The Predictive Maintenance Workflow
Module 5 · Predictive Maintenance and Anomaly Detection

The Predictive Maintenance Workflow

Alerts, work orders, and human-in-the-loop decisions — how digital twin predictions connect to real maintenance operations.

The Three Maintenance Strategies

Before digital twins, facilities operated under one of two strategies:

Reactive maintenance — fix it when it breaks. Low upfront cost, high downtime cost. An unplanned failure in a production line can cost $50,000–$500,000 per hour.

Preventive maintenance — replace components on a fixed schedule (every 6 months, every 10,000 cycles). Predictable but wasteful — components are often replaced when they still have 40–60% of their useful life remaining.

Predictive maintenance — replace components when degradation data indicates failure is imminent. Maximum component utilization, minimum unplanned downtime. Enabled entirely by the digital twin’s anomaly detection and RUL estimation.


The End-to-End Workflow

Digital Twin
    │
    │  Anomaly detected: joint2 torque drift
    │  RUL estimate: 58 ± 14 hours
    ▼
Alert Management System
    │
    │  Severity: WARNING
    │  Recommended action: inspect joint2 bearing
    │  Suggested window: next maintenance shift
    ▼
CMMS (Computerized Maintenance Management System)
    │
    │  Work order created: WO-2024-8847
    │  Assigned to: Maintenance Team B
    │  Parts: 6205-2RS bearing × 2, bearing puller kit
    │  Scheduled: Wednesday 02:00–04:00
    ▼
Maintenance Execution
    │
    │  Technician inspects joint2
    │  Bearing shows pitting on inner race (confirmed)
    │  Bearing replaced
    ▼
Twin Feedback Loop
    │
    │  Post-replacement baseline re-established
    │  RUL clock restarted for joint2
    │  Maintenance event logged for model refinement
    ▼
Model Improvement
    │
    │  Actual failure time vs. predicted failure time recorded
    │  Degradation model parameters updated
    │  Anomaly detection thresholds re-calibrated

The CMMS Integration

A Computerized Maintenance Management System (CMMS) — systems like IBM Maximo, SAP PM, or Fiix — manages work orders, parts inventory, and technician scheduling.

The digital twin integrates with the CMMS via API, converting predictions into actionable work orders:

def create_maintenance_work_order(alert: MaintenanceAlert, rul: RULEstimate):
    cmms_client.create_work_order(
        asset_id=alert.asset_id,
        description=f"Predictive alert: {alert.anomaly_type} detected on {alert.joint_id}",
        priority=determine_priority(rul.rul_hours),
        required_parts=parts_lookup(alert.joint_id, alert.anomaly_type),
        latest_start=datetime.now() + timedelta(hours=rul.confidence_interval[0]),
        instructions=generate_inspection_procedure(alert.joint_id),
        source="digital_twin",
        prediction_confidence=rul.confidence_interval,
    )

def determine_priority(rul_hours):
    if rul_hours < 24:
        return "CRITICAL"
    elif rul_hours < 72:
        return "HIGH"
    elif rul_hours < 168:
        return "MEDIUM"
    else:
        return "LOW"

Closing the Loop: Feeding Outcomes Back

The prediction is only half the system. What actually happened is equally important:

  • Technician inspects joint2: confirmed bearing failure → prediction was correct
  • Technician inspects joint5: no defect found → false positive
  • Component runs another 200 hours before failure → RUL was underestimated

Each outcome is logged and used to refine the degradation model:

  • True positives: validate the anomaly detection threshold
  • False positives: tighten the threshold or add a confirmation window
  • RUL errors: adjust the degradation rate model

Over time, the system becomes more accurate — the twin learns the specific degradation behavior of each asset in this facility, under these operating conditions.


Human-in-the-Loop Design

Autonomous maintenance scheduling is technically possible but operationally risky. Best practice: the twin recommends, the human decides.

Key design principles:

  • Explain predictions — show the technician the residual trend that triggered the alert, not just the alert
  • Show confidence — present the RUL confidence interval, not a single number
  • Enable override — operators must be able to dismiss or escalate alerts with a documented reason
  • Track overrides — when an operator dismisses a warning and the component fails anyway, that’s data

Human oversight is not a limitation on the twin’s capability — it’s a quality control layer that keeps the system honest and builds operator trust over time.


🧠 Knowledge Check

Why is it important to log the outcome of every digital twin maintenance prediction, including false positives?

Reveal Answer & Explanation

The twin’s anomaly detection and RUL models are calibrated on historical data. Without outcome feedback, those models can’t improve — the system is forever as accurate as it was on day one.

False positives in particular are valuable: they reveal where the model’s sensitivity is miscalibrated, where normal operating variation is being mistaken for degradation. Each logged false positive is a training example for threshold refinement. Over months, a system that tracks outcomes becomes significantly more precise than one that generates alerts and never learns from them.


What’s Next

Module 6 brings everything together — real-world digital twin deployments across industries, and the open research problems that make this field one of the most active in engineering today.