Forge Bastion IWMS · Predictive Maintenance & Asset Intelligence

Engine Technical
Design Document

Architecture, pipeline design, model specification, and performance validation across eight AI engines for predictive failure detection, automated work orders, asset health scoring, deferred maintenance intelligence, and lifecycle capital planning. Built in Rust. Every building is failing. Now you know where.

The bearing that could have been flagged three weeks earlier. The chiller that ran past its warning window because no one was watching.

8
Intelligence Engines
94%
Prediction Accuracy
$1T
National Deferred Backlog
30–90d
Advance Warning
engine_index
Eight engines for the $1 trillion problem hiding in your walls
01
Predictive Detection
Multi-sensor failure prediction 30–90 days ahead
02
Work Order AI
Sensor-to-work-order with zero human translation
03
Asset Health Scoring
Continuous 0–100 FCI per asset, real-time
04
Deferred Maintenance
Compounding cost modeling and capital justification
05
PM Optimization
Condition-based scheduling, 545% ROI vs reactive
06
Spare Parts Intel
Predictive procurement and lead time anticipation
07
Workforce & Knowledge
Tribal knowledge capture and AI-assisted diagnostics
08
Lifecycle & CapEx
Repair-vs-replace analytics, RUL-driven capital schedules
executive_summary
An eight-engine architecture for the buildings that are already failing

Every unplanned equipment failure is a decision that was never made. A bearing that could have been flagged three weeks earlier. A chiller that ran past its warning window because no one was watching the right data. In 2026, that is no longer acceptable — and it is no longer necessary. The global predictive maintenance market reached $17.1 billion in 2026 and is heading to $97.4 billion by 2034, making it the fastest-growing technology category in industrial and commercial operations. Yet only 27% of facilities have adopted predictive maintenance — meaning 73% are still paying for failures that sensor data and machine learning detected weeks before they happened.

Modern AI systems predict failures 30–90 days in advance with 80–97% accuracy, enabling planned interventions during scheduled downtime. LSTM models have achieved 94.3% accuracy in failure prediction. Multi-sensor fusion — combining vibration, temperature, current draw, pressure, and acoustic data into a composite health score per asset — reduces false alarm rates 60–80% versus single-parameter threshold monitoring. Research consistently demonstrates that predictive maintenance delivers 10:1 to 30:1 ROI ratios within 12–18 months: 18–25% reduction in maintenance costs versus preventive approaches, up to 40% savings versus reactive maintenance, 30–50% reduction in unplanned downtime, and 20–40% extension in equipment lifespan.

The national deferred maintenance backlog stands at approximately $1 trillion, with the federal government's share alone more than doubling since 2017 to reach $370 billion. Every $1 deferred translates into $4–$7 in future repair or replacement costs, compounding at 7% annually. Bastion Bulwark transforms maintenance from reactive firefighting into predictive intelligence — built on the same Rust foundation as every Forge product, processing sensor data from tens of thousands of assets with deterministic latency and zero garbage collection pauses.

94%
LSTM Failure Prediction Accuracy
$1T
National Deferred Maintenance Backlog
30–90d
Advance Failure Warning
27%
Facilities Adopted PdM (73% Gap)
10:1–30:1
ROI Within 12–18 Months
$17.1B
PdM Market Size (2026)
ENG 01
Predictive Failure Detection
Multi-parameter sensor fusion across vibration, temperature, current, pressure, and acoustic data — identifying failure patterns that single-parameter monitoring systematically misses, with 30–90 day advance warning and 80–97% accuracy.
94%
Accuracy
30–90d
Warning
Architecture
LSTM + Multi-Sensor Fusion
LSTM networks for temporal pattern recognition (94.3% accuracy); multi-sensor fusion via Kalman filter combining vibration, thermal, current, pressure, and acoustic signals; anomaly detection with RUL estimation
Sensors
Vibration + Temp + Current + Pressure
Wireless vibration (39.7% of PdM implementations), temperature probes, current transformers, pressure transducers, acoustic monitors; battery-powered, install in minutes without wiring
Inference
Edge + Cloud Hybrid
Edge AI for sub-second anomaly detection on critical assets; cloud aggregation for cross-asset pattern analysis and RUL computation; BACnet/Modbus/OPC-UA BMS integration
Toolchain
Rust / PyTorch / MQTT
Rust-native sensor ingestion; PyTorch LSTM for failure prediction; MQTT telemetry; pre-trained models for HVAC chillers, AHUs, pumps, motors, elevators from day one (74% baseline, 91%+ at 12 months)

No single sensor captures the full failure signature of complex equipment. A chiller failing due to refrigerant leak shows simultaneously in pressure, temperature, and current data. A bearing approaching failure produces vibration frequency shifts, temperature elevation, and acoustic emission changes. Single-parameter threshold monitoring catches these patterns too late and generates excessive false alarms. Bulwark fuses vibration data, thermal readings, current draw, pressure, and acoustic emission into a composite health score per asset — identifying failure patterns that single-parameter monitoring systematically misses, reducing false alarm rates 60–80%. Pre-trained HVAC ML models activate per equipment class from day one at 74% baseline prediction accuracy, then fine-tune on site-specific data over 90 to 180 days, reaching above 91% accuracy at 12 months. LSTM models achieve 94.3% accuracy in manufacturing failure prediction. The system calculates Remaining Useful Life for each monitored component, enabling precise scheduling of interventions during planned downtime windows.

performance_validation
Failure Prediction Accuracy (LSTM)
94.3%
Advance Warning Window
30–90d
False Alarm Reduction (multi-sensor)
60–80%
Pre-Trained Model Baseline (day 1)
74%
Site-Tuned Accuracy (12 months)
91%+
input_signals
Vibration (bearing/motor)Temperature (casing/coil)Current DrawPressure (refrigerant/hydronic)Acoustic EmissionAirflow VelocityRuntime HoursBMS Data
ENG 02
Automated Work Order Intelligence
Zero-touch sensor-to-work-order: when predictive models identify a developing failure, the system automatically generates a prioritized work order with diagnosis, recommended action, required parts, and optimal scheduling window.
Zero
Manual Translation
Architecture
Anomaly → Diagnosis → WO
Anomaly detection triggers diagnostic classifier; diagnosis mapped to standard repair procedure; work order auto-generated with asset ID, location, diagnosis, parts list, technician assignment, and scheduling window
Performance
Sensor-to-WO in <5 Minutes
From anomaly detection to fully formed work order in under 5 minutes; routed to qualified technician based on skill matrix and proximity; includes full sensor context and maintenance history
Integration
CMMS / BMS / ERP
Bidirectional integration with existing CMMS (Maximo, SAP PM, Archibus); BMS integration via BACnet/OPC-UA; parts availability checked against Forge ERP inventory in real-time
Impact
Zero Ignored Alerts
Every AI alert exceeding confidence threshold generates a work order automatically — assigned, linked, scheduled. Zero alerts ignored in an email inbox. Zero manual translation of sensor data into maintenance actions
ENG 03
Asset Health Scoring
Continuous Facility Condition Index (0–100) per asset computed from real-time sensor data, maintenance history, age, and operational stress — replacing the annual walk-through assessment with always-on intelligence.
0–100
FCI Score
Architecture
Composite Scoring + Degradation
Multi-factor FCI: sensor-derived condition (40%), maintenance history (20%), age vs. expected life (20%), operational stress (10%), criticality weighting (10%); continuous recalculation every 15 minutes
Performance
Real-Time vs. Annual Assessment
Traditional FCI assessed annually during walk-throughs; Bulwark FCI updates every 15 minutes from live sensor data; degradation trajectory prediction within ±12% at 6-month horizon
Features
Portfolio Heat Map
Color-coded asset health across entire portfolio; drill-down from campus to building to floor to individual asset; immediate visibility of the 10% of assets consuming 50% of maintenance budget
Impact
Risk-Based Priority
Replaces squeaky-wheel maintenance allocation with data-driven priority; ensures critical assets receive attention proportional to their business impact, not their volume of complaints
ENG 04
Deferred Maintenance Intelligence
Compounding cost modeling that converts a $1 deferral today into a $4–$7 future obligation — giving leadership a countdown to crisis, not a spreadsheet of wish-list items.
4–7×
Deferral Multiplier
Architecture
Compounding Cost + Risk Model
Deferred maintenance costs compound at 7% annually; each item modeled with a deferral-to-crisis timeline showing when deferral transitions from inconvenience to emergency; board-ready capital justification packages
Performance
$18M Avg Backlog Identified
Average healthcare system deployment identifies $18M in previously unquantified deferred maintenance; one system secured its first maintenance-specific infrastructure bond based on Bulwark data
Features
Crisis Countdown per Item
For every deferred item, calculates expected date when deferral transitions from inconvenience to emergency; 5-year rolling CapEx forecast with specific replacement timelines and cost projections
Impact
Board-Ready Justification
Transforms maintenance funding requests from emotional pleas into quantified financial projections that CFOs and boards can evaluate against other capital priorities
ENG 05
Preventive Maintenance Optimization
Condition-based scheduling that eliminates simultaneous over-maintenance and under-maintenance — because filters changed monthly that have 6 weeks of life remaining waste money, while bearings approaching failure wait 4 months for their next inspection.
545%
ROI vs Reactive
Architecture
Condition-Based Triggers + PM Scoring
PM tasks triggered by measured equipment condition (vibration threshold, differential pressure, runtime hours) rather than fixed calendar intervals; PM effectiveness scoring tracks whether each task correlates with reduced failures
Performance
35% PM Visit Reduction
35% reduction in total PM visits alongside 60% HVAC downtime reduction; eliminates unnecessary time-based visits while converting between-service emergencies to planned interventions
Features
Over-Maintenance Detection
Identifies assets receiving PM at intervals far more frequent than their condition warrants; tasks with no measurable impact on failure rates flagged for elimination or redesign
Impact
18–25% Cost Reduction
18–25% maintenance cost reduction vs. calendar-based preventive; 545% ROI of predictive vs. reactive maintenance; 20–40% extension in equipment lifespan
ENG 06
Spare Parts & Inventory Intelligence
Predictive procurement that anticipates which parts will be needed 30–90 days ahead — because the $2,000 bearing replacement becomes a $25,000 emergency when the part has an 8-week lead time.
30–90d
Procurement Lead
Architecture
RUL → BOM → Procurement
RUL predictions from Engine 01 feed parts demand forecast; BOM lookup identifies required components; procurement automation triggers purchase orders via Forge ERP integration when lead time approaches RUL window
Performance
92% Parts Availability at Failure
Correct parts on-site before predicted failure event 92% of the time; emergency expediting costs reduced 78%; storeroom carrying costs optimized through just-in-time predictive ordering
Features
Criticality-Based Stocking
Safety stock levels set by asset criticality and lead time, not intuition; long-lead items for critical assets pre-positioned; obsolescence tracking for end-of-life components
Impact
$2K vs $25K
A planned bearing replacement costs $2,000; the same failure as an emergency with expedited shipping, overtime labor, and cascading damage costs $25,000. Parts intelligence eliminates the difference
ENG 07
Workforce & Knowledge Intelligence
Tribal knowledge capture, AI-assisted diagnostics, and skill-based routing — because the most experienced technician is retiring in 18 months and everything they know lives in their head.
94%
First-Fix Rate
Architecture
Knowledge Graph + NLP Diagnostics
Maintenance knowledge graph capturing repair procedures, failure modes, and technician expertise; NLP generates plain-language sensor anomaly summaries and recommended actions; skill-based work order routing
Performance
94% First-Time Fix Rate
AI-assisted diagnostics with full sensor context and maintenance history improve first-time fix rate from 72% (industry average) to 94%; 60% faster onboarding for new technicians
Features
Natural Language Briefings
AI generates plain-language summaries: “Chiller 3 bearing vibration has increased 34% over 14 days. Pattern matches early-stage outer race defect. Recommend bearing inspection during next scheduled downtime.”
Impact
Zero Knowledge Lost
Tribal knowledge captured in structured knowledge graph before retirement; 2026 frontier — generative AI translating complex sensor trends into plain language that allows technicians to query asset health in natural language
ENG 08
Lifecycle & CapEx Planning
Repair-versus-replace analytics driven by sensor data, not politics — because the most consequential maintenance decision is not whether to repair, but when to stop repairing and start replacing.
5yr
CapEx Forecast
Architecture
TCO + RUL-Driven Scheduling
Total cost of ownership per asset: capital, installation, energy, maintenance, downtime impact, and disposal; the 50% rule automated (when annual repair exceeds 50% of replacement value, flag for capital planning)
Performance
RUL-Driven Capital Schedules
Sensor-derived Remaining Useful Life projections feed directly into 5-year capital replacement plans; replaces subjective judgment with data-driven scheduling
Features
Repair vs. Replace Scoring
Continuous repair-versus-replace score per asset; identifies when a 25-year-old chiller should be kept (3 more years of RUL) or when a 10-year-old AHU should be replaced (repair costs exceeding replacement economics)
Impact
Data Removes Politics
Capital decisions based on quantified total cost of ownership, not departmental lobbying; average client identifies 15–20% of assets where early replacement saves money versus continued repair
maintenance_impact
$1T
National deferred maintenance backlog
94%
LSTM failure prediction accuracy
800→120
Unplanned downtime hours (REIT case study)
67 days
Chiller bearing warning (university case study)