. The mean standard deviation of the thermal resistance of each heat exchange edge under steady-state conditions is about 0.15 to 0.2. The mean plus one standard deviation is taken as the judgment benchmark. Those skilled in the art can make calibration adjustments according to the heat exchange characteristics of the specific equipment. The node attribute offset of a fine-scale reference unit in the rotor material area is calculated to be 15, which is greater than 10 and is marked as a distorted node. The calculation result of the bidirectional heat exchange edge thermal resistance increase between two adjacent sets of material units is 0.6, which is greater than 0.4 and is marked as a distorted connection edge.

[0067] This step uses a quantified threshold to differentiate between normal process fluctuations and high-risk heat exchange imbalances, filtering out a large number of minor parameter deviations that have no regulatory significance. It retains only the heat units and heat exchange interfaces with significant heat exchange anomalies to form a set of distortion carriers, reducing the amount of invalid data computation in subsequent generative adversarial network perturbation decomposition and collaborative weight matrix solution steps.

[0068] The fine-scale material units inside the internal mixing chamber of rubber and plastics compounding correspond to concentrated shear heat generation areas. Even slight parameter deviations can cause process defects such as localized scorching of rubber and plastics and a decrease in mixing uniformity, making them significantly more sensitive to heat transfer imbalance. The heat transfer rate of the empty units at the corners of the coarse-scale chamber is slow, and the process risk corresponding to the same parameter difference is lower. If the same weight is uniformly used to complete the distortion, the quantitative amplitude of small distortions in material hot spots will be weakened, resulting in insufficient weight allocation of subsequent actuator control. At the same time, node offset distortion and thermal resistance increase distortion belong to two completely independent distortion causes: internal heat transfer imbalance and inter-unit heat transfer obstruction. Without classification labels, the subsequent generative adversarial network will be unable to decompose different types of disturbance sources by channel.

[0069] The thermal unit with a node attribute offset greater than the first preset threshold represents a region of disordered energy diffusion. The imbalance in this type of region is caused by the temperature of the rubber and plastic and the flow velocity of the medium inside the unit deviating from the steady-state range, resulting in local material heat transfer disorder. The heat transfer interface with a thermal impedance increase greater than the second preset threshold represents a region of high thermal resistance barrier. The imbalance in this type of interface is caused by the difference in thermal impedance between adjacent thermal units exceeding the steady-state allowable range, which obstructs the heat transfer path across units. The distortion weight is a correction coefficient that characterizes the sensitivity of heat transfer imbalance of thermal units at different scales. According to the division scale (fine scale, medium scale, coarse scale) of the thermal unit to which the distortion carrier belongs, the corresponding weight coefficient is obtained by looking up the table, where the fine scale is 1.8, the medium scale is 1, and the coarse scale is 0.7. The weighted distortion amplitude is the standardized distortion index obtained by multiplying the original parameter difference by the distortion weight of the matching unit scale.

[0070] The weighted distortion amplitude is A = D × W, where A represents the weighted distortion amplitude corresponding to a single distortion carrier; D represents the original parameter difference corresponding to that distortion carrier, and the node carrier is substituted into D. node Substitute D into the heat exchange side carrier Z That is, when the distortion carrier is a node (a region of disordered energy diffusion), D is taken as the attribute offset D of that node. node When the distortion carrier is a connecting edge type (high thermal resistance barrier region), D is taken as the increase in thermal resistance of the connecting edge. Z W represents the distortion weighting coefficient that matches the thermal unit scale to which the distortion carrier belongs.

[0071] The distortion weighting coefficient of the fine-scale material thermal unit is fixed to a value higher than that of the coarse-scale empty unit. After the difference is multiplied by the weight to obtain the weighted distortion amplitude, a classification label is added to each distortion carrier to distinguish between node distortion of disordered energy diffusion and heat transfer edge distortion of high thermal resistance barrier. All distortion carrier information carrying thermal unit scale label, distortion type label and weighted distortion amplitude is uniformly summarized and integrated to form topological distortion features.

[0072] For example, the distortion weight coefficient for fine-scale material thermal units is set to 1.8, while the distortion weight coefficient for coarse-scale corner thermal units is set to 0.7. The coarse-scale unit weight of 0.7 is used as a baseline. Due to the concentrated shear heat generation and high sensitivity to heat exchange imbalance, the fine-scale unit weight is increased to 1.8 (approximately 2.6 times that of the coarse-scale unit) according to the process risk amplification ratio. This ratio is based on the ratio of temperature runaway risk between hot spots and empty corner areas in the rubber and plastics mixing process. Those skilled in the art can adjust this ratio according to the specific material formulation and equipment structure. For a fine-scale distortion node with an original node offset of 15, the weighted distortion amplitude is 15 × 1.8 = 27. For a coarse-scale distortion node with an original node offset of 14, the weighted distortion amplitude is 14 × 0.7 = 9.8. Both types of distortion carriers are assigned classification labels and simultaneously stored in the final topological distortion features.

[0073] This step allocates distortion weight coefficients differently based on the division scale of thermal units, amplifies the quantitative amplitude corresponding to the distortion in the hot spot area of ​​rubber and plastic materials at the fine scale, and matches the objective process characteristics of the high sensitivity of hot spots in the internal mixing process. It adds classification labels to the two types of distortion carriers, clearly distinguishing between two independent distortion causes: heat transfer disorder within the unit and heat transfer barriers between units. This provides a clear classification basis for the subsequent generation of adversarial networks to decompose disturbance sources from different sources. The distortion classification labels support the independent decomposition of the dual-channel disturbance contribution, and the differentiated weighted distortion amplitude serves as the core evaluation criterion for the attention mechanism to determine the distortion suppression and regulation effectiveness of each actuator, achieving a precise match between the actual severity of distortion and the actuator's control intensity.

[0074] S3. Based on the topological distortion characteristics, the contribution of the disturbance source to the topological distortion is separated by generative adversarial network, and the regulatory effectiveness of the actuator in suppressing the topological distortion is determined by attention mechanism, generating a cooperative weight matrix.

[0075] Specifically, generating the collaborative weight matrix includes: The topological distortion features are input into a generative adversarial network, and the proportion of distortion coverage corresponding to each type of disturbance source is statistically analyzed to separate the contribution of each disturbance source to the topological distortion. The contribution of the disturbance source, the spatial coordinates of the thermal unit corresponding to the topological distortion, and the weighted distortion amplitude are input into the attention mechanism. Based on the spatial coverage relationship, the distortion region of the actuator is determined, and the actual suppression capability of the actuator against the topological distortion is determined one by one to obtain the regulation efficiency of the actuator. The basic weight is obtained by multiplying the contribution of the same thermal unit and actuator pairing with the regulation efficiency, and then generating a collaborative weight matrix with thermal units as rows and actuators as columns.

[0076] Among them, the number of thermal units that undergo topological distortion corresponding to the disturbance source under the two types of distortion carriers, namely the energy disorder diffusion region and the high thermal resistance barrier region, is counted respectively, and the proportion of distortion coverage is determined as the contribution of the corresponding disturbance source. The attention mechanism specifically prioritizes matching the spatial coverage relationship between the actuator and the corresponding thermal unit of the topological distortion. The adjustment efficiency of the same actuator for the configuration of fine-scale thermal units is greater than that for coarse-scale thermal units.

[0077] In the internal mixing process of rubber and plastics, there are disturbances caused by fluctuations in material feed flow rate, drift in jacket heating power, and deviation in ambient temperature within the chamber. Different disturbance sources correspond to two types of distortion carriers: disordered energy diffusion regions and high thermal resistance barrier regions. If a single channel is used to analyze all the distortion data without distinction, the impact of different disturbances will be superimposed and confused, making it impossible to accurately separate the degree of heat transfer imbalance caused by each type of disturbance. The method of uniformly calculating the distortion range cannot distinguish the disturbance causes corresponding to the two different types of distortion: heat transfer disorder within the unit and heat transfer obstruction between units. Consequently, the subsequent actuator control intensity allocation will lack accurate data basis.

[0078] Generative Adversarial Networks (GANs) are analytical networks capable of identifying distorted samples and separating disturbance features. The training set for GANs consists of historical distorted samples labeled according to disturbance source type (feed velocity fluctuations, jacket power drift, and ambient temperature shifts). The generator learns to simulate the distortion patterns generated by each disturbance source, while the discriminator learns to distinguish the types of distortion sources. After training convergence, the network parameters are solidified for online disturbance separation. A disturbance source refers to a process variable that can cause the heat transfer parameters of the mixing chamber to deviate from the steady-state range, including three categories: rubber and plastic feed velocity, mixing jacket heating power, and workshop ambient temperature. The distortion coverage ratio refers to the proportion of the number of thermal units induced by a single disturbance source to the total number of distorted units of the same type of distortion carrier. The disordered energy diffusion region is the carrier of heat transfer imbalance within a unit caused by excessive node parameter offset, while the high thermal resistance barrier region is the carrier of cross-unit heat transfer obstruction caused by excessive increase in thermal resistance at the connection edge. The contribution value is used to characterize the weight of the influence of a single disturbance source on the corresponding type of topological distortion.

[0079] Specifically, after the topological distortion features are fed into the generative adversarial network (GAN), the network internally splits the data into two subsets based on the distortion's built-in classification identifier: a subset of regions with disordered energy diffusion and a subset of regions with high thermal resistance. These two subsets of data are independent and perform perturbation matching and unit count separately; the perturbation contribution is C=N. s / N all In the formula, C represents the contribution of a certain disturbance source to a certain type of distortion, and N... s N represents the total number of thermal units that induce distortion from this disturbance source. all This represents the total number of all distorted thermal units under this type of distorted carrier.

[0080] For example, in the current sampling period of the internal mixer, there are a total of 20 distorted heat units on the high thermal resistance barrier-type distortion carrier, among which 12 distorted units are induced by the jacket power drift disturbance. Substituting into the calculation, the contribution of this disturbance source to such distortion is 0.6; there are a total of 15 distorted heat units on the energy disordered diffusion-type distortion carrier, and 9 distorted units are induced by the feed flow rate fluctuation disturbance, and the calculated contribution is 0.6.

[0081] The entire statistics and calculation process is only carried out for distortion carriers corresponding to real physical heat transfer, distortions derived from virtual heat transfer edges are not included in the statistical scope, and there is no data operation on auxiliary topology with no physical meaning.

[0082] In this step, disturbance unit statistics and proportion calculation are performed for two types of distortion carriers, realizing complete separation of disturbance sources corresponding to different heat transfer imbalance types, avoiding mutual overlapping interference of distortion effects of multiple types of disturbances. The contribution only represents the independent effect level of a single disturbance, objectively reflects the actual influence degree of various process disturbances on the heat transfer imbalance of internal mixing, and eliminates the problem of quantitative distortion of disturbance influence caused by unified global statistics.

[0083] The heating, cooling and medium flow regulating actuators matched with the rubber and plastic internal mixer all have a fixed spatial radiation range. A single actuator can only exert heat transfer regulation effect on heat units within the spatially overlapping area. Without spatial matching screening, actuators with no regulation effect will be included in the calculation, resulting in invalid effectiveness values; shear heat generation of fine-scale material heat units is concentrated, and even tiny distortion can cause process defects such as rubber and plastic scorching and uneven mixing. The heat transfer sensitivity of coarse-scale vacant corner units is lower, and the required regulation intensity is significantly different under the same distortion degree. Calculating regulation effectiveness with a unified standard will weaken the regulation priority of actuators in material hot spot areas.

[0084] The attention mechanism refers to an operation module that quantifies the regulation adaptation ability in multiple dimensions according to distortion position, distortion severity, unit scale and the spatial range of the actuator; the spatial coordinates of heat units are three-dimensional positioning identifiers of each distortion carrier; the weighted distortion amplitude is a distortion quantification index after completing scale weighting; the spatial coverage relationship is used to determine whether there is an overlap between the radiation regulation range of the actuator and the spatial area of the distorted heat unit; the regulation effectiveness represents the adaptation ability of a single actuator to suppress heat transfer imbalance for a specified distorted unit, and the effectiveness assignment benchmark corresponding to fine-scale heat units is higher than that of coarse-scale heat units.

[0085] Specifically, three types of input data (disturbance contribution, three-dimensional spatial coordinates of the distorted thermal unit, and weighted distortion amplitude) are simultaneously fed into the attention mechanism. First, each group of distorted carriers is matched with the corresponding disturbance contribution, three-dimensional spatial coordinates, and weighted distortion amplitude to complete the position binding. In the calculation process, the fixed radiation control space interval of the actuator is compared with the coordinate interval of the distorted thermal unit. Pairings with only spatial overlap are retained and their effectiveness is calculated. Actuators without spatial overlap are directly eliminated and no corresponding effectiveness value is generated. For the retained effective pairings, an effectiveness benchmark is assigned according to the thermal unit scale to which the distorted carrier belongs. Fine-scale material units use a high benchmark value interval, and coarse-scale vacant units use a low benchmark value interval. The higher the weighted distortion amplitude, the higher the corresponding allocated regulation effectiveness value.

[0086] For example, the fine-scale distortion thermal unit in the rotor region of the mixing chamber has spatial overlap with the corresponding jacket heating actuator, and the paired actuator is assigned a high reference regulation efficiency value of 0.82; the coarse-scale distortion thermal unit at the corner of the chamber has minimal spatial overlap with the same actuator, and is assigned a low reference regulation efficiency value of 0.35.

[0087] Throughout the operation, the adjustment and adaptation logic is distinguished between two types of distortion carriers. For cross-unit heat transfer distortion with high thermal resistance barriers, the actuator that adjusts the medium flow rate in both directions will further improve the performance benchmark to match the control requirements of bidirectional thermal resistance imbalance in the mixing process.

[0088] This step involves pre-selecting and filtering actuators that have no actual control function, thus reducing the data volume of subsequent weight matrix calculations. Performance benchmarks are set based on the differences in thermal unit scales to meet the high control priority requirements of hot spots in rubber and plastic materials. Simultaneously, performance values ​​are corrected by combining weighted distortion amplitude, achieving a matching logic where the more severe the distortion imbalance, the higher the quantitative value of the corresponding actuator's adjustment and adaptation capability. This objectively restores the true control and adaptation characteristics of the internal mixer actuators.

[0089] Relying solely on the disturbance contribution can only reflect the strength of the disturbance's impact, while relying solely on the actuator's regulation efficiency can only reflect the equipment's control and adaptation capabilities. Since these two types of parameters exist independently, they cannot form a unified standard for control allocation constraints. By coupling and multiplying the two to obtain the basic weight, the degree of disturbance impact and the equipment's control and adaptation level are taken into account simultaneously, so that actuators with more severe distortion, higher disturbance contribution, and stronger adaptability are allocated greater control weights.

[0090] The basic weights are the control allocation constraint coefficients obtained by coupling calculations of disturbance contribution and regulation efficiency under a single thermal unit and single actuator pairing; the collaborative weight matrix is ​​a two-dimensional structured dataset, with each row corresponding to a thermal unit in the mixing chamber and each column corresponding to a control actuator, and the cells inside the matrix storing the basic weight values ​​corresponding to the pairing relationship; the basic weight is W. base =C×E, where Wbase The base weight represents the pairing, C represents the contribution of the disturbance source corresponding to the distortion carrier, and E represents the regulation efficiency of the corresponding actuator matching the thermal unit.

[0091] All thermal units and actuators that have passed the spatial coverage screening are effectively paired. For each pairing, the contribution and regulation efficiency values ​​of the binding are retrieved and substituted into the formula to perform multiplication to obtain the basic weight. Invalid pairings without spatial overlap are filled with zero values ​​in the corresponding matrix cells, indicating that the actuator does not participate in the coordinated regulation of the corresponding thermal unit. After all pairing calculations are completed, all basic weights are integrated according to the arrangement rules of row corresponding to thermal units and column corresponding to actuators to generate a coordinated weight matrix. The number of rows and columns of the matrix is ​​consistent with the total number of thermal units in the mixing chamber and the total number of regulating actuators.

[0092] For example, the disturbance contribution of a fine-scale material distortion thermal unit is 0.6, and the matching actuator regulation efficiency is 0.82. Substituting these values ​​into the calculation, the basic weight is 0.492. The contribution of a coarse-scale corner distortion thermal unit is 0.5, and the matching actuator regulation efficiency is 0.35. The calculated basic weight is 0.175. The collaborative weight matrix carries the coupling control constraint coefficients for each thermal unit and each actuator, and synchronously marks the thermal unit scale and distortion type identifiers corresponding to each cell, which facilitates differentiated calls in subsequent control and solution stages.

[0093] This step obtains the basic weights by coupling and multiplying contribution and regulation efficiency, and simultaneously integrates the dual constraints of the strength of disturbance impact and the actuator's control adaptation capability, objectively reflecting the priority control allocation ratio of the corresponding actuator for the corresponding distortion region; a unified mapping relationship is established across the entire domain, and the pairing constraints of all hot units and actuators are centrally integrated, eliminating the risk of computational logic disorder caused by scattered numerical retrieval and matching; the collaborative weight matrix serves as the input carrier for solving the multi-actuator collaborative control instructions, and the fixed row and column correspondence of units and actuators within the matrix provides a basis for the allocation ratio for topology difference splitting and itemized control requirements summarization. The basic weights within the matrix determine the allocation ratio of control intensity for different distortion regions and different actuators, while the scale and distortion type identifiers carried by the matrix provide a preliminary judgment basis for subsequent adjustment of the reverse compensation instruction intensity.

[0094] S4. Using the collaborative weight matrix as a constraint and the topological difference between the predicted topology map and the standard thermal equilibrium topology map as the objective, calculate and execute the collaborative instructions of the actuator. When the rate of change of the topological difference is detected to be greater than the change threshold, generate a reverse compensation instruction according to the direction of change to execute the hedging control.

[0095] Specifically, generating the reverse compensation instruction includes: By comparing the predicted topology map with the standard thermal balance topology map, the topology differences corresponding to the thermal units are obtained. The topology differences of each thermal unit are then split according to the weight ratio of the actuators in the collaborative weight matrix to obtain the individual actuator's sub-item control requirements for the thermal unit. The individual control requirements of all thermal units matched by the same actuator are summarized, the coordinated instructions of the actuator are obtained and executed; The adjusted topology difference is obtained and the rate of change of the topology difference is calculated. When the rate of change is greater than the change threshold, a reverse compensation instruction is generated according to the direction of change of the topology difference. This instruction is then superimposed on the current cooperative instruction and output to perform hedging control.

[0096] Among them, when generating the reverse compensation command, the cooperative weight that matches the thermal unit that generates the topological difference is retrieved. The larger the cooperative weight, the greater the control force of the corresponding reverse compensation command. Reverse compensation is initiated only when the rate of change of topological differences exceeds the change threshold for multiple consecutive sampling periods.

[0097] The internal mixing chamber for rubber-plastic blending contains multi-scale thermal units, equipped with multiple heating, cooling, and medium flow rate regulating actuators. Each actuator only has a regulating effect on a portion of the spatial thermal units. The regulation allocation ratio of different actuators for the same thermal unit is limited by the cooperative weight matrix obtained in the previous steps. If the overall topological differences are not split according to the weight ratio and a uniform regulation amount is directly assigned to a single actuator, the problem of insufficient regulation force of actuators in severely distorted areas and ineffective action of actuators in non-distorted areas will occur, which cannot match the differentiated weight constraints obtained in the previous steps based on the disturbance contribution and unit scale.

[0098] The predicted topology is the future thermal impedance topology of the mixing chamber obtained by spatiotemporal neural network deduction; the standard thermal balance topology is the baseline topology for offline solidification storage under long-term stable and undisturbed operating conditions of the mixing machine; the topology difference refers to the total comprehensive deviation of all node attributes and bidirectional heat transfer thermal impedance of the same three-dimensional coordinate thermal unit between the predicted topology and the standard topology; the collaborative weight matrix is ​​a two-dimensional constraint matrix with rows corresponding to thermal units and columns corresponding to actuators, and the values ​​of the matrix cells represent the weight allocation ratio of the corresponding actuator for the corresponding thermal unit; the sub-item control requirement refers to the basic control amplitude that a single actuator needs to output for a single thermal unit.

[0099] Specifically, the three-dimensional spatial coordinates of the thermal unit are used as the matching identifier. All node parameters and thermal impedance differences of the bidirectional physical heat exchange edge of the units with the same coordinates in the two topologies are summarized and summed to obtain the total topological difference corresponding to the thermal unit. The weight values ​​of all actuators corresponding to the thermal unit in the collaborative weight matrix are retrieved, and the proportion of the weight of a single actuator in the total weight of all effective actuators in the unit is calculated. The total topological difference is then divided based on this proportion to obtain the sub-item control requirements of the actuator for the current thermal unit.

[0100] Topological differences are broken down into sub-items. In the formula This represents the sub-item control requirements of the j-th actuator for the i-th thermal unit. This represents the total topological difference corresponding to the i-th thermal unit. is the basic weight value in the i-th row and j-th column of the collaborative weight matrix, and m represents the total number of actuators that have a regulatory effect on the i-th thermal unit.

[0101] For example, the total topological difference of a fine-scale material thermal unit in a mixing chamber is calculated to be 32. There are three effective control actuators for this unit, and the corresponding weights in the collaborative weight matrix are 0.492, 0.261, and 0.113, respectively. The sum of the weights of the three actuators is 0.866. Substituting these values ​​into the calculation, the corresponding sub-control requirements for the three actuators are 18.18, 9.64, and 4.18, respectively. The matrix weight of the actuators without spatial control coverage is zero. After splitting, the sub-control requirements are directly set to zero and do not participate in the subsequent collaborative instruction aggregation calculation.

[0102] This step completes the topological difference ratio decomposition based on the previously formed collaborative weight matrix, inherits the dual differential control constraints of disturbance contribution and hot unit scale, and matches the sub-item control requirements obtained by the decomposition with the spatial adaptability and distortion impact of each actuator. From the data level, it avoids the problems of insufficient correction in hot areas and excessive control in inert areas caused by uniform control amplitude, and simultaneously eliminates the invalid calculation amount of actuators without control function, reducing the data processing scale of the subsequent instruction aggregation stage.

[0103] A single internal mixing control actuator simultaneously covers multiple thermal units of different sizes and distortion levels within the space. Each thermal unit generates individual control requirements. The scattered individual values ​​cannot be converted into unified control commands that the equipment can recognize. It is necessary to integrate and summarize all individual requirements corresponding to the same actuator to form a control output that takes into account the deviation of the heat exchange area it covers. If individual requirements are issued directly without summarization, it will cause the actuator to operate in multiple segments and the control output to fluctuate frequently, which is not conducive to the stable control of rubber and plastic mixing temperature.

[0104] The coordinated instruction refers to a unified control output instruction that can be recognized by a single heating, cooling, and medium flow regulation actuator. The instruction value is obtained by accumulating the sub-item control requirements of all thermal units covered by the actuator. Specifically, according to the actuator number, all sub-item control requirements corresponding to all thermal units are traversed, and all sub-item control requirements belonging to the same actuator are arithmetically summed. The summed value is converted into the power and medium flow regulation range of the corresponding equipment, forming a coordinated instruction, which is synchronously transmitted to the corresponding actuator to complete the basic control action.

[0105] The formula for solving cooperative instructions is: In the formula This represents the value of the coordination instruction corresponding to the j-th actuator. The sub-item control requirements are allocated to the j-th actuator for the i-th thermal unit, where n represents the total number of thermal units that the j-th actuator can cover for control.

[0106] For example, the No. 1 jacket heating actuator of the internal mixer simultaneously covers six fine-scale material heating units. The sub-item control requirements of the six heating units allocated to the equipment are 18.14, 7.36, 5.81, 3.95, 2.67, and 1.92, respectively. The sum of all sub-item values ​​yields a total value of 39.85 for the coordinated instruction. This value is converted into a heating power adjustment instruction and sent to the jacket actuator to complete the basic temperature control.

[0107] The entire summary calculation only adds up the valid values ​​of the sub-item control requirements that are not zero. Invalid sub-item requirements with zero weight are not included in the summation and do not consume computing resources.

[0108] This step involves a unified accumulation and summary of the control requirements for all areas covered by a single actuator, ensuring that the actuator can complete the coordinated control of the entire coverage area in a single operation, thus avoiding frequent fluctuations in the mixing temperature caused by multiple segmented controls. The entire process uses the sub-item data after the coordinated weighting, retains the differentiated control priorities of different thermal units, and achieves synchronous and balanced correction of heat exchange deviations in multiple areas.

[0109] The rubber and plastic internal mixing process exhibits a heat transfer lag effect. After the basic coordination command is issued and executed, the topological deviation in some areas shows a continuous unidirectional expansion trend, and the basic coordination command alone cannot suppress the continuous deterioration of the deviation. The instantaneous single sampling cycle change rate exceeding the standard is only a short-term disturbance to the process, while continuous multi-cycle exceeding the standard indicates that the heat transfer imbalance has a continuous deterioration trend. Instantaneous trigger compensation will cause the actuator output to oscillate frequently, damaging the control equipment and destroying the uniformity of rubber and plastic mixing. At the same time, the topological difference has two types of change: continuous increase and continuous decrease. It is necessary to match the corresponding reverse adjustment force to offset the deviation change trend, and the compensation force needs to match the coordination weight corresponding to the distorted area. The higher the weight, the stronger the distortion sensitivity, and the required reverse compensation force should be increased accordingly.

[0110] The adjusted topological difference refers to the updated comprehensive heat transfer deviation obtained by re-comparing the predicted topology with the standard topology after executing the coordinated command to complete the control action; the rate of change of topological difference refers to the ratio of the difference between the topological difference values ​​of two consecutive sampling periods to the sampling interval; the change threshold is the quantitative judgment benchmark for distinguishing between normal deviation fluctuations and a continuous deterioration trend of deviation; the direction of change refers to the two evolution trends of topological difference over time: continuous increase or continuous decrease; the reverse compensation command is an auxiliary control command to offset the unidirectional evolution trend of deviation; the coordinated weight refers to the matrix basic weight value corresponding to the thermal unit that generates topological difference, and the weight value is positively correlated with the compensation control intensity; multiple sampling periods refer to continuous and uninterrupted fixed sampling time sequence units, and the compensation logic is only allowed to be activated if the rate of change exceeds the threshold for multiple consecutive periods.

[0111] The specific fixed sampling interval is consistent with the sampling period of the preceding topology generation. After each execution of the cooperative instruction, the updated topology differences of all thermal units in the entire domain are recalculated. The topology difference values ​​of two adjacent samples of the same thermal unit are extracted and substituted into the rate of change formula V=(D t2 -D t1 ) / Δt, where V represents the rate of change of topological difference, D t1 D represents the topological difference at the previous sampling time. t2 The topological difference at the current sampling time is represented by Δt, which is a fixed time interval between two samplings. After the rate of change is calculated, the value is compared with the preset change threshold. The number of sampling periods that continuously exceed the threshold is recorded simultaneously. Only when multiple consecutive sampling periods meet the condition that the rate of change is greater than the change threshold is it determined to be a condition where the deviation continues to worsen. The direction of change is distinguished by the positive or negative value of V. A positive V indicates that the topological difference continues to expand, generating a reverse compensation command to reduce the heat transfer deviation. A negative V indicates that the topological difference continues to narrow, generating a reverse compensation command to slightly increase the heat transfer deviation.

[0112] The change threshold is determined based on the long-term statistical distribution of the topological difference change rate under steady-state operating conditions of the internal mixer. The mean change rate under steady-state conditions is approximately 0.2 to 0.3. The mean plus one standard deviation is taken as the threshold judgment benchmark to distinguish between normal deviation fluctuations and a continuous deterioration trend of deviation. Those skilled in the art can adjust it according to specific process requirements.

[0113] During the compensation instruction generation stage, the corresponding cooperative weight value of the distorted thermal unit is retrieved. The larger the weight value, the greater the adjustment range of the matched reverse compensation instruction. After the compensation instruction is formed, it is superimposed with the currently executing cooperative instruction. The superimposed total instruction is used as the latest output control quantity of the actuator to complete the hedging control.

[0114] For example, the calculated values ​​of the topological difference change rate of a fine-scale material thermal unit in the mixing chamber for four consecutive sampling cycles are 0.72, 0.78, 0.81, and 0.75, respectively, all of which are greater than the preset change threshold of 0.5, meeting the condition for continuous exceeding the standard in multiple cycles; the corresponding collaborative weight value of this thermal unit is 0.492, which matches and generates a reverse compensation command with a larger amplitude, superimposed on the current jacket heating actuator collaborative command to complete the offset and curb the trend of continuous expansion of the temperature difference in the material area; the corresponding collaborative weight of the coarse-scale vacant thermal unit in the chamber is 0.175, which generates a compensation command with a smaller amplitude under the same change rate conditions.

[0115] This step employs a continuous multi-sampling cycle exceeding-limit judgment rule to filter out single-time rate of change exceeding limits caused by instantaneous process disturbances, avoiding frequent oscillations in actuator control output. It accurately identifies the direction of deviation evolution based on the positive and negative values ​​of topological differences, generating reverse compensation commands that match the trend to achieve targeted hedging. The reverse compensation strength is bound to the collaborative weight of the thermal unit, achieving a differentiated control effect where the compensation strength is higher in hot-spot areas of fine-scale materials and smoother in inert coarse-scale areas, taking into account both the heat-sensitive process characteristics of rubber and plastic materials and the stability of equipment operation. The control commands are directly sent to the actuators to complete temperature control. The control actions output by the entire process correct the topological deviation across the entire mixing chamber. Simultaneously, the real-time topological difference after compensation serves as the original base data for the next round of topological distortion feature extraction, entering the next control cycle.

[0116] Example 2 like Figure 3 As shown, this embodiment provides a process temperature control system based on multi-parameter coupling, including a topology module, a prediction module, a generation module, and a control module.

[0117] The topology module is used to acquire the sensing data of the current process environment, and generates a thermal impedance topology map with thermal units in the process space as nodes, heat transfer relationships between thermal units as edges, and sensing data as node attributes. The prediction module is used to input the thermal impedance topology map into the spatiotemporal graph neural network, predict the evolution trend of the thermal impedance topology map, obtain the predicted topology map, and compare it with the standard thermal equilibrium topology map to extract topological distortion features. The generation module, based on topological distortion features, separates the contribution of perturbation sources to topological distortion through generative adversarial networks, and determines the regulatory effectiveness of the actuator in suppressing topological distortion through an attention mechanism, thereby generating a collaborative weight matrix. The control module is used to calculate and execute the actuator's cooperative instructions with the cooperative weight matrix as a constraint and the topological difference between the predicted topology map and the standard thermal equilibrium topology map as the objective. When the rate of change of the topological difference is detected to be greater than the change threshold, a reverse compensation instruction is generated according to the direction of change to execute the hedging control.

[0118] The above description is merely a preferred embodiment of this application. The scope of protection of this application is not limited to the above embodiments. All technical solutions falling within the scope of this application's concept are within the scope of protection of this application. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of this application should also be considered within the scope of protection of this application.

Claims

1. A process temperature control method based on multi-parameter coupling, characterized in that, include: Acquire the sensing data of the current process environment, and generate a thermal impedance topology map with thermal units in the process space as nodes, heat transfer relationships between thermal units as edges, and sensing data as node attributes. The thermal impedance topology map is input into the spatiotemporal graph neural network to predict the evolution trend of the thermal impedance topology map, and the predicted topology map is obtained. It is then compared with the standard thermal equilibrium topology map to extract topological distortion features. Based on the characteristics of topological distortion, the contribution of the disturbance source to the topological distortion is separated by generative adversarial network, and the regulatory effectiveness of the actuator in suppressing topological distortion is determined by attention mechanism, generating a cooperative weight matrix. With the collaborative weight matrix as a constraint and the topological difference between the predicted topology map and the standard thermal equilibrium topology map as the objective, the collaborative instructions of the actuator are calculated and executed. When the rate of change of the topological difference is detected to be greater than the change threshold, a reverse compensation instruction is generated according to the direction of change to execute the hedging control.

2. The process temperature control method based on multi-parameter coupling according to claim 1, characterized in that, The generated thermal impedance topology includes: Acquire the current process environment sensing data and perform preprocessing; The heat load gradient in the process space is calculated based on the preprocessed sensing data. The heat load gradient is then compared with the gradient threshold group and divided into multiple thermal units. Using thermal units as nodes, connection edges are constructed based on the heat transfer relationship between thermal units, and node attributes are determined based on preprocessed sensing data. Perform connectivity checks on all nodes and connecting edges. After adding virtual heat transfer edges to isolated nodes identified by the checks, generate a thermal impedance topology diagram.

3. The process temperature control method based on multi-parameter coupling according to claim 2, characterized in that, The heat load gradient is the difference in heat change per unit distance within the process space, formed by temperature, medium flow rate and equipment output power. The heat units are divided into different scales according to the gradient threshold group, and each heat unit is filled with only the same heat exchange medium. The heat transfer relationship is divided into two heat transfer directions: heat inflow and heat outflow. The difficulty of heat transfer in the two heat transfer directions corresponds to different thermal resistance values. The connectivity verification checks the physical heat transfer correlation of the thermal units to identify isolated nodes, matches the heat transfer intensity of isolated nodes with the heat transfer level of historical steady-state conditions, and adds virtual heat transfer edges to isolated nodes.

4. The process temperature control method based on multi-parameter coupling according to claim 3, characterized in that, The extracted topological distortion features include: After separately separating the node attributes and the thermal impedance of the connecting edges in the historical thermal impedance topology, they are spliced ​​together in chronological order to form a time series diagram sample sequence. By reading the parameter variation patterns in the time series sample sequence through the spatiotemporal graph neural network, the trend of node attributes and thermal impedance of connecting edges within a preset time period is deduced along the time axis to obtain the predicted topology graph. Establish a correspondence between the nodes and connecting edges in the predicted topology and the nodes and connecting edges corresponding to the same thermal unit spatial coordinates in the standard thermal equilibrium topology; Calculate the parameter difference between the corresponding node attributes and the connecting edges according to the comparison relationship, select the nodes and connecting edges with parameter differences greater than the preset threshold as distortion carriers, and summarize them to obtain the topological distortion features.

5. The process temperature control method based on multi-parameter coupling according to claim 4, characterized in that, The parameter difference includes the offset of node attributes and the increase in thermal resistance of the connecting edge. Nodes with an offset greater than a first preset threshold represent regions of disordered energy diffusion, which are regions where the temperature and medium flow rate in the thermal unit deviate from the steady-state range, causing local heat transfer imbalance. Connecting edges with an increase in thermal resistance greater than a second preset threshold represent regions of high thermal resistance barriers, which are heat transfer interfaces where the thermal resistance difference between adjacent thermal units exceeds the steady-state range.

6. The process temperature control method based on multi-parameter coupling according to claim 5, characterized in that, The generated collaborative weight matrix includes: The topological distortion features are input into a generative adversarial network, and the proportion of distortion coverage corresponding to each type of disturbance source is statistically analyzed to separate the contribution of each disturbance source to the topological distortion. The contribution of the disturbance source, the spatial coordinates of the thermal unit corresponding to the topological distortion, and the weighted distortion amplitude are input into the attention mechanism. Based on the spatial coverage relationship, the distortion region of the actuator is determined, and the actual suppression capability of the actuator against the topological distortion is determined one by one to obtain the regulation efficiency of the actuator. The basic weight is obtained by multiplying the contribution of the same thermal unit and actuator pairing with the regulation efficiency, and then generating a collaborative weight matrix with thermal units as rows and actuators as columns.

7. The process temperature control method based on multi-parameter coupling according to claim 6, characterized in that, The number of thermal units that undergo topological distortion corresponding to disturbance sources under the two types of distortion carriers, namely the energy disorder diffusion region and the high thermal resistance barrier region, is counted separately. The proportion of distortion coverage is determined and used as the contribution of the corresponding disturbance source. The attention mechanism prioritizes matching the spatial coverage relationship between the actuator and the corresponding thermal unit of the topological distortion. The adjustment efficiency of the same actuator for the configuration of fine-scale thermal units is greater than that for coarse-scale thermal units.

8. The process temperature control method based on multi-parameter coupling according to claim 7, characterized in that, The generation of reverse compensation instructions includes: By comparing the predicted topology map with the standard thermal balance topology map, the topology differences corresponding to the thermal units are obtained. The topology differences of each thermal unit are then split according to the weight ratio of the actuators in the collaborative weight matrix to obtain the individual actuator's sub-item control requirements for the thermal unit. The individual control requirements of all thermal units matched by the same actuator are summarized, the coordinated instructions of the actuator are obtained and executed; The adjusted topology difference is obtained and the rate of change of the topology difference is calculated. When the rate of change is greater than the change threshold, a reverse compensation instruction is generated according to the direction of change of the topology difference. This instruction is then superimposed on the current cooperative instruction and output to perform hedging control.

9. The process temperature control method based on multi-parameter coupling according to claim 8, characterized in that, When generating a reverse compensation command, the cooperative weight that matches the thermal unit that caused the topological difference is retrieved. The larger the cooperative weight, the greater the control force of the corresponding reverse compensation command. Reverse compensation is initiated only when the rate of change of topological differences exceeds the change threshold for multiple consecutive sampling periods.

10. A process temperature control system based on multi-parameter coupling, used to implement the process temperature control method based on multi-parameter coupling as described in any one of claims 1-9, characterized in that, include: Topology module, prediction module, generation module, and control module; The topology module is used to acquire the sensing data of the current process environment, and generates a thermal impedance topology map with thermal units in the process space as nodes, heat transfer relationships between thermal units as edges, and sensing data as node attributes. The prediction module is used to input the thermal impedance topology map into the spatiotemporal graph neural network, predict the evolution trend of the thermal impedance topology map, obtain the predicted topology map, and compare it with the standard thermal equilibrium topology map to extract topological distortion features. The generation module, based on topological distortion features, separates the contribution of perturbation sources to topological distortion through generative adversarial networks, and determines the regulatory effectiveness of the actuator in suppressing topological distortion through an attention mechanism, thereby generating a collaborative weight matrix. The control module is used to calculate and execute the actuator's cooperative instructions with the cooperative weight matrix as a constraint and the topological difference between the predicted topology map and the standard thermal equilibrium topology map as the objective. When the rate of change of the topological difference is detected to be greater than the change threshold, a reverse compensation instruction is generated according to the direction of change to execute the hedging control.

Images