AlarmLogic — Product Description

Read-only industrial alarm intelligence & causal situation assessment. Commercial product description · 2026-09-19

---

1. Executive Summary

AlarmLogic is a read-only, non-intrusive alarm intelligence appliance for process industries. During an upset, hundreds of alarms hit a console in minutes; only one is the cause. AlarmLogic groups the consequence flood onto the physical P&ID topology, pins the probable initiating event with a millisecond evidence chain, and reconstructs the incident on a synchronized replay timeline — without ever writing to the DCS, the SIS, or any controller.

It is deployed in the Level 3 / 3.5 OT DMZ (Purdue model), sells as per-site subscriptions through the Sensifai Industrial AI Marketplace, and targets the refinery, petrochemical, LNG, offshore, power, and specialty-chemical segments where a single avoided trip justifies the annual license many times over.

---

2. The Problem

Operators can process roughly one alarm per minute. Floods deliver a hundred. Under ANSI/ISA-18.2 the sustainable steady-state rate is 1–2 alarms per 10 minutes; an abnormal state above ~10 per 10 minutes is a flood. When the annunciator saturates, operators stop diagnosing and start silencing — and the single actionable alarm disappears into the noise.

The industry's catastrophic incidents share this signature: Texaco Milford Haven (275 alarms in the final 11 minutes) and BP Texas City (a stale, unmanaged list masking a tower overfill that killed 15 people). The culprit is structural: chattering, fleeting, stale, and consequential alarms all arrive with equal visual prominence, and correlation-based analytics cannot distinguish cause from cascade.

A small fraction of tags (typically under 5%) produces 70–80% of annunciations — the "bad actors." They are candidates for rationalization, not for censorship; what the console needs is intelligent presentation grounded in physical causality, not another black-box score.

---

3. The Solution

AlarmLogic never suppresses a physical alarm — that would violate IEC 61511 independence. It adds an advisory intelligence layer beside the DCS:

1. Ingest (read-only). OPC UA Part 9 alarms & conditions, OPC UA DataChange (PV/SP/OP), 1 ms SOE records, MQTT (Sparkplug-style topics), and DEXPI P&ID imports cross a unidirectional Level 3→3.5 boundary through purpose-built collectors converging on one typed ingest service. There is no write path.
2. Reason (deterministically). PrefixSpan flood mining, Smith–Waterman sequence alignment, DEXPI P&ID reachability, and telemetry change-points produce explained hypotheses — no correlation matrices, no LLM in the causal path.
3. Present (with evidence). Consequence alarms stay fully active in the native DCS journal but are visually de-emphasized under the initiating event; every hypothesis ships a decomposed confidence and a millisecond evidence chain for human judgment, streamed live to the control-room console (SSE).

---

4. Key Capabilities

---

5. How It Works

---

6. Safety, Compliance & Independence

Zero write-back is structural, not configurable. AlarmLogic has no OPC UA write client, no DCS register API, and no control endpoint. It performs zero autonomous suppression and zero dynamic priority writes — actions explicitly categorized in the blueprint as EXTREME SAFETY RISK / REJECT.

• SIS & fire-and-gas points bypass all de-emphasis logic and stay locked at the highest visual tier.
• Cognitive de-emphasis only: consequence alarms remain fully active and recordable in the native DCS; AlarmLogic grays/group-orders them in the advisory view (the blueprint's LOW RISK / RECOMMENDED tier).
• Explainable uncertainty, never a bare percentage: S = 0.35·temporal + 0.30·topological + 0.20·telemetry + 0.15·data-quality.
• Standards: ANSI/ISA-18.2, IEC 62682, EEMUA 191, ISA/IEC 62443, IEC 61511/61508, OSHA 1910.119, API RP 1167, ASM Consortium, ISA-5.5, DEXPI (ISO 15926).

---

7. Technology & Architecture

The domain layer is pure and dependency-free; infrastructure ports (marketplace, identity, storage) are swappable — which is what enables the fully-offline mock mode for air-gapped sites (MARKETPLACE_MODE=mock).

---

8. Commercial Model & Pricing

Sold via the Sensifai Industrial AI Marketplace — per-site subscriptions with a prepaid-wallet option.

Every analysis action emits an idempotent metered-usage event (identifier minted once, reused on retry, never re-minted) batched to the marketplace. The user dashboard displays live subscription status, free-tier consumption, and the prepaid wallet balance to six decimal places — never a static number.

---

9. Competitive Positioning

AlarmLogic's differentiation is safety-first independence: while incumbents write suppression into the controller, AlarmLogic deliberately does not — which removes the LOPA-revalidation and change-management friction that stalls incumbent deployments, and earns operator trust with explainable evidence instead of black-box scores.

---

10. Target Customers & Go-to-Market

Primary segments (blueprint's top-ten ranking): petroleum refining, petrochemical/chemical, LNG & gas processing, offshore/FPSO, power generation. Expansion: specialty/batch chemicals, mining & metals, pulp & paper, pharma (21 CFR Part 11 audit trails), water utilities.

Buying centers: Process Safety Manager, Automation/Controls Engineer, Plant Manager, Reliability Director, and the insurance/regulatory reviewer they answer to.

Entry path: a shadow-mode pilot ingesting a single unit's historian (zero control risk, no recertification) over the bundled OPC UA / SOE / MQTT collectors, producing the ISA-18.2 dashboard and bad-actor list in the first week — a concrete, day-one deliverable that justifies the subscription before any causal feature is even exercised. The marketplace live mode remains credential-dependent; every collector has been proven against real in-process wire peers (node-opcua server, aedes broker, SOE file) ahead of first site hookup.

---

11. Business Case & Unit Economics

Modeled per typical continuous-processing facility (e.g., a 250 kbpd refinery or 1.2 MTPA ethylene cracker):

At an enterprise subscription of $120k–$180k per site per year, the payback is 1.2–3.5 months under the Expected Model — a capital-allocation argument plant leadership can take to the board unchanged.

---

12. Roadmap

Principles locked for all releases: zero write-back, zero suppression, explainable reasoning, and marketplace-native commercial reality.

---

AlarmLogic — advisory intelligence, not control.