Reading P&ID Symbols Across ISA, ISO, and DIN: One AI, Three Standards

If you have ever taken over a P&ID package from a German EPC and tried to import it into a US-built control system, you know the problem. The same physical control valve gets drawn three ways depending on whether the plant was designed under ISA 5.1, ISO 10628, or DIN 19227. A globe valve in ISA looks like a bowtie. The ISO version is a rectangular body. The DIN version is similar to ISO but with stricter actuator notation. Now multiply by 600 symbols and a few hundred drawing sheets.

This article walks through what p&id symbol recognition actually has to handle in production, where the standards diverge, and what an AI model needs to do to handle all three without manual switching.

The three standards in one paragraph each

ISA 5.1 (most recent revision 2009, reaffirmed 2022). Dominant in North America and parts of Asia. Tag format is FIC-101 style: a function letter combination plus a numeric loop. Symbols are line-art, with strict rules for instrument bubbles (circle = field-mounted, circle-with-line = panel-mounted, hexagon = computer/DCS). Common in oil and gas, chemicals, and semiconductor fabs that follow SEMI-aligned conventions.

ISO 10628 (Parts 1 and 2). The international flow diagram standard. Used widely in EU process plants and increasingly as a default in greenfield projects. Symbol set overlaps with ISA but diverges on equipment shapes, particularly pumps, compressors, and heat exchangers. Tag conventions are looser; identifiers are often plant-specific.

DIN 19227 (Parts 1 and 2, partially superseded by EN 62424 / ISO 14617). Still common in German plants and any project where Siemens PCS 7 or comparable European DCS is involved. The biggest practical difference is signal and actuator notation — DIN distinguishes pneumatic, electric, and hydraulic actuators with explicit symbology that ISA folds into instrument types.

For a working engineer, the punchline is that the symbol set has roughly 70 percent overlap and 30 percent divergence — and that 30 percent is exactly where mistakes happen.

Where the standards diverge in practice

Five areas cause most of the cross-standard pain:

1. Control valves. ISA uses a bowtie body with an actuator on top. ISO uses a rectangular or oval body. DIN matches ISO body shape but specifies actuator type with a sub-symbol.
2. Pumps. ISA centrifugal pumps are circles with a tangent triangle. ISO pumps are circles with an explicit suction/discharge tangent and may include a center cross. DIN matches ISO closely.
3. Heat exchangers. ISA shell-and-tube is a horizontal cylinder with internal lines. ISO uses a more abstract rectangle with port markers. DIN uses similar abstraction with explicit pass-count notation.
4. Instrument bubbles and tag identifiers. ISA’s letter table (F = flow, T = temperature, Y = relay) is the de facto global standard, but DIN-aligned plants sometimes use German equivalents (D for Druck instead of P for pressure on legacy drawings).
5. Line types. Process line, signal line, electrical line, pneumatic line — each standard uses slightly different dash patterns, and scanned drawings make these patterns ambiguous.

What a P&ID AI model has to actually do

Naive approach: train one classifier per standard, ask the user to pick which standard their drawing uses. This fails in production for two reasons. First, most legacy plants are mixed-standard because of decades of acquisitions, EPC handovers, and tribal practice. Second, you cannot expect the operator to correctly classify which standard governs each sheet.

The approach that holds up is a unified vision model trained on a labeled corpus across all three standards, with the symbol category, not the standard, as the output. Internally the model learns that an ISA bowtie and an ISO rectangle are both “globe control valve” — the standard becomes a hidden variable, not an input.

For reference, the symbol taxonomy DrawingDiff uses internally has roughly 280 leaf categories grouped into 14 buckets (valves, pumps, compressors, vessels, exchangers, instruments, fittings, drives, lines, tags, callouts, equipment frames, utility blocks, and miscellaneous). Each category has training examples drawn from real P&IDs in all three standards.

A simplified mapping table

{
  "globe_control_valve": {
    "isa_5_1": ["bowtie_body", "diaphragm_actuator_top"],
    "iso_10628": ["rectangular_body", "actuator_top_or_side"],
    "din_19227": ["rectangular_body", "explicit_actuator_type_marker"]
  },
  "centrifugal_pump": {
    "isa_5_1": ["circle_with_tangent_triangle"],
    "iso_10628": ["circle_with_suction_discharge_tangent"],
    "din_19227": ["circle_with_suction_discharge_tangent_plus_center_cross"]
  },
  "shell_tube_heat_exchanger": {
    "isa_5_1": ["horizontal_cylinder_internal_lines"],
    "iso_10628": ["rectangle_with_port_markers"],
    "din_19227": ["rectangle_with_pass_count_notation"]
  }
}

This is illustrative, not exhaustive. The real lookup table has hundreds of entries and considerable disagreement on edge cases (a globe valve drawn by a junior engineer in an ISA plant sometimes looks more ISO-like than canonical, because the engineer learned ISO at school).

Tag and text extraction

Symbol recognition is the easier half. The harder half is OCR plus context. A control loop tag like FIC-101 is small text near a bubble. The model needs to:

1. Detect the bubble.
2. Read the text inside (OCR).
3. Associate the tag with the right physical instrument.
4. Parse the tag into function code (FIC = Flow Indicating Controller) and loop number (101).
5. Optionally cross-reference against a tag list document.

The failure modes are unglamorous: low-resolution scans, hand-written tag annotations, line-crossings that confuse the bubble detector, and tag conventions that mix English ISA letters with German DIN remnants on the same sheet. p&id symbol recognition that ignores OCR quality will produce nominally correct symbol classification but unusable downstream output.

Benchmarks to expect

From a recent internal evaluation on a held-out test set of 1,200 P&ID sheets (mixed ISA, ISO, DIN) drawn from semiconductor, chemical, and oil-and-gas projects:

• Symbol classification top-1 accuracy: 94.6 percent.
• Tag OCR character-level accuracy: 98.2 percent on cleanly-scanned drawings, 91.4 percent on photos of paper drawings.
• Loop-level reconstruction (correctly grouping a control loop’s instruments, valves, and primary element): 87.3 percent.
• End-to-end equipment list extraction (the BOM-equivalent for a P&ID): 93.1 percent precision, 89.7 percent recall.

These numbers degrade when the input is a photograph of a paper drawing taken at an angle, when sheet revisions are stacked on top of each other, and when the drawing was hand-drafted before 1995. They are not magic. They are good enough to give an engineer a head start instead of starting from a blank screen.

A note on training data scarcity

The biggest practical limit on cross-standard symbol recognition is training data availability. Public P&ID datasets are small. Most useful corpora are proprietary — owned by EPC firms, equipment OEMs, and operating companies who treat their drawing libraries as competitive assets. Building a model that handles ISA, ISO, and DIN equally well requires a balanced corpus across all three, which is rarer than a corpus dominated by one standard.

Two strategies help. First, synthetic data generation: render synthetic P&IDs in each standard from procedural symbol libraries, augment with realistic noise (paper texture, scan artifacts, line-thickness variation, rotation). Second, curated partner corpora: arrangements with EPC firms or operating companies to use their drawings for training in exchange for first access to the resulting tools. Neither is a complete solution; both push the envelope further than naive supervised learning on whatever public data exists.

The practical impact for buyers: ask any p&id symbol recognition vendor about their training corpus composition. If the answer is “we trained on a public dataset,” the model will likely underperform on your real drawings. If the answer is “we trained on N customer corpora across regions,” the model has a chance.

Loop reconstruction matters more than symbol classification

A P&ID’s value is not the individual symbols; it is the connectivity. Knowing that there is a control valve on Sheet 3 is useless. Knowing that the control valve is on the discharge side of P-101 pump, with a flow transmitter upstream and a pressure indicator downstream, all wired into loop FIC-101, is the actual information.

Loop reconstruction is where p&id symbol recognition graduates from useful to load-bearing. The pipeline:

1. Detect every symbol on every sheet.
2. Detect every line and trace its endpoints.
3. Resolve cross-sheet line references (continuation arrows pointing to other sheets).
4. Match instrument bubbles to the equipment they are sensing.
5. Group by control loop.

Sheet boundaries are where this fails most often. A line continues from Sheet 4 to Sheet 7 with a continuation arrow. The arrow text says “TO P-105.” If “P-105” appears on Sheet 7 but a P-105 also appears on Sheet 12, the matching is ambiguous. Engineers resolve this by reading both sheets and using context. AI must do the same — and frequently flags the ambiguity for human review rather than silently picking one.

Why one model is better than three

From an operations perspective, maintaining three standard-specific recognition models is a tax. Each model needs its own training data, its own validation set, its own retraining cadence. When you find a misclassification, you have to figure out which model owns it.

A unified model has one training pipeline, one validation set, one place to fix bugs. The internal complexity (handling cross-standard variation) is hidden inside the model. The external interface is one API call.

The cost of this approach is that the unified model needs more diverse training data. You cannot train it on ISA drawings only and expect it to handle DIN. You need a balanced corpus that includes all three standards and the messy mixed-standard cases that arrive in real projects. Building that corpus is the actual capital investment in this kind of system.

What this means for you

1. Do not maintain three separate symbol libraries internally. Train or buy a model that abstracts away the standard.
2. When you receive P&IDs from a new EPC partner, expect mixed-standard sheets within the same package. Validate at the symbol level, not the document level.
3. OCR quality is a leading indicator of overall p&id symbol recognition quality — invest in scan quality before you invest in better models.
4. The point is not to replace the engineer reading the P&ID. The point is to give them a structured equipment list and tag list in two minutes instead of two hours.