Topological layout optimization method for heat pipe network compensator based on topological representation learning

Abstract

This invention relates to the field of multi-agent system technology, specifically to a method for optimizing the topology layout of thermal pipeline compensators based on topology representation learning. The method includes acquiring heterogeneous features such as pipeline topology, coking temperature variation sequence, vibration matrix, creep strain, and spatial obstacle repulsion, and embedding these features into graphs. A spatiotemporal graph neural network is used to aggregate the feature graphs, extracting the pipeline stress evolution trajectory and generating stress representation vectors related to each node. A three-party intelligent agent decision-making system consisting of the furnace, pipeline compensation, and heat exchanger is established, executing game-theoretic decisions under the guidance of spatial repulsion. A reinforcement learning mechanism is used to calculate the global stiffness matrix perturbation and dynamically adjust the compensation weights of each node until displacement absorption and interface stress meet the steady-state threshold. Finally, the installation coordinates are determined based on the weights, the number of bellows layers is matched, and the output pre-tension offset is calculated. This invention achieves automatic optimization of pipeline layout through topology representation and game theory.

Description

Technical Field

[0001] This invention relates to the field of multi-agent system technology, specifically to a method for optimizing the topology layout of thermal pipeline compensators based on topology representation learning. Background Technology

[0002] With the deepening application of industrial digital transformation and computer-aided design technology, the layout design of thermal pipeline networks for complex industrial scenarios (such as ethylene cracking furnace areas) has shifted from traditional manual drawing to automated wiring and component configuration based on numerical simulation and algorithm-driven methods. In the design logic of thermal pipeline networks, compensators, as core components that absorb thermal displacement and balance internal forces of the system, directly determine the safety and economy of the overall engineering design based on the quality of their layout scheme.

[0003] In the process of auxiliary design of pipeline networks in modern factories, the design space often faces multiple superimposed constraints: on the one hand, existing infrastructure, steel structures, and process equipment form stringent three-dimensional geometric barriers, constituting a complex non-convex search space; on the other hand, unique operating conditions such as coking in pyrolysis furnaces introduce time-fluctuating unsteady thermal loads, causing the stress distribution within the pipeline network topology to exhibit high nonlinearity and time lag. Furthermore, strong mechanical coupling exists within the pipeline network system, and layout changes at a certain local location can generate global stiffness matrix disturbances through the topological path.

[0004] Existing design-aided methods often struggle to balance computational efficiency and design accuracy when dealing with high-dimensional, multi-constraint combinatorial optimization problems. As the scale of the pipeline network increases or the constraints become extreme (such as the coexistence of extremely narrow spaces and variable load conditions), the design system is prone to getting trapped in local optima and struggles to accurately characterize the mechanical evolution under dynamic conditions. This results in the generated layout scheme often failing to meet the standards in the actual mechanical steady-state verification stage, increasing the cost of design iteration.

[0005] To address this, a method for optimizing the topology layout of thermal pipeline compensators based on topology representation learning is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method for optimizing the topology layout of thermal pipeline compensators based on topology representation learning. Through topology representation and game theory, automatic optimization of the pipeline layout is achieved. This includes acquiring heterogeneous features such as pipeline topology, coking temperature variation sequence, vibration matrix, creep strain, and spatial obstacle repulsion, and embedding these features into a graph. A spatiotemporal graph neural network is used to aggregate the feature graphs, extract the pipeline stress evolution trajectory, and generate stress representation vectors related to each node. A three-party intelligent decision-making system consisting of the furnace body, pipeline compensation, and heat exchanger is established to execute game theory decisions under the guidance of spatial repulsion. A reinforcement learning mechanism is used to calculate the global stiffness matrix perturbation and dynamically adjust the compensation weights of each node until displacement absorption and interface stress meet the steady-state threshold. Finally, the installation coordinates are determined based on the weights, the number of bellows layers is matched, and the output pre-tension offset is calculated.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A topology layout optimization method for thermal pipeline compensators based on topology representation learning includes: Obtain the pipeline topology, coking temperature change sequence, furnace body expansion field, vibration matrix, creep strain, and repulsive force of spatial obstacles; Based on the pipeline network topology, a graph structure containing nodes and edges is constructed. The clearing temperature change sequence is embedded as the node temporal feature, and the vibration matrix is ​​mapped as the edge stiffness attribute to complete the embedding of topological heterogeneous features. A spatiotemporal graph neural network is used to aggregate the heterogeneous feature graph, extract the stress evolution trajectory of the pipeline network under temperature change cycle, and generate the stress characterization vector of each node time-related. A decision-making system is established, consisting of three intelligent agents: furnace body, pipeline compensation, and heat exchanger. The furnace body expansion field, creep strain, and interface stress limit are input into the corresponding intelligent agents as boundary displacement, attenuation parameters, and stress constraints, respectively. The decision-making system calls the stress characterization vector and combines it with the repulsive force of spatial obstacles to perform game-theoretic decision-making, calculates the disturbance of the selected position to the global stiffness matrix of the system, and iteratively adjusts the compensation weights of each node until the displacement absorption and interface stress meet the steady-state threshold. The installation coordinates are determined in three-dimensional space based on the iterated compensation weights, and the number of bellows layers is matched based on the characterization vector. The pre-tension offset is calculated and output using creep strain.

[0008] Preferably, the process of acquiring the pipeline topology, coking temperature change sequence, furnace body expansion field, vibration matrix, creep strain, and spatial obstacle repulsion includes: extracting the centerline coordinates, pipe diameter, wall thickness parameters, and topological connection relationships of the pipelines in the pyrolysis furnace area; combining the material linear expansion coefficient to generate the pipeline topology; extracting the temperature sampling time sequence including the heating section, the isothermal coking section, and the cooling section from the production control system to form a coking temperature change sequence reflecting the thermal shock cycle of the pipeline; acquiring non-uniform axial displacement vector data at coordinate points at different heights of the furnace wall, and using the displacement vector data as the dynamic displacement boundary at the beginning of the pipeline. The conditions are as follows: the expansion field of the furnace body is obtained; the rotational speed and vibration amplitude of the moving equipment are collected and the mechanical excitation frequency is calculated. Combined with the natural frequency of the pipeline network, a numerical matrix characterizing the coupling strength between excitation and resonance is constructed as the vibration matrix; based on the high-temperature creep rate curve of the material and the cumulative operating hours of the pipeline, the irreversible permanent deformation within the preset service period is calculated as the creep strain; the existing infrastructure coordinate boundary in the pyrolysis furnace area is identified and the minimum physical distance of the pipeline sampling point relative to the coordinate boundary is calculated. The minimum physical distance is mapped to a potential energy weight with a spatial gradient to construct the spatial obstacle repulsion force.

[0009] Preferably, the process of embedding the heterogeneous topological features includes: mapping the pipe segment intersections, furnace connection points, and heat exchanger inlet and outlet points in the pipe network topology to a set of nodes in a graph structure, and mapping the physical pipe segments connecting each node to a set of edges in a graph structure; performing sliding window sampling on the coking temperature change sequence to generate temperature state vectors corresponding to each node at different time steps, and assigning the temperature state vectors as time-series feature attributes to the corresponding nodes in the node set; extracting the excitation frequency components of each edge in the vibration matrix, calculating the dynamic response disturbance coefficient of the excitation frequency components to the inherent stiffness of the pipe segment, and mapping the dynamic response disturbance coefficient to the weight attribute of each edge in the edge set, so as to realize the feature fusion of the graph structure in the spatial topology and dynamic load dimensions.

[0010] Preferably, the process of generating time-related stress characterization vectors for each node includes: mapping physical nodes and pipe segments in the pipeline network topology to vertices and edges of a graph structure, respectively; processing the clearing temperature change sequence into dynamic temperature feature vectors for each node using a sliding window method; extracting the dynamic response coefficients of each pipe segment as edge weights using the vibration matrix to align physical attributes to the graph feature space; using a graph convolution operator to transfer features along the pipeline network topology path; using the dynamic response coefficients on the edges to weight and aggregate the temperature features of adjacent nodes to extract the physical transmission characteristics of thermal stress in the pipeline network under a specific spatial configuration; inputting the spatially aggregated feature stream into a time-series processing unit to capture the stress response differences of the pipeline network at different stages of heating, isothermal decoking, and cooling; extracting the stress accumulation effect and hysteresis features that change over time; and fusing spatial topology features and time-series features through an attention mechanism to compress the complex stress fluctuations within the entire operating cycle into a high-dimensional hidden state sequence to generate stress characterization vectors.

[0011] Preferably, the process of establishing a decision-making system composed of three intelligent agents—furnace body, pipeline compensation, and heat exchanger—includes: establishing independent physical intelligent agents representing the furnace body, pipeline compensation, and heat exchanger respectively; converting the furnace body expansion field into the active displacement output of the furnace body intelligent agent, converting creep strain into the long-term performance degradation coefficient of the pipeline compensation intelligent agent, and using the interface force limit as the mandatory force constraint boundary of the heat exchanger intelligent agent; having the pipeline compensation intelligent agent call the stress characterization vector, and searching for a candidate set of compensator installation coordinates under the guidance of the potential energy constructed by the repulsive force of spatial obstacles; each intelligent agent, according to its own constraint conditions... The system engages in collaborative game theory among candidate nodes, iteratively evolving with the common goal of minimizing global system stress and achieving interface stress compliance. For each agent's collaboratively selected compensator installation location, the system calculates in real time the disturbance to the global stiffness matrix of the pipeline system caused by the intervention at that location. By analyzing the impact of the disturbance on the system's vibration frequency and stress redistribution, a stiffness feedback command is generated. Based on the stiffness feedback command, the system dynamically adjusts the compensation weights of each candidate node and drives the agent group to re-execute the game decision through a reinforcement learning mechanism until the displacement absorption of the pipeline system and the stress at the heat exchanger interface both meet the preset steady-state threshold.

[0012] Preferably, the process of driving the group of intelligent agents to re-execute the game decision through reinforcement learning mechanism includes: mapping the stress representation vector to the environmental state of the decision system, and constructing a physical search space composed of active displacement source, target force boundary and potential energy barrier field according to the furnace body expansion field, interface force limit and spatial obstacle repulsion; establishing a decision model based on non-cooperative game to drive the furnace body, pipeline compensation and heat exchanger intelligent agents to perform coordinate optimization actions in the physical search space; each intelligent agent takes minimizing the interface force residual and maximizing the compensator absorption efficiency as conflict objectives, and converges the search interval by dynamically adjusting the compensation weight of each node; For each candidate coordinate selected by each agent, the disturbance energy generated on the global stiffness matrix of the pipeline system after the compensation component is introduced at that coordinate is calculated. If the disturbance energy causes the system fundamental frequency shift or the energy distribution entropy value to exceed the preset safety threshold, the deviation is converted into a negative incentive signal and fed back to the decision system. A reward function with global elastic potential energy balance as the core is constructed. The reward function is corrected using the negative incentive signal, and the compensation weight of each node is updated using the gradient descent method. The agent group is driven to re-execute the game action through the redistribution of weights until the system displacement absorption and interface force synchronously enter a steady state, and the globally optimal layout scheme is output.

[0013] Preferably, the process of calculating and outputting the pre-tension offset using creep strain includes: identifying the distribution of compensation weights after iteration, extracting the node numbers of the local extreme points of the weights, mapping the node numbers back to the three-dimensional spatial centerline coordinates of the pipeline topology, and determining the geometric center installation position of the compensator; inputting the stress characterization vector into a preset fatigue life assessment model, and extracting the equivalent stress amplitude and cycle frequency characteristics of the pipeline under full-cycle descaling temperature change; searching in a preset compensator parameter library based on the characteristics, and matching the number of bellows layers and single-layer wall thickness compatible with the stress evolution trajectory; extracting the cumulative irreversible displacement component of the creep strain within a preset service period, calculating the reverse compensation vector according to the direction and magnitude of the displacement component, and defining the reverse compensation vector as the pre-tension offset during the installation stage; associating the installation coordinates, the number of bellows layers, and the pre-tension offset to generate an engineering instruction set containing spatial positioning data, component structural parameters, and construction installation offset guidance values.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By mapping the dynamic response coefficients extracted from the vibration matrix as edge weights to the pipeline network topology graph, deep coupling modeling of mechanical excitation frequency and structural natural frequency in the graph representation space is realized. This allows the sensitivity of different spatial configurations to the excitation of moving equipment to be perceived in the initial modeling stage. This heterogeneous feature embedding logic can effectively identify resonance risk areas that may be caused by changes in pipeline network layout, thereby completing the collaborative prediction of "stress-vibration" multi-physics interference in the early stage of layout search.

[0015] 2. By employing a spatiotemporal graph neural network to deeply aggregate and extract the coking temperature variation sequence, the physical properties of the pipeline network were precisely aligned from a static geometric space to a temporal dynamic feature space. This processing method effectively captures the stress hysteresis characteristics and cumulative loss effects generated during the unique heating, coking, and cooling cycles of the pyrolysis furnace. It provides underlying characterization data with deep mechanical awareness for the accurate selection of subsequent compensators, thereby reducing design deviations caused by model simplification under complex operating conditions.

[0016] 3. By converting the perturbation energy of the global stiffness matrix of the pipeline system into a negative excitation signal in the reinforcement learning mechanism, a physically self-consistent closed-loop decision constraint is constructed. This constraint can monitor in real time the fluctuations in the system's natural frequency and energy distribution caused by multiple agents during the optimization process, effectively avoiding the technical defects of traditional heuristic algorithms that are prone to getting trapped in local optima or causing instability of the system's fundamental frequency when dealing with large-scale nonlinear constraints. This reward and punishment function based on the global elastic potential energy balance forces the agent group to achieve global optimal convergence of displacement absorption efficiency through weight redistribution, under the premise of satisfying the stringent interface force boundaries. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the topology layout optimization method for thermal pipeline compensators based on topology representation learning according to the present invention. Figure 2 This is a schematic diagram of the process for generating time-related stress characterization vectors for each node according to the present invention; Figure 3 This is a schematic diagram illustrating the process of driving a group of intelligent agents to re-execute game decision-making through a reinforcement learning mechanism, as described in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figures 1 to 3 This invention provides a method for optimizing the topology layout of thermal pipeline compensators based on topology representation learning. The technical solution is as follows: Example

[0020] A method for optimizing the topology layout of thermal pipeline compensators based on topology representation learning, specifically as follows: Figure 1 As shown, it includes: Obtain the pipeline topology, coking temperature change sequence, furnace body expansion field, vibration matrix, creep strain, and repulsive force of spatial obstacles; Based on the pipeline network topology, a graph structure containing nodes and edges is constructed. The clearing temperature change sequence is embedded as the node temporal feature, and the vibration matrix is ​​mapped as the edge stiffness attribute to complete the embedding of topological heterogeneous features. A spatiotemporal graph neural network is used to aggregate the heterogeneous feature graph, extract the stress evolution trajectory of the pipeline network under temperature change cycle, and generate the stress characterization vector of each node time-related. A decision-making system is established, consisting of three intelligent agents: furnace body, pipeline compensation, and heat exchanger. The furnace body expansion field, creep strain, and interface stress limit are input into the corresponding intelligent agents as boundary displacement, attenuation parameters, and stress constraints, respectively. The decision-making system calls the stress characterization vector and combines it with the repulsive force of spatial obstacles to perform game-theoretic decision-making, calculates the disturbance of the selected position to the global stiffness matrix of the system, and iteratively adjusts the compensation weights of each node until the displacement absorption and interface stress meet the steady-state threshold. The installation coordinates are determined in three-dimensional space based on the iterated compensation weights, and the number of bellows layers is matched based on the characterization vector. The pre-tension offset is calculated and output using creep strain.

[0021] Furthermore, the process of acquiring the pipeline topology, coking temperature change sequence, furnace body expansion field, vibration matrix, creep strain, and spatial obstacle repulsion includes: extracting the centerline coordinates, pipe diameter, wall thickness parameters, and topological connection relationships of the pipelines in the pyrolysis furnace area; combining the material linear expansion coefficient to generate the pipeline topology; extracting the temperature sampling time sequence including the heating section, the isothermal decoking section, and the cooling section from the production control system to form a coking temperature change sequence reflecting the thermal shock cycle of the pipeline; acquiring non-uniform axial displacement vector data at coordinate points at different heights of the furnace wall, and using the displacement vector data as the dynamic displacement boundary at the beginning of the pipeline. The conditions are as follows: the expansion field of the furnace body is obtained; the rotational speed and vibration amplitude of the moving equipment are collected and the mechanical excitation frequency is calculated. Combined with the natural frequency of the pipeline network, a numerical matrix characterizing the coupling strength between excitation and resonance is constructed as the vibration matrix; based on the high-temperature creep rate curve of the material and the cumulative operating hours of the pipeline, the irreversible permanent deformation within the preset service period is calculated as the creep strain; the existing infrastructure coordinate boundary in the pyrolysis furnace area is identified and the minimum physical distance of the pipeline sampling point relative to the coordinate boundary is calculated. The minimum physical distance is mapped to a potential energy weight with a spatial gradient to construct the spatial obstacle repulsion force.

[0022] The system first connects to the plant's physical information model (such as PDMS or a standard 3D design database), extracts the underlying geometric description file of the target pipeline, and after parsing the file, automatically identifies the 3D coordinate nodes of the centerline of the quench oil outlet pipeline, and simultaneously reads the nominal diameter (e.g., 600 mm), wall thickness (e.g., 10 mm), and pipe fitting connection relationships of the pipe segment. Subsequently, the system matches the 304L stainless steel material parameters used in the pipe segment from the material property library, focusing on extracting its coefficient of linear expansion in the 0-900 degree Celsius range. Finally, the system constructs a graph model using the centerline 3D coordinate nodes as node set V and the pipe segment connection relationships as edge set E; it extracts the cross-sectional moment of inertia and elastic modulus of each pipe segment to construct the node compliance matrix, and maps it together with the coefficient of linear expansion to form the initial feature vector of the node, transforming the connection relationships into an N×N dimension adjacency matrix, thereby forming a digital pipeline topology with embedded physical attributes.

[0023] By accessing the real-time data interface of the factory's distributed control system, historical operating records of the pyrolysis furnace for nearly 12 months were retrieved. The trigger point for the decoking operation was automatically located using slope detection, and a complete decoking cycle segment was captured. This segment was finely divided into three physical time periods: the first period was the heating period, recording the temperature change trajectory from 450 degrees Celsius to 600 degrees Celsius; the second period was the isothermal decoking period, recording the stable state with fluctuations of around 10 degrees Celsius around 600 degrees Celsius, typically lasting 12 to 24 hours; the third period was the cooling period, recording the cooling process after decoking to the rated operating temperature of 850 degrees Celsius. By sampling these segments at equal intervals with a 5-minute cycle, a decoking temperature change sequence reflecting heat load fluctuations was generated.

[0024] Considering the macroscopic deformation of the pyrolysis furnace body as a giant high-temperature vessel, the system imports a pre-generated finite element analysis model of the furnace body and reads the thermal displacement vector data of the outer side of the furnace wall under rated load, paying particular attention to the vertical height coordinates of the pipeline connection flanges. Through spatial linear interpolation calculations, the non-uniform axial displacement values ​​of the flange connection points varying with height are obtained. For example, the displacement at a height of 12 meters is 15 millimeters, while the displacement at 15 meters is 18 millimeters. This set of displacement vectors varying with spatial position is directly defined as the forced displacement constraint at the starting end of the pipe gallery, i.e., the furnace body expansion field.

[0025] Vibration monitoring sensors installed on the bearing housings of the quench pump and induced draft fan collect the operating frequency (e.g., 25 Hz or 50 Hz) and vibration amplitude of the dynamic equipment in real time. Simultaneously, the system calculates the natural frequency of the pipeline structure under the current support span. The system uses the absolute value of the difference between the excitation frequency and the natural frequency as the reciprocal to calculate the coupling strength between the two. The closer the two are, the higher the value of the matrix component corresponding to that pipe segment, indicating a greater risk of resonance. The coupling strength values ​​of each pipe segment along the entire line are filled into the adjacency matrix to generate a vibration matrix characterizing the dynamic sensitivity. Specifically, a preset small positive smoothing factor is introduced during the coupling strength calculation to avoid numerical overflow. The calculated coupling strength of each pipe segment is normalized using Min-Max and then used as weights in the edge feature matrix of the graph structure to characterize the response gain of the spatial layout to dynamic excitation.

[0026] For pipe sections operating in high-temperature zones above 500 degrees Celsius for extended periods, a time-cumulative material deformation correction is introduced. This involves referencing the experimental curve of the high-temperature creep rate of the pipe section's material and combining it with the cumulative operating time of the furnace unit recorded at the factory (e.g., 20,000 hours). The creep rate over this operating period is numerically integrated to calculate the irreversible permanent deformation caused by the high-temperature load. This deformation, typically measured in millimeters, is used as creep strain to correct the zero-point position of the compensator's initial layout, ensuring that the compensator remains in an effective operating range after long-term operation. The creep strain is then converted into a preset displacement vector for the corresponding node. During the layout optimization phase, this preset displacement vector is used to correct the zero-point position of the compensator's initial layout. The initial coordinates of the free end of the compensator are superimposed to offset the center point of the compensator's full-cycle working envelope, thereby ensuring that the thermal elongation of the compensator in the later stages of service remains within the rated effective displacement range of the bellows. The specific generation process of the preset displacement vector is as follows: obtain the creep strain of each pipe segment node, calculate the cumulative permanent deformation of each candidate installation point relative to the fixed support by accumulating the node displacements; define the cumulative permanent deformation as the modulus of the preset displacement vector, and specify its direction as the direction away from the fixed support along the pipeline axis; before the layout optimization action is executed, automatically superimpose the preset displacement vector onto the initial geometric coordinates of the compensator to form a pre-offset installation command.

[0027] Retrieve 3D scan point cloud of the plant area or 3D blueprint of the steel structure, and use a geometric envelope algorithm to identify the spatial boundary coordinates of existing infrastructure, maintenance platforms and support beams. Calculate the minimum Euclidean distance from discrete sampling points on the pipeline centerline to the above boundary coordinates, preset a 500 mm avoidance threshold, and convert the physical distance into weights through a nonlinear mapping function: when the distance is greater than the threshold, the weight value is 0; when the distance decreases, the weight value increases sharply inversely; when the distance is less than the 200 mm critical value required for installation, the weight is assigned a maximum value to simulate a "space no-go zone". This set of weights based on distance constitutes the spatial obstacle repulsion force guiding the algorithm to avoid obstacles.

[0028] Furthermore, the process of embedding the heterogeneous topological features includes: mapping the pipe segment intersections, furnace connection points, and heat exchanger inlet and outlet points in the pipe network topology to a set of nodes in a graph structure, and mapping the physical pipe segments connecting each node to a set of edges in a graph structure; performing sliding window sampling on the coking temperature change sequence to generate temperature state vectors corresponding to each node at different time steps, and assigning the temperature state vectors as time-series feature attributes to the corresponding nodes in the node set; extracting the excitation frequency components of each edge in the vibration matrix, calculating the dynamic response disturbance coefficient of the excitation frequency components to the inherent stiffness of the pipe segment, and mapping the dynamic response disturbance coefficient to the weight attributes of each edge in the edge set, so as to realize the feature fusion of the graph structure in the spatial topology and dynamic load dimensions.

[0029] First, the acquired digital file of the pipeline topology is parsed, and node and edge extraction operations are performed. Geometric key points in the pipeline are automatically identified, and the three-way junction of the quench oil outlet pipeline, the furnace outlet flange connection point, and the inlet and outlet flange points of the quench heat exchanger are defined as node sets in the graph structure. Each node is assigned a unique numerical index, and its 3D spatial coordinates in the plant coordinate system are recorded. Then, the physical pipe segments connecting these nodes are defined as edge sets. For example, if the furnace outlet flange is defined as node 1, and the other end of the first straight pipe segment it connects to is node 2, then this pipe segment is mapped as an edge connecting node 1 and node 2. Finally, an adjacency matrix describing the global connectivity of the pipeline network is generated, completing the initial transformation from physical geometry to mathematical topology.

[0030] To enable static graph nodes to sense dynamic temperature fluctuations, a sliding window sampling is performed on the coking temperature change sequence carried by each node. A sliding window with a length of 60 minutes is set, and the window slides on the time-series curve of the coking temperature change sequence with a step size of 5 minutes. Within each sliding window, 12 equally spaced temperature sampling points are extracted and transformed into a 12-dimensional numerical vector, which is defined as the temperature state vector of that node. Subsequently, this vector is used as a multi-dimensional feature attribute of the node and assigned to the corresponding node in the node set. In this way, each topological node carries dynamic heat load characteristics that reflect the changes in its location with the coking cycle.

[0031] The physical properties in the edge set are modified using the vibration matrix to reflect the influence of dynamic loads on structural stability. First, the main excitation frequency data of each physical pipe segment (i.e., each edge) is extracted from the vibration matrix. Then, the natural frequency of each pipe segment under the current constraints is called, and the ratio of the excitation frequency to the natural frequency is calculated. A mapping logic is set: when the ratio is in the resonance sensitive range of 0.8 to 1.2, a dynamic response disturbance coefficient greater than 1.2 is calculated; when the ratio is far from the resonance range, the coefficient tends to 1.0. This coefficient reflects the severity of the change in the equivalent stiffness of the pipe section caused by the vibration of the moving equipment. Finally, the calculated dynamic response disturbance coefficients of each pipe section are mapped to the weight attributes of the corresponding edges in the graph structure. Specifically, the calculation of the dynamic response disturbance coefficient adopts a piecewise mapping logic: First, 1.5 is set as the maximum disturbance gain value. When the frequency ratio is 1, the coefficient is directly assigned the maximum value of 1.5. When the frequency ratio is between 1 and 1.2 or between 0.8 and 1, linear attenuation calculation is performed according to the degree of offset from the center point 1, so that the ratio is closer to 1.0 the closer it is to the interval boundary. When the ratio exceeds the interval of 0.8 to 1.2, the coefficient is fixed at 1.0. In this way, the continuous frequency ratio is transformed into a numerical weight with physical penalty significance.

[0032] After independently calculating node features and edge weights, the final encapsulation of heterogeneous features is performed. This involves structurally aligning the node tensor containing temperature time-series features with the edge weight matrix containing dynamic response perturbation coefficients. Specifically, a two-dimensional node feature matrix is ​​created, where each row represents a pipeline node, and the 12 columns in each row correspond to the temperature features sampled by the sliding window. Simultaneously, an adjacency matrix corresponding to the pipeline topology is created. In this adjacency matrix, the Boolean value '1' representing connection relationships is replaced with the calculated dynamic response perturbation coefficients, thus constructing a heterogeneous feature space. This allows for subsequent information processing along the pipeline topology path. During transmission, the thermal stress transfer between nodes is not only affected by the temperature gradient, but also modulated by the edge weights (i.e., vibration disturbance coefficients). The modulation process is realized through the message passing mechanism of the graph neural network: when performing feature updates, for each target node, all neighboring nodes connected to it are automatically retrieved, and the temperature feature vector of each neighboring node is multiplied with the dynamic response disturbance coefficient connecting the path. This makes the temperature change features carried by the high vibration risk pipe section amplified proportionally when transmitted to the target node. Finally, all modulated features are weighted and summed to generate a comprehensive representation vector containing three information: spatial topology, temperature change time series and vibration disturbance.

[0033] By deeply coupling the pipeline topology with dynamic temperature and vibration load in the graph representation space, a unified digital mapping of multi-source heterogeneous data is achieved, accurately representing the nonlinear correlation between spatial configuration and variable operating conditions, improving the ability to capture the thermal stress characteristics of the resonance region, providing feature input with physical perception depth for subsequent layout optimization, and effectively avoiding design deviations caused by the isolation of multi-physics field data.

[0034] Further, the process of generating time-related stress characterization vectors for each node includes mapping physical nodes and pipe segments in the pipeline network topology to vertices and edges of a graph structure, processing the descaling temperature change sequence into dynamic temperature feature vectors for each node using a sliding window method, and extracting the dynamic response coefficients of each pipe segment as edge weights using the vibration matrix to align physical attributes to the graph feature space; using a graph convolution operator to transfer features along the pipeline network topology path, and using the dynamic response coefficients on the edges to weight and aggregate the temperature features of adjacent nodes to extract the physical transmission characteristics of thermal stress in the pipeline network under a specific spatial configuration; inputting the spatially aggregated feature stream into a time-series processing unit to capture the stress response differences of the pipeline network at different stages of heating, isothermal descaling, and cooling, and extracting the stress accumulation effect and hysteresis features that change over time; and fusing spatial topology features and time-series features through an attention mechanism to compress the complex stress fluctuations within the entire operating cycle into a high-dimensional hidden state sequence to generate stress characterization vectors. The specific process is as follows: Figure 2 As shown.

[0035] After feature embedding is completed, the feature map containing node temperature feature vectors and edge weight attributes is input into a preset graph convolution operator to perform multi-layer spatial convolution operations. In each layer of calculation, the target node actively aggregates the feature information of its directly adjacent nodes through topological connections. In specific implementation, instead of simply averaging the temperature vectors of adjacent nodes, the dynamic response coefficients corresponding to the adjacency matrix are used as the transmission gain. If the dynamic response coefficient of a certain pipe segment is high, it means that the path is more prone to fatigue coupling under vibration load. At this time, the temperature features transmitted by the path will be amplified proportionally. Through 3 to 5 layers of convolution stacking, the feature vector of each node not only contains local thermodynamic information, but also integrates the spatial transmission characteristics of thermal stress of the entire pipeline topology under a specific geometric configuration. Specifically, the aggregation process adopts a weighted summation operator. The updated value of the node feature is equal to the sum of the products of the feature vectors of its neighboring nodes and the corresponding edge dynamic response coefficients. After each layer of convolution, nonlinear mapping is performed through a linear rectifier unit, and layer normalization is performed to ensure that the stability of the numerical space is maintained while amplifying the features of high resonance paths.

[0036] The spatially aggregated feature streams enter the time-series processing unit sequentially. To accurately identify the mechanical differences at different stages under the decoking condition, the sampling sequence is divided into three logical intervals: heating, isothermal decoking, and cooling. The time-series processing unit uses an internal gating mechanism to perform differential calculations on the feature vectors of different stages to identify the instantaneous impact of the temperature change rate on the pipe stress. Specifically, the differential calculation refers to calculating the difference between the temperature feature vectors of adjacent time steps and using this difference vector as an auxiliary input, concatenating it with the original feature vector, and inputting it into the gating loop unit. The temperature change rate is quantified by calculating the magnitude of the difference vector, thereby dynamically adjusting the activation level of the internal update gate. During drastic changes such as heating and cooling, the update weight of the gating loop unit is automatically increased to capture the stress response lag caused by the thermal inertia of the material. In the isothermal decoking section, the focus is on extracting the cumulative stable features of stress under long-term high-temperature loads. This path ensures that the model can perceive the stress evolution trend of the physical pipe network under the entire cycle of operation due to the accumulation of time.

[0037] The self-attention operator is used to perform deep fusion on the feature matrix after spatial aggregation and temporal capture to calculate an importance score for each time step in the entire working cycle. For specific moments with drastic temperature fluctuations or tight spatial constraints, the attention mechanism will assign them extremely high weight scores. In this process, the static spatial obstacle repulsion data is mapped to spatial attention weights through a fully connected layer and then added element-wise to the attention score matrix generated by the temporal features. This means that when the temperature change trajectory at a certain moment is coupled with the spatially highly sensitive region in the feature space, the comprehensive attention weight at that moment will gain a non-linear gain. Subsequently, the feature flow is weighted and summed according to these weight scores, compressing the originally lengthy full-cycle dynamic features into a digital sequence of fixed length (128 dimensions in this embodiment), which is the final stress characterization vector.

[0038] The specific architecture of the spatiotemporal graph neural network was selected as ST-GCN (Spatiotemporal Graph Convolutional Network), which consists of three graph convolutional layers and two temporal processing layers stacked together. In the initial stage of data input, in order to match the network dimension and extract deep features, a fully connected layer is first used to map the original 12-dimensional node temperature feature vector to a higher-dimensional space of 64 dimensions. The graph convolutional layer uses an operator based on spectral domain convolution. By setting the convolution kernel order, the network can effectively capture the stress transmission relationship across nodes in the pipeline topology and gradually expand the spatial feature dimension to 128 dimensions.

[0039] In the specific feature transfer process, each graph convolution layer weights and aggregates the temperature features of adjacent nodes according to the dynamic response perturbation coefficients defined in the adjacency matrix. This design ensures that in vibration risk areas with high frequency coupling, the transfer weight of thermal stress is automatically amplified, thus reflecting the interference between physical fields. After completing the spatial dimension feature aggregation, the data stream enters the temporal processing unit composed of a bidirectional long short-term memory network in time step order. This unit contains two hidden layers. The forward network is specifically responsible for capturing the stress accumulation effect of the pipeline from the heating to the isothermal decoking stage, while the backward network is responsible for capturing the stress release and hysteresis process during the cooling stage. The features of the two are concatenated to form a complete temporal feature description.

[0040] To further extract core stress features from massive amounts of data, a multi-head self-attention mechanism was introduced, configured with eight attention heads. This mechanism concatenates and fuses the temporal encoding output by BiLSTM with the spatial topological features extracted through graph convolution. By calculating global dependency weights, it automatically identifies the key stress points that have the greatest impact on system strength throughout the entire operating cycle. Through transformation and weighted summation of the Query, Key, and Value matrices, the system compresses the complex full-cycle stress fluctuation information into a fixed 128-dimensional high-dimensional hidden state sequence, which is the final generated stress representation vector.

[0041] The training process of this model adopts a supervised learning approach. The labeled data is mainly generated in advance by finite element simulation software. For different pipeline network topology samples, corresponding clearing temperature change sequences and physical constraints are applied in the simulation software. The full-time stress response curves of each node calculated by the software are used as true value labels to guide the learning mapping of the neural network. During the training phase, the mean square error is used as the loss function to measure the deviation between the model output and the simulation true value. The Adam optimizer is used to iterate the parameters with an initial learning rate of 0.001 until the model loss on the validation set reaches a stable convergence state.

[0042] By integrating spatiotemporal convolution aggregation and multi-dimensional attention compression mechanisms, nonlinear fusion of the complex topological heterogeneity of the pipeline network and the dynamic load time series was achieved. The stress hysteresis effect and vibration stiffness disturbance generated by the temperature rise and fall impact during the coking cycle of the pyrolysis furnace were captured. The lengthy physical evolution process was highly condensed into a standardized stress characterization vector, providing an evaluation benchmark with mechanical logic depth for the subsequent layout game of multi-agent systems, and improving the overall steady-state reliability of the optimization scheme.

[0043] Furthermore, the process of establishing a decision-making system composed of three intelligent agents—furnace body, pipeline compensation, and heat exchanger—includes: establishing independent physical intelligent agents representing the furnace body, pipeline compensation, and heat exchanger respectively; converting the furnace body expansion field into the active displacement output of the furnace body intelligent agent, converting creep strain into the long-term performance degradation coefficient of the pipeline compensation intelligent agent, and using the interface force limit as the mandatory force constraint boundary of the heat exchanger intelligent agent; the pipeline compensation intelligent agent calling the stress characterization vector, guided by the potential energy constructed by the repulsive force of spatial obstacles, searching for a candidate set of compensator installation coordinates; and each intelligent agent, according to its respective constraint conditions... The system engages in collaborative game theory among candidate nodes, iteratively evolving with the common goal of minimizing global system stress and achieving interface stress compliance. For each agent's collaboratively selected compensator installation location, the system calculates in real time the disturbance to the global stiffness matrix of the pipeline system caused by the intervention at that location. By analyzing the impact of the disturbance on the system's vibration frequency and stress redistribution, a stiffness feedback command is generated. Based on the stiffness feedback command, the system dynamically adjusts the compensation weights of each candidate node and drives the agent group to re-execute the game decision through a reinforcement learning mechanism until the displacement absorption of the pipeline system and the stress at the heat exchanger interface both meet the preset steady-state threshold.

[0044] First, three logical entities with independent decision-making power and physical attribute definitions are initialized in the decision space. For the agent representing the furnace body, the pre-generated furnace body expansion field data is mapped to its active displacement output command to simulate the forced thrust of the furnace body on the starting end of the pipe gallery when heated at high temperature. For the agent representing pipeline compensation, the calculated creep strain value is converted into a long-term performance degradation coefficient through a linear mapping relationship. This coefficient is used to attenuate and correct the effective stroke of the compensator during the decision-making process, simulating the degradation of the compensation capability in the later stages of equipment service. Specifically, the linear mapping relationship is set as follows: the attenuation coefficient equals 1 minus the ratio of creep strain to the preset material limit creep value. When calculating the compensator selection, the pipeline compensation agent uses the attenuation coefficient as a multiplication factor on the rated effective displacement of the bellows, thereby forcibly compressing its selectable displacement compensation range in the early stage of decision-making. For the agent representing the heat exchanger, the allowable stress and allowable moment values ​​of its physical interface are set as mandatory force constraint boundaries, serving as the bottom line judgment conditions in the game process.

[0045] The pipeline compensation agent invokes the stress characterization vector generated in the preceding steps in real time. This vector provides a sensitivity heatmap of the global stress distribution. The agent determines the initial site selection priority of the compensator based on the high-amplitude regions in the stress vector. In this process, a digital potential energy field constructed by the repulsive force of spatial obstacles is introduced as a movement guide: when the agent probes a certain installation coordinate, it automatically reads the potential energy weight value of that coordinate point; if the weight value is greater than 0, the agent will perform obstacle avoidance action according to the potential energy gradient direction, that is, shift to the direction with lower potential energy (larger space clearance). Through this path, the pipeline compensation agent initially selects a set of candidate compensator installation coordinates that take into account both mechanical sensitivity and spatial feasibility.

[0046] Each agent performs a cooperative game against the generated candidate set. At this stage, each agent has a different priority: the furnace agent requires the pipeline to absorb the maximum active displacement; the heat exchanger agent requires its interface stress residual to always remain within a preset deviation range of 0.03%; and the pipeline compensation agent pursues the homogenization of global stress distribution. The game system employs a model combining non-cooperative game theory and cooperative optimization, driving the three parties to exchange parameters within the candidate coordinate space. By calculating the system's global elastic potential energy under each layout scheme, the candidate interval is iteratively narrowed. When a certain position... When a punctuation point simultaneously satisfies the conditions of the heat exchanger not exceeding the limit and the furnace displacement being effectively absorbed, that point is selected as the equilibrium point of the game in the current round. The game process is implemented using the alternating direction multiplier method framework: In each iteration, the coordinates of the furnace and the heat exchanger are first fixed, and the pipeline compensation agent solves for the temporary optimal point for the stress minimization objective; then, the point is fixed, the heat exchanger agent verifies the interface force vector, and the point coordinates are corrected by a penalty term through Lagrange multipliers, until the objective function deviation of the three agents converges in the process of minimizing the global elastic potential energy.

[0047] For the selected coordinate point in the game, the system calls the finite element kernel to calculate in real time the disturbance to the global stiffness matrix of the pipeline system after the compensator is inserted at that location. By extracting the changes in the eigenvalues ​​of the stiffness matrix, the system analyzes whether the disturbance causes the system's natural frequency to shift towards the excitation frequency of the moving equipment, and whether it causes a secondary concentration of stress in the local area. Based on the analysis results, the system generates corresponding stiffness feedback instructions: if the disturbance causes the fundamental frequency shift to exceed the safety threshold, negative feedback is generated, and the compensation weight of the candidate node is reduced through the reinforcement learning mechanism. The reinforcement learning model updates the policy weights of each node according to the reward function (composed of the degree of stress minimization and the force achievement rate), driving the group of agents to re-execute the game cycle.

[0048] The above game and feedback logic is repeated. In each iteration, the installation position is optimized by redistributing weights. As the number of reinforcement learning training steps increases, the compensation weights of each node gradually converge to the optimal solution. When the system determines that the overall displacement absorption of the pipeline network reaches more than 99% of the active displacement of the furnace body, and the stress on the heat exchanger interface is stable within the steady-state threshold range, the iteration process terminates. Finally, the optimal installation coordinates after the game of multiple physical quantities, the optimal compensation capacity allocation scheme under the coordinates, and the corresponding stiffness matrix distribution data are output, thus completing the global precise balance of the pipeline network compensator layout.

[0049] This embodiment provides a clear logical definition of the game model of the three-party intelligent agents. In this system, the strategy of the furnace body intelligent agent is directly driven by the displacement data generated by the furnace body expansion field; the strategy of the pipeline compensation intelligent agent is manifested in the allocation of compensation weights for each node along the entire line, and its goal is to find the optimal compensator installation position; the heat exchanger intelligent agent is responsible for maintaining the safety of the interface stress and setting the allowable stress of the material as an insurmountable boundary constraint.

[0050] The core of the game lies in the mutual checks and balances of the payoff functions of the three parties: the furnace agent aims to maximize displacement absorption, i.e., minimize the displacement residuals not absorbed by the compensator; the pipeline compensation agent aims to equalize the global stress of the system, reducing the overall stress level by favoring high-stress nodes; while the heat exchanger agent's payoff depends on whether the force vector at the interface is within the limit, and exceeding the limit will generate a strong negative signal. This conflict of interest prompts the three parties to continuously adjust the node weights in the physical search space, seeking a Nash equilibrium point that can take into account the interests of all parties.

[0051] The specific iterative solution employs an alternating direction optimization logic. In each round of the game, the pipeline compensation agent first attempts a preliminary set of weights based on the current stress distribution. Subsequently, the system calls the finite element calculation engine to simulate the mechanical state under this layout and feeds the results back to the furnace and heat exchanger agents. If the stress on the heat exchanger interface exceeds the limit or the displacement absorption is insufficient, each agent will fine-tune the weight parameters of each node using the gradient descent method. This process will continue to cycle until the preset steady-state threshold is met, i.e., the displacement absorption rate reaches more than 99%, and the safety factor of the heat exchanger interface stress meets the margin requirement of more than 1.5 times.

[0052] After the game finally converges, the system extracts the maximum value from the weight allocation vector of each node. Since only nodes that can effectively alleviate system stress and do not violate spatial constraints will accumulate higher weights during the iteration process, the optimal installation coordinates of the compensator can be accurately located in three-dimensional space by identifying the index of the node with the highest weight.

[0053] By constructing a collaborative game mechanism among the three intelligent entities of the furnace body, pipelines, and heat exchangers, highly coupled physical constraints are transformed into logically autonomous decision-making objectives. Utilizing closed-loop drive through reinforcement learning and stiffness feedback, the global stress of the system can be dynamically distributed while meeting the rigidity limit of the heat exchanger interface. This avoids the dilemma of "paying attention to one thing but losing sight of another" in local optimization in traditional design and ensures the structural stability of the design scheme under long-term service and complex operating conditions. This provides a technical guarantee with deep logical support for the inherent safety of pipe corridors in large industrial plants.

[0054] Furthermore, the process of driving the agent group to re-execute the game decision through reinforcement learning includes: mapping the stress representation vector to the environmental state of the decision system, and constructing a physical search space composed of active displacement sources, target force boundaries, and potential energy barrier fields based on the furnace expansion field, interface force limits, and spatial obstacle repulsion; establishing a decision model based on non-cooperative game theory to drive the furnace, pipeline compensation, and heat exchanger agents to perform coordinate optimization actions within the physical search space; each agent uses minimizing the interface force residual and maximizing the compensator absorption efficiency as conflict objectives, and converges the search interval by dynamically adjusting the compensation weights of each node; for For each candidate coordinate selected by each agent, the disturbance energy generated on the global stiffness matrix of the pipeline system after the compensation component is introduced at that coordinate is calculated. If the disturbance energy causes the system's fundamental frequency shift or energy distribution entropy to exceed a preset safety threshold, the deviation is converted into a negative incentive signal and fed back to the decision system. A reward function based on the global elastic potential energy balance is constructed, and the reward function is corrected using the negative incentive signal. The compensation weights of each node are updated using the gradient descent method. The agent group is driven to re-execute the game action through the redistribution of weights until the system displacement absorption and interface force synchronously enter a steady state, and the globally optimal layout scheme is output. The specific process is as follows: Figure 3 As shown.

[0055] First, a mapping from physical parameters to algorithm features is performed. The stress characterization vector, which contains full-condition information and is generated in the previous steps, is directly defined as the environmental state perceived by the reinforcement learning agent. Simultaneously, by integrating the active displacement vector of the furnace expansion field, the allowable force value of the heat exchanger interface, and the potential energy distribution of the repulsive force of spatial obstacles, a digital search space with physical boundaries is constructed. In this space, the furnace expansion field is set as the active displacement source driving the deformation of the pipeline network, the interface limit is set as the force boundary that must be forcibly satisfied, and the repulsive potential energy field serves as an invisible barrier to guide the agent to avoid interference. Together, they form a three-dimensional decision-making environment with multi-dimensional physical constraints.

[0056] Within the constructed search space, a non-cooperative game model is initiated, involving the furnace, pipeline compensation, and heat exchanger. Each agent performs coordinate optimization actions driven by the environmental state: the pipeline compensation agent, as the primary executor, attempts to alter the stiffness distribution of the pipeline network by tentatively moving the compensator's installation coordinates within the search space. During this process, each agent has clearly conflicting objectives: the heat exchanger agent strictly pursues minimizing the interface stress residual, while the pipeline compensation agent aims to maximize its displacement absorption efficiency. Through this game mechanism, the three agents mutually restrain each other through continuous action interactions, gradually converging the search interval towards a mechanically feasible intersection region. In the specific game process, a centralized training distributed execution framework is adopted, using a global commentator network to simultaneously evaluate the action value of the three agents. A Pareto front search mechanism is introduced to force the search for the Pareto optimal solution among multiple conflicting objectives, ensuring that the interface stress does not exceed the limit and that the compensation efficiency is the highest.

[0057] For each candidate coordinate selected in the game, the underlying mechanics kernel is invoked to perform real-time evaluation. By intervening a compensation component at the current layout node, the global stiffness matrix of the pipeline system is reconstructed. Subsequently, the difference in matrix eigenvalues ​​before and after the intervention is calculated to obtain the disturbance energy caused by the layout change. The impact of the disturbance energy on the system's fundamental frequency is monitored, and an energy distribution entropy value is introduced to quantify the degree of stress disorder within the pipeline network. Specifically, the calculation logic of the energy distribution entropy value is as follows: First, the pipeline model is discretized into several independent calculation units, and the elastic potential energy value carried by each calculation unit is obtained in real time. Then, the system calculates the percentage weight of the elastic potential energy of each calculation unit to the total elastic potential energy of the system. Subsequently, information entropy quantification is performed: the percentage weight of each unit is multiplied by its own logarithm, and the product results of all calculated units are negatively accumulated to synthesize an entropy value characterizing the global energy evolution state. This entropy value can intuitively quantify the uniformity of the pipeline stress distribution: when the calculated entropy value is high, it means that the energy proportion of each unit tends to be consistent, that is, the overall stress distribution of the pipeline tends to be balanced and stable; conversely, if the entropy value is low, it means that the system energy is highly concentrated in a few specific units, that is, it is determined that there is a significant risk of local stress concentration in the pipeline. If the calculated offset of the system fundamental frequency exceeds the preset safe frequency range (e.g., offset exceeds 2 Hz), or the entropy value of the stress distribution exceeds the preset stability threshold, it is determined that the current candidate coordinates have a risk of causing system instability or local resonance.

[0058] The physical evaluation results are fed back into the reinforcement learning reward mechanism to construct a reward function centered on global elastic potential energy equilibrium. This function evaluates the merits of different layout schemes. If a fundamental frequency shift or entropy exceeding the limit is detected in the preceding steps, a corresponding negative incentive signal is generated and applied as a penalty to the reward function. Specifically, when candidate coordinates lead to a deterioration in mechanical performance, the score of the reward function will drop significantly. Through this feedback path, complex changes in physical stiffness are transformed into a "good or bad evaluation" that the agent can recognize, thereby forcibly driving the game process towards a direction with global mechanical stability. The generation path of the negative incentive signal is as follows: a penalty function based on safety deviation is established. When the fundamental frequency shift exceeds a 2 Hz threshold, the penalty term increases exponentially with the deviation value and deducts from the reward function in a range of -10 to -100, thereby rapidly suppressing high-risk actions during policy gradient updates.

[0059] To enable the reinforcement learning agent to perceive the mechanical environment of the pipeline network, this embodiment defines the environmental state of the decision-making system as a multi-dimensional feature vector. This vector is composed of five key parts: first, the 128-dimensional stress representation vector generated in the previous steps, which provides the depth features of the global stress distribution; second, the compensation weight distribution of each node in the pipeline network, reflecting the current layout state; third, the first 5 eigenvalues ​​of the global stiffness matrix, which characterize the natural frequency characteristics of the system; and finally, the furnace expansion field vector and the heat exchanger interface force vector are added to reflect the real-time state of external boundary displacement and force constraints.

[0060] The agent's actions are defined as incremental fine-tuning of the compensation weights of each node in the pipeline network. Specifically, in each training step, the agent outputs an adjustment vector consistent with the total number of nodes, where each element represents the increase or decrease in the weight of the corresponding node. To ensure system stability, the adjustment range of the weights is strictly limited to a small continuous interval (e.g., between -0.1 and 0.1), and after each adjustment, the system performs normalization to ensure that the sum of the weights of all nodes remains constant.

[0061] To guide the agent towards the global optimum, this embodiment constructs a reward function centered on mechanical steady-state dynamics. This reward function consists of a weighted sum of four indices: the first is a stress distribution reward, which encourages a more uniform stress distribution by calculating the negative root mean square of the global stress; the second is a displacement absorption reward, used to measure the adequacy of the furnace expansion displacement absorbed by the compensator; the third is a stress safety penalty, where the reward score drops significantly if the stress at the heat exchanger interface exceeds the allowable limit; and the fourth is a stiffness stability reward, the aforementioned negative excitation signal, used to suppress any actions that could lead to system fundamental frequency instability or resonance risk. By adjusting the weight coefficients of each index, the system forces the agent to maximize displacement absorption efficiency while satisfying the hard boundary of interface stress.

[0062] At the algorithmic level, the system employs a proximal policy optimization algorithm to train the decision network. This algorithm, through a policy pruning mechanism, effectively prevents drastic oscillations in the mechanical scheme caused by excessive parameter updates during training. Within the physics search space, the agent gradually learns how to balance complex nonlinear mechanical constraints through thousands of rounds of game sampling and weight updates. As training converges, the system can automatically output a set of compensator layout schemes that achieve the most balanced global elastic potential energy and the safest interface stress.

[0063] The construction of the global stiffness matrix involves first discretizing the pipeline network topology into several spatial beam elements, and then using the finite element method to construct the basic stiffness matrix. During modeling, each pipe segment is treated as a Timoshenko beam element with six degrees of freedom (three translations and three rotations), and its mechanical properties are determined by the material's elastic modulus, shear modulus, and the moment of inertia of the cross-section, which is determined by the pipe diameter and wall thickness. By assembling the local stiffness matrices of all pipe segments according to their topological connections, a global stiffness matrix describing the entire pipeline network's mechanical framework is formed.

[0064] When the agent determines to intervene in the compensator at a certain node, it mathematically represents the compensator as an equivalent six-degree-of-freedom compliance matrix with specific compliance characteristics. This compliance matrix precisely defines the stiffness contribution of the bellows in axial tension / compression, lateral shear, and rotational directions. The system directly superimposes the equivalent stiffness submatrix of the compensator into the row and column sub-blocks of the corresponding node in the global stiffness matrix, thereby correcting the physical stiffness changes caused by layout variations in real time.

[0065] To achieve real-time perception during the decision-making process, the disturbance is defined as the eigenvalue shift of the stiffness matrix before and after correction, i.e., the change in the system's natural frequencies. Considering the enormous computational cost of matrix decomposition for large-scale pipeline networks, this embodiment employs the Lanczos iterative algorithm to quickly extract only the top 10 natural frequencies that have the greatest impact on system stability. By monitoring changes in the fundamental frequency, the system can instantly determine whether the current layout will cause resonance risks between the pipeline network and moving equipment. If the detected fundamental frequency shift exceeds the safety threshold of 2Hz, the system will convert this deviation into a negative excitation signal and force it back to the reinforcement learning loop, driving the layout scheme to converge towards a more stable direction.

[0066] After obtaining the corrected reward score, the decision-making system uses gradient descent to update the node compensation weights. Specifically, the node compensation weights are defined as the probability distribution parameters for installing compensation components on each candidate node. The gradient direction of the global reward function is calculated, and the installation probability of each node is adjusted using the Softmax activation function, so that the selection probability of high-reward (low elastic potential energy, low interface stress) nodes continuously increases in the iteration. The gradient direction of the reward score relative to the installation weight of each node is calculated, and the installation priority weight of each node is fine-tuned according to a preset step size. Through this weight redistribution process, the system forces the intelligent agent group to avoid high-risk areas in the next iteration and instead explore new coordinates with more balanced elastic potential energy distribution and safer interface stress. The system repeats the above game and feedback process until the displacement absorption of the pipeline network and the stress on the heat exchanger both enter the preset steady-state range. At this time, the system outputs the globally optimal layout scheme that has been dynamically verified under all operating conditions.

[0067] By introducing the perturbation energy of the global stiffness matrix of the pipeline network into the feedback loop of reinforcement learning, a deep transformation from physical constraints to algorithmic reward mechanisms is achieved. The game direction of the intelligent agent group is forcibly constrained by "negative incentives". This weight update path based on elastic potential energy balance ensures that the automatically generated layout scheme has extremely high global mechanical steady-state reliability while meeting the interface force limit.

[0068] Further, the process of calculating and outputting the pre-tension offset using creep strain includes: identifying the distribution of compensation weights after iteration, extracting the node numbers where the local extreme points of the weights are located, mapping the node numbers back to the three-dimensional spatial centerline coordinates of the pipeline network topology, and determining the geometric center installation position of the compensator; inputting the stress characterization vector into a preset fatigue life assessment model, and extracting the equivalent stress amplitude and cycle frequency characteristics of the pipeline network under full-cycle descaling temperature change; searching in a preset compensator parameter library based on the characteristics, and matching the number of bellows layers and single-layer wall thickness compatible with the stress evolution trajectory; extracting the cumulative irreversible displacement component of the creep strain within a preset service period, calculating the reverse compensation vector according to the direction and magnitude of the displacement component, and defining the reverse compensation vector as the pre-tension offset during the installation stage; associating the installation coordinates, the number of bellows layers, and the pre-tension offset to generate an engineering instruction set containing spatial positioning data, component structural parameters, and construction installation offset guidance values.

[0069] First, the final state matrix output by reinforcement learning is frozen. A full scan of the compensation weights of all nodes in the pipeline network is performed. A numerical threshold (e.g., 0.85) is set, and nodes with weights exceeding this threshold are identified. The weights of adjacent nodes are compared, and local extreme points are extracted using a non-maximum suppression algorithm. The final winning node number is locked (e.g., Node_ID: 1024). Then, the initially constructed pipeline topology database is called, and the node number is mapped back to three-dimensional spatial data in the factory's absolute coordinate system through a hash index. The geometric coordinates (X, Y, Z) of the centerline of the pipe segment where the node is located are accurately obtained, and these coordinates are fixed as the geometric center installation point of the compensator.

[0070] The stress characterization vector corresponding to the installation site is extracted (this vector has previously aggregated the clearing temperature change sequence and vibration characteristics through a spatiotemporal graph neural network). This high-dimensional vector is input into a preset fatigue life assessment model. The model decrypts the vector and reconstructs the full-time stress waveform at this location during a future major overhaul cycle. It automatically calculates the equivalent stress amplitude (e.g., 180 MPa) and the corresponding thermal cycle frequency. Then, it accesses the preset compensator engineering parameter library through an SQL interface, using the calculated stress amplitude and cycle number as search keys to match the bellows structure parameters that meet the fatigue life requirements. Specifically, it determines the number of bellows layers (e.g., from a single layer to a double layer structure) and the single-layer wall thickness (e.g., 1.5 mm) to ensure that the natural frequency of the component is compatible with the characteristics of the piping system.

[0071] Using the creep strain data calculated in the preceding steps, the system specifically extracts the cumulative irreversible displacement vector of the piping system at the installation point at the end of a preset service life (e.g., 100,000 hours). For example, if it is calculated that due to high-temperature material aging, the piping system will naturally produce a permanent elongation of +25mm in the positive axial direction, based on the "cold tightening" principle, the system calculates the reverse compensation value of this displacement vector. The system then defines the pre-tension offset amount for the installation stage, requiring that the compensator be pre-stretched by 25mm in the negative axial direction during cold installation.

[0072] The calculation of the cumulative irreversible displacement vector adopts a piecewise accumulation method. First, the creep strain value (i.e., elongation per unit length) of each discrete node on the pipeline centerline is obtained. Then, it is multiplied by the physical length of the pipe segment to which each node belongs to obtain the absolute elongation of each pipe segment. Finally, along the pipeline route, starting from the fixed support end, the absolute elongations of each pipe segment are vectorically superimposed to calculate the total cumulative displacement of the compensator installation point relative to the fixed end.

[0073] Finally, the system's data processor encapsulates the above-mentioned solution information in a structured manner. The geometric center coordinates (X, Y, Z), component structural features (double-layer / 1.5mm / 304L), and construction pre-offset value (Axial: -25mm) are combined to generate a standard JSON-formatted engineering instruction set. This instruction set is directly output to the digital delivery platform to guide on-site construction personnel to select the specified compensator based on the coordinates after pipeline cutting, and to perform precise pre-stretching operations before welding to counteract creep elongation that will occur during future operation.

[0074] To accurately predict the permanent deformation of the pipeline over its tens of thousands of hours of service, this embodiment introduces the Norton-Bailey creep constitutive equation from materials science. Instead of simply providing a fixed value, it dynamically calculates the creep rate of each pipe segment based on the full-time-domain stress curve and operating temperature predicted by the aforementioned spatiotemporal neural network. Considering that the pyrolysis furnace pipeline operates in a high-temperature, high-pressure environment for extended periods, the preset service life (e.g., 100,000 hours) is discretized into several calculation time steps. The irreversible plastic strain generated at each stage under the current stress and temperature level is accumulated to obtain the total creep strain over the entire lifespan.

[0075] After obtaining the creep strain rate of each pipe segment, it is converted into a specific physical displacement. In the specific calculation, the original length of each physical pipeline segment is multiplied by its corresponding total creep strain to obtain the axial elongation of that segment due to material aging. Next, the system utilizes the topological connections of the pipeline network to perform vector accumulation from the fixed support end of the piping system along the radial compensator installation point along the central line. This means that the cumulative displacement at the compensator location is actually the sum of the elongations of all pipe segments in front of it; the resulting three-dimensional displacement vector is the cumulative irreversible displacement component at that point.

[0076] The determination of the final pre-tension offset follows the "cold tightening" compensation principle in mechanics. The system inverts the calculated cumulative irreversible displacement vector and defines it as the reverse compensation vector. For example, if the prediction shows that the pipeline will elongate axially by 25mm due to creep at the end of its service life, it is required that the compensator be pre-stretched 25mm in the opposite direction before welding and fixing in a "cold" environment during the installation phase. This operation ensures that as the pipeline gradually heats up and creeps over the next few years, the compensator can gradually return from the pre-stretched state to near zero, thus ensuring that it maintains optimal displacement absorption capacity throughout the later stages of its service life.

[0077] The preset fatigue life assessment model adopts a five-stage pipeline architecture, which consists of the following layers from input to output: input layer, time-domain decoding layer, hysteresis energy mapping layer, multi-physics damage accumulation layer, and post-processing layer.

[0078] The input layer, located at the very front of the model, is responsible for cleaning and enhancing the raw input data to eliminate dimensional differences and inject physical context.

[0079] This layer not only receives the stress representation vector from the upstream graph neural network, but also establishes an independent material property embedding interface. It automatically reads the pipe material code of the current node (such as 304L or Incoloy800H), converts it into a high-dimensional material embedding vector, and concatenates this vector with the stress representation vector. To address the problem of large stress value ranges under different working conditions (from a few megapascals to hundreds of megapascals), this layer adopts the adaptive Z-Score algorithm to map the concatenated hybrid feature vector to the standard normal distribution range, eliminate numerical singularities, and provide a stable input distribution for subsequent neural network layers.

[0080] Specifically, the material embedding vector is not a randomly generated value, but a feature sequence constructed based on physical properties. The system has a pre-set standard material property database. By reading the material code (e.g., 304L), it retrieves five physical parameters of the material at different temperatures: elastic modulus, yield strength, Poisson's ratio, coefficient of thermal expansion, and hardening index. After normalizing these five parameters, they are arranged in order to form a five-dimensional feature vector, which serves as the material embedding vector. This vector is then concatenated with the stress characterization vector, enabling subsequent calculations by the neural network to perceive the mechanical stiffness and hardening characteristics of the specific material.

[0081] The temporal decoding layer takes over the preprocessed feature stream and uses a long short-term memory network to restore the static vector to a dynamic physical process.

[0082] By utilizing the forgetting gate mechanism of a Long Short-Term Memory (LSTM) network, high-frequency noise in the input features is filtered out, focusing on reconstructing the baseline thermal stress waveform sequence corresponding to the coking cycle (heating-isothermal-cooling) of the pyrolysis furnace. In parallel, high-frequency components are extracted from the features, and combined with the input edge weight attributes (dynamic response perturbation coefficients), a high-frequency vibration acceleration envelope superimposed on the baseline waveform is reconstructed. This step achieves precise decoupling of the thermal load and mechanical vibration signal in the time domain.

[0083] The restoration process employs a copy-reconstruction decoding logic. First, the network layer copies the input static mixed feature vector along the time dimension, with the number of copies equal to the sampling time steps of the descaling cycle (e.g., 1000 copies), forming an initial time-series matrix. Subsequently, a long short-term memory network is used to perform stepwise regression calculations on this matrix. It is particularly important to note that the network's weight parameters are learned during the training phase using a large amount of finite element simulation data. Specifically, during training, the actual full-time-domain stress waveform calculated by finite element software is used as the label, forcing the neural network to learn the mapping relationship from the feature vector to the complete waveform.

[0084] The hysteresis energy mapping layer transforms the signal flow into a concrete material mechanical response, with the core being the construction of a virtual energy dissipation model.

[0085] Based on the peaks and troughs of the baseline thermal stress waveform sequence, and combined with the constitutive parameters of the material injected in the first layer, closed hysteresis loops are dynamically plotted in a virtual stress-strain coordinate system. The plotting process follows the multiplication law of material mechanics. First, the loading path is determined using the material's stress-strain skeleton curve. When the stress waveform reverses and declines from its peak, it does not simply return along the original path, but instead doubles the size of the original skeleton curve to describe the shape of the unloading path. Through this geometric mapping, the stress waveform points that change with time are mapped one by one to the stress-strain coordinate system, naturally forming one or more closed loop trajectories (i.e., hysteresis loops). Subsequently, a polygon area algorithm is used to calculate the geometric area enclosed by this closed loop trajectory. This area is physically equivalent to the plastic deformation energy dissipated by the material in one cycle, thus achieving the goal of accurately quantifying thermal fatigue damage using geometric area.

[0086] Multiphysics damage accumulation layer: This layer integrates multi-dimensional damage mechanisms based on the linear cumulative damage law.

[0087] The plastic strain energy density calculated by the hysteresis energy mapping layer is mapped to the low-cycle thermal fatigue damage value; at the same time, the high-frequency vibration acceleration envelope of the time-domain decoding layer is processed by the rainflow counting method and mapped to the high-cycle vibration fatigue damage value; externally input creep strain data is introduced to calculate the creep damage factor at high temperature; the above three damage indicators (thermal fatigue, vibration fatigue, and creep) are weighted and summed to generate the original cumulative damage index of the node throughout its entire life cycle.

[0088] The post-processing layer, located at the end of the model, is responsible for converting abstract numerical indices into executable engineering instructions and eliminating computational jitter.

[0089] To prevent frequent changes in component selection between different specifications due to minor fluctuations in values, this layer introduces hysteresis comparator logic. Only when the change in the original cumulative damage index exceeds a preset dead zone threshold (e.g., 5%) will the selection level be changed. The continuous damage index is discretized into standard engineering levels (in this embodiment: Level I - no compensation required, Level II - standard compensation, Level III - enhanced compensation, Level IV - path reconstruction).

[0090] Finally, the quantified engineering grade, recommended component parameters (number of layers, wall thickness), and estimated remaining lifespan are packaged into a standard JSON or XML data packet and output as the final engineering instruction set to the downstream digital design system.

[0091] Example 2: This embodiment applies the aforementioned optimization method based on topological representation learning to the quench oil outlet pipe gallery system of a large ethylene cracking furnace area. This scenario involves ultra-high temperature (up to 900°C), severe spatial obstacles (due to the non-convex space formed by the existing steel structure and maintenance platform), and high-frequency vibration interference generated by moving equipment (such as large quench oil pumps).

[0092] First, the system connects to the factory's PDMS 3D design database via an industrial digital interface to extract the underlying geometric description of the target pipeline. For the quench oil outlet pipeline, the system analyzes the 3D coordinates of the centerline, pipe diameter (e.g., 600mm), wall thickness parameters (e.g., 10mm), and topological connections, and matches the linear expansion coefficient of 304L stainless steel in the high-temperature range (0-900℃) to construct the pipeline network topology. Simultaneously, sampling sequences including heating, isothermal decoking, and cooling sections are extracted from the DCS system to form a decoking temperature change sequence reflecting the thermal shock cycle of the pipeline network. In addition, the vibration amplitude of the quench pump at a 50Hz operating frequency is collected, and a vibration matrix characterizing the coupling strength between excitation and resonance is constructed by combining it with the natural frequency of the pipeline network. Finally, the spatial boundary of existing infrastructure around the pipe gallery is identified through a geometric envelope algorithm, and an obstacle repulsive potential energy field with spatial gradient is constructed.

[0093] The heterogeneous feature maps are input into a spatiotemporal graph neural network. The graph convolution operator is used to transmit thermal stress features along the topological path. The dynamic response coefficients on the edges are used to amplify and aggregate the features of high vibration risk paths. The spatially aggregated feature stream enters the temporal processing unit. The gating mechanism identifies the stress response differences at different stages of heating, isothermal and cooling. In particular, the stress hysteresis features caused by the thermal inertia of the material are captured. Finally, the spatial and temporal features are fused through the attention mechanism, and the complex stress evolution trajectory is compressed into a 128-dimensional high-dimensional hidden state sequence to generate the stress representation vector of the node.

[0094] A decision-making system is established, consisting of three intelligent agents: the furnace body (active displacement source), pipeline compensation (coordinate optimization), and the heat exchanger (stress constraint). Guided by a sensitivity heatmap provided by a stress representation vector, the pipeline compensation agent performs obstacle avoidance actions in conjunction with a spatial repulsive potential energy field, searching for candidate coordinate sets for the compensator installation. A reward function is constructed with global elastic potential energy equilibrium as its core. The reward value is composed of the interface stress residual (which must meet a deviation of 0.03%), displacement absorption efficiency, and energy distribution entropy. Before calculating the total reward value, each sub-item index is normalized. The disturbance energy generated on the global stiffness matrix after the compensator intervention is calculated in real time. If the disturbance energy causes the system fundamental frequency shift to exceed 2Hz or causes local stress concentration (reduced energy distribution entropy), a negative excitation signal is generated and fed back to the decision-making system. The gradient descent method is used to update the compensation weights of each node, driving the agent group to re-execute the game decision until the system displacement absorption and interface stress synchronously enter a steady state.

[0095] Based on the iterative weight distribution, the final winning node (e.g., Node_ID: 1024) is locked using a non-maximum suppression algorithm, and the compensator installation location is determined. Then, the fatigue life assessment model of the five-stage pipeline architecture is entered: the 304L material code is read and converted into a material embedding vector containing five parameters, including the elastic modulus, and adaptive Z-Score normalization is performed. The temporal decoding layer generates a full-time-domain stress waveform through a "copy-reconstruction" logic; the hysteresis energy mapping layer draws a virtual hysteresis loop according to the "multiplication rule," calculates its enclosing area to quantify the plastic strain energy density, and, based on the assessed stress amplitude (e.g., 180 MPa) and plastic damage value, matches and selects a double-layer, 1.5 mm thick bellows from the parameter library; the cumulative creep elongation (e.g., 25 mm) at this location within a 100,000-hour service cycle is extracted, the reverse compensation vector is calculated, and it is defined that a pre-stretch of 25 mm in the negative axial direction is required during the installation stage.

[0096] The system's data processor encapsulates all the calculation information in a structured manner, generates a standard set of engineering instructions, and outputs them to the delivery platform. The output is a JSON file containing {"Target_Node": "1024", "Coordinates": {"X": 42560", "Y": 15820", "Z": 12500}, "Specs": {"Material": "304L", "Layers": 2, "Thickness": "1.5mm"}, and "Installation": {"Pre_Stretch": "-25.0mm", "Action": "Pre-stretching in the axial negative direction of the welding front edge"}}.

[0097] Target_Node (1024) represents the installation node number selected by the algorithm in the pipeline topology model. The system scans the distribution of compensation weights after iteration and uses the non-maximum suppression algorithm to extract the node with the highest local weight and determine it as the optimal location for compensator installation.

[0098] Coordinates(X, Y, Z) maps the node number back to the three-dimensional spatial centerline coordinates in the factory's absolute coordinate system. This set of coordinates precisely indicates the installation position of the compensator's geometric center in the real physical space, ensuring that the component can be accurately placed in the mechanically sensitive area.

[0099] The Specs section defines the hardware specifications of the position compensator, which are derived from the fatigue life assessment model. The system inputs the stress characterization vector generated by this node into the model to extract the equivalent stress amplitude and cycle frequency over the entire operating cycle.

[0100] Based on these characteristics, the system automatically retrieves and matches the most compatible number of bellows layers (Layers: 2) and single-layer wall thickness (Thickness: 1.5mm) from the preset parameter library. For example, the double-layer structure here is to ensure that the compensator can still achieve the expected service life under severe temperature change cycles such as coking in the cracking furnace.

[0101] Pre_Stretch (-25.0mm) represents the pre-stretch offset during the installation phase. It is calculated using creep strain, which predicts that the pipeline will experience approximately 25mm of irreversible permanent elongation due to material aging during the tens of thousands of hours of high-temperature operation in the future.

[0102] Action specifies the construction procedures: before welding, workers must perform reverse pre-tensioning of the compensator in the negative axial direction. This cold tightening operation can offset future creep displacement in advance, ensuring that the compensator remains within the effective displacement absorption range throughout the later stages of the pipeline's life.

[0103] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for topology layout optimization of heat pipe network compensator based on topology representation learning, characterized in that, include: Obtain the pipeline topology, coking temperature change sequence, furnace body expansion field, vibration matrix, interface stress limit, creep strain, and repulsive force of spatial obstacles; Based on the pipeline network topology, a graph structure containing nodes and edges is constructed. The clearing temperature change sequence is embedded as the node time-series feature, and the vibration matrix is ​​mapped as the edge stiffness attribute to complete the embedding of topological heterogeneous features and obtain a heterogeneous feature map. A spatiotemporal graph neural network is used to aggregate the heterogeneous feature map, extract the stress evolution trajectory of the pipeline network under temperature change cycle, and generate the stress characterization vector related to the time sequence of each node. A decision-making system is established, consisting of three intelligent agents: furnace body, pipeline compensation, and heat exchanger. The furnace body expansion field, creep strain, and interface stress limit are input into the corresponding intelligent agents as boundary displacement, attenuation parameters, and stress constraints, respectively. The decision-making system calls the stress characterization vector and combines it with the repulsive force of spatial obstacles to perform game-theoretic decision-making, calculates the disturbance of the selected position to the global stiffness matrix of the system, and iteratively adjusts the compensation weights of each node until the displacement absorption and interface stress meet the steady-state threshold. The installation coordinates are determined in three-dimensional space based on the iterated compensation weights, and the number of bellows layers is matched based on the characterization vector. The pre-tension offset is calculated and output using creep strain.

2. The topology characterization learning based compensation topology layout optimization method of the heat pipe network according to claim 1, characterized in that, The process of acquiring the pipeline topology, coking temperature change sequence, furnace body expansion field, vibration matrix, creep strain, and spatial obstacle repulsion includes: extracting the centerline coordinates, pipe diameter, wall thickness parameters, and topological connection relationships of the pipelines in the pyrolysis furnace area; combining this with the material linear expansion coefficient to generate the pipeline topology; extracting the temperature sampling time sequence, including the heating section, the isothermal coking section, and the cooling section, from the production control system to form a coking temperature change sequence reflecting the thermal shock cycle of the pipeline; acquiring non-uniform axial displacement vector data at coordinate points at different heights of the furnace wall, and using the displacement vector data as the dynamic displacement boundary condition at the beginning of the pipeline. The furnace body expansion field is obtained; the rotational speed and vibration amplitude of the moving equipment are collected and the mechanical excitation frequency is calculated. Combined with the natural frequency of the pipeline network, a numerical matrix characterizing the coupling strength between excitation and resonance is constructed as the vibration matrix; based on the high-temperature creep rate curve of the material and the cumulative operating hours of the pipeline, the irreversible permanent deformation within the preset service period is calculated as the creep strain; the existing infrastructure coordinate boundary in the pyrolysis furnace area is identified and the minimum physical distance of the pipeline sampling point relative to the coordinate boundary is calculated. The minimum physical distance is mapped to a potential energy weight with a spatial gradient to construct the spatial obstacle repulsion force.

3. The method for optimizing the topology layout of thermal pipeline compensators based on topology representation learning according to claim 1, characterized in that, The process of embedding the heterogeneous topological features includes: mapping the pipe segment intersections, furnace connection points, and heat exchanger inlet and outlet points in the pipe network topology to a set of nodes in a graph structure, and mapping the physical pipe segments connecting each node to a set of edges in the graph structure; performing sliding window sampling on the coking temperature change sequence to generate temperature state vectors corresponding to each node at different time steps, and assigning the temperature state vectors as time-series feature attributes to the corresponding nodes in the node set; extracting the excitation frequency components of each edge in the vibration matrix, calculating the dynamic response disturbance coefficient of the excitation frequency components to the inherent stiffness of the pipe segment, and mapping the dynamic response disturbance coefficient to the weight attributes of each edge in the edge set, thereby realizing the feature fusion of the graph structure in the spatial topology and dynamic load dimensions.

4. The method for optimizing the topology layout of thermal pipeline compensators based on topology representation learning according to claim 1, characterized in that, The process of generating time-related stress characterization vectors for each node includes mapping physical nodes and pipe segments in the pipeline network topology to vertices and edges of a graph structure, processing the descaling temperature change sequence into dynamic temperature feature vectors for each node using the sliding window method, and extracting the dynamic response coefficients of each pipe segment as edge weights using the vibration matrix to align physical attributes to the graph feature space; using a graph convolution operator to transfer features along the pipeline network topology path, and using the dynamic response coefficients on the edges to weight and aggregate the temperature features of adjacent nodes to extract the physical transmission characteristics of thermal stress in the pipeline network under a specific spatial configuration; inputting the spatially aggregated feature stream into a time-series processing unit to capture the stress response differences of the pipeline network at different stages of heating, isothermal descaling, and cooling, and extracting the stress cumulative effect and hysteresis features that change over time; By fusing spatial topological features and time series features through an attention mechanism, the complex stress fluctuations within the entire working cycle are compressed into a high-dimensional hidden state sequence, generating a stress representation vector.

5. The method for optimizing the topology layout of thermal pipeline compensators based on topology representation learning according to claim 1, characterized in that, The process of establishing a decision-making system composed of three intelligent agents—furnace body, pipeline compensation, and heat exchanger—includes the following: establishing independent physical intelligent agents representing the furnace body, pipeline compensation, and heat exchanger respectively; converting the furnace body expansion field into the active displacement output of the furnace body intelligent agent, converting creep strain into the long-term performance degradation coefficient of the pipeline compensation intelligent agent, and using the interface stress limit as the mandatory stress constraint boundary of the heat exchanger intelligent agent; the pipeline compensation intelligent agent calling the stress characterization vector, guided by the potential energy constructed by the repulsive force of spatial obstacles, searching for a candidate set of compensator installation coordinates; and each intelligent agent adjusting its own constraints according to its own conditions. The candidate set engages in collaborative game theory, iteratively evolving with the common goal of minimizing global system stress and achieving interface stress standards. For the compensator installation location selected collaboratively by each agent, the disturbance amount generated by the intervention at that installation location on the global stiffness matrix of the pipeline system is calculated in real time. By analyzing the impact of the disturbance on the system vibration frequency and stress redistribution, stiffness feedback instructions are generated. The compensation weights of each candidate node are dynamically adjusted according to the stiffness feedback instructions. The agent group is driven to re-execute the game decision through a reinforcement learning mechanism until the displacement absorption of the pipeline system and the stress at the heat exchanger interface both meet the preset steady-state threshold.

6. The method for optimizing the topology layout of thermal pipeline compensators based on topology representation learning according to claim 5, characterized in that, The process of driving the agent group to re-execute game decisions through reinforcement learning includes: mapping the stress representation vector to the environmental state of the decision system, and constructing a physical search space composed of active displacement sources, target force boundaries, and potential energy barrier fields based on the furnace expansion field, interface force limits, and spatial obstacle repulsion; establishing a decision model based on non-cooperative game theory to drive the furnace, pipeline compensation, and heat exchanger agents to perform coordinate optimization actions within the physical search space; each agent uses minimizing the interface force residual and maximizing the compensator absorption efficiency as conflicting objectives, and converges the search interval by dynamically adjusting the compensation weights of each node; for For each candidate coordinate selected by each agent, the disturbance energy generated on the global stiffness matrix of the pipeline system after the compensation component is introduced at that coordinate is calculated. If the disturbance energy causes the system fundamental frequency shift or the energy distribution entropy value to exceed the preset safety threshold, the deviation is converted into a negative incentive signal and fed back to the decision system. A reward function with the global elastic potential energy balance as the core is constructed. The reward function is corrected using the negative incentive signal, and the compensation weight of each node is updated using the gradient descent method. The agent group is driven to re-execute the game action through the redistribution of weights until the system displacement absorption and interface force synchronously enter a steady state, and the globally optimal layout scheme is output.

7. The method for optimizing the topology layout of thermal pipeline compensators based on topology representation learning according to claim 1, characterized in that, The process of calculating and outputting the pre-tension offset using creep strain includes: identifying the distribution of compensation weights after iteration, extracting the node numbers of the local extreme points of the weights, mapping the node numbers back to the three-dimensional spatial centerline coordinates of the pipeline network topology, and determining the geometric center installation position of the compensator; inputting the stress characterization vector into a preset fatigue life assessment model, and extracting the equivalent stress amplitude and cycle frequency characteristics of the pipeline network under full-cycle descaling temperature change; searching in a preset compensator parameter library based on the characteristics, and matching the number of bellows layers and single-layer wall thickness compatible with the stress evolution trajectory; extracting the cumulative irreversible displacement component of the creep strain within a preset service period, calculating the reverse compensation vector based on the direction and magnitude of the displacement component, and defining the reverse compensation vector as the pre-tension offset during the installation stage; associating the installation coordinates, the number of bellows layers, and the pre-tension offset to generate an engineering instruction set containing spatial positioning data, component structural parameters, and construction installation offset guidance values.

Images