Skip to content

Predictive Maintenance: Topology-Aware Failure Forecasting

Industry: Industrial Manufacturing | Role: VP Reliability, Plant Manager | Time to value: 2-3 hours

The Problem

Your CMMS lists 500 assets due for maintenance. Your team has capacity to service ten this week. Pick the wrong ten and the eleventh fails in production, takes out three downstream assets through process-line dependencies, and turns a planned maintenance window into a multi-day unplanned outage. Sensor-based predictive maintenance tools score each asset in isolation; they do not know that asset 47's failure cascades into the entire downstream subsystem.

The Traditional Approach

Reliability engineers run vibration analysis, temperature trending, and oil sampling per asset, then build a ranked list in a spreadsheet or specialist PdM tool. CMMS handles work-order scheduling but knows nothing about sensor data. Process-line topology lives in a P&ID drawing nobody has digitized. Calibration of the underlying ML scores against actual outcomes happens annually, if at all -- most teams ship raw classifier outputs and rely on engineer judgment to interpret them. Cascade-aware scheduling -- servicing the asset whose failure takes down the most downstream value -- is informal.

With Uni

The notebook ingests AI4I 2020 sensor data plus a synthesized process-line topology directly into Uni's graph. A failure-likelihood classifier runs as a Locy neural predicate registered via LocyConfig.register_classifier. The classifier output is calibrated in-language using CALIBRATE ... METHOD platt_scaling against the dataset's ground-truth actual_failed labels, then validated with VALIDATE on Brier score and accuracy — the notebook demonstrates the full calibrate/validate loop end to end (on this small slice, Platt scaling does not necessarily improve Brier). Component-level risk composes through FOLD MNOR(1.0 - c.health) over a HAS_PART child set; line-level reliability composes through FOLD MPROD(1.0 - failure_likelihood(...)) across UPSTREAM_OF chains using inline classifier invocation, and a recursive upstream_reaches rule plus IS NOT stratified negation derive blast-radius and healthy-asset sets. The EXPLAIN RULE trace surfaces a classifier call's NeuralProvenance (model name and raw score) for audit.

What You'll See

  • Calibrated per-asset failure probabilities from a registered Python classifier (in production, an ONNX-exported XGBoost or similar)
  • Line-level joint reliability rollups computed by composing per-asset probabilities through FOLD MPROD(1.0 - failure_likelihood(...)) across UPSTREAM_OF chains
  • Component-level risk aggregation via FOLD MNOR(1.0 - c.health) over a HAS_PART child set
  • Calibration delta: raw Brier vs Platt-calibrated Brier on the labeled dataset
  • Validation report (VALIDATE) with Brier score and accuracy
  • Audit trail (EXPLAIN RULE with NeuralProvenance) on demand: the rule bindings and the classifier call's model name and raw score

Why It Matters

Unplanned outages at a typical continuous-process plant cost $50k-200k per hour. Reducing false-positive maintenance alerts and false-negative missed failures directly protects throughput. Calibrated probabilities plus topology-aware composition (line-level reliability built from per-asset risks) change which ten assets get serviced this week.

Data: AI4I 2020 Predictive Maintenance Dataset (UCI #601, CC BY 4.0) -- 10k machine instances, 14 sensor features, 5 failure modes. Process-line topology is synthesized for the notebook and clearly marked as such.

Run the notebook →