Viscous chemical reactor — the agitator motor current tells the batch's story before any alarm does.

The problem

By the time the warning is a stopped blade, the batch is already gone.

Blade seizure in a viscous chemical reactor almost never arrives all at once. It is a ramp: viscosity rises non-linearly during one stage of the batch, the motor responds by drawing more current to hold rpm, the drive starts limiting torque, and if nobody reacts, the next snapshot is a seized blade, mass baked onto the jacket wall and a batch out of specification.

The problem is not that the information does not exist — the current is in the drive, the rpm are in SCADA, the recipe is in the MES or in a lab spreadsheet. The problem is that those three realities live on islands, and the operator has to decide while looking at a mimic screen that shows jacket temperature and little else:

1. The normal current signature of each recipe is not something anyone knows by heart — it shifts with the product, the charge, the starting temperature.
2. The SCADA alarm fires when the fixed threshold is crossed. By then the reactor no longer recovers by opening solvent.
3. The veteran's knowledge ("with this product, once you're 70% through the addition you drop 5 rpm") lives in one head, not in the system.

The operator does what they can with what they see. The shift supervisor runs over once the alarm is already sounding. The batch is lost not for lack of data, but for lack of judgement in time — and in chemicals that means product out of specification, a forced reactor clean-out and hours of washing with expensive solvent.

How it fits into the IRIS system

iLEAN Edge doesn't add another panel — it seals the gap between the drive, SCADA and the recipe.

The piece that solves the pain is iLEAN Edge: a physical terminal installed next to the agitator cabinet that sees the reactor the way a veteran with three shifts of experience on that product would. It carries a neural network trained on the normal current signature per recipe, and cross-references it in real time with rpm, estimated torque and jacket temperature.

Edge sees the current. Connect reads the recipe wherever it lives — MES, SCADA or a lab spreadsheet. The agent cross-references the drift with the batch stage and prepares the correction. The person signs off — the reactor never restarts on its own.

The three iLEAN pieces applied to agitator motor current:

• Edge — a terminal on the shop floor, next to the drive, reading RMS current per phase, active power and estimated torque. The CNN learns the normal shape of the current for each recipe and reactor — not a fixed threshold, but a signature. When the signature drifts, it raises an alert on the operator panel and, depending on configuration, a dry contact to SCADA. It works with no network: if the plant loses WiFi, Edge keeps watching the current and warning locally, because what is critical cannot depend on connectivity.
• Connect — captures the batch recipe whether it comes from the vertical MES, from SCADA, or from a lab sheet in a shared folder that the R&D lead updates whenever a new product comes in. And it also captures what arrives from outside — the email from the raw-material supplier saying this lot is 4% more viscous, the message from the previous night's shift lead with a change of order — at second zero, without anyone forwarding anything.
• Agent — cross-references the drifting current signature, the batch stage (are we in the critical addition?), the reactor's history with that recipe and the lot's raw material. If warranted, it proposes to the operator: drop 5 rpm, open solvent addition, or wait. It does not lower rpm by itself: the human signature stays in the middle. The management team sets the level of autonomy per recipe.

See the full IRIS architecture →

Before and after

Reactor without Edge vs. reactor with iLEAN Edge on the motor current

Aspect	Reactor with SCADA + fixed alarm	With iLEAN Edge on the agitator current
Signal that triggers the reaction	Fixed current threshold, the same for every recipe	Current signature learned per recipe and per reactor
Moment of the warning	Once the threshold is crossed — blade close to seizing	When the drift begins — margin left to correct
Decision to lower rpm / add solvent	Operator's judgement, with no recipe context	Agent proposal with stage + recipe + history, person signs off
New recipe from the lab	Forgotten spreadsheet, reaches the panel late	Connect captures it at second zero
Operation without network	Local alarm yes, context no	Edge keeps reading current and warning on cabinet power
Batch traceability	Rebuilt by hand for the audit	Per-batch dossier — current signature, decisions, human sign-offs

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We set it out so the committee has an order of magnitude; we refine it during the diagnostic.

• Specialty chemicals plant, viscous batch reactors (resins, polymers, paints), a history of one or two batches lost per quarter to blade seizure or viscous deviation.
• Edge pilot on one reactor (clamp-on current sensor + integration with the SCADA/MES holding the recipe + operator panel). First value expected within a few weeks: detection of the drifting signature before the current alarm threshold.
• Indicative payback between 4 and 9 months, depending on the frequency of documented incidents and the average cost of a lost batch (raw material + cleaning hours + reactor occupancy).
• Expected reduction in batches lost to viscous deviation of ≥ 30% in the first year, depending on product and process maturity. The hard lever is a single batch recovered — it pays for the pilot.

And the process manager's fair objection

"What if the AI proposes lowering rpm when it shouldn't and wrecks my yield?" — the proposal never executes itself. iLEAN works on an anchored task: it reads a physical signal that exists (current) and compares it with the learned signature of the product. On anchored tasks, the best models brought error below 1.5% ^[1]. And even so, the operator and the shift supervisor see the proposal, see the signature, and sign off. The three safety rings exist precisely for this: the reactor does not restart on its own, nor lower rpm on its own, except where you decide it should.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.