AI for oil and gas: a practical plan for refinery reliability

In a refinery, unplanned downtime goes straight to the bottom line. Lost hours on a unit that runs around the clock add up quickly, and the International Society of Automation has written about the scale of these losses across large manufacturers.

Reliability separates the best-run refineries from the rest. They operate at a higher Overall Equipment Effectiveness (OEE) than the average plant, and the difference shows in throughput, product quality and safety. Reliability is a driver of operating performance as much as a maintenance metric.

Pressure from environmental, social and governance (ESG) standards and tightening safety requirements raises the stakes. Regulators and customers expect higher operational integrity, which makes reliability a plant-wide concern rather than the maintenance department's alone.

Sustained improvement is still hard to achieve. Data sits in separate systems, teams work to different priorities, and maintenance cultures stay reactive. Most efforts produce incremental gains that fade. This article sets out a path to reduce downtime, spend maintenance budget where it matters, and keep improving.

Five moves to start this quarter

A full reliability program takes time and cultural change, but a few focused actions can start reducing downtime now. Each targets part of Overall Plant Reliability (OPR).

1. Run a critical-asset Pareto review. Use criticality analysis on the past year's outage data to identify the assets that cause the most downtime, and direct maintenance effort there first.
2. Set up a 24-hour root cause loop. A dedicated cross-functional team investigates every unplanned outage within a day of the equipment restarting, rather than weeks later. A fast root cause analysis (RCA) cycle speeds problem-solving and shortens mean time to repair (MTTR).
3. Tune operator rounds with failure logs. Adjust round checklists to look for the early warning signs that have preceded failures in the past, starting with the most common failure modes, so issues are caught before they escalate.
4. Digitize downtime tagging. A simple tagging application lets operators record equipment ID, failure mode and root cause with standardized codes on a mobile device. Clean, consistent classification makes trends visible and improvement measurable.
5. Hold a daily reliability review. A short meeting each day, around 15 minutes, reviews recent outages and emerging risks and tracks corrective actions. It builds accountability, reduces repeat faults and breaks down silos between teams.

Seven steps to a reliability-centered program

These steps are meant to make reliability gains last at plant scale.

Step 1: Agree on the metrics, OPR and OEE

Before improving reliability, the organization has to agree on what success looks like. Confusion between Overall Equipment Effectiveness (OEE) and Overall Plant Reliability (OPR) often holds progress back.

OEE measures the effectiveness of individual equipment by combining availability, performance and quality into a single percentage. It is an asset-level measure and does not reflect how integrated units interact.

OPR takes the plant-wide view. It reflects the combined effect of asset availability, product quality and process rate across multiple units. OPR sits at the top of a KPI tree that cascades down to area OEE and then to asset statistics such as mean time between failures (MTBF) and MTTR.

Teams often work to conflicting KPIs. Operators push throughput and OEE, maintenance tracks failure rates and repair times, and engineering focuses on process constraints and design changes. Without shared metrics, priorities clash and effort scatters. A well-structured KPI tree connects these views, so that maintenance work on MTBF visibly contributes to area OEE, which in turn raises plant OPR.

Common pitfalls distort the measurement:

• Micro-stops shorter than five minutes go unrecorded and inflate availability.
• Failure tagging is vague, with "other" or "unknown" dominating the records.
• Speed losses are misclassified as quality defects, which hides the root cause.

Reliable measurement needs reliable sources: distributed control system (DCS) historian logs, computerized maintenance management system (CMMS) records, operator logbooks, lab quality data, maintenance schedules and safety system alerts.

Step 2: Build the data backbone and failure history

Many refineries work with data that is incomplete, inconsistent or inaccurate, and every decision built on it inherits those errors.

A minimum viable data set includes operating run logs, maintenance histories, sensor streams and operator observations. Without that connected view, degradation trends and repeating failure patterns are nearly impossible to see.

A focused data hygiene sprint is often the best starting point:

• Catalog the data sources across the DCS, CMMS, lab systems and operator notes.
• Standardize naming conventions.
• Filter out noisy or faulty sensor readings.
• Validate the cleaned data with frontline teams.

Useful signals often sit unused in DCS historian archives. A slow drop in pump discharge pressure can foreshadow mechanical wear weeks before a shutdown. Thermal patterns in heat exchangers signal fouling long before efficiency falls off.

Treat sensor data with care. Instruments overdue for calibration and flatlined readings must be flagged, or they drive false alarms and misguided interventions. The cleaner and more complete the data backbone, the more confidently a plant can apply closed-loop optimization.

Step 3: Prioritize assets by criticality and risk

Failures do not carry equal risk. Some assets fail often but cause little disruption. Others fail rarely but can stop an entire process unit for days. The effort should go to the assets that threaten the plant's results.

Reliability-centered maintenance (RCM) frameworks guide this by evaluating each piece of equipment on the likelihood of failure and its consequences across operational, safety, environmental and financial dimensions.

A criticality matrix makes the priorities clear. High-probability, high-consequence assets need immediate attention, while low-risk equipment can be run to failure. Identifying the critical subset lets maintenance programs put resources where they count.

For critical equipment, failure mode and effects analysis (FMEA) identifies the failure modes, their effects and how each can be detected. That analysis informs inspection schedules, condition monitoring plans and design improvements. Cross-functional workshops with operations, maintenance and engineering improve the quality of the risk assessment and the buy-in that follows.

Step 4: Move from reactive to preventive and predictive maintenance

Reactive maintenance, where repair follows failure, is costly and disruptive. Preventive maintenance (PM) schedules inspections and replacements by elapsed time or usage, which can create unnecessary work or still miss failures.

Predictive maintenance (PdM) uses real-time data and analytics to detect early signs of degradation and schedule the intervention when it is needed. It extends asset life, reduces unexpected downtime and controls maintenance cost.

Failure patterns guide the strategy. The bathtub curve describes typical lifecycle failure rates: early infant mortality, then a stable period of random failures, then a wear-out phase. Weibull analysis turns historical failure data into models that set PM intervals. Instead of fixed schedules, teams adjust frequencies to the data and can often extend intervals without losing reliability.

PdM technologies give a closer view of equipment condition:

• Vibration analysis detects bearing wear or misalignment before it becomes audible.
• Infrared thermography reveals electrical and mechanical hot spots invisible to the eye.
• AI-based anomaly detection finds patterns across many sensors at once.

Over-maintenance also backfires. Excess interventions introduce human error, consume spares and stretch maintenance windows into downtime of their own. The right balance comes from the data.

Step 5: Add AI and closed-loop control

AI turns raw sensor and operating data into earlier, more specific warnings. Models trained on historical data detect multi-sensor anomaly patterns that single-tag alarms miss, and real-time monitoring flags deviations early enough to intervene before a failure.

Closed-loop systems go a step further. Connected to the distributed control system (DCS), programmable logic controllers (PLC) or supervisory control and data acquisition (SCADA) platforms, they adjust process parameters to hold the unit in its target operating window while operators keep authority, which reduces process variability and supports throughput.

AI also speeds root cause analysis, because models relate failure precursors to outcomes faster than manual review. That speed helps teams move from reactive fixes to prevention.

Models drift as the plant changes, so their predictions degrade unless they are retrained on fresh data. Effective programs build feedback loops in which operations and maintenance teams validate model outputs continuously. Before deploying AI, assess readiness: data quality, integration with the control systems, and organizational alignment.

Step 6: Build shared ownership and a reliability culture

Lasting reliability depends on shared accountability across operations, maintenance and engineering:

• Operators are the frontline custodians, responsible for routine care and early detection of anomalies.
• Maintenance teams carry out the optimized preventive and predictive work.
• Engineering identifies root causes, drives design improvements and supports continuous learning.

A RACI matrix (responsible, accountable, consulted, informed) clarifies roles and closes gaps. Regular cross-functional reliability reviews give teams a place to share lessons, update practices and reinforce goals. Training in reliability principles and AI literacy gives people the skills to use new tools well, and incentives tied to OPR improvement, such as bonuses, keep reliability a priority.

Step 7: Monitor, analyze and iterate

Continuous improvement keeps the program on track. Monitor leading indicators, such as early warning signals, alongside lagging indicators, such as downtime hours.

Weibull plots show failure patterns and how well maintenance is working. Plan-do-check-act (PDCA) cycles formalize iteration and adjust strategy to the data. Quarterly reliability health checks benchmark performance against industry standards and internal goals, and dashboards keep teams informed.

Troubleshooting persistent reliability problems

Some problems resist quick fixes. These come up often:

Know when to escalate a persistent problem to engineering or outside specialists, so it is resolved in time.

How Imubit applies this

In refining, Imubit applies Map, Model and Control to reliability and margin. Map traces lost availability and margin to the recurring decisions behind them, across operations, maintenance and engineering. Model uses Deep Learning Process Control® (DLPC®) to learn a nonlinear dynamic model of each unit from its own history, guided by first principles. Control takes the critical decisions to closed loop, writing setpoints to the existing APC or DCS minute by minute, inside limits set by plant engineers. On one crude furnace at one refinery, after the 2020 IMO specification change, fouling prediction and a crude diet study delivered a 33% reduction in furnace downtime.