Tool life prediction without sensors: what your SCRAP data already knows

The skeptic's first question

Every CAM engineer who hears "we predict remaining tool life" asks the same thing: without a sensor, how?

The honest answer is that sensors give you a more precise estimate on the tools that have one. They also cost more to install than most of the end-mills in your crib cost to buy. For the 90% of tools that never get retrofit telemetry, the question becomes: can you get a useful estimate from the data you already keep?

This post walks through how Niyamis TMS does that. The headline is: a rolling statistical model over SCRAP notes and TAKE/RETURN cadence gets you a life estimate with a confidence band. Good enough to drive a replacement decision a shift before the part fails. Not good enough — and we don't pretend otherwise — to predict the exact cycle number the flute will chip on.

The three inputs

TMS's life predictor reads three streams, all of which your shop already produces:

1. TAKE / RETURN events. Every time a tool leaves the crib and comes back, you have a service window. The window includes job context (which part, which machine, which operator), so the predictor doesn't average apples and oranges.
2. SCRAP notes. When a tool is scrapped, the operator almost always writes a note. "Chipped flute, replaced mid-run." "Worn, normal end-of-life." "Broke on first cut, bad insert." The notes are noisy, but they're also remarkably consistent per alloy / job pattern.
3. Part counts from jobs. How many parts ran between the TAKE and the SCRAP? Some shops log this explicitly. Others infer it from production order counts. TMS supports both.

None of these require new hardware. They require you to already do the thing good shops already do: log transactions and write scrap notes.

The model

The model is intentionally boring. TMS doesn't claim to train a neural network on your eighteen months of SCRAP notes. A neural network on that volume of data would overfit dramatically. Instead:

1. Bucket. Tools are bucketed by SKU, alloy, and machine class. A 10 mm end-mill on aluminium on a 5-axis machine is a different bucket from the same end-mill on stainless on a 3-axis.
2. Rolling average. For each bucket, compute a rolling average of parts-per-tool across the most recent N service windows. N adapts to your volume — shops with 200 windows/bucket use N = 30; shops with 40 use N = 10.
3. Scrap-note classification. Notes get parsed into one of four classes: normal wear, breakage, geometry change (someone reground), other. Only normal wear windows feed the life estimate. Breakages and reground tools are treated as sample noise, not as data points about the population.
4. Confidence band. TMS publishes an 80% confidence interval around the point estimate. The width of the band tells you how much to trust it.

What you see in the UI

Open a tool in service, and the Life Predictor shows:

• Parts produced so far in the current service window.
• Expected remaining life (median) for this SKU / alloy / machine bucket.
• Confidence band — typically ±25% of the point estimate with 100+ prior windows; wider when data is thin.
• Top three SCRAP note themes from the bucket, so the supervisor knows what kind of failure to expect.

The dashboard also surfaces the fifteen tools most likely to hit end-of-life in the next two shifts. That's the actionable view for the toolroom supervisor on a Monday morning: what to stage, what to re-sharpen, what to reorder.

A worked example

A Swiss-type lathe running a medical-device part on 316 stainless uses a specific 3 mm drill. Over the last 18 months, the shop's SCRAP records show 140 service windows for this SKU/alloy/machine combination. The mean life is 410 parts, with 80% of windows falling between 320 and 480.

Today's tool has produced 380 parts so far. TMS flags it as likely to reach end-of-life within one shift, with the operator's SCRAP note patterns from the last 30 windows suggesting normal wear rather than breakage. The supervisor stages a replacement on the adjacent shelf. When the tool comes out at 425 parts, the changeover costs the shop ninety seconds instead of an unplanned stoppage mid-cycle.

Where this breaks down (honest version)

Three known weak spots:

1. First forty windows are rough. Below ~40 prior service windows in a bucket, the confidence band is wide enough that the estimate is directional rather than actionable. TMS shows the amber confidence indicator and tells you to treat the estimate as a lower bound.
2. New alloys throw the model off. If the shop starts machining titanium for the first time, the existing SCRAP history doesn't apply. The predictor marks the new bucket as learning until it has enough data.
3. Regrinds confuse parts counts. A reground tool that re-enters service with a new geometry is a new tool, not a continuation of the old one. The SCRAP-note classifier catches most of this, but edge cases require the toolroom to flag regrinds in the crib — which TMS makes a first-class event.

Closing

The pitch isn't that statistical models beat sensors. The pitch is that the data you already keep gets you 80% of the way there for 0% of the retrofit cost — and for the tools that matter most (end-mills, drills, taps at spindle), 80% is more than enough to change how the toolroom plans its week.

Want to see this on your own SCRAP records? The thirty-minute discovery call ends with a yes or a no on whether your data's shaped well enough to run it.

— Philippe, Niyamis