Predictive Maintenance

Predictive Maintenance gives every asset a 0–100 health score, a level (Healthy / Watch / At risk / Critical), a list of ranked risks and recommended actions, plus a predicted failure window when there's enough history. It runs nightly on every asset in your fleet and is also available on-demand from any asset detail page.

How to refresh an alert

Open any asset detail page and click 'Predict health' (or the badge if a cached alert already exists). The server gathers signals (failure count, MTBF, MTTR, availability, anomalous meter readings, meter trend regressions), runs heuristic scoring, then asks the LLM for a narrative summary and risk polish. Results are persisted (append-only with TTL — default 7 days) so the trend is visible over time at /predictive.

Signals used

Failure count and MTBF/MTTR (filtered to CORRECTIVE + EMERGENCY work orders per the standard reliability convention), availability % from AssetDowntime spans, anomaly burst (recent vs 60-day baseline of out-of-range / regressive meter readings), and per-meter linear regression (slope + r²) projecting days-until-threshold for monotonic meters.

Strict tenant scoping

Every context query is filtered by companyId. Asset from another tenant returns 404 (never 403, to avoid leaking existence). Cron isolates each tenant in its own try/catch — one failure never blocks the global scan.

Permissions and tier gating

• Feature: PREDICTIVE_MAINTENANCE — Enterprise tier only. Lower tiers see an upgrade card on /predictive.
• Permission: ai:predict — seeded for ADMIN by default. Grant via /people/roles for any other custom role.
• Alerts are advisory — verify against your maintenance procedures before acting.

Nightly auto-scan

By default the cron runs at 03:00 UTC for every active Enterprise tenant. It scans up to 50 eligible assets per tenant (those with ≥3 failures OR ≥10 meter readings), persists a heuristic-only alert for each, then re-runs the worst 10 with the LLM narrative. Disable globally by setting the SystemSetting 'predictive.cron_enabled' to 'false' via /admin. Adjust the alert TTL with 'predictive.ttl_days' (default 7).

Cost

Heuristic-only pass: free (pure stats). LLM narrative pass: typically <$0.005 per asset using the default 'google/gemini-2.5-flash-lite'. Per-tenant per-night cost ceiling: ~$0.05 (10 narrated alerts) regardless of fleet size. Admin can swap models per-feature at /admin/ai-usage.