Industrial AI development
AI for oil and gas operations
Plan software for equipment monitoring, production surveillance, inspection review and field document search. Define the operational decision, required data and acceptance criteria before selecting a model.
TensorBlue develops AI applications and integrations. For an oil and gas engagement, the proposal should identify the operator’s engineering responsibilities, system interfaces and evidence required to progress from an offline pilot to a monitored deployment.
Choose a workflow with a measurable decision
Predictive maintenance
Start with one equipment class, such as pumps or compressors. Combine timestamped vibration, temperature and operating-load signals with maintenance work orders. A model should prioritize investigation; a reliability engineer decides whether to inspect or intervene.
Pilot evaluation
Measure useful warning time, missed events and false alerts per asset-day. Compare with current threshold alarms, and split evaluation by time and asset so the same failure does not appear in training and testing.
Production surveillance
Join well tests, pressure, flow and equipment-state records to flag changes that deserve an engineer’s attention. Distinguish sensor faults, planned shutdowns and changing operating regimes from model anomalies. Start with recommendations visible in a review dashboard.
Pilot evaluation
Use historical replay and a prospective observation period. Track accepted recommendations, operator review time and production measures adjusted for operating conditions. A forecast improvement alone does not establish additional production.
Inspection and pipeline review
Organize inspection reports, images and defect records for search and reviewer triage. Establish asset identifiers, inspection dates and expert labels before training an image model. Route uncertain or unfamiliar defects to qualified inspection personnel.
Pilot evaluation
Evaluate missed defects and false positives by defect category and inspection conditions. Keep model confidence separate from an engineering integrity assessment. Automated image classification does not establish that a pipeline is safe to operate.
Field document copilot
Build retrieval over approved manuals, procedures and maintenance records. Return the document, revision and passage behind an answer. Enforce source permissions, show missing evidence and let users report incorrect answers.
Pilot evaluation
Test questions with known answers and questions the system should decline. Measure citation correctness, answer support and access-control behavior. Operational procedures and emergency actions remain subject to the operator’s approved process.
Data readiness before model development
Create an asset-to-signal map and document units, sampling rates, clock drift, outages and changes to sensors or equipment. Work orders can contain delayed entries or inconsistent failure descriptions; assess how accurately they label the prediction target. Preserve raw observations alongside cleaning rules so engineers can reproduce an alert.
Use historical periods that cover relevant operating modes. A model trained on one field, equipment version or product mix may need separate evaluation before use elsewhere. Agree on the handling of missing or stale data, and show those conditions in the interface rather than silently treating them as normal readings.
Integration with operational technology
Map the approved route from a historian or data export to ingestion, quality checks, model evaluation and a review dashboard. Define network boundaries, service access, retention, audit logs and deployment ownership. Account for intermittent connectivity if inference or document access needs to operate at the edge.
NIST’s operational technology guidance addresses security alongside OT performance, reliability and safety requirements. Use it as a planning reference with the operator’s security team. The first pilot can be read-only; adding control-system writes is a separate scope with engineering authorization, tested operating limits and rollback procedures.
Read NIST SP 800-82 Rev. 3A staged delivery and acceptance plan
- Scope and data assessment. Choose an asset group or document collection, the decision owner and the existing baseline. Deliver a data-quality report, interface map and proposed evaluation criteria.
- Offline evaluation. Build a reproducible baseline and candidate model. Deliver held-out results, error analysis, limitations and examples for domain review. Stop or revise the pilot if the data cannot support the intended decision.
- Observation and integration. Connect approved data feeds and run in shadow mode. Record alerts or recommendations without changing operations, then compare model behavior with engineer review and actual events.
- Controlled rollout. Release an agreed advisory workflow after acceptance. Define monitoring, escalation, model updates, support responsibilities and conditions for disabling the system.
Evaluate value and the full cost
For maintenance, track avoided interruptions only where the intervention and attribution are documented. For a document copilot, measure time spent finding and verifying information with representative users. Include false-alert review, missed-event consequences, data engineering, hosting and ongoing maintenance in the economic assessment. Establish which metrics justify continuation before reviewing results.
As an external industry example, SLB and ADNOC describe an AI production-system optimization deployment using real-time operational data. That announcement shows one approach at a different organization; it does not establish TensorBlue project experience or predict the return from your pilot.
Read the SLB and ADNOC announcementPrepare a project brief
Bring the target workflow, available data sources, existing software interfaces, deployment constraints and the person who will accept the result. Share a description of the data first; sensitive operational records should use an agreed transfer process after scope and access are established.
