ACCURACY
Direct accuracy gains
Verified F1 +21%p with Knowledge-Infused Loss (KIL). Domain knowledge compensates for scarce training data.
Products
NUBISON Ontology
Process Ontology turns domain knowledge into a structure computers can understand,
giving the business a persistent knowledge asset and giving AI accuracy and explainability.
Two core challenges facing
industrial sites
They look like different problems, but the root cause is the same.
Business perspective
Domain knowledge disappearsWhen veteran engineers retire, core know-how is lost
- Even when documented, it's hard to search and reuse
- Equipment change histories depend on personal experience
- Different standards per plant limit integrated analysis
- New hires take a long time to adapt to the field
AI perspective
AI cannot explain "why"Limited training data caps accuracy
- Black-box models make results hard to interpret and validate
- Excessive false alarms erode field trust
- Environmental changes create heavy retraining burdens
- No explanation of the causes behind anomalies
The shared root of both problems
Domain knowledge does not exist in a form computers can understand
Our approach
Process Ontology structures domain knowledge and converts it into a digital asset that both the business and AI can use together.
The semantic layer
that connects domain knowledge and AI
Turn field know-how into a digital asset and use it to boost AI accuracy and explainability.
Field authors
Structured input of expert domain knowledge
- Normal operating criteria per stage
- Possible anomaly types
- Equipment-to-sensor relationships
- Cause-effect relationships (RCA)
New product
Process
Process
Ontology
Convert domain knowledge into
executable digital assets
AI consumes
Higher ML model accuracy and explainability
- Advanced training with domain constraints
- Fewer unnecessary false alarms
- Natural-language cause explanations generated automatically
- Transfer learning across new domains
Data → Meaning → Action
From data to meaning, from meaning to AI value
The Semantic Layer between data and AI transforms field knowledge into a form AI can understand.
Existing product
Datalake
What happened?"What happened"
- Raw data — sensors, video, logs
- Store · search · record
New product
Process Ontology
Why did it happen?"What it means"
- Domain knowledge and context
- Semantic meaning · relationships · explainability
Existing product
ML Platform
How should we respond?"How to respond"
- AI model training · deployment
- Prediction · inference · automation
Beyond model performance —
strengthening AI operations
FALSE ALARM
Block false alarms
The Ontology automatically filters predictions that violate domain rules, restoring field trust.
EXPLAINABILITY
Explainability
Every prediction ships with an automatic answer to "why," meeting safety and certification requirements.
DOMAIN ADAPTATION
Instant domain transfer
Handle new plants by swapping the Ontology plugin — very little model retraining needed.
DATA EFFICIENCY
Data efficiency
Operates with normal-operation data alone, solving the scarcity of labeled anomaly data.
EDGE EFFICIENCY
Edge efficiency
State-based gating runs costly analyses only on qualified segments, cutting GPU usage by 75%.
Improve cost, accuracy, and
operational efficiency at once
The impact of asset-based knowledge and AI advancement.
AI detection accuracy
+21%p
F1 72% → 93%
With Knowledge-Infused Loss
False alarms
50%↓
Field trust restored
Automatic filtering of domain-violating predictions
GPU usage
75%↓
State-gating impact
Precise analysis only on qualified spans
New-line effort
85%↓
Instant plugin swap
Reuse of shared Ontology
Onboarding time
6 wks
6 months → 6 weeks
Instantly searchable process manuals
Improve AI model accuracy and
explainability
One Ontology, three mechanisms that lift AI performance across training, inference, and output.
Training timeTraining
Higher model accuracy
L = WMSE + λ · P(violation)
Penalize predictions that violate the domain so the model learns physically consistent outputs from the start.
F1 +21%p · NSF-MAP validated
Inference timeInference
Automatic false-alarm blocking
if anomaly ∉ allowed[state]: block
The Ontology automatically filters out predictions that cannot happen in the current stage.
51 / 801 NoNose false predictions blocked
Output timeOutput
Explainability secured
explain(pred, state, sensor) → text
Add Ontology context to prediction results to auto-generate natural-language explanations for safety and certification audits.
ROUGE-1 0.91 · expert consensus
Validated
quantitative value
Validated on the rocket assembly line at USC Future Factories Lab and across four peer-reviewed papers.
F1 Score improvement
+21%p
72% → 93%
Knowledge-Infused Loss significantly boosts anomaly-prediction accuracy.
NSF-MAP · IJCAI 2025
RCA accuracy
94%
Expert-consensus alignment 0.91
Automate Root Cause Analysis to trace incidents down to root causes.
CausalTrace · AAAI 2026
Automated false positives blocked
51/ 801
NoNose false predictions
The Ontology validates and blocks anomaly predictions that cannot occur in the given stage.
AssemAI · ICMLA 2024Data becomes knowledge,
knowledge becomes action
The moment Process Ontology is added, the full AI cycle — data → knowledge → action — is complete.
WHATDatalake
What happened
Preserve sensors, video, and logs in chronological order
WHYProcess Ontology
Why it happened
Add domain meaning, normal-state criteria, and causality to the data
HOWML Platform
How to respond
Ontology-enhanced models predict, classify, and automate
Integrated Cycle
Integrated cycle
- 1
Datalake collects new data
- 2
Ontology assigns stage and sensor meaning
- 3
ML models predict with Ontology context
- 4
Predictions automatically carry Ontology explanations
- 5
Results and feedback flow back to Datalake
After step 5, the cycle loops back to step 1
Swap the Ontology alone —
ready for any industry
Change the industry-specific Ontology on the same platform to jointly create value for the field and AI.
Rotating machine diagnostics
Motors · pumps · compressors · turbines
Field
Turn ISO 10816 standards into a company knowledge asset
AI
Automate RCA from bearing-defect frequencies
Predictive maintenance +30–50%Catastrophic failures −70%
Assembly line quality
Automotive · Electronics · Precision Machinery
Field
Standardize normal patterns for 21 cycle states
AI
Stage-gating saves GPU 75% + F1 +21%p
New-line rollout: 6 months → 1 monthFalse alarms −60%
Process industry
Power · Chemical · Refining · Cement
Field
Standardize plants around Recipe and thermodynamic criteria
AI
Mass · Energy conservation checks / Digital Twin integration
Avoid catastrophic failuresAutomate safety and certification audits
Progressive adoption strategy — expand while securing value at each phase
Start small and stack the next step on top of validated value.
- Phase 13 months
PoC with 1 domain
Minimal ontology + explanation API
FieldFirst knowledge asset
AIAutomatic explanations
- Phase 23–6 months
Enable validation
Change tracking + auto-block off-spec predictions
FieldOperate audit logs
AIFalse alarms −50%
- Phase 36–12 months
Integrate KIL training
Enforce domain constraints during retraining
FieldStandardization complete
AIF1 +15–20%p
- Phase 46–12 months
Multi-plant expansion
LLM-based ontology auto-generation
FieldMulti-plant sharing
AIRCA automated