Intelligent optimization method for equipment layout on host island based on 3D modeling

Abstract

This invention provides a method for intelligently optimizing the layout of host island equipment based on 3D modeling. This method involves extracting global device feature data through feature recognition and semantic segmentation, constructing a device relationship diagram coupled with multiple physical fields, determining the device layout order using a graph attention network, discretizing the layout space using an octree structure, and optimizing device position and posture using reinforcement learning. Finally, verification is achieved through 3D dynamic simulation. This method enables intelligent layout of host island equipment, improves space utilization and maintenance convenience, and reduces thermodynamic interference between devices.

Description

Technical Field

[0001] The present invention relates to three-dimensional modeling technology, and in particular to a host island equipment layout intelligent optimization method based on three-dimensional modeling. Background Art

[0002] Currently, the main host island equipment layout technology mainly includes rule-based layout methods, optimization algorithm-based layout methods, and artificial intelligence-assisted layout methods. The existing host island equipment layout technology has the following defects and shortcomings:

[0003] Existing technologies struggle to fully account for the multi-physics coupling relationships between devices, particularly the interactions between fluid dynamics and thermodynamics. This can lead to layout solutions that may suffer from operational inefficiencies or safety hazards. The thermodynamic, fluid dynamics, and other physical relationships between devices are complex and mutually influential, yet traditional approaches often only consider geometric constraints and simple physical connections, ignoring the deeper physical coupling effects.

[0004] The representation and optimization methods for equipment layout spaces are not precise and flexible enough. Existing technologies typically use uniform grid division or simplified geometric representations, which are difficult to adapt to the precise representation of equipment of different scales and the efficient utilization of complex spaces. This limitation is particularly evident for equipment and spaces with irregular shapes.

[0005] Lack of a dynamic assessment mechanism for equipment maintenance requirements. Existing layout methods mostly consider only static spacing requirements, making it difficult to simulate and assess dynamic operational space requirements during equipment maintenance, such as the disassembly path for large equipment and the workspace for maintenance personnel. This makes subsequent equipment maintenance difficult, increasing operating costs and safety risks. Summary of the Invention

[0006] The embodiment of the present invention provides a host island equipment layout intelligent optimization method based on three-dimensional modeling, which can solve the problems in the prior art.

[0007] A first aspect of an embodiment of the present invention provides a method for intelligently optimizing host island equipment layout based on three-dimensional modeling, comprising:

[0008] Establish a three-dimensional model of the host island and apply feature recognition and semantic segmentation to extract global feature data of the device, including the geometric and physical features of the device and the constraint relationships between the devices;

[0009] Based on the global feature data, a device relationship graph for multi-physics coupling is constructed, in which the devices are nodes, the physical associations between the devices are edges, and the edge attributes include fluid properties and heat transfer properties. The graph attention network is used to analyze the influence weights between the devices to obtain the device layout order;

[0010] An octree structure is used to adaptively discretize the equipment layout space into a multi-resolution three-dimensional grid. A reinforcement learning method with a multi-scale optimization strategy is applied to optimize the equipment's spatial position. The equipment's translation and rotation parameters are used as the action space, with collision avoidance, clearance maintenance, and maintenance accessibility as optimization objectives. The optimal position and attitude parameters are obtained through iterative calculation.

[0011] The equipment layout sequence, optimal position and posture parameters are used as a layout plan, which is verified through three-dimensional dynamic simulation to evaluate the spatiotemporal envelope requirements of equipment maintenance operations; after verification, it is confirmed as the final equipment layout plan.

[0012] In an optional embodiment,

[0013] Based on the global feature data, a multi-physics field coupled device relationship graph is constructed, where the devices are nodes and the physical associations between devices are edges. The edge attributes include fluid properties and heat transfer properties. The graph attention network is used to analyze the influence weights between devices to obtain the device layout order, which includes:

[0014] Constructing a node feature vector based on the geometric features and physical features, wherein the node feature vector includes spatial dimension parameters, position and posture parameters, and operation parameters of the device;

[0015] Determine the physical association between devices based on the constraint relationships between them, and use the pipeline connection relationships and support structure relationships as edges to establish topological connections between devices;

[0016] Constructing edge attributes based on fluid characteristics and heat transfer characteristics between devices, including: constructing a pressure loss matrix of fluid characteristics based on the Darcy-Weisbach equation, constructing a thermal resistance matrix of heat transfer characteristics based on thermal conductivity and Nusselt number, calculating the physical field strength of the fluid field and thermal field according to equipment operating parameters, normalizing the pressure loss matrix and thermal resistance matrix, and weightedly fusing them according to the ratio of physical field strengths to obtain a comprehensive feature matrix; performing a consistency check on the comprehensive feature matrix based on mass conservation and energy conservation, and extracting the edge feature vectors to form an edge attribute feature set;

[0017] The graph attention network is used to calculate the influence weights between devices, and the node feature vectors, topological connections and edge attribute feature sets are integrated into the edge type perception mechanism and gated fusion to obtain the device layout order.

[0018] In an optional embodiment,

[0019] The graph attention network is used to calculate the influence weights between devices. The node feature vectors, topological connections, and edge attribute feature sets are integrated into the edge type perception mechanism and gated fusion to obtain the device placement order, which includes:

[0020] Divide the equipment into functional types and construct a type embedding matrix, which is then concatenated with the node feature vector to obtain the equipment feature vector.

[0021] Constructing an edge type parameter matrix for the physical associations between different types of devices, calculating an edge type perception attention score based on the device feature vector and the edge type parameter matrix in combination with the topological connection, and weighting the score with the edge attribute feature vector to obtain an edge type feature;

[0022] Constructing a relationship conversion matrix to perform nonlinear conversion on the edge type features to obtain conversion features, training a multilayer perceptron based on the conversion features and device feature vectors of adjacent devices to obtain gating coefficients, and using the gating coefficients to adaptively fuse different types of conversion features to obtain inter-device influence weights;

[0023] Constructing a node-level attention heat map based on the influence weights between the devices, and performing perturbation analysis on the device feature vectors to obtain feature sensitivity;

[0024] The device placement order is determined based on the influence weights between devices, the influence index of the node-level attention heat map, and the feature sensitivity.

[0025] In an optional embodiment,

[0026] The equipment layout space is adaptively discretized into a multi-resolution three-dimensional grid using an octree structure. A reinforcement learning method with a multi-scale optimization strategy is applied to optimize the equipment spatial position. The equipment's translation and rotation parameters are used as the action space, and collision avoidance, clearance maintenance, and maintenance accessibility are the optimization goals. The optimal position and posture parameters obtained by iterative calculation include:

[0027] The octree structure is used to adaptively discretize the equipment layout space, and recursive subdivision is performed in the equipment boundary area and the equipment gap area to obtain a multi-resolution three-dimensional grid.

[0028] Constructing a state space for reinforcement learning based on the multi-resolution three-dimensional grid, the state space including position and attitude parameter states, grid occupancy state, and constraint state, and an action space including translation and rotation parameters of the device, wherein a movement step size in the action space is proportional to a grid size at the corresponding grid level;

[0029] A reinforcement learning method is applied to optimize equipment positions on multi-resolution three-dimensional grids. The reinforcement learning method adopts a multi-scale optimization strategy. At the coarse grid level, a global position search is performed using a reward function that includes a collision penalty term and a gap reward term to obtain the initial equipment layout position. At the fine grid level, a local fine optimization is performed by adding a maintenance accessibility assessment term to the reward function, and the initial optimal position and posture parameters are iteratively obtained.

[0030] Constraint verification is performed on the multi-resolution three-dimensional grid based on the initial optimal position and attitude parameters, including: checking pipeline connection constraints, equipment spacing constraints and support structure constraints; verifying the accessibility of maintenance channels, lifting paths and operating spaces, and generating the final optimal position and attitude parameters after verification.

[0031] In an optional embodiment,

[0032] The weight adjustment methods for each item in the reward function include:

[0033] A hierarchical adaptive reward mechanism is designed to determine the weight of the reward function according to the degree of constraint violation during the optimization process: when the collision volume between devices is greater than the first preset threshold, the weight of the collision penalty item is adjusted to a preset multiple of the current weight; when the equipment gap is less than the second preset threshold, the weight of the gap reward item is adjusted to a preset multiple of the current weight; when the number of accessible paths in the maintenance channel is less than the third preset threshold, the weight of the maintenance accessibility assessment item is adjusted to a preset multiple of the current weight; when the degree of violation of each constraint is reduced to below the preset proportion of the corresponding preset threshold, the corresponding weight is restored to the initial setting value.

[0034] In an optional embodiment,

[0035] The equipment layout sequence, optimal position, and posture parameters are used as the layout plan. Three-dimensional dynamic simulation is used to verify and evaluate the spatiotemporal envelope requirements of equipment maintenance operations, including:

[0036] Constructing a three-dimensional geometric structure of the device layout scene based on the device layout sequence, optimal position, and posture parameters, where the optimal position parameter is the coordinate value of the device in three-dimensional space, and the posture parameter is the rotation matrix of the device in three-dimensional space;

[0037] Acquiring a device maintenance operation trajectory, dividing the device maintenance operation trajectory into an operation sequence, and establishing a human joint motion parameter set and a tool operation parameter set based on the operation sequence;

[0038] Calculating the motion trajectory during the maintenance operation based on the human joint motion parameter set and the tool operation parameter set, and generating a spatiotemporal envelope of the maintenance operation;

[0039] Performing interference detection on the maintenance operation space-time envelope surface and the three-dimensional geometric structure, and calculating a minimum distance value and a collision volume value during the maintenance operation;

[0040] The accessibility of the maintenance operation is determined according to the minimum distance value and the collision volume value, and a feasibility assessment result of the equipment layout plan is generated.

[0041] In an optional embodiment,

[0042] Calculating the motion trajectory during the maintenance operation based on the human joint motion parameter set and the tool operation parameter set, and generating the spatiotemporal envelope of the maintenance operation includes:

[0043] Constructing a velocity field function of the maintenance operation trajectory according to the human joint motion parameter set and the tool operation parameter set, wherein the velocity field function includes a maintenance difficulty function, a curvature term, and a global constraint term, wherein the maintenance difficulty function is related to the joint motion range and the tool operation torque;

[0044] Driving the level set evolution based on the velocity field function to generate an initial space-time envelope surface;

[0045] Establishing a multi-scale uncertainty distribution for the operation trajectory, including: establishing a Gaussian distribution based on the standard deviation of the human joint positions and the standard deviation of the tool positions, calculating the joint probability distribution at each sampling moment, the joint probability distribution characterizing the spatial distribution of the human-tool system at that moment, and accumulating the uncertainty characteristics of the complete operation trajectory in the time dimension;

[0046] The expansion coefficient of the initial spatiotemporal envelope surface is dynamically adjusted according to the uncertainty characteristics. The expansion coefficient is composed of a reference value and an uncertainty correction value. The uncertainty correction value is the product of the larger value of the joint position standard deviation and the tool position standard deviation and the proportional coefficient related to the operation difficulty, so as to generate the final maintenance operation spatiotemporal envelope surface.

[0047] According to a second aspect of an embodiment of the present invention, an electronic device is provided, including:

[0048] processor;

[0049] a memory for storing processor-executable instructions;

[0050] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0051] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0052] The present invention proposes an intelligent optimization method for host island equipment layout based on three-dimensional modeling, which realizes efficient automation of equipment layout and effectively solves many technical problems existing in traditional equipment layout.

[0053] By constructing a multi-physical field coupled device relationship diagram and applying a graph attention network to analyze the influence weights between devices, the present invention can scientifically and rationally determine the device layout order, fully consider the physical correlation between devices, avoid the problem of unreasonable layout caused by subjective experience, and improve the equipment operation efficiency and safety.

[0054] The present invention adopts an adaptive multi-resolution three-dimensional grid with an octree structure and a reinforcement learning method to achieve intelligent optimization of the spatial position of equipment. This not only ensures that the equipment layout meets multiple constraints such as collision avoidance, clearance maintenance, and maintenance accessibility, but also evaluates the spatiotemporal envelope of maintenance operations through a three-dimensional dynamic simulation verification mechanism, greatly improving the reliability and practicality of the equipment layout plan. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 Schematic diagram of the process of the intelligent optimization method for host island equipment layout based on three-dimensional modeling according to an embodiment of the present invention;

[0056] Figure 2 A comparison chart of equipment layout effects with different layout sequence methods;

[0057] Figure 3 The spatial envelope interference detection result diagram for equipment layout maintenance operation. DETAILED DESCRIPTION

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0059] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0060] Figure 1 FIG. 1 is a flow chart of an intelligent optimization method for host island equipment layout based on three-dimensional modeling according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0061] Establish a three-dimensional model of the host island and apply feature recognition and semantic segmentation to extract global feature data of the device, including the geometric and physical features of the device and the constraint relationships between the devices;

[0062] Based on the global feature data, a device relationship graph for multi-physics coupling is constructed, in which the devices are nodes, the physical associations between the devices are edges, and the edge attributes include fluid properties and heat transfer properties. The graph attention network is used to analyze the influence weights between the devices to obtain the device layout order;

[0063] An octree structure is used to adaptively discretize the equipment layout space into a multi-resolution three-dimensional grid. A reinforcement learning method with a multi-scale optimization strategy is applied to optimize the equipment's spatial position. The equipment's translation and rotation parameters are used as the action space, with collision avoidance, clearance maintenance, and maintenance accessibility as optimization objectives. The optimal position and attitude parameters are obtained through iterative calculation.

[0064] The equipment layout sequence, optimal position and posture parameters are used as a layout plan, which is verified through three-dimensional dynamic simulation to evaluate the spatiotemporal envelope requirements of equipment maintenance operations; after verification, it is confirmed as the final equipment layout plan.

[0065] In an optional embodiment, a multi-physics field coupled device relationship graph is constructed based on the global feature data, wherein the devices are nodes, the physical associations between the devices are edges, and the edge attributes include fluid properties and heat transfer properties. The device layout order is obtained by analyzing the influence weights between the devices using a graph attention network, including:

[0066] Constructing a node feature vector based on the geometric features and physical features, wherein the node feature vector includes spatial dimension parameters, position and posture parameters, and operation parameters of the device;

[0067] Determine the physical association between devices based on the constraint relationships between them, and use the pipeline connection relationships and support structure relationships as edges to establish topological connections between devices;

[0068] Constructing edge attributes based on fluid characteristics and heat transfer characteristics between devices, including: constructing a pressure loss matrix of fluid characteristics based on the Darcy-Weisbach equation, constructing a thermal resistance matrix of heat transfer characteristics based on thermal conductivity and Nusselt number, calculating the physical field strength of the fluid field and thermal field according to equipment operating parameters, normalizing the pressure loss matrix and thermal resistance matrix, and weightedly fusing them according to the ratio of physical field strengths to obtain a comprehensive feature matrix; performing a consistency check on the comprehensive feature matrix based on mass conservation and energy conservation, and extracting the edge feature vectors to form an edge attribute feature set;

[0069] The graph attention network is used to calculate the influence weights between devices, and the node feature vectors, topological connections and edge attribute feature sets are integrated into the edge type perception mechanism and gated fusion to obtain the device layout order.

[0070] For example, a precise 3D model of the mainframe island is created by combining BIM technology with 3D laser scanning. Deep learning frameworks such as PointNet++ are used to perform feature recognition on point cloud data, combined with a 3D semantic segmentation network based on the U-Net architecture to extract equipment geometric features, including size, shape, and spatial topology. Equipment CAD models are imported to extract physical characteristics, including parameters such as weight. Connection constraints such as pipe connection points and cable interfaces, as well as operational constraints such as maintenance space and installation distances, are automatically identified and extracted.

[0071] The node feature vector contains the device's spatial dimensions, position, and attitude parameters, as well as operating parameters. Spatial dimensions include the device's length, width, and height. For example, the main transformer's spatial dimensions are 5.6 meters long, 3.2 meters wide, and 4.8 meters high. Position parameters are expressed as the position of the device's center point in the three-dimensional coordinate system, such as (x=10.5, y=8.2, z=0.0). Attitude parameters are expressed in Euler angles, such as (pitch=5°, yaw=0°, roll=0°). Operating parameters include the device's rated power, flow rate, pressure, and temperature. For example, a pump's rated power is 250 kW, flow rate is 400 cubic meters per hour, outlet pressure is 0.6 MPa, and operating temperature is 85°C. These parameters are organized into a 128-dimensional feature vector and normalized to the [0, 1] range using the MinMaxScaler method to ensure effective fusion of features of different dimensions.

[0072] The pipeline connection relationship is determined based on the pipe diameter, medium type and connection method. For example, the DN200 steam pipe connection between two heat exchangers uses saturated steam as the medium and flange connection as the connection method. The support structure relationship is determined based on the structural mechanics model, including foundation support and hanging support, such as fixed support between the pump and the foundation, and elastic hanger support between the pipeline and the structural beam. The connection relationship is identified by matching the interface points of the three-dimensional model, and then the constraint type is determined based on the pipeline layout specifications and structural force calculations. The spatial tree index structure is used to accelerate the interface point search, and the connection tolerance is set to 5mm. The support mechanics calculation considers both static and dynamic load conditions, and the load safety factor is 1.2. Topological connections are stored in the form of an adjacency matrix, 1 indicates the existence of a connection, and 0 indicates the absence of a connection.

[0073] For fluid properties, a pressure loss matrix is ​​constructed based on the Darcy-Weisbach equation. For each pair of connected devices, the pressure loss in the pipes between them is calculated. For example, for a DN150 pipe connecting a pump and a heat exchanger, with a length of 15 meters, considering a pipe friction coefficient of 0.018, a flow velocity of 2.5 m / s, and a fluid density of 998 kg / m³, the calculated pressure loss is 35 kPa. For heat transfer properties, a thermal resistance matrix is ​​constructed based on thermal conductivity and Nusselt number. For example, if the thermal conductivity between the two devices is 45 W / (m·K), the convective heat transfer coefficient is 2800 W / (m²·K), the pipe wall thickness is 12 mm, and the pipe length is 20 meters, the calculated thermal resistance is 0.12 K / W. The physical field strengths of the fluid and thermal fields are calculated based on the device operating parameters. For example, the fluid field strength is calculated as 0.7 based on flow rate and pressure, and the thermal field strength is calculated as 0.4 based on temperature difference and heat flow. The pressure loss matrix and thermal resistance matrix are normalized separately, ensuring that their values ​​fall within the range [0, 1]. They are then weighted and fused according to the ratio of physical field strengths, with a fluid field weight of 0.7 and a thermal field weight of 0.4, to produce a comprehensive feature matrix. Each element of this matrix represents the combined physical coupling strength between devices. The comprehensive feature matrix is ​​subjected to consistency checks based on the conservation of mass and energy to ensure the correctness of the physical relationships. Specifically, the system verifies the fluid mass balance of each node, requiring the difference in the total amount of fluid inflow and outflow across all devices to be less than 1%. Heat transfer is also subject to an energy balance check, requiring the difference in the total amount of heat inflow and outflow to be less than 2%. This consistency check identifies edge relationships that conform to physical laws, extracts information such as the comprehensive strength value, connection type identifier, and physical parameter differences between each pair of connected devices, and constructs a 32-dimensional edge attribute feature vector, forming an edge attribute feature set.

[0074] The graph attention network is used to calculate the influence weights between devices and determine the order of device layout.

[0075] The present invention realizes the precise modeling and quantitative analysis of the physical associations of host island equipment, overcoming the problem of insufficient expression of complex coupling relationships between devices in traditional methods; constructs a device relationship graph reflecting the coupling of multiple physical fields, and realizes the intelligent calculation of device influence weights through the graph attention network, which can capture the implicit associations between devices and automatically generate a reasonable equipment layout sequence, effectively improving the rationality of equipment layout and engineering implementation efficiency, and providing intelligent support for the optimized layout of host island equipment.

[0076] In an optional implementation, a graph attention network is used to calculate the influence weights between devices, and the node feature vectors, topological connections, and edge attribute feature sets are integrated into the edge type perception mechanism and gated fusion to obtain the device placement order including:

[0077] Divide the equipment into functional types and construct a type embedding matrix, which is then concatenated with the node feature vector to obtain the equipment feature vector.

[0078] Constructing an edge type parameter matrix for the physical associations between different types of devices, calculating an edge type perception attention score based on the device feature vector and the edge type parameter matrix in combination with the topological connection, and weighting the score with the edge attribute feature vector to obtain an edge type feature;

[0079] Constructing a relationship conversion matrix to perform nonlinear conversion on the edge type features to obtain conversion features, training a multilayer perceptron based on the conversion features and device feature vectors of adjacent devices to obtain gating coefficients, and using the gating coefficients to adaptively fuse different types of conversion features to obtain inter-device influence weights;

[0080] Constructing a node-level attention heat map based on the influence weights between the devices, and performing perturbation analysis on the device feature vectors to obtain feature sensitivity;

[0081] The device placement order is determined based on the influence weights between devices, the influence index of the node-level attention heat map, and the feature sensitivity.

[0082] For example, the main island equipment can be divided into 10 functional categories, including pumps, heat exchangers, storage devices, and valves. For each functional type of equipment, a 16-dimensional type embedding vector is constructed. For example, the type embedding vector for pumps is [0.8, 0.2, 0.3, ..., 0.5], and the type embedding vector for heat exchangers is [0.2, 0.9, 0.1, ..., 0.4]. The type embedding vector is obtained using a pre-trained equipment type classification network. This network adopts a three-layer fully connected structure with 64, 32, and 16 hidden units, respectively, and uses the ReLU activation function. It is obtained by jointly training the equipment CAD model features and parameter labels. For example, for the safety valve in the main steam piping system, its node feature vector is 128-dimensional and its type embedding vector is 16-dimensional. The two are concatenated to obtain a 144-dimensional equipment feature vector. The concatenation operation uses a vector concatenation method to ensure the effective integration of type information with the inherent characteristics of the equipment.

[0083] In the host island system, physical connections between devices primarily include fluid connections, thermal connections, mechanical connections, and electrical connections. For each connection type, an 8×8 parameter matrix is ​​constructed to capture the characteristic interaction patterns of the different connection types. For example, the parameter matrix for fluid connections focuses on flow and pressure characteristics, while the parameter matrix for thermal connections focuses on temperature and heat flow characteristics. Specifically, for an edge e connecting two devices, the device feature vectors h_i and h_j of the two devices are taken, the edge type is t, and the corresponding edge type parameter matrix is ​​W_t. Matrix multiplication is performed to calculate the edge type-aware attention score a_ij = (the transpose of h_i·W_t·h_j). The softmax function is then used to normalize the attention scores of all edges connected to device i. The normalized attention scores are then multiplied by the edge attribute feature vector to obtain the edge type feature. For example, for the fluid connection edge between the main pump and the heat exchanger, the calculated attention score is 0.28. This is multiplied by the 32-dimensional edge attribute feature vector of the edge to obtain the weighted edge type feature. In the actual system, a total of four edge type parameter matrices are constructed, occupying 256 parameters, which can effectively distinguish the influence patterns of different types of edges.

[0084] A 32×32 transformation matrix is ​​designed for each edge type. Matrix multiplication is used to transform edge type features into a new feature space. Nonlinearity is introduced using the LeakyReLU activation function to obtain the transformed features. For example, after relational transformation, the fluid connection edge type features highlight pressure gradient and flow balance characteristics. A two-layer multilayer perceptron (MLP) with 16 hidden units and 1 output unit is trained based on the transformed features and the feature vectors of adjacent devices to learn the importance of different types of transformation features. The output of the MLP is processed with a Sigmoid function to obtain gating coefficients in the range [0, 1]. For example, the gating coefficients for the fluid connection and thermal connection between a pair of devices are 0.75 and 0.45, respectively, indicating that the fluid connection has a greater impact on the relationship between the devices. The gating coefficients are used to weight the different types of transformation features and average them to obtain the overall impact weight between the devices.

[0085] The device placement order is determined using a weighted ranking method, comprehensively considering three factors: inter-device influence weights, node-level attention values, and feature sensitivity. For each device, a comprehensive score is calculated: score = 0.5 × node-level attention value + 0.3 × maximum influence weight + 0.2 × weighted sum of feature sensitivity. The node-level attention value, or influence metric in the node-level attention heatmap, is a numerical indicator generated by aggregating the influence weights of all edges connected to that node. It represents the overall importance of a device within the entire system network. For example, the node-level attention value of the main pump is 4.28, indicating that it has high connection importance within the entire device system. The maximum influence weight refers to the maximum strength of the association between the current device and any other device, reflecting the critical connection between the devices. For example, the influence weight between the main heat exchanger and the main pump is 0.82, which is higher than the connection weights with other devices, so its maximum influence weight is 0.82. The weighted sum of feature sensitivity considers the impact of each device parameter on the system. The sensitivity of each device feature dimension is multiplied by the corresponding weight coefficient and then summed. After using this method to score 10 devices on a host island, the main pump received a comprehensive score of 3.92, the main heat exchanger received a score of 3.61, and the pressure regulator received a score of 3.22. The final equipment layout order was determined by sorting them from high to low based on the comprehensive scores, ensuring that key and closely related equipment were prioritized, reducing constraint conflicts during subsequent layout.

[0086] Figure 2 This chart compares the device placement results of different placement sequence methods. The horizontal axis represents placement steps (1-9), and the vertical axis represents the cumulative number of constraint conflicts (0-60). The four methods in the chart include: the present invention (edge-type perceptual graph attention network), represented by circular data points and a solid line; the degree centrality heuristic method, represented by square data points and a solid line; the expert rule-of-experience method, represented by triangular data points and a solid line; and the conventional topological sorting method, represented by square data points and a dashed line. The results clearly show that after completing nine placement steps, the present invention's method only generated 27 constraint conflicts, significantly outperforming the other three methods. The degree centrality heuristic method generated 42 conflicts, the expert rule-of-experience method generated 47 conflicts, and the conventional topological sorting method generated the most, 51 conflicts. The chart shows that the present invention's method maintains a low conflict growth rate starting from the first step, and the gap with the other methods gradually widens as the number of placement steps increases. The present invention's method particularly demonstrates a clear advantage in the later stages of placement (steps 7-9), with the number of conflicts growing relatively slowly, while the other methods exhibit a more abrupt increase. This shows that the device layout order optimization method based on the edge type perception graph attention network proposed in this invention can more effectively consider various physical associations between devices, thereby generating a more reasonable layout order and significantly reducing constraint conflicts.

[0087] The present invention effectively solves the modeling problem of multiple types of physical associations in the host island device layout by integrating the graph attention network and the edge type perception mechanism; by introducing type embedding and relationship conversion matrix, the expression ability of device functional characteristics and association types is enhanced; the gated fusion mechanism is used to realize the adaptive integration of different physical associations, overcoming the limitation of the single physical relationship expression in traditional methods; combined with node attention and feature sensitivity analysis, the layout order takes into account both system structure importance and parameter sensitivity.

[0088] In an optional embodiment, an octree structure is used to adaptively discretize the equipment layout space into a multi-resolution three-dimensional grid. A reinforcement learning method with a multi-scale optimization strategy is applied to optimize the equipment spatial position. The translation and rotation parameters of the equipment are used as the action space, and collision avoidance, clearance maintenance, and maintenance accessibility are optimization objectives. The optimal position and posture parameters obtained by iterative calculation include:

[0089] The octree structure is used to adaptively discretize the equipment layout space, and recursive subdivision is performed in the equipment boundary area and the equipment gap area to obtain a multi-resolution three-dimensional grid.

[0090] Constructing a state space for reinforcement learning based on the multi-resolution three-dimensional grid, the state space including position and attitude parameter states, grid occupancy state, and constraint state, and an action space including translation and rotation parameters of the device, wherein a movement step size in the action space is proportional to a grid size at the corresponding grid level;

[0091] A reinforcement learning method is applied to optimize equipment positions on multi-resolution three-dimensional grids. The reinforcement learning method adopts a multi-scale optimization strategy. At the coarse grid level, a global position search is performed using a reward function that includes a collision penalty term and a gap reward term to obtain the initial equipment layout position. At the fine grid level, a local fine optimization is performed by adding a maintenance accessibility assessment term to the reward function, and the initial optimal position and posture parameters are iteratively obtained.

[0092] Constraint verification is performed on the multi-resolution three-dimensional grid based on the initial optimal position and attitude parameters, including: checking pipeline connection constraints, equipment spacing constraints and support structure constraints; verifying the accessibility of maintenance channels, lifting paths and operating spaces, and generating the final optimal position and attitude parameters after verification.

[0093] For example, taking an equipment layout space that is 10 meters long, 8 meters wide, and 6 meters high as an example, the entire space is initially divided into a 0th-level grid with a size of 10×8×6 meters. Then, adaptive subdivision is performed according to the preset subdivision criteria. The subdivision criteria can be set as follows: if the current grid intersects with the device boundary, or is located between two devices and the distance between them is less than a preset threshold, the grid is subdivided into 8 equal subgrids. For example, the 0th-level grid containing the device boundary is subdivided for the first time to obtain 8 1st-level subgrids with a size of 5×4×3 meters; the 1st-level grid that still contains the device boundary or the device gap area is further subdivided to obtain a 2nd-level subgrid with a size of 2.5×2×1.5 meters. In actual applications, for complex equipment, up to 5 levels of subdivision can be performed, and the minimum grid size can reach 1 / 32 of the initial grid size to ensure a sufficiently detailed description of the key areas.

[0094] The state space for reinforcement learning is constructed based on a multi-resolution three-dimensional grid. The position and attitude parameter states record the current position coordinates (x, y, z) and rotation angles (α, β, γ) of each device. The grid occupancy state records the occupancy of each grid cell in the multi-resolution grid, with a value of 0 indicating idle and a positive integer indicating occupied by a device, with the occupancy value equal to the device number. The constraint state records the satisfaction of various constraints, including distance constraints, connection constraints, and support constraints. The action space includes the translation and rotation parameters of the device. Translational actions are defined as movements along the x, y, and z axes, with a step size proportional to the grid size of the current grid level. For example, at the second grid level, the step size is 2.5 meters in the x direction, 2 meters in the y direction, and 1.5 meters in the z direction. Rotational actions are defined as rotations around the x, y, and z axes, with a step size of 30 degrees at the coarse grid level and 10 degrees at the fine grid level.

[0095] Reinforcement learning was applied to optimize device placement on a multi-resolution three-dimensional grid. This method employed a multi-scale optimization strategy and consisted of two phases: coarse and fine optimization. A global position search was performed on the coarse grid level (layers 0-2), guided by a reward function consisting of a collision penalty and a gap reward. The collision penalty was set to a penalty of -100 for detected collisions between devices or between devices and their boundaries. The gap reward was set to a positive reward, up to +50, related to the spacing between devices when the distance between them was greater than the minimum safe distance (set at 0.5 meters) but less than the maximum allowed distance (set at 2 meters). Training was performed using a Q-learning algorithm with a learning rate of 0.1 and a discount factor of 0.9. After 500 iterations of the coarse optimization phase, a globally optimal initial device placement was achieved.

[0096] Then, local fine optimization is performed on the fine grid level (layers 3-5), and a maintenance accessibility evaluation item is added to the reward function. The maintenance accessibility evaluation uses the A* path planning algorithm to check the feasible path from the entry point to each equipment maintenance point. If the path exists and the length is less than the preset threshold, a positive reward is given. The reward value is inversely proportional to the path length, up to +80; if there is no feasible path, a penalty of -50 is given. For example, for a valve device that requires regular maintenance, its maintenance point is set to be located on the north side of the equipment. The path length from the entrance to the maintenance point is required to be no more than 15 meters, and the path width is not less than 0.8 meters. By setting the moving step size of the fine grid level to 0.3 meters and the rotation step size to 5 degrees, fine adjustments are made and a total of 1000 iterations are performed to finally obtain the initial optimal position and posture parameters.

[0097] Constraint verification is performed on a multi-resolution 3D mesh based on the initial optimal position and posture parameters. Constraint verification includes two aspects: functional constraint verification and accessibility verification. Functional constraint verification includes checking pipe connection constraints, equipment spacing constraints, and support structure constraints. Pipe connection constraints require that pipes can be routed between the interface points of connected equipment, and the number of pipe bends does not exceed the preset value of three. Equipment spacing constraints require that the minimum distance between equipment is no less than a safe clearance of 0.5 meters. Support structure constraints require that heavy equipment (weighing over 500 kg) has sufficient support area at the bottom (at least 80% of the equipment bottom area). Accessibility verification includes verifying the accessibility of maintenance access, lifting paths, and operating spaces. Maintenance access verification requires a passageway with a width of at least 0.8 meters from the entry point to the maintenance point of each equipment. Lifting path verification requires unobstructed vertical lifting space above large equipment (over 1.5 meters in any dimension). Operating space verification requires that an operating space of at least 1 meter by 1 meter be reserved around equipment requiring manual operation.

[0098] After the above verification, if it is found that a constraint is not satisfied, return to the third step for local adjustment; if all constraints are satisfied, the final optimal position and posture parameters are generated to complete the equipment layout optimization process.

[0099] The present invention realizes adaptive discretization of equipment layout space through octree structure, creates multi-resolution three-dimensional grid, and concentrates computing resources in key areas; combines the multi-scale optimization strategy of reinforcement learning, realizes global position search at the coarse grid level, and performs fine optimization at the fine grid level, effectively balancing the optimization range and accuracy, while integrating multi-objective constraints such as collision avoidance, gap maintenance and maintenance accessibility, and guiding layout optimization through hierarchical reward function; overcomes the problems of low computational efficiency and difficulty in handling complex constraints of traditional layout methods, while ensuring the rationality of layout, significantly improves layout efficiency, reduces manual intervention, and provides new ideas for the intelligent layout of industrial equipment.

[0100] In an optional embodiment, the method for adjusting the weights of each item in the reward function includes:

[0101] A hierarchical adaptive reward mechanism is designed to determine the weight of the reward function according to the degree of constraint violation during the optimization process: when the collision volume between devices is greater than the first preset threshold, the weight of the collision penalty item is adjusted to a preset multiple of the current weight; when the equipment gap is less than the second preset threshold, the weight of the gap reward item is adjusted to a preset multiple of the current weight; when the number of accessible paths in the maintenance channel is less than the third preset threshold, the weight of the maintenance accessibility assessment item is adjusted to a preset multiple of the current weight; when the degree of violation of each constraint is reduced to below the preset proportion of the corresponding preset threshold, the corresponding weight is restored to the initial setting value.

[0102] For example, in the initial phase, initial weights are set for each item in the reward function. Specifically, the initial weight of the collision penalty term can be set to 10.0, the initial weight of the equipment clearance reward term can be set to 5.0, and the initial weight of the maintenance accessibility assessment term can be set to 8.0. These initial weights can be adjusted appropriately based on the specific application scenario to reflect the relative importance of different constraints.

[0103] The hierarchical adaptive weight adjustment mechanism mainly includes the following specific steps:

[0104] Adjusting the weight of the inter-device collision penalty: After each layout evaluation, the sum of the collision volumes between all device pairs in the current layout is calculated. If the sum of the collision volumes is greater than a first preset threshold (e.g., 100 cubic centimeters), the collision penalty weight is adjusted to a preset multiple of the current weight. This preset multiple can be set to 2.0, meaning that when a severe collision is detected, the collision penalty weight is increased from its current value to twice its current value. For example, if the current weight is 10.0, it will be adjusted to 20.0, thereby increasing the algorithm's propensity to avoid device collisions.

[0105] Adjusting the weight of the device clearance bonus: When evaluating the current layout solution, calculate the minimum clearance distance between all adjacent device pairs. If the minimum clearance distance is less than a second preset threshold (e.g., 50 mm), adjust the clearance bonus weight to a preset multiple of the current weight, which can be set to 2.0.

[0106] To adjust the weight of the maintenance accessibility assessment item: Count the number of accessible paths for the maintenance corridor in the current layout plan. If the number of accessible paths is less than a third preset threshold (for example, for a system with 10 key maintenance points, the threshold can be set to 7), the maintenance accessibility assessment item weight is adjusted to a preset multiple of the current weight, which can be set to 2.0.

[0107] In the actual implementation process, it is also necessary to define the judgment criteria for the reduction of constraint violation degree. When the violation degree of each constraint is reduced to below the preset proportion of the corresponding preset threshold, the corresponding weight will be restored to the initial setting value. Specifically:

[0108] When the collision volume between devices decreases to less than 30% of the first preset threshold (100 cubic centimeters), that is, less than 30 cubic centimeters, the collision penalty weight is restored to the initial value of 10.0;

[0109] When the minimum gap between devices increases to more than 150% of the second preset threshold (50 mm), that is, greater than 75 mm, the gap bonus weight is restored to the initial value of 5.0;

[0110] When the number of accessible paths to the maintenance channel increases to more than 120% of the third preset threshold (7 paths), that is, greater than or equal to 9 paths, the weight of the maintenance accessibility evaluation item is restored to the initial value of 8.0.

[0111] To ensure algorithm robustness, the weight adjustment process should also include upper and lower bounds. For example, the upper bound for the collision penalty weight could be set to 5 times the initial weight (i.e., 50.0), while the upper bounds for the clearance reward and accessibility assessment weights could be set to 3 times the initial weights (i.e., 15.0 and 24.0, respectively). Furthermore, the lower bounds for all weights should not be lower than their initial values.

[0112] Through a hierarchical adaptive reward mechanism, the optimization algorithm of the present invention can dynamically adjust the weights of various evaluation indicators according to the constraint violations at different stages, effectively improving the efficiency of layout optimization and the quality of results. It is suitable for equipment layout optimization problems in complex engineering environments, especially for scenarios where multiple constraints need to be met simultaneously.

[0113] In an optional embodiment, the equipment arrangement sequence, optimal position and posture parameters are used as an arrangement scheme, and the space-time envelope requirements for evaluating equipment maintenance operations are evaluated through three-dimensional dynamic simulation.

[0114] Constructing a three-dimensional geometric structure of the device layout scene based on the device layout sequence, optimal position, and posture parameters, where the optimal position parameter is the coordinate value of the device in three-dimensional space, and the posture parameter is the rotation matrix of the device in three-dimensional space;

[0115] Acquiring a device maintenance operation trajectory, dividing the device maintenance operation trajectory into an operation sequence, and establishing a human joint motion parameter set and a tool operation parameter set based on the operation sequence;

[0116] Calculating the motion trajectory during the maintenance operation based on the human joint motion parameter set and the tool operation parameter set, and generating a spatiotemporal envelope of the maintenance operation;

[0117] Performing interference detection on the maintenance operation space-time envelope surface and the three-dimensional geometric structure, and calculating a minimum distance value and a collision volume value during the maintenance operation;

[0118] The accessibility of the maintenance operation is determined according to the minimum distance value and the collision volume value, and a feasibility assessment result of the equipment layout plan is generated.

[0119] For example, a three-dimensional geometric structure of the equipment layout scenario is constructed based on the equipment layout order, optimal position, and attitude parameters. Equipment layout is arranged in descending order of importance. For example, in the main engine island, the order is "main pump → main heat exchanger → pressurizer → condenser → feedwater pump." The optimal position parameters for each device are determined by the position of the device's center point in the three-dimensional coordinate system. These parameters consist of three components (x, y, z) and are expressed in meters. For the main pump, for example, its optimal position parameters are (12.45, 8.72, 1.50), indicating that the main pump's center is located at the corresponding position in the coordinate system with the lower left corner of the layout area as the origin. The attitude parameters use a rotation matrix to represent the device's orientation, determined by rotation angles α, β, and γ around the three coordinate axes. The rotation matrix is ​​a 3×3 matrix, derived by rotating the device by an angle α around the z-axis, by an angle β around the y-axis, and by an angle γ around the x-axis. For example, the rotation angles of the main pump are (0, 0, 45°), indicating a 45-degree rotation around the z-axis only. Based on these parameters, a complete 3D scene model is constructed using CAD modeling software or a virtual reality engine. This model includes the geometry, relative positions, and piping connections of all equipment, with millimeter-level accuracy. Technically, parametric modeling is employed, using the equipment's dimensions, positional coordinates, and rotation angles as input parameters to automatically generate the 3D geometry. The model encompasses the equipment's structure, maintenance areas, and connecting piping, down to the level of detail needed to represent key features such as the operating panel and maintenance interfaces.

[0120] Maintenance operation trajectories are captured in two ways: first, through motion capture of on-site maintenance personnel during their operations, using inertial sensors or optical tracking systems to record the sequence of movements they perform as they complete specific maintenance tasks. Second, ergonomics experts construct standardized maintenance action models based on the standard operating procedures specified in the maintenance manual. For example, routine maintenance on a heat exchanger involves: approaching the equipment → opening the access door → inspecting internal components → replacing the seals → closing the access door → exiting the equipment. This process is divided into 25 basic actions, forming an operation sequence. For each action, a human joint motion parameter set is established, containing the position and angle information of 23 key joints to reflect the maintenance personnel's posture changes. For example, during the "opening the access door" operation, the maintenance personnel's wrist, elbow, and shoulder joint angle changes, as well as hand trajectory and other parameters are recorded. A tool operation parameter set is also established to record the position, posture, and motion trajectory of the maintenance tool during the operation. For example, when using a wrench, the wrench's position coordinates, rotation angle, and force direction are recorded. These parameters are collected through a motion capture system with a sampling frequency of 60Hz. The data accuracy is better than 1 cm and 1 degree, ensuring accurate reflection of actual operating characteristics.

[0121] The human motion trajectory and tool motion trajectory are combined to generate a complete set of maintenance operation motion trajectories. Based on the motion trajectory set, the spatial envelope of the trajectory points is calculated to generate the spatiotemporal envelope of the maintenance operation.

[0122] Interference detection is performed between the maintenance operation's space-time envelope and the 3D geometric structure. This interference detection utilizes a hierarchical collision detection algorithm, consisting of two phases: coarse and fine detection. The coarse detection phase uses an axial bounding box technique to simplify the equipment and the space-time envelope into axial bounding boxes. By comparing the overlap of the bounding boxes, the system quickly eliminates unlikely collisions. The fine detection phase uses the separating axis theorem or the GJK algorithm for precise collision detection, calculating the interference between the space-time envelope and the 3D equipment model. Two key data sets are recorded during the calculation process: minimum distance and collision volume. The minimum distance represents the minimum clearance distance between the maintenance operation's space-time envelope and surrounding equipment, measured in meters. The collision volume represents the overlap volume between the space-time envelope and the equipment model, measured in cubic meters. For example, during maintenance operations on a main pump, the detection revealed that the minimum distance between the space-time envelope and the adjacent pipe support was 0.08 meters, which was below the safety clearance standard. Interference with a valve control box resulted in a collision volume of 0.015 cubic meters. These data intuitively reflect the spatial constraints and potential collision risks associated with maintenance operations.

[0123] Maintenance accessibility is determined based on pre-set evaluation criteria, including minimum clearance requirements and maximum permissible collision volume. Typically, the minimum clearance requirement is set at 0.15 meters to account for human motion errors and operator comfort, while the maximum permissible collision volume is set at 0 cubic meters, theoretically preventing any collisions. These criteria can be adjusted in practice based on project requirements. If the minimum distance is less than the minimum clearance requirement, or the collision volume is greater than the maximum permissible collision volume, the current maintenance operation is deemed unreachable and the equipment layout plan is deemed unfeasible. The system generates a detailed evaluation report, including the identification of the problematic equipment, the specific location of the collision or insufficient clearance, and a severity score. If the evaluation results do not meet the pre-set requirements, the system returns to the equipment placement optimization step for re-iteration. Specifically, the position or posture parameters of the equipment in the collision area are adjusted, the weight of the maintenance space in the optimization objective is increased, and the placement optimization algorithm is re-run. For example, to address the issues identified during the main pump maintenance operation, the system automatically adjusted the pipe support outward by 0.1 meters and raised the installation height of the valve control box by 0.3 meters, followed by re-verification. When the evaluation results meet the preset requirements, the final equipment layout plan is confirmed and output, and a complete layout drawing and 3D model including equipment location, posture, connection relationship and maintenance space are generated.

[0124] Figure 3This graph shows the results of the spatiotemporal envelope interference detection for equipment layout maintenance operations, comparing the performance of three different methods (the present invention, a traditional geometric constraint method, and a manual layout method) during equipment maintenance operations. The graph is divided into two parts: the left side shows the "minimum interference distance for maintenance operations" for each method, and the right side shows the "collision volume for maintenance operations." The left graph shows that the present invention method achieves the longest interference distance for all five equipment types (main pump, heat exchanger, pressurizer, valve group, and storage tank). For main pump maintenance operations, the present invention achieves a minimum interference distance of 37.5 cm, while the traditional geometric constraint method achieves 26.8 cm and the manual layout method only 19.3 cm. The most significant improvement is for storage tanks, where the present invention method achieves an interference distance of 44.2 cm, 16.1 cm higher than the manual layout method. The right graph compares the collision volumes of the various methods. The present invention method achieves the smallest collision volume for all equipment, particularly for pressurizers, where the present invention achieves a collision volume of only 378 cubic centimeters, compared to 1123 cubic centimeters for the traditional method and 1758 cubic centimeters for the manual layout method. On storage tank equipment, the proposed method reduced the collision volume to 158 cubic centimeters, a 78.8% reduction compared to traditional methods and an 88.5% reduction compared to manual placement. Overall, the proposed method reduced the collision volume by an average of 78.6%, significantly improving the safety and convenience of maintenance operations. These results demonstrate that the proposed three-dimensional dynamic simulation verification method, based on equipment placement sequence, optimal position, and posture parameters, can effectively improve the feasibility of equipment maintenance operations and reduce interference risks.

[0125] This invention achieves a precise assessment of the feasibility of equipment maintenance operations through three-dimensional dynamic simulation and space-time envelope surface analysis. Compared to traditional static spatial analysis methods, this approach considers the spatial requirements of maintenance personnel during dynamic operations, improving the accuracy and reliability of the assessment. Quantitative analysis of minimum distance and collision volume values ​​provides clear improvement directions for equipment layout optimization, effectively addressing the issue of insufficient maintenance space in densely populated industrial equipment environments. This invention not only improves the rationality of equipment layout but also significantly reduces the safety risks and time costs of maintenance operations, providing strong support for the full lifecycle design of complex industrial systems.

[0126] In an optional embodiment, calculating the motion trajectory during the maintenance operation based on the human joint motion parameter set and the tool operation parameter set to generate the maintenance operation spatiotemporal envelope surface includes:

[0127] Constructing a velocity field function of the maintenance operation trajectory according to the human joint motion parameter set and the tool operation parameter set, wherein the velocity field function includes a maintenance difficulty function, a curvature term, and a global constraint term, wherein the maintenance difficulty function is related to the joint motion range and the tool operation torque;

[0128] Driving the level set evolution based on the velocity field function to generate an initial space-time envelope surface;

[0129] Establishing a multi-scale uncertainty distribution for the operation trajectory, including: establishing a Gaussian distribution based on the standard deviation of the human joint positions and the standard deviation of the tool positions, calculating the joint probability distribution at each sampling moment, the joint probability distribution characterizing the spatial distribution of the human-tool system at that moment, and accumulating the uncertainty characteristics of the complete operation trajectory in the time dimension;

[0130] The expansion coefficient of the initial spatiotemporal envelope surface is dynamically adjusted according to the uncertainty characteristics. The expansion coefficient is composed of a reference value and an uncertainty correction value. The uncertainty correction value is the product of the larger value of the joint position standard deviation and the tool position standard deviation and the proportional coefficient related to the operation difficulty, so as to generate the final maintenance operation spatiotemporal envelope surface.

[0131] For example, a velocity field function is constructed based on a set of human joint motion parameters and a set of tool operation parameters. This function describes the motion trend of the maintenance operation in three-dimensional space. The velocity field function consists of three components: a maintenance difficulty function, a curvature term, and a global constraint term. The maintenance difficulty function is directly related to the joint range of motion and the tool operating torque. It is calculated as the weighted sum of the degree to which the joint angle deviates from the comfortable position and the corresponding joint importance weight, plus the weighted sum of the tool operating torque and the tool importance weight. For example, the shoulder joint is assigned a weight of 0.3, the elbow 0.25, and the wrist 0.2. When the shoulder joint angle deviates by 45 degrees from the comfortable angle, the difficulty value for that joint is 0.3 × (45 / 90) = 0.15. The curvature term smoothes the trajectory to prevent abrupt changes and is calculated as the curvature of the trajectory curve at each point multiplied by a smoothing coefficient. The global constraint term accounts for environmental constraints such as narrow passages or fixed obstacles and is represented by a distance field. The closer the distance to the obstacle, the greater the constraint value. The three components are combined in a weighted ratio of 6:3:1 to form the complete velocity field function.

[0132] The level set evolution is driven by the velocity field function to generate an initial space-time envelope. The level set method is an implicit surface representation technique that achieves surface evolution by solving partial differential equations. Technically, the maintenance operation space is first discretized into a three-dimensional grid with a mesh resolution of 10 cm, covering the maintenance area. For example, for a heat exchanger, the maintenance area is set to a spatial extent of 3 meters by 2 meters by 2.5 meters. An initial level set function is defined on this grid. Typically, a sphere centered at the maintenance worker's starting position is selected as the initial level set, with a radius of 0.5 meters to represent the space occupied by the human body. Then, based on the velocity field function constructed above, the level set evolution equation is solved iteratively. During the iteration process, the level set function expands along the velocity field, expanding more slowly in areas with high maintenance difficulty and more rapidly in areas with less difficulty. The iteration step size is set to 0.1, and the maximum number of iterations is 200. By visualizing the zero isosurface of the level set function, the initial space-time envelope is obtained. This envelope reflects the approximate spatial extent of the maintenance operation but does not yet account for operational uncertainty. The calculation results of a main pump maintenance operation show that the initial space-time envelope surface presents an irregular shape, shrinking in narrow spaces and expanding in areas with sufficient operating space, with a total volume of approximately 7.5 cubic meters.

[0133] A multi-scale uncertainty distribution for the manipulation trajectory is established. Uncertainty primarily stems from the natural variability of human motion, the precision limitations of tool manipulation, and environmental factors. For each joint, the standard deviation of its position is calculated by analyzing historical manipulation data or the differences between repeated manipulations. For example, the standard deviation of shoulder joint position is 2 cm, that of the elbow is 3 cm, that of the wrist is 4 cm, and that of the fingers is 5 cm. This standard deviation increases with joint flexibility. For tool position, the standard deviation is determined based on tool type and precision requirements. For example, the standard deviation for precision tools is 1 cm, for general maintenance tools is 3 cm, and for heavy tools is 5 cm. A Gaussian distribution model is constructed based on these standard deviations, assuming that the variability of joint and tool positions follows a normal distribution. For each sampling moment, the joint probability distribution of the joint and tool positions is calculated, characterizing the possible spatial distribution of the human-tool system at that moment. Technically, a Monte Carlo method is employed to generate 1,000 random samples that conform to the joint distribution at each moment to simulate possible positional variability. In the temporal dimension, the spatial distributions of all sampling moments are cumulatively superimposed to obtain the uncertainty characteristics of the complete manipulation trajectory. This accumulation adopts the maximum value strategy, that is, for each point in space, the maximum value of the probability distribution at all times is taken as the final uncertainty value of the point.

[0134] The expansion coefficient determines the extent to which the spatiotemporal envelope surface expands relative to the initial trajectory, directly affecting its size and shape. The expansion coefficient consists of a baseline value and an uncertainty correction value. The baseline value is set based on the safety level of the maintenance operation. For example, for general maintenance operations, it is set to 1.2, indicating a 20% basic expansion; for precision maintenance operations, it is set to 1.1; and for heavy or high-risk maintenance operations, it is set to 1.3. The uncertainty correction value is the product of the larger of the joint position standard deviation and the tool position standard deviation and the operation difficulty-related scaling factor. The operation difficulty-related scaling factor is determined by the average value of the maintenance difficulty function. The coefficient increases with difficulty and typically ranges from 0.5 to 2.0. For example, in a precision valve adjustment operation, due to high operational accuracy requirements, the tool position standard deviation of 1 cm is smaller than the finger joint position standard deviation of 4 cm. Therefore, the larger value of 4 cm is used. The operation difficulty is assessed as 0.7 (medium difficulty), corresponding to a difficulty-related scaling factor of 1.2. The uncertainty correction value is 4 cm × 1.2 = 4.8 cm. The final expansion coefficient is the baseline value of 1.1 plus the uncertainty correction of 4.8 cm and is applied to the initial spatiotemporal envelope surface. Each point on the initial envelope is expanded along the normal direction by an expansion coefficient, with the expansion distance proportional to the uncertainty characteristic of that point. Regions of high uncertainty (such as hand manipulation) are expanded more, while regions of low uncertainty (such as standing) are expanded less. Through this dynamic adjustment, the generated spatiotemporal envelope of maintenance operations more accurately reflects actual maintenance needs, avoiding space waste or safety risks.

[0135] This invention combines ergonomic principles with computational geometry methods to accurately generate the space-time envelope of maintenance operations by constructing velocity field functions, level set evolution, and uncertainty analysis. Compared to traditional static space reservation methods, this invention considers the dynamic operational characteristics of the human body and uncertainty factors, and can more accurately reflect actual maintenance needs. By introducing a maintenance difficulty function and a dynamic expansion coefficient, the space-time envelope ensures sufficient space in critical operating areas while avoiding excessive space reservation in non-critical areas, achieving a balance between space utilization and maintenance safety.

[0136] According to a second aspect of an embodiment of the present invention, an electronic device is provided, including:

[0137] processor;

[0138] a memory for storing processor-executable instructions;

[0139] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0140] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0141] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent optimization method for host island equipment layout based on 3D modeling, characterized in that: include: Establish a three-dimensional model of the host island and apply feature recognition and semantic segmentation to extract global feature data of the device, including the geometric and physical features of the device and the constraint relationships between the devices; Based on the global feature data, a device relationship graph for multi-physics coupling is constructed, in which the devices are nodes, the physical associations between the devices are edges, and the edge attributes include fluid properties and heat transfer properties. The graph attention network is used to analyze the influence weights between the devices to obtain the device layout order; An octree structure is used to adaptively discretize the equipment layout space into a multi-resolution three-dimensional grid. A reinforcement learning method with a multi-scale optimization strategy is applied to optimize the equipment's spatial position. The equipment's translation and rotation parameters are used as the action space, with collision avoidance, clearance maintenance, and maintenance accessibility as optimization objectives. The optimal position and attitude parameters are obtained through iterative calculation. The equipment layout sequence, optimal position and posture parameters are used as a layout plan, which is verified through three-dimensional dynamic simulation to evaluate the spatiotemporal envelope requirements of equipment maintenance operations; after verification, it is confirmed as the final equipment layout plan.

2. The method according to claim 1, characterized in that Based on the global feature data, a multi-physics field coupled device relationship graph is constructed, in which the devices are nodes and the physical associations between devices are edges. The edge attributes include fluid properties and heat transfer properties. The graph attention network is used to analyze the influence weights between devices to obtain the device layout order, which includes: Constructing a node feature vector based on the geometric features and physical features, wherein the node feature vector includes spatial dimension parameters, position and posture parameters, and operation parameters of the device; Determine the physical association between devices based on the constraint relationships between them, and use the pipeline connection relationships and support structure relationships as edges to establish topological connections between devices; Constructing edge attributes based on fluid characteristics and heat transfer characteristics between devices, including: constructing a pressure loss matrix of fluid characteristics based on the Darcy-Weisbach equation, constructing a thermal resistance matrix of heat transfer characteristics based on thermal conductivity and Nusselt number, calculating the physical field strength of the fluid field and thermal field according to equipment operating parameters, normalizing the pressure loss matrix and thermal resistance matrix, and weightedly fusing them according to the ratio of physical field strengths to obtain a comprehensive feature matrix; performing a consistency check on the comprehensive feature matrix based on mass conservation and energy conservation, and extracting the edge feature vectors to form an edge attribute feature set; The graph attention network is used to calculate the influence weights between devices, and the node feature vectors, topological connections and edge attribute feature sets are integrated into the edge type perception mechanism and gated fusion to obtain the device layout order.

3. The method according to claim 2, characterized in that The graph attention network is used to calculate the influence weights between devices. The node feature vectors, topological connections, and edge attribute feature sets are integrated into the edge type perception mechanism and gated fusion to obtain the device placement order, which includes: Divide the equipment into functional types and construct a type embedding matrix, which is then concatenated with the node feature vector to obtain the equipment feature vector. Constructing an edge type parameter matrix for the physical associations between different types of devices, calculating an edge type perception attention score based on the device feature vector and the edge type parameter matrix in combination with the topological connection, and weighting the score with the edge attribute feature vector to obtain an edge type feature; Constructing a relationship conversion matrix to perform nonlinear conversion on the edge type features to obtain conversion features, training a multilayer perceptron based on the conversion features and device feature vectors of adjacent devices to obtain gating coefficients, and using the gating coefficients to adaptively fuse different types of conversion features to obtain inter-device influence weights; Constructing a node-level attention heat map based on the influence weights between the devices, and performing perturbation analysis on the device feature vectors to obtain feature sensitivity; The device placement order is determined based on the influence weights between devices, the influence index of the node-level attention heat map, and the feature sensitivity.

4. The method according to claim 1, wherein The equipment layout space is adaptively discretized into a multi-resolution three-dimensional grid using an octree structure. A reinforcement learning method with a multi-scale optimization strategy is applied to optimize the equipment spatial position. The equipment's translation and rotation parameters are used as the action space, and collision avoidance, clearance maintenance, and maintenance accessibility are the optimization goals. The optimal position and posture parameters obtained by iterative calculation include: The octree structure is used to adaptively discretize the equipment layout space, and recursive subdivision is performed on the equipment boundary area and the equipment gap area to obtain a multi-resolution three-dimensional grid. Constructing a state space for reinforcement learning based on the multi-resolution three-dimensional grid, the state space including position and attitude parameter states, grid occupancy state, and constraint state, and an action space including translation and rotation parameters of the device, wherein a movement step size in the action space is proportional to a grid size at the corresponding grid level; A reinforcement learning method is applied to optimize equipment positions on multi-resolution three-dimensional grids. The reinforcement learning method adopts a multi-scale optimization strategy. At the coarse grid level, a global position search is performed using a reward function that includes a collision penalty term and a gap reward term to obtain the initial equipment layout position. At the fine grid level, a local fine optimization is performed by adding a maintenance accessibility assessment term to the reward function, and the initial optimal position and posture parameters are iteratively obtained. Constraint verification is performed on the multi-resolution three-dimensional grid based on the initial optimal position and attitude parameters, including: checking pipeline connection constraints, equipment spacing constraints and support structure constraints; verifying the accessibility of maintenance channels, lifting paths and operating spaces, and generating the final optimal position and attitude parameters after verification.

5. The method according to claim 4, characterized in that The weight adjustment methods for each item in the reward function include: A hierarchical adaptive reward mechanism is designed to determine the weight of the reward function according to the degree of constraint violation during the optimization process: when the collision volume between devices is greater than the first preset threshold, the weight of the collision penalty item is adjusted to a preset multiple of the current weight; when the equipment gap is less than the second preset threshold, the weight of the gap reward item is adjusted to a preset multiple of the current weight; when the number of accessible paths in the maintenance channel is less than the third preset threshold, the weight of the maintenance accessibility assessment item is adjusted to a preset multiple of the current weight; when the degree of violation of each constraint is reduced to below the preset proportion of the corresponding preset threshold, the corresponding weight is restored to the initial setting value.

6. The method according to claim 1, characterized in that The equipment layout sequence, optimal position, and posture parameters are used as the layout plan. Three-dimensional dynamic simulation is used to verify and evaluate the spatiotemporal envelope requirements of equipment maintenance operations, including: Constructing a three-dimensional geometric structure of the device layout scene based on the device layout sequence, optimal position, and posture parameters, where the optimal position parameter is the coordinate value of the device in three-dimensional space, and the posture parameter is the rotation matrix of the device in three-dimensional space; Acquiring a device maintenance operation trajectory, dividing the device maintenance operation trajectory into an operation sequence, and establishing a human joint motion parameter set and a tool operation parameter set based on the operation sequence; Calculating the motion trajectory during the maintenance operation based on the human joint motion parameter set and the tool operation parameter set, and generating a spatiotemporal envelope of the maintenance operation; Performing interference detection on the maintenance operation space-time envelope surface and the three-dimensional geometric structure, and calculating a minimum distance value and a collision volume value during the maintenance operation; The accessibility of the maintenance operation is determined according to the minimum distance value and the collision volume value, and a feasibility assessment result of the equipment layout plan is generated.

7. The method according to claim 6, characterized in that Calculating the motion trajectory during the maintenance operation based on the human joint motion parameter set and the tool operation parameter set, and generating the spatiotemporal envelope of the maintenance operation includes: Constructing a velocity field function of the maintenance operation trajectory according to the human joint motion parameter set and the tool operation parameter set, wherein the velocity field function includes a maintenance difficulty function, a curvature term, and a global constraint term, wherein the maintenance difficulty function is related to the joint motion range and the tool operation torque; Driving the level set evolution based on the velocity field function to generate an initial space-time envelope surface; Establishing a multi-scale uncertainty distribution for the operation trajectory, including: establishing a Gaussian distribution based on the standard deviation of the human joint positions and the standard deviation of the tool positions, calculating the joint probability distribution at each sampling moment, the joint probability distribution characterizing the spatial distribution of the human-tool system at that moment, and accumulating the uncertainty characteristics of the complete operation trajectory in the time dimension; The expansion coefficient of the initial spatiotemporal envelope surface is dynamically adjusted according to the uncertainty characteristics. The expansion coefficient is composed of a reference value and an uncertainty correction value. The uncertainty correction value is the product of the larger value of the joint position standard deviation and the tool position standard deviation and the proportional coefficient related to the operation difficulty, so as to generate the final maintenance operation spatiotemporal envelope surface.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images