EI Sentinel^™

1 · Product Overview

Product

EI Sentinel — pipeline leak detection & localization platform

Application

Computational Pipeline Monitoring (CPM) per API 1130; Pipeline Leak Detection Program (PLDP) per API RP 1175

Coverage

Hazardous liquids (crude, NGL, Y-grade, refined products) + natural gas gathering & transmission

Detection method

Cross-stream Negative Pressure Wave (NPW) analysis via deep ML inference; per-segment model

Detection latency

< 60 s typical (NPW onset to operator alert)

Localization accuracy

± 25 ft typical (NPW arrival-time differential, product-dependent)

Form factor

Sealed field enclosure + AWS-hosted cloud platform (HaaS — no on-premises servers required)

Minimum deployment

2 stations per pipeline segment (upstream + downstream); 20 mile maximum station spacing

2 · Field Device Hardware (EI Sentinel station)

Compute platform

Raspberry Pi 4B (4 GB RAM) standard; image is universal-compatible across Pi 3B+, Pi 4, Pi Zero 2W, Pi 5

Operating system

Raspberry Pi OS Lite, 32-bit Bookworm; Linux kernel 6.6+ LTS; headless (no GUI)

Pressure transducer

Keller series digital pressure sensor, USB interface, factory calibration; 100 Hz continuous sample rate

Pressure resolution

Sub-PSI deviation sensitivity (demonstrated detection at Δ -0.4 PSI single-event)

Pressure range

Configurable 0–3000 PSI absolute (selected per pipeline MOP)

Sensor temperature

Reported by transducer; 5-minute update cadence

GPS receiver

I²C / UART GPS module with PPS output; NTP-disciplined via GPSD + PPS for sub-millisecond timestamping

Vibration sensor

USB MEMS accelerometer (3-axis); environmental context for alert disambiguation

Display

OLED status panel — uptime, cellular signal, GPS fix, alert state at a glance

Hardware watchdog

BCM2835 (Pi 3/4) or BCM2712 (Pi 5); 15-second timeout, auto-reboot on hang

Field connections

1× pressure tap (¼″ NPT) · 2× LTE main antennas · 1× LTE diversity · 1× GPS antenna · 1× Ethernet (to modem) · 1× DC power

3 · Power & Environmental

Input voltage

12 – 24 VDC

Power consumption

Typical 5 W; peak 12 W during modem boot + cellular tower acquisition

Solar-compatible

Yes — direct off 12 V solar/battery rail; no inverter required

Operating temperature

-20 °C to +60 °C (-4 °F to +140 °F) standard; extended range available

Storage temperature

-30 °C to +70 °C

Humidity

0–95% non-condensing

Enclosure

NEMA 4X / IP66 sealed industrial polycarbonate; gland-sealed entries

Mounting

DIN-rail interior / bolt-on exterior; fits standard valve cabinets

Vibration / shock

IEC 60068-2-6 / -27 compliant for fixed-installation industrial environments

Surge protection

Cellular + Ethernet + DC power surge suppressors in-enclosure

4 · Communications & Backhaul

Primary backhaul

4G LTE cellular via Sierra Wireless RV55 (ALEOS) or Cradlepoint industrial router

Carriers

Verizon, AT&T (US); roaming SIMs available for cross-border / remote

Cellular antennas

2× LTE main + 1× diversity (MIMO 2×2); external mast-mount or panel

Auto-recovery

Ping watchdog every 120 s (8.8.8.8 / 8.8.4.4); modem power-cycle after 5 failures; full device reboot after 50 consecutive failures

Telemetry protocol

MQTT over TLS 1.2+; topic-per-device authentication

Telemetry rate

100 sample/s pressure + 1 sample/min metadata + 0.2 sample/min device health

Average bandwidth

~12–18 MB / day / station (steady-state, compressed)

Store-and-forward

Local SQLite buffer for offline operation; auto-backfill on connection restore (roadmap — v2.7)

Remote management

MQTT command channel — reboot, config update, diagnostics request, model reload

Time sync

GPS PPS + NTP server fallback; sub-millisecond cross-station timestamp accuracy for NPW localization

Backup ingress

Ethernet WAN supported (where available); cellular remains primary

5 · Machine-Learning Algorithm

Algorithm class

EI Combined v3 — cross-stream deep encoder-decoder with Time-delay Embedding Differential (TiED) layer

Inputs

Upstream pressure stream + downstream pressure stream + segment metadata (product, flow direction, geometry)

Output

Predicted downstream pressure on every 10 ms sample; deviation from prediction drives alert state

Training infrastructure

PyTorch on NVIDIA RTX 3090 / 3080 GPUs; export to TensorFlow inference graph via numpy weight migration

Training data window

14-day minimum; 60-day standard for stable model

Per-segment models

One dedicated model per pipeline segment — learns pump-curve, viscosity, geometry, and flow direction unique to that line

Inference

TensorFlow 2.x on AWS c5.xlarge model nodes; Docker Swarm orchestrated; per-segment service isolation

Inference latency

< 50 ms per sample evaluation; pipeline of evaluations runs continuously per stream

Threshold formula

threshold = max(5×MAE, 1.5×MaxErr, 2×P95, floor=2.0) — per-segment, calibrated against training-time accuracy

Median model accuracy

0.219 PSI average prediction error on strongest segments

Worst-case accuracy

7.59 PSI max prediction error on most-difficult segments (vs 64 PSI on legacy single-stream models)

Retraining cadence

Monthly full retrain + weekly fine-tune; gate against current production performance; automatic rollback on regression

Shadow-mode validation

New model runs model_silent=1 for 24 h minimum before live alerts enabled

6 · Alert Types & Pattern Detection

PCR (Pressure Change Rate)

Primary leak signature — model prediction error exceeds learned threshold. Includes magnitude, duration, GPS estimate.

UOC (Unplanned Operating Condition)

Pressure trajectory outside learned envelope but not consistent with leak signature; operator review required.

DF (Data Fault)

Sensor or device-health alert — out-of-range readings, communication loss, GPS fix lost, NTP drift.

Pattern anomaly

Time-of-day + GPS clustering analyzer identifies repeating sub-threshold events suggestive of scheduled extraction (hot-tap theft) or developing defect pattern.

Cooldown / suppression

Configurable per-analyst cooldown (default 60 s) to prevent alert storming during ongoing events.

Operational-transient absorption

Pump starts, valve openings, pig launches, batch changes — learned by the model and absorbed into the prediction envelope, not flagged.

7 · SCADA & System Integration

Modbus TCP

Standard holding-register set: alarm state, current pressure, segment health, alert magnitude/duration/coordinates. Custom register mapping available.

OPC UA

OPC Unified Architecture server endpoint; SecurityPolicy Basic256Sha256 supported; certificate-based auth.

REST API

HTTPS JSON endpoints — pull alert history, segment configuration, model status, device health.

Webhook

POST-JSON to operator-defined URL on PCR alerts (PagerDuty, Slack, Opsgenie, custom).

Email / SMS

Per-analyst distribution list; alert magnitude + GPS in body.

CSV / Parquet export

Historical alert log + pressure traces, on-demand or scheduled; suitable for IMP audit binders.

Pipeline route data

KMZ file upload — operator-provided pipeline centerlines; renders alert GPS on the operator's actual route map.

Identity / SSO

Role-based dashboard auth; SAML / OAuth2 integration available for enterprise SSO (add-on).

8 · Cloud Infrastructure (AWS, us-east-1)

Application cluster

5× t3.large EC2 nodes, Docker Swarm orchestrated; dashboard, alert generation, SCADA endpoints, RabbitMQ message bus

Model inference cluster

7× c5.xlarge EC2 nodes; one Docker service per pipeline segment running TensorFlow inference

MQTT broker

Amazon MQ — Apache ActiveMQ, TLS-only, single-tenant broker per customer-class

Production database

AWS RDS MariaDB 10.11.15, db.m5.large class; leakmon-main — 2 TB allocated

Telemetry database

AWS RDS MariaDB 10.11.15, db.m5.large class; leakmon-raw — 3.8 TB allocated; high-volume time-series

Model storage

AWS EFS — shared model-weight volume mounted to all inference nodes (general-purpose bursting throughput)

Load balancers

ALB (dashboard, MQTT brokers), NLB (raw MQTT ingress), CLB (Modbus endpoints) — all TLS-terminated

DNS / certificates

AWS Route 53 + ACM TLS certs, automatic renewal

Telemetry ingestion rate

~1.28 billion pressure samples / day across the fleet (steady-state, all customers)

Region

us-east-1 primary; disaster-recovery posture is point-in-time RDS snapshots + EFS backup

9 · Device Operations & Self-Healing

First-boot provisioning

Auto-detect serial + MAC; POST to provisioning API; receive MQTT credentials + device-name assignment; one-time, self-disable

Sensor auto-detect

USB enumeration discovers Keller transducer + vibration sensor at boot; no manual configuration

Supervisor

Linux supervisord manages the leakmon application — auto-restart, 10-second grace, exponential backoff

Hardware watchdog

15-second BCM watchdog timeout; auto-reboots the Pi on kernel hang or stuck process

Cellular watchdog

120-second ping interval; modem power-cycle after 5 failures; full reboot after 50 failures

Log rotation

10 MB max + 3 backup files; prevents disk exhaustion across long deployments

Time discipline

NTP with GPS PPS reference; secondary NTP server fallback; cross-station time drift < 1 ms typical

Security posture

SSH key-only authentication; default passwords disabled at image build; MQTT credentials provisioned per-device

Update mechanism

MQTT-delivered config + software updates; atomic apply with rollback on health-check failure

10 · Regulatory Compliance Mapping

49 CFR § 195.134

CPM equipment installation on new/replaced hazardous-liquid pipelines

49 CFR § 195.444

CPM operation in compliance with API 1130 — algorithm class, alarm scheme, performance test, record-keeping

49 CFR § 195.452

Integrity management on Could-Affect-HCA segments — per-segment continuous monitoring

49 CFR § 195.50

Reportable accident definition — faster detection narrows release magnitude below reporting threshold

49 CFR Part 192

Natural gas transmission / gathering — extension applicable per PIPES Act of 2020 final rule

API 1130

CPM Algorithm Description + performance-test record + maintenance procedure + alarm management — full conformance package provided

API RP 1175

PLDP management lifecycle artifacts — strategy, methods, performance management, training/competency records

API RP 1162

Public-awareness program supplement (alert routing into operator's existing emergency response protocol)

PIPES Act 2020

Compliant with PHMSA Final Rule extending leak detection to gas gathering & transmission

State overlays

TX RRC §3.107 · CA OSFM Hazardous Liquid LDP · NY 16 NYCRR Part 754 · all addressed

11 · Performance & Detection

Pressure sample rate

100 Hz continuous per station

Detection latency (NPW)

< 60 s typical from leak onset to operator alert

Detection latency (synthetic)

38 s demonstrated on injected NPW signature at Δ -0.95 PSI

GPS localization accuracy

± 25 ft typical envelope (product-speed dependent)

Wave-speed constants

Natural gas ~350 m/s · NGL ~1000 m/s · refined products ~1100 m/s · crude ~1200 m/s

False-alarm reduction

99.8% measured across multi-operator deployment (vs legacy SCADA threshold systems)

Operator alert load (post-deploy)

Single-digit per day per segment (typical), down from thousands per day on legacy CPM

Sub-PSI sensitivity

Single-event detection demonstrated at Δ -0.4 PSI magnitude

Pattern-detection capability

Recurring sub-threshold events surfaced via time-of-day + GPS clustering — applies to hot-tap detection & developing defects alike

Alert acknowledgment

Full audit trail with operator, timestamp, response notes; exportable for PHMSA audit binder

12 · Deployment Timeline

Hardware delivery

2 weeks after PO (US continental); pre-flashed and pre-provisioned

Per-station installation

~1 day (operator-installable); no mod/demob charges

First MQTT publish

Within 4 minutes of power-on; visible on dashboard immediately

Training data window

14 days minimum; 60 days for a stable v3 model

Per-segment model training

3–5 days on EI-side GPU infrastructure

Shadow-mode validation

24 hours minimum on production data before live alerts enabled

SCADA integration

1–3 days depending on Modbus register-map requirements

Total time-to-live

~30 days from PO to operational alerting

13 · Support, Warranty & SLA

Service model

Hardware-as-a-Service — Endpoint Integrity owns and replaces all field hardware over the term

Standard response

Business-hours (Mon–Fri 08:00–18:00 CT): 4-hour first-response on PCR-level alerts

24/7 SLA

Add-on tier: 30-minute first-response on PCR-level alerts; on-call sales engineering

Hardware replacement

Next-business-day shipment on confirmed device failure; on-site replacement option in enterprise tier

Software updates

Included; pushed via MQTT with atomic apply + rollback

Quarterly accuracy report

Standard deliverable — per-segment model accuracy, alert summary, recommended threshold adjustments

Annual on-site review

Add-on tier — engineering visit, IMP-document refresh, audit-prep walkthrough

Service term

1, 3, and 5-year subscription terms; 10% and 15% volume discounts on 3-yr / 5-yr respectively

14 · Pricing — Hardware-as-a-Service

Includes hardware (EI Sentinel device), cellular connectivity, AWS cloud infrastructure, per-segment ML model training + inference, dashboard access, SCADA integration (Modbus/OPC UA/REST), email/SMS alerts, business-hours support. Multi-year discounts: 10% on 3-yr, 15% on 5-yr. Volume discounts above 6 stations baked into Pipeline tier. Add-ons: 24/7 SLA (+$3.5k/station/year) · white-glove deployment ($5k/site) · custom Modbus register mapping ($7.5k one-time) · API 1130 algorithm description package ($9.5k one-time, included with 3-yr commit).

15 · Ordering & Contact

Pre-sales engineering

info@endpointintegrity.com · trace evaluations returned within 48 h

Sales

Tel: +1 (772) 402-4417

Procurement / RFP

Security questionnaire + SOC 2 / ISO 27001 attestation status available under NDA

Customer references

Tier-1 midstream operator references available under mutual NDA

Web

endpointintegrity.com · /spec-sheet · /compliance

Customer dashboard

dashboard.endpointintegrity.com