SynthPID: P&ID digitization from Topology-Preserving Synthetic Data

1. Problem Statement

The digitization of Piping and Instrumentation Diagrams (P&IDs)—raster images encoding plant components and connections into structured process graphs—is critical for regulatory compliance, predictive maintenance, and digital twins. However, automating this process is severely bottlenecked by data scarcity:

• Proprietary Nature: Engineering drawings contain sensitive process information, preventing external sharing.
• Limited Benchmarks: The entire research community relies on a single public benchmark, PID2Graph OPEN100, which contains only 12 annotated real-world images.
• Failure of Existing Synthetic Data: Prior attempts to generate synthetic training data using template-based methods (randomly placing symbols on a canvas) failed. Models trained on such data achieved only ~33% edge detection accuracy on real images. The authors argue this failure is structural, not visual: random generators produce unrealistic graph topologies (e.g., uniform clustering, narrow degree distributions) that do not reflect the complex flow dynamics of real process plants.

2. Methodology

The authors propose SynthPID, a novel approach to generating synthetic data that preserves the topological structure of real P&IDs, combined with a patch-based deep learning architecture.

A. Topology-Preserving Generation Pipeline

Instead of generating from scratch, the pipeline uses 12 real P&IDs as structural seeds:

1. Extraction: The physical component graph is extracted from a seed image, removing auxiliary nodes (connectors and crossings) to create a "collapsed" graph of meaningful components.
2. Perturbation:
   • Spatial: Node positions are displaced by small random offsets while strictly preserving adjacency (topology).
   • Symbolic: Symbols are replaced with randomly sampled templates from the same class to introduce visual variety without changing component types.
   • Routing: Pipe connections are re-routed using Manhattan routing to reflect new positions.
3. Rendering & Filtering: The layout is rendered into high-resolution PNGs. To ensure diversity, candidates are filtered using:
   • Weisfeiler-Lehman (WL) Hash: To reject structurally identical graphs.
   • Perceptual Hash: To reject visually near-duplicate images.
4. Output: This process yields SynthPID, a corpus of 665 synthetic P&IDs with full graph-level annotations.

B. Model Architecture: Patch-Based Relationformer

Due to the high resolution of P&IDs (often >7000px) and the small size of symbols (often 8px), full-image processing is unreliable.

• Patch Extraction: Images are split into overlapping 1500×1500 px patches (stride 750px) to bring symbols to a detectable scale (50–200px).
• Backbone: The model uses Relationformer (based on Deformable DETR), which employs a shared relation token to jointly predict bounding boxes, classes, and edges in a single forward pass.
• Post-Processing: Predictions from patches are merged using:
  • Border confidence attenuation.
  • Non-Maximum Suppression (NMS) and Weighted Box Fusion (WBF).
  • Reconnection of cut edges via border nodes.

3. Key Contributions

1. SynthPID Dataset: The first publicly available, topology-preserving synthetic P&ID dataset (665 images) with full graph annotations, generated from real seeds.
2. Generation Code & Models: Release of the generation pipeline and trained weights.
3. Empirical Evidence: Controlled experiments proving that generation quality (topology) is the primary driver of synthetic-to-real performance, rather than model architecture choice.
4. Rigorous Benchmarking: Establishment of a reproducible evaluation protocol on OPEN100 using 5-fold cross-validation with 3 random seeds (15 total runs per configuration).

4. Experimental Results

The model was evaluated on the PID2Graph OPEN100 benchmark (12 real images) using Edge mAP (mean Average Precision for edge detection).

Key Findings:

• SynthPID vs. Template: Training on SynthPID alone achieves 63.8% edge mAP, a 30 percentage point improvement over template-based synthetic data (~33%).
• SynthPID vs. Oracle: The gap between training on 665 synthetic images (E2) and the "Oracle" trained on real data (E1) is only 5.6 pp, despite E2 seeing zero real P&IDs.
• Superiority over Modular Baseline: The SynthPID-only model outperforms the modular baseline trained on 60 real images by 18 percentage points.
• Scaling Analysis: Performance gains flatten after ~400 images, indicating that seed diversity (the number of unique real topologies used) is the binding constraint, not the volume of synthetic images.

5. Significance and Implications

• Solving the Data Bottleneck: The paper demonstrates that high-quality digitization is possible without access to proprietary real-world training data, provided the synthetic data preserves the structural topology of real systems.
• Shift in Paradigm: It challenges the assumption that visual realism is the most important factor in synthetic data. Instead, topological fidelity (correct degree distributions, clustering, and flow directionality) is the critical factor for transfer learning.
• Practical Application: Industrial sites can now use a small set of existing (but non-disclosed) drawings as "seeds" to generate a massive training corpus, enabling the deployment of competitive P&ID digitizers without the legal and logistical hurdles of sharing sensitive data.
• Future Directions: The authors suggest that further improvements will come from increasing the diversity of seed diagrams (e.g., from different plant types) rather than simply generating more images from the same seeds.

In conclusion, SynthPID bridges the gap between synthetic and real data for P&ID digitization by prioritizing structural correctness over visual randomness, achieving near-oracle performance with zero real training data.