Abstract
Ship pipe routing is a complex task that requires placing pipes with diverse functions within limited spaces. The pipe routes must avoid interference with equipment, follow ship structures such as decks and pillars, and comply with design regulations, routing efficiency, and space utilization requirements. Designers repeatedly revise pipe routes while considering pipe interference, routing priorities among pipes, and equipment connection constraints. Expert knowledge accumulated from past project data plays a critical role in this process. Depending on the design objective, designers create pipe routes that follow ship structures, maintain sufficient clearance from equipment, or emphasize shorter and simpler routes with fewer bends. However, consistently applying such expert knowledge is time-consuming, as many competing constraints must be considered simultaneously. To reduce this manual burden, automated pipe routing methods have been widely proposed. Many of them rely on cost functions defined by rules or heuristic pathfinding algorithms. However, design tendencies in real ship pipe routing often depend on context and are implicitly applied by designers. Therefore, they are difficult to quantify and represent consistently as rules or cost functions.
To overcome this limitation, we proposed a data-driven ship pipe routing method that learns expert knowledge directly from data using a diffusion model based on a Transformer architecture. The surrounding equipment, ship structures, and start and goal points are provided as input to support the model’s spatial context awareness. The model learns design tendencies that embody expert knowledge from past project data, without manual rule crafting or cost function tuning. The proposed method consists of three parts: representation of the routing environment, generation of pipe waypoints, and postprocessing into a rectilinear route. To represent the routing environment, the surrounding equipment, ship structures, boundaries, and the start and goal points are encoded into tokens. An encoder based on a Transformer architecture relates these tokens via attention, allowing the model to consider the overall spatial context simultaneously. Guided by this spatial context, the model predicts the pipe route’s waypoints. Pipe routing is inherently multimodal, since several feasible pipe routes can exist under the same start and goal points. Therefore, the waypoints are generated through a model refinement process rather than a single deterministic prediction. Starting from randomly initialized points, the model progressively refines their positions and forms a route that reflects the learned design tendencies. The model is trained on pipe routes collected from spaces with diverse equipment arrangements. The training objectives guide the refinement direction, preserve the waypoint order, match the overall route shape, and discourage interference with equipment. Since the predicted route may not fully satisfy orthogonality, postprocessing is applied to further ensure orthogonality and remove redundant points.
The proposed method was applied to routing environments with surrounding equipment and ship structures. Given information about the target pipe and the surrounding space, the model predicted pipe waypoints and generated routes that reflected the learned design tendencies. The predicted routes were then converted into rectilinear pipe routes through postprocessing. The results were compared with those of a graph search method under different conditions. The proposed method generated routes through a fixed number of refinement steps, rather than repeated search based on manually defined cost functions. In addition, when trained on data with different design tendencies, such as shortest-route and space-utilization preferences, the model generated routes with the corresponding characteristics.
In conclusion, this study proposed a data-driven ship pipe routing method that incorporates expert knowledge using a diffusion model based on Transformer architecture. By learning design tendencies from past project data and predicting pipe waypoints through iterative refinement, the method reduces reliance on manually defined rules and on cost function tuning. Future work will extend the method to more diverse routing environments and incorporate expert knowledge from real design records.
To overcome this limitation, we proposed a data-driven ship pipe routing method that learns expert knowledge directly from data using a diffusion model based on a Transformer architecture. The surrounding equipment, ship structures, and start and goal points are provided as input to support the model’s spatial context awareness. The model learns design tendencies that embody expert knowledge from past project data, without manual rule crafting or cost function tuning. The proposed method consists of three parts: representation of the routing environment, generation of pipe waypoints, and postprocessing into a rectilinear route. To represent the routing environment, the surrounding equipment, ship structures, boundaries, and the start and goal points are encoded into tokens. An encoder based on a Transformer architecture relates these tokens via attention, allowing the model to consider the overall spatial context simultaneously. Guided by this spatial context, the model predicts the pipe route’s waypoints. Pipe routing is inherently multimodal, since several feasible pipe routes can exist under the same start and goal points. Therefore, the waypoints are generated through a model refinement process rather than a single deterministic prediction. Starting from randomly initialized points, the model progressively refines their positions and forms a route that reflects the learned design tendencies. The model is trained on pipe routes collected from spaces with diverse equipment arrangements. The training objectives guide the refinement direction, preserve the waypoint order, match the overall route shape, and discourage interference with equipment. Since the predicted route may not fully satisfy orthogonality, postprocessing is applied to further ensure orthogonality and remove redundant points.
The proposed method was applied to routing environments with surrounding equipment and ship structures. Given information about the target pipe and the surrounding space, the model predicted pipe waypoints and generated routes that reflected the learned design tendencies. The predicted routes were then converted into rectilinear pipe routes through postprocessing. The results were compared with those of a graph search method under different conditions. The proposed method generated routes through a fixed number of refinement steps, rather than repeated search based on manually defined cost functions. In addition, when trained on data with different design tendencies, such as shortest-route and space-utilization preferences, the model generated routes with the corresponding characteristics.
In conclusion, this study proposed a data-driven ship pipe routing method that incorporates expert knowledge using a diffusion model based on Transformer architecture. By learning design tendencies from past project data and predicting pipe waypoints through iterative refinement, the method reduces reliance on manually defined rules and on cost function tuning. Future work will extend the method to more diverse routing environments and incorporate expert knowledge from real design records.