Ship cable layout design method based on physical-topological graph neural network

Abstract

The invention discloses a ship cable layout design method based on a physical-topological graph neural network, and relates to the technical field of crossing of ship engineering design and artificial intelligence. The method comprises the following steps: firstly, constructing a physical-topological heterogeneous enhancement graph comprising an electrical entity topological edge and an electromagnetic interference virtual exclusion edge; secondly, circuit logic connection features and physical field interference features are extracted in parallel through a heterogeneous graph neural network, and adaptive fusion of the features is achieved through an electromagnetic potential energy attention mechanism; then, introducing a physical rejection loss function to guide model pre-training, and internalizing the electromagnetic compatibility constraint into a geometric distance on a characteristic manifold; and finally, mapping the fusion features into cabin coordinates through a space decoder, and directly generating a layout scheme. According to the method, the problem of conflict between physical isolation and wiring cost in ship cable laying and equipment layout is effectively solved, a design mode is spanned from manual trial and error to intelligent generation, and the compliance and economical efficiency of a layout scheme are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of marine engineering design and artificial intelligence, and in particular relates to a method for designing ship cable layout based on physical-topology graph neural networks. Background Technology

[0002] As the global shipping industry accelerates its transformation towards green and intelligent technologies, modern ships are gradually evolving from traditional mechanical propulsion to integrated electric propulsion systems. Integrated electric systems achieve comprehensive energy management for both propulsion and daily loads through unified power distribution across the entire ship, significantly improving energy efficiency and maneuverability. Therefore, given the thousands of electrical nodes within the system, including generator sets, switchboards, frequency converters, propulsion motors, and numerous sensors, the scientific planning of the ship's cable network system is crucial to handling the exponentially increasing system size and extremely complex layout topology.

[0003] It is worth noting that within the limited space of a ship's cabins, the design of cable laying and equipment layout presents a severe challenge of "multi-objective conflict" due to the unique complexity of ship cable networks. On the one hand, as a key aspect of shipbuilding cost control, designers need to optimize equipment locations as much as possible to shorten the total cable length; on the other hand, high-voltage interference sources and low-voltage sensitive equipment coexist in the ship's electrical network, and electromagnetic compatibility standards must be strictly followed between the two. This contradictory requirement of "both compact to save cables and dispersed to prevent interference" makes the solution space for the layout design problem extremely fragmented and constrained, placing higher demands on intelligent layout design strategies for ship circuits.

[0004] For a long time, the layout design of ship cable networks has mainly relied on manual interactive design by designers using CAD (Computer-Aided Design) software. This method not only heavily depends on the personal experience of engineers, but also often gets bogged down in a tedious iterative process of "layout-verification-modification" when dealing with a massive number of equipment nodes, resulting in extremely low efficiency. To overcome the limitations of manual design, in recent years, academia and industry have begun to try to introduce metaheuristic algorithms, represented by the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) proposed by Professor Deb, attempting to automate the layout through evolutionary search strategies. However, when faced with the high-dimensional complexity and strict electromagnetic compatibility constraints of modern ship cable network systems, such algorithms have exposed obvious "physical semantic blind spots" in practical applications: due to the inherent lack of perception of physical field effects and the reliance on penalty functions to passively handle constraints, random searches driven solely by mathematical fitness functions often lead to blind iteration directions and low convergence efficiency. Therefore, existing technologies are still insufficient to meet the dual stringent requirements of engineering compliance and intelligent design for modern ship cable laying and equipment layout design.

[0005] The Physics-Topology Graph Neural Network (PNN) is a novel intelligent algorithm architecture that combines the powerful non-Euclidean data processing capabilities of graph neural networks with physical field constraints. Addressing the issue that existing graph neural network models focus only on logical topological connections while neglecting Euclidean spatial physical field effects, this architecture introduces a mechanism capable of sensing the "virtual repulsive force" between high- and low-voltage equipment. It can not only handle complex logical connections but also guide physical layout through feature representations, effectively avoiding the problem of solutions generated solely based on connection relationships failing to be implemented due to violations of safety spacing constraints. It is highly suitable for applications in the design of integrated circuits or cable laying and equipment layout with complex physical constraints.

[0006] This invention addresses the challenges of cable laying and equipment layout design in modern integrated electric propulsion vessels, where limited cabin space presents challenges related to wiring costs and electromagnetic compatibility (EMC) safety, as well as the physical semantic blind spots and convergence difficulties inherent in traditional optimization algorithms. It proposes a ship cable layout design method based on a physics-topology graph neural network. This method constructs a heterogeneous graph containing physical topological edges and virtual repulsion edges. It then uses a two-stream heterogeneous graph neural network to extract topological and physical features in parallel, and uses a physical repulsion loss function to guide model pre-training. This generates node embedding vectors with embedded physical priors, and directly generates layout schemes through spatial decoding. This solves the problem of traditional methods struggling to balance cable length and electromagnetic interference, achieving a leap from "manual trial and error" to "intelligent generation" in ship cable laying and equipment layout. Summary of the Invention

[0007] Given the increasing complexity of integrated shipboard electrical systems, and addressing the resulting engineering challenges such as large-scale layouts, severe electromagnetic interference between high- and low-voltage equipment, and stringent cable cost control, this invention proposes a shipboard cable layout design method based on a physical-topological graph neural network. This method constructs a heterogeneous graph containing physical topological edges and virtual exclusion edges, and utilizes a pre-training mechanism to internalize physical isolation criteria. This improves the layout conflicts and inefficiencies caused by the lack of physical constraint awareness in shipboard cable network design, achieving low-cost, highly compliant, and automated generation of shipboard circuit layout schemes.

[0008] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: S1: Read the ship's cable network schematic diagram and cabin environment data, construct a heterogeneous graph containing physical topological edges and virtual exclusion edges, and synchronously encode logical connections and physical interference constraints into the graph structure. S2: Utilize a dual-stream heterogeneous graph neural network to extract topological and physical features in parallel, and then adaptively fuse them through an electromagnetic potential energy attention mechanism to generate node embedding vectors with embedded physical priors. S3: Construct a composite loss function that includes topological reconstruction and physical repulsion, perform self-supervised pre-training on the model, and internalize the physical isolation criterion into the model parameters by minimizing the physical repulsion loss; S4: Extract node embedding vectors using a pre-trained model, and project them onto the ship's cabin physical coordinate system through a spatial decoding and mapping module to generate a layout scheme that meets the requirements of isolation and partitioning.

[0009] Furthermore, in step S1, the process of constructing the physical-topological heterogeneous enhancement graph includes: firstly, defining the power supply equipment, load equipment, control equipment, and switching equipment in the ship cable network system as a set of heterogeneous nodes in the graph structure. V Secondly, based on the cable network schematic diagram, a set of entity topology edges representing the actual cable connection relationships is established. E topo This process forms a logical mode. Finally, based on electromagnetic compatibility standards, the potential interference intensity between the high-voltage interference source node and the low-voltage sensitive node is calculated. When the interference intensity exceeds a preset threshold, a virtual repulsion edge is established between the two nodes, forming a physical mode, thus completing the graph. G= ( V, E topo , E virt ) The construction.

[0010] Furthermore, regarding the graph G The node feature matrix in the algorithm uses the multi-dimensional attribute vectors it contains to define the engineering semantic information contained in the electrical equipment nodes. Each equipment node corresponds to a feature vector, which includes the equipment's functional type code, rated power, voltage level, and electromagnetic susceptibility level. The dimension of the generated node feature vector is mapped through the embedding layer. ,in D It is the feature dimension required for model processing, and this feature vector serves as the input benchmark for subsequent neural networks.

[0011] Furthermore, in step S1, when establishing a virtual repulsion edge, the physical potential energy value calculated based on the minimum safe isolation distance required by the design specifications is assigned to the virtual repulsion edge as a weight attribute; this weight attribute quantifies the degree of physical repulsion that must be maintained between nodes. The larger the weight, the farther the distance that the two nodes need to be isolated in the physical layout.

[0012] Furthermore, in step S2, the constructed heterogeneous enhancement graph is input into the dual-flow heterogeneous graph neural network model, which includes parallel topology-aware flow channels and physical field-aware flow channels. The topology-aware flow channels are used to perform feature aggregation on the physical topology edges through graph convolution to learn the functional cooperation relationship between devices. The physical field-aware flow channels are used to propagate features on the virtual repulsion edges through potential energy diffusion to simulate the repulsive force in the physical field and transmit the high potential features of the interference source to the sensitive device.

[0013] Furthermore, in step S2, when introducing the electromagnetic potential energy attention mechanism to fuse dual-flow features, the process includes: calculating the sum of the weights of all virtual repulsion edges connected to each node, which is defined as the potential perturbation degree of the node; generating a dynamic attention mask based on the perturbation degree, and performing a weighted summation of the physical field-aware flow features and the topology-aware flow features; for nodes with high perturbation degree, automatically increasing the weight of the physical flow features, and for ordinary nodes with low perturbation degree, automatically increasing the weight of the topology flow features, and finally generating a node embedding vector with embedded physical priors.

[0014] Furthermore, in step S3, the process of constructing the composite loss function includes: defining a topology reconstruction loss term and a physical exclusion loss term; wherein the topology reconstruction loss term is used to supervise the model to retain the logical connection structure of the circuit in the feature space, and the physical exclusion loss term is used to supervise the model to satisfy the physical isolation constraint; the gradient of the composite loss function with respect to the network weight parameters is calculated through the backpropagation algorithm to drive the model parameter update until the embedded vector output by the model simultaneously satisfies the connection compactness and isolation compliance.

[0015] Furthermore, in step S3, a large-interval hinge loss function is used as the physical repulsion loss term. Specifically, it is defined as follows: the Euclidean distance between node pairs with virtual repulsion edges in the feature space is used as a metric, and a safe isolation threshold is set in the feature space. When the distance between the node pairs in the feature space is less than this threshold, the loss function generates an exponentially increasing penalty value; when the distance is greater than the threshold, the loss value is zero. This mechanism forces the model to "learn" to push away strongly interfering nodes and sensitive nodes in the feature space during the pre-training phase.

[0016] Furthermore, in step S4, the spatial decoding mapping module is constructed using a multilayer perceptron or a manifold learning decoder, which establishes a nonlinear mapping relationship from the high-dimensional graph feature space to the two-dimensional physical space of the ship's cabin.

[0017] Furthermore, in step S4, the process of generating the layout scheme includes: inputting the node embedding vector output in step S2 into the spatial decoding mapping module, and directly regressing to predict the centroid coordinates of each equipment node in the cabin coordinate system; since the physical repulsion loss has been internalized by step S3, the output coordinate points naturally present the form of automatic clustering of equipment where high-voltage equipment clusters and low-voltage sensitive equipment clusters are automatically separated in space and electrically connected.

[0018] Furthermore, it also includes a post-processing step for the layout scheme: using the centroid coordinates generated in step S4 as a reference, an envelope rectangle is generated in combination with the actual physical dimensions of the equipment, a collision detection algorithm based on non-overlapping constraints is executed, the coordinates of the equipment that have physical interference are fine-tuned, and finally a conflict-free ship circuit layout design drawing is output.

[0019] Compared with the prior art, the present invention has the following advantages: (1) Breakthrough in physical perception bottleneck: This invention solves the problem that traditional graph neural networks can only process topological connections and cannot perceive Euclidean space physical constraints by constructing a heterogeneous graph containing virtual repulsion edges, enabling the algorithm to understand the "repulsion" relationship between high-voltage and low-voltage equipment.

[0020] (2) Significantly improve design efficiency: By introducing a pre-training mechanism that internalizes physical constraints, the cumbersome electromagnetic compatibility specifications are transformed into the parameter "intuition" of the neural network. The generated layout scheme has extremely high compliance, avoiding a large number of invalid random searches in the infeasible solution space by traditional solution methods, and greatly shortening the layout design time.

[0021] (3) Balancing cost and safety: The dual-flow architecture and attention fusion mechanism of this invention can dynamically balance the contradictory goals of "shortening cable length" and "maintaining isolation distance", providing an end-to-end intelligent solution with strong engineering practicality for ship cable laying and equipment layout design. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of a ship cable layout design method based on a physical-topology graph neural network proposed in this invention; Figure 2 This is a schematic diagram illustrating the construction principle of the physical-topological heterogeneous enhancement graph described in this invention, showing the coexistence structure of physical topological edges and virtual repulsion edges; Figure 3 This is a network architecture diagram of the dual-flow heterogeneous graph neural network proposed in this invention, which shows the connection relationship between topology-aware flow, physical field-aware flow and electromagnetic potential energy attention mechanism. Figure 4 is a schematic diagram comparing the expected effects before and after the optimization of the ship cable layout generated by the method of the present invention. Detailed Implementation

[0023] To more clearly illustrate the purpose and technical solutions of the embodiments of the present invention, the present invention will be described below in conjunction with the accompanying drawings.

[0024] This invention addresses the electromagnetic compatibility (EMC) layout challenges arising from the coexistence of high-voltage interference sources and low-voltage sensitive equipment in the increasingly complex integrated power systems of large ships, as well as the shortcomings of traditional graph algorithms that focus only on logical topology and lack physical spatial awareness. By introducing "virtual exclusion edges" to construct a physical-topology graph and combining it with a two-stream heterogeneous graph neural network architecture, a ship cable layout design method based on a physical-topology graph neural network is proposed. The following is a detailed description of this ship cable layout design method.

[0025] Please see Figure 1 , Figure 1 This is a flowchart illustrating a ship cable layout design method based on a physical-topology graph neural network, as described in this invention. This method constructs a heterogeneous graph containing physical topology edges and virtual repulsion edges, designs an electromagnetic potential energy attention mechanism and a pre-training strategy incorporating physical repulsion losses, and dynamically balances the compactness of circuit connections with the security of physical isolation. This solves the problems of high search blindness and low engineering compliance rates in traditional layout methods, significantly improving the automation level of ship cable network layout design.

[0026] like Figure 1 As shown, the present invention provides a ship cable layout design method based on a physical-topology graph neural network, which specifically includes the following steps: Step 1: Heterogeneous Map Construction Steps: This step aims to transform unstructured ship cable network principle data into a graph representation model with physical field awareness capabilities through data fusion and structure mapping. This process achieves mathematical dimensionality reduction of complex engineering constraints by deeply coupling the circuit's "logical state" with the spatial "physical state." The specific implementation process includes the following sub-steps: (1) Data parsing and semantic definition of heterogeneous nodes First, the system reads the original layout drawing of the vessel to be designed and the associated equipment technical specifications. This invention abstracts each physical entity in the cable network as a heterogeneous node in a graph network. To accurately preserve the engineering functional attributes of the equipment, a global node set is defined. V It is the union of multiple heterogeneous subsets.

[0027] Let the set of nodes be V The mathematical definition of is as follows: (1) in: VThis represents the total set of all controlled and undeployed nodes within the ship's cable network; V src It represents a collection of power energy source nodes, including active devices such as main generators, emergency power supplies, and energy storage batteries; V load This represents a set of energy-consuming load nodes, including high-power propulsion motors, deck machinery, and daily power distribution loads. V ctrl It represents a collection of signal control and sensing nodes, encompassing a central control cabinet, various sensors, and communication base stations; V sw This represents a set of logic switching and overload protection nodes, covering circuit breakers, busbars, and contactors on the distribution board.

[0028] For sets V Each node in v i This invention constructs a multidimensional initial feature matrix. X The feature vector of its individual node x i Defined as: (2) in: x i Representing the i The original multimodal feature vectors of each device node; T i This represents the functional category identifier of the device, and typically uses one-hot encoding to characterize the node's role in the system; P i The nominal rated power of the equipment is used to measure the heat load and energy weight of the node during operation; U i The rated operating voltage of the equipment is the main criterion for determining high and low voltage isolation constraints; Represents the electromagnetic susceptibility level, used to characterize the device's tolerance limit to external electromagnetic interference; C i The thermal radiation coefficient of the device is used to account for heat dissipation spacing constraints in subsequent auxiliary layouts.

[0029] (2) Logical mode construction: Entity edge generation based on connection graph constraints Subsequently, based on electrical connection logic, the system establishes a set of entity topological edges representing the directions of energy flow and signal flow. E topo .

[0030] The mapping rules for entity edges are defined as follows: (3) like i, j If there is an electrical connection between them, then the corresponding logical adjacency matrix is... Defined as: (4) If there is no electrical connection, the value is 0.

[0031] in: E topo It represents the set of logical modal edges composed of physical wires or busbars; e ij Represents the connection node v i and v j The rightful side; f conn Represents the join decision function; The sparse adjacency matrix representing the logical topology; This represents the edge weight, and its value is proportional to the cross-sectional area and current intensity of the connecting cable.

[0032] (3) Physical mode construction: virtual edge generation based on EMI potential field The core of this step lies in transforming the electromagnetic compatibility constraints that originally existed in two-dimensional space into a penalty structure within a graph network by establishing a virtual topology.

[0033] First, for each pair of nodes ( v i ,v j ), calculate the mutual interference and repulsion force index. F rep : (5) in: F rep ( i, j ) represents a node i With nodes j The virtual repulsive force index between them; K This represents the environmental medium attenuation constant and is related to the shielding effectiveness of the compartment. U i , U j These represent the absolute values ​​of the operating voltages at the two nodes, respectively. S i 、S j These represent the electromagnetic shielding levels of the two nodes, respectively. The constraint coefficient is determined by the isolation distance table in the national ship inspection standards.

[0034] Set critical interference threshold Construct a set of virtual repulsion edges E virt : (6) in: E virt This represents a set of virtual repulsive edges composed of non-contact constraints; e virt This represents a virtual constraint edge created due to safety clearance requirements; This represents the maximum permissible unit interference tolerance of the system, serving as the trigger condition for establishing virtual edges.

[0035] Please see Figure 2 The topology diagram after completion is as follows: Figure 2 As shown.

[0036] Step 2: Dual-stream feature extraction and potential energy attention fusion After constructing the physical-topological heterogeneous graph, this embodiment employs a parallel two-stream heterogeneous graph neural network architecture to perform deep processing on the graph data. The specific implementation process mainly includes the following three sub-stages: (1) Topology-aware flow channel feature aggregation This channel is based on entity topology edges. E topo By utilizing multi-layer graph convolutional networks to aggregate neighbor information, functionally cooperating devices are clustered in the feature space. Let... If the input features are for layer l, then the output calculation formula for layer l+1 is: (7) in: This is the node feature matrix output by the (l+1)th layer topology-aware flow; This represents a non-linear activation function; in this embodiment, the ReLU function is preferred. The logical adjacency matrix after adding self-loops; for The corresponding degree matrix is ​​used for normalization; is the trainable weight matrix for the topological flow channel.

[0037] (2) Potential energy diffusion in the physical field sensing flow channel This channel is based on the virtual repulsion edge E. virt A potential energy diffusion operator is designed to simulate the repulsive force of a physical field, diffusing the "high potential" characteristic of the interference source to sensitive devices. The interlayer propagation formula is defined as: (8) in: Output the feature matrix for the (l+1)th layer physical flow, encoding the force state of the nodes; A virt This is a weighted adjacency matrix based on virtual repulsion edges, where each element represents a potential energy weight. D virt This is a virtual graph degree matrix used to prevent feature value explosion; The weight matrix and bias vector of the physical flow channel; It is a leakage activation function used to preserve weak interference signals.

[0038] (3) Electromagnetic potential energy attention mechanism and fusion Given the different dependencies of devices on the two types of features, an adaptive attention mechanism is introduced.

[0039] First, compute nodes i Potential disturbance S i That is, the sum of the weights of all the virtual edges it connects to: (9) Next, the perturbation degree is mapped to normalized attention coefficients. : Finally, the final node embedding vector is generated. Z i : (10) in: is the fusion weight coefficient for physical flow features, with a value range of (0, 1); Zi Embed vectors for the final fused nodes; This is a weighted summation operation on the eigenvectors.

[0040] This mechanism ensures that nodes ( ) are protected under strong interference environments. Physical isolation constraints are prioritized for nodes, while ordinary nodes ( Prioritize the compactness of topology connections.

[0041] Step 3: Internalization of physical constraints - pre-training This step is the core of the invention's "intelligent obstacle avoidance" mechanism. To enable the two-stream heterogeneous graph neural network to automatically sense and avoid electromagnetic interference, this invention constructs a composite loss function that includes topological reconstruction and physical repulsion, and pre-trains the model using a self-supervised learning paradigm.

[0042] (1) Construction of composite loss function Define the total loss function L total This is a weighted sum of the topology branch loss and the physical branch loss: (11) in: L total The overall objective function for model training; The penalty weight coefficient for physical constraints forces the model to prioritize satisfying safety constraints. The regularization coefficient is . This is a set of network parameters.

[0043] (2) Mathematical definition of Hinge Loss To transform the engineering specification that "high-voltage equipment must be kept away from sensitive equipment" into a mathematically differentiable constraint, this invention employs a margin-based hinge loss. This loss is applied to the set of virtual repulsion edges. Evirt Each pair of nodes in ( u, v Define its physical repulsion loss. L phy for: (12) in, Z u , Z v These represent the embedding vectors of the interference source node and the sensitive node in the feature space, respectively. This is a preset feature space security isolation threshold.

[0044] The physical meaning of this formula is: when the feature vector output by the network... Z u , Z v The distance between them is less than the safety threshold At this point, the loss function produces a positive penalty value. During backpropagation, this gradient will force... Z u , Z v By moving in the opposite direction on the feature manifold, "forced isolation at the feature level" is achieved.

[0045] (3) Topology reconstruction loss Meanwhile, binary cross-entropy loss is used as the topology reconstruction loss. L topo This ensures that the feature vectors retain the circuit's connection structure information. (13) Through joint optimization, the model will simultaneously possess the dual characteristics of "compact connectivity" and "compliant isolation" upon convergence.

[0046] Step 4: Spatial Decoding and Layout Generation This step utilizes the pre-trained model to establish a mapping from the high-dimensional feature space to the two-dimensional physical space of the cabin, enabling the constructive generation of layout schemes. A spatial decoder composed of multilayer perceptrons is then constructed. The fusion node embedding vector output in step S2 is used to embed the nodes into the fusion node embedding vector. Z i As input, the centroid coordinates C of the direct regression prediction device in the ship's cabin coordinate system are... i = ( x i , y i Due to the input vector Z i The physical repulsion loss has been internalized in step S3, and the set of coordinates output by the decoder is now complete. C The system will automatically cluster devices that naturally exhibit a spatially separated distribution pattern of "high-voltage equipment clusters and low-voltage equipment clusters" and have close electrical connections. This output can serve as a layout scheme that meets both electromagnetic compatibility and wiring cost constraints without the need for complex post-processing iterations.

[0047] To further illustrate the application potential and expected effects of the layout design method based on physical-topology graph neural network proposed in this invention, this embodiment uses a typical cable network equipment compartment on a ship as an application scenario.

[0048] This application scenario selects a complex cable network system, including the main frequency converter, propulsion control unit, cooling pump group, transformer and various sensors, as the planning object. The system also contains high-voltage interference source nodes and low-voltage sensitive nodes, and there are complex cable connection relationships and strict electromagnetic compatibility constraints between the nodes.

[0049] Figure 4 shows a before-and-after comparison of the expected device layout designed using the method of the present invention.

[0050] As shown in Figure 4(a), in the initial state without the adoption of this method, due to the lack of perception of physical field constraints, high-voltage equipment and low-voltage sensitive equipment often present a mixed distribution, resulting in many potential electromagnetic interference hazards and cable detours. However, as shown in Figure 4(b), after applying the method of this invention, it is expected to achieve a significant optimization effect in space: the high-voltage equipment cluster represented by the gray area and the low-voltage sensitive equipment cluster represented by the white area in the figure automatically form a clear physical isolation zone in space, and various types of equipment exhibit obvious clustering characteristics based on functional logic.

[0051] In summary, by introducing a physical constraint internalization mechanism, this invention aims to solve the conflict between "physical isolation" and "wiring cost" in ship cable laying and equipment layout. It is expected to enable a leap from "manual trial and error" to "intelligent generation" in the design mode, and significantly improve the layout compliance and design efficiency of complex ship cable network systems.

[0052] This embodiment proposes a ship cable layout design method based on a physical-topology graph neural network. By comparing and analyzing traditional manual design modes, conventional optimization algorithms (such as NSGA-II), and the intelligent design logic proposed in this invention, it can be seen that: when ship electrical equipment is laid out using the method proposed in this invention, electromagnetic interference conflicts between high- and low-voltage equipment can be effectively avoided in theory, achieving compliant distribution of equipment in physical space; at the same time, thanks to the aggregation effect of topology-aware flow on connection relationships, the total length of cable connections is expected to be significantly shortened and the layout speed improved, thereby enhancing the engineering feasibility and economic benefits of ship cable laying and equipment layout schemes.

[0053] The above content describes the technical concept of this invention. Those skilled in the art can make various corresponding changes, modifications, simplifications, and combinations based on the technical solutions and concepts described above, and all such changes, modifications, simplifications, and combinations are included within the protection scope of the claims of this invention.

Claims

1. A method for designing ship cable layout based on a physical-topological graph neural network, characterized in that: The method includes the following steps: S1: Read the schematic diagram of the ship's cable network and the cabin environment data, construct a heterogeneous graph containing physical topology edges and virtual exclusion edges, and synchronously encode logical connections and physical interference constraints into the graph structure; S2: Utilize a dual-stream heterogeneous graph neural network to extract topological and physical features in parallel, and then adaptively fuse them through an electromagnetic potential energy attention mechanism to generate node embedding vectors with embedded physical priors. S3: Construct a composite loss function that includes topological reconstruction and physical repulsion, perform self-supervised pre-training on the model, and internalize the physical isolation criterion into the model parameters by minimizing the physical repulsion loss; S4: Extract node embedding vectors using a pre-trained model, and project them onto the ship's cabin physical coordinate system through a spatial decoding and mapping module to generate a layout scheme that meets the requirements of isolation and partitioning.

2. The ship cable layout design method based on physical-topology graph neural network according to claim 1, characterized in that: In step S1, the construction process of the physical-topological heterogeneous enhancement graph is as follows: Define the graph structure as ,in V A set of device nodes; Based on the logical connection relationships in the cable network schematic, construct a set of entity topology edges. This forms a logical mode; Calculate the high-voltage interference source nodes according to electromagnetic compatibility standards. u With low-pressure sensitive nodes v Potential interference strength between ,when Exceeding the preset threshold At the node u, v Establish virtual exclusion edges and will be based on safe distance Calculated potential energy value Assign weight attributes to the edge to form a physical mode.

3. The ship cable layout design method based on physical-topology graph neural network according to claim 1, characterized in that: In step S2, the working process of the two-stream heterogeneous graph neural network model includes: (1) Feature extraction stage: The topology-aware flow channel uses graph convolution operators along the topological edges of entities. Aggregate the functional semantic information of neighboring nodes The physical field sensing flow channel utilizes the potential energy diffusion operator only at the virtual repulsion edge. Uppropagation interference characteristics are used to obtain physical field characteristics. ; (2) Feature fusion stage: computing nodes i Total potential weight of all associated virtual repulsion edges To measure the degree of disturbance, an attention mask is generated accordingly. According to the formula: ； The two-stream features are weighted and summed to generate the final node embedding vector. Z i .

4. The ship cable layout design method based on physical-topology graph neural network according to claim 1, characterized in that: In step S3, the composite loss function includes a physical repulsion loss term. Constructed using hinge loss: Define a pair of nodes that have a virtual exclusion edge. u, v Euclidean distance in feature space ; Using this as a metric, a safe isolation distance threshold is set. ; And construct the loss function: When the distance between node pairs is less than the threshold A positive penalty value is generated at each time, and the network weights are updated through the backpropagation algorithm until the condition is met. Physical isolation constraints.

5. The ship cable layout design method based on physical-topology graph neural network according to claim 1, characterized in that: In step S4, the spatial decoding and layout generation process includes: Construct a nonlinear mapping function to establish the projection relationship from the multidimensional graph feature space to the two-dimensional cabin physical space; The node embedding vector output in step S2 Z Input this mapping function and output the physical centroid coordinates of each device node. ,because Z The physical repulsion loss has been internalized through step S3, and the output coordinate set is as follows. C It naturally presents a spatial separation between high-voltage and low-voltage equipment.

6. The ship cable layout design method based on a physical-topology graph neural network according to claim 5, characterized in that, It also includes post-processing steps for the layout scheme: Using the centroid coordinates generated in step S4 C i Based on the actual length and width dimensions of the equipment ( l i , w i , h i Generate the envelope rectangle B i ; Execute a collision detection algorithm based on non-overlapping constraints. If the envelope rectangles of the two devices satisfy the intersection condition... Then, the coordinates are fine-tuned to finally generate a ship circuit layout design drawing that satisfies the conflict-free constraint.

Images