Find the last cheap maintenance window.
Condition monitoring tells you a reading has passed a limit. AnantState shows how long an asset has been heading there, and when the last inexpensive window to fix it is.
In plain termsKnow the last cheap moment to maintain a machine.
The window before the limit
An illustration of one asset, from normal wear to an unplanned stop.
A threshold says when a limit has been passed. It does not say how long the asset has been heading there, or which window was the last cheap one. Pricing the intervention against the cost of waiting is what makes condition-based maintenance something a planner can defend over a calendar.
One asset, from wear to a stop · illustrative
- Week 0: Normal wearBehavior matches the expected path
- Week 3: Starts to driftVisible against the expected path, no alarm
- Week 6: Last cheap windowInspect now: a planned, low-cost stop
- Week 9: The limit is crossedWhat a threshold alarm sees
- Week 10: Unplanned stopThe expensive way to find out
- Phase: Cheap to actfrom Starts to drift to Last cheap window
- Events are shown in order, not to scale.
What moves in an asset
| State | Baseline drivers | Unexpected drivers |
|---|---|---|
| Condition and degradation | Duty cycle, runtime hours, ambient conditions | Bearing wear, contamination, misalignment |
| Yield and quality | Recipe, feedstock, throughput | Drift in a process parameter, feedstock variation |
| Unplanned-stop risk | Maintenance interval, age | Cascading failure, operator variance |
| Energy posture | Load profile, tariff, weather | Equipment inefficiency, leakage |
This is the domain closest to classical condition monitoring, and where the approach differs most: the physics-informed baseline is available and the residual is small. A model asked to correct a known degradation curve has a much easier job than one asked to invent it.
A worked example
Forty machines, a quarter of them inspected early, against the cost of a stop.
Round, illustrative figures for machines: they show the arithmetic, leave out the subscription, and are for you to change in the calculator.
| Input | Value | What it stands for |
|---|---|---|
| Machines in scope | 40 | Everything the team is responsible for |
| Value at stake per one | $60,000 | The cost of one unplanned stop |
| Cost of the action per one | $2,500 | A planned inspection or early maintenance stop |
| Share acted on | 25% | The machines furthest along a degradation path: 10 of 40 |
| Modeled, USD | Pessimistic | Expected | Optimistic |
|---|---|---|---|
| Share of value at stake avoided | 10% | 20% | 35% |
| Cost to act, in total | $25,000 | $25,000 | $25,000 |
| Avoided loss | $60,000 | $120,000 | $210,000 |
| Net position | $35,000 | $95,000 | $185,000 |
Break-even is 4.2%: each action costs $2,500 against $60,000 at stake, so it has to avoid that share of the loss to pay for itself.
Across an estate
- “Which of 400 assets gets the next maintenance crew?” is answered in one queue ordered by value at stake.
- Counterfactual replay turns the post-mortem on a stoppage into a measurement.
- Yield and energy are state too, using the same machinery, because a drifting process and a wearing bearing are the same problem shape.
Integration reality check
- Plant data is frequently network-segmented. The read-mostly posture on the integrations page is why this is deployable in an operational-technology environment at all.
- Latency needs vary from milliseconds (interlocks) to hours (maintenance planning). We target the planning and prioritization layer.
What exists today
The industrial world exists as a catalog template (the catalog carries equipment maintenance, detect to repair and design to produce) and as one of the six domains compiled into the engine. It is not a live simulated world.
Price the maintenance window
Bring one asset class and its maintenance cost. We will show the cost of waiting against the cost of acting now.