> ## Documentation Index
> Fetch the complete documentation index at: https://docs.appliedaifoundation.org/llms.txt
> Use this file to discover all available pages before exploring further.

# Remaining Useful Life (RUL) Prediction

> How per-lamp failure probability and RUL hours are predicted and surfaced in the dashboard.

## What RUL is

**Remaining Useful Life** is the predicted number of operating hours before a UV lamp is expected to fail or require replacement. It is stored in the `bwts_iot_predictions` table and surfaced in the Predictive Maintenance tab.

A lamp with an RUL of 0 should be treated as imminently failed. A lamp with an RUL of 2,800 hours has most of its rated 3,000-hour life still ahead of it.

## The prediction record schema

Each prediction record contains:

```json theme={"system"}
{
  "timestamp": "2026-04-30T00:00:00Z",
  "component_id": "LAMP_01",
  "component_type": "UV_LAMP",
  "predictions": {
    "remaining_useful_life_hours": 450,
    "failure_probability": 0.62,
    "efficiency_percent": 81.3
  },
  "current_state": {
    "runtime_hours": 2150,
    "efficiency_percent": 81.3,
    "status": "OPERATIONAL"
  }
}
```

| Field                                     | Description                                                         |
| ----------------------------------------- | ------------------------------------------------------------------- |
| `component_id`                            | Lamp identifier, e.g. `LAMP_01` through `LAMP_16`                   |
| `component_type`                          | Always `UV_LAMP` for lamp predictions                               |
| `predictions.remaining_useful_life_hours` | Predicted hours of useful life remaining                            |
| `predictions.failure_probability`         | Float 0.0–1.0; probability of failure within the prediction horizon |
| `predictions.efficiency_percent`          | Predicted efficiency from the model                                 |
| `current_state.runtime_hours`             | Cumulative hours the lamp has operated                              |
| `current_state.status`                    | `OPERATIONAL`, `DEGRADED`, or `FAILED`                              |

## How predictions are generated

Predictions are pre-computed by a separate ML pipeline — **not by the dashboard**. The pipeline analyses each lamp's runtime history and efficiency degradation curve and writes prediction records to the database. The dashboard reads results via `/api/predictions`.

## Dashboard deduplication logic

The `/api/predictions` endpoint may return multiple prediction records per lamp (one per prediction run). The Predictive Maintenance tab applies this deduplication:

1. Fetch up to 100 prediction records
2. Filter to `component_type === 'UV_LAMP'`
3. For each `component_id`, keep only the record with the most recent `timestamp`
4. Sort remaining records by `failure_probability` descending (most at-risk first)

This ensures exactly one row per lamp regardless of how many historical prediction records exist.

## Failure probability thresholds

| Failure probability | Risk level | Row colour | Action                                                   |
| ------------------- | ---------- | ---------- | -------------------------------------------------------- |
| ≥ 70%               | Critical   | Red        | Plan replacement at next port call                       |
| 50–69%              | High       | Orange     | Elevated risk — schedule replacement within 2 port calls |
| 30–49%              | Moderate   | Yellow     | Monitor — include in next routine inspection             |
| \< 30%              | Good       | Green      | No action needed                                         |

## Lamp lifetime reference

| Threshold           | Value       | Source                            |
| ------------------- | ----------- | --------------------------------- |
| Rated lifetime      | 3,000 hours | `THRESHOLDS.LAMP_RUNTIME.MAX`     |
| Maintenance warning | 2,500 hours | `THRESHOLDS.LAMP_RUNTIME.WARNING` |

When `current_state.runtime_hours` approaches 2,500h, failure probability rises sharply on the degradation curve. Lamps beyond 2,500h with failure\_probability ≥ 0.5 should be treated as high priority for replacement.

## Cost implication

The Predictive Maintenance tab's cost comparison ($18,500 reactive vs $5,000 predictive annually) is based on the following logic:

* **Reactive**: Emergency lamp replacements require expedited parts, unplanned dry-dock time, and elevated labour costs
* **Predictive**: Scheduling replacements based on RUL during planned port calls allows bulk procurement and standard crew schedules

## References

* [Health Score Calculation](/methodology/health-score) — how lamp health feeds into the composite score
* [Predictive Maintenance Tab](/product/predictive-maintenance-tab) — where this data is displayed
* `/api/predictions` — the endpoint serving this data