Ambaprasad Ganapathi
Hegde

Layer 1 — Templates: abstract test cases per library component, parameterized ({POS_A}, {SENSOR_1}), stored suite-side.

Layer 2 — Information models: each component type carries a model of its signals, positions and behavior — the suite knows every object the PLC can use.

Layer 3 — Population engine: binds templates to real control data (symbols, I/O, instance names) of the PLC under test at runtime.

Layer 4 — Execution: drives the S7 PLC and the virtual model (fe.screen-sim) simultaneously; asserts in both directions — actuate virtual/observe PLC, actuate PLC/observe virtual.

≈22% lower testing cost per PLC and ≈28% shorter testing cycles, measured across productive use. Test knowledge moved out of experts' heads into a reusable suite — the same test case now validates every station that uses the component. Patent-pending; rolled out to international Gen6 battery sites.

MCP server: embedded in the test suite, exposing its execution tools (move, rotate, check_overlap…) to the LLM.

Skill files: describe every component type and tool, so the model knows what exists and how to drive it.

Entity resolution: free-text "gantry G02", "robot IR1", "pick position" mapped to concrete suite objects.

Execution & evidence: sequence runs in the VC environment with automated collision checks; a recording captured straight from the virtual model window accompanies every verdict, so the engineer verifies at a glance.

Scenario definition went from hand-built sequences to a typed sentence; while one scenario executes, the engineer is already formulating the next. Validation throughput rises without adding headcount, and video evidence makes results reviewable by anyone.

Local hosting: model runs on-machine — engineering data never leaves the premises.

Vector DB: library elements, planning data and control data embedded for semantic retrieval.

Skill routing: SKILL.MD files describe each data source and when to consult it; the model routes queries accordingly.

Grounded synthesis: answers cite their sources — EPLAN revision, control data version, test plan station.

Change tracing that used to mean manual archaeology across EPLAN exports and code diffs now takes one question. Test-plan triage ("which function group at station 20 needs attention?") comes back with failure counts and a recommendation. Also lowers the onboarding barrier for new engineers.

Export: both TIA projects converted to XML in full.

Chunking: XMLs broken into comparable units (FBs, networks).

Deterministic diff: exact structural comparison — no AI in the loop where precision matters.

LLM classification: structured diffs fed to the model, which assigns change categories and relevance, producing the high/medium/low report.

Master/Copy Station alignment reviews that took days of side-by-side reading collapse into a generated report; engineers start from the handful of high-priority changes instead of thousands of XML lines. Deterministic where it must be, AI where it helps.

Ruleset: house standards formalized as JSON rule definitions.

Parser: control code decomposed into checkable units.

Rule engine: each unit validated against every applicable rule; violations collected with locations into a report.

Standards compliance became a batch job instead of a review meeting — consistent, repeatable, and fast enough to run on every change. Reliable code and data quality downstream in simulation and commissioning.

Connectivity: OPC UA server/client linking the physical line's PLC with Plant Simulation.

Two-way binding: state flows physical→virtual; commands flow both directions.

Dashboard: web KPI dashboard for energy metrics and shop-floor information with dual-channel interaction.

A working research foundation for shop-floor development — live energy monitoring, process insight, and a hands-on digital twin for students and researchers.