How to verify the accuracy of estimated CAM program time ART-142

The most common error is comparing different intervals

CAM may show 42 minutes of “cutting time.” The operator records 58 minutes from starting to look for the sheet until handing over sorted parts. Both figures can be correct, but they measure different things.

Record the boundaries before testing:

| Interval | What may be included | |---|---| | NC / program time | Execution of the approved program from start to finish | | Machine sheet cycle | Loading to the working position, program, automatic transitions, unloading — depending on configuration | | Order production time | Preparation, waiting, cycle, sorting, inspection | | Lead time | From accepting the order until all operations are complete |

For CAM calibration, compare the forecast with the closest possible machine interval and track auxiliary operations separately. Otherwise the technologist will start “correcting” speeds in the program while the difference is created by material search or sorting.

What CAM usually includes

Depending on the system and postprocessor, the estimate can include:

• contour motion at the assigned speed;
• piercing and approaches;
• idle moves between contours;
• lead-in and lead-out;
• marking or other specified operations;
• mode changes between geometry types;
• automatic sequencing;
• individual machine delays, if described in the model.

Official SigmaNEST materials state automatic cutting-time calculation during import and retention of toolpath and machine data. Hypertherm describes simulation, sequencing, feedrate, piercing techniques and detailed reports for ProNest. TRUMPF separately offers time and cost calculation based on its machines’ technological data. This confirms that estimation is possible, but does not guarantee that the local database matches the actual configuration.

What often remains outside the forecast

Actual time can include:

• waiting for a crane, forklift or automated warehouse;
• material and documentation checks;
• permitted routine operator actions;
• a stop to inspect the first part;
• cleaning or removing a problematic fragment;
• restart after a warning;
• manual sorting;
• waiting for a technologist’s decision;
• unplanned maintenance;
• network or organizational delays.

These events should not be hidden inside a “CAM coefficient.” They need their own reason codes. The NC-cycle forecast then stays technically clean, while production planning adds a real allowance for auxiliary work.

Prepare a control sample

One ideal demonstration sheet does not characterize the system. The sample should represent production:

• thin, medium and thick materials in the enterprise’s real range;
• short and long contours;
• different quantities of internal cutouts;
• serial and mixed nests;
• typical sheet formats;
• recurring and new parts;
• programs with marking, where used;
• different operator shifts without personal rating.

An initial pilot can use a limited but representative group of correctly recorded programs. Its size is not universal: it depends on material and geometry diversity, process stability and the accuracy required for the particular decision. Accumulate a larger history for stable coefficients. Rare emergency events should be analysed separately: they matter to reliability, but distort a typical forecast.

Record the data version

Comparison loses meaning if the technologist changes a nest or mode while the file remains under the same name. The minimum test record is:

• ART or internal analysis ID;
• order and nest ID;
• part revision;
• CAM and postprocessor version;
• machine identifier;
• material, thickness and format;
• name/version of the approved technology set;
• estimated time;
• actual time;
• boundaries of both intervals;
• stops and their causes;
• date and shift.

Do not copy secret parameters, serial numbers or unsafe settings into a public report. Stable references to controlled records are sufficient for internal traceability.

Key accuracy indicators

Calculate for every program:

Absolute deviation = actual time − estimated time.

Relative deviation = (actual − estimated) / estimated × 100%.

The sign matters. If actual time is consistently higher, the model underestimates time or the actual interval contains additional events. If actual time is lower, the database may be conservative or the machine may operate differently from its description.

Do not limit review to an average percentage. Positive and negative errors can cancel each other out. It is useful to view:

• median deviation;
• mean absolute error;
• the 80th or 90th percentile of absolute error for planning reserve;
• error by material, thickness, geometry type and machine;
• share of programs within an acceptable band.

The enterprise defines the acceptable band according to purpose. It can be wider for a preliminary quotation than for dispatching a tightly scheduled shift.

Decompose deviations into causes

A general “add 15%” coefficient becomes obsolete quickly. It is better to classify:

| Group | Example question | |---|---| | Geometry | Does the error grow with the number of small contours? | | Technological model | Are machine events and permitted modes described correctly? | | Sequencing | Do CAM and actual NC have the same path? | | Machine | Is there a systematic difference on one configuration? | | Material | Does a particular group more often require control pauses? | | Organization | Did waiting or sorting enter the actual time? | | Data | Are different program revisions being compared? |

If divergence occurs only in programs with many pierces, correcting the overall cutting speed will not solve it. If the deviation is the same in minutes for every program, a fixed machine event may be unaccounted for. These are hypotheses for the responsible technologist to verify, not setup instructions.

Separate normal pauses from exceptional events

Planning needs two layers:

1. Base program forecast — recurring events CAM can model. 2. Operational reserve — normal auxiliary work that is not part of NC.

Emergency stops, material defects or a service incident should not be spread over every program by one identical coefficient. Analyse them as availability losses or risk. This borders on OEE, but the full OEE calculation belongs to ART-146.

Calibrate without unsafe shortcuts

After analysis, standards can be updated, but only in a controlled process:

1. Identify a group where error is stable. 2. Verify versions and time boundaries. 3. Find the cause, not merely the percentage. 4. Prepare the change in a test environment or database copy. 5. Verify it on control parts according to enterprise rules. 6. Obtain approval from the responsible specialist. 7. Assign a version and effective date. 8. Repeat plan/actual after implementation.

Universal speeds, gas parameters, focus positions or other technological values must not be published. They depend on the particular source, head, machine, material, gas and manufacturer documentation.

How to use accuracy in costing

If median error is small but actual time is much higher in 10% of complex programs, use a complexity class or statistical reserve for quotations rather than one average coefficient. A price forecast should be conservative enough to cover normal variability, but not conceal a systemic data problem.

For a recurring part, the best forecast is often the latest stable actual result for the same revision and comparable conditions. CAM remains necessary for new orders, scenario comparison and change verification.

Control after a CAM or machine update

An update to CAM, the postprocessor, controller or technology database creates a new time-model version. Do not transfer the old coefficient automatically. After a change, choose a short approved set of control programs, retain the previous forecast, produce a new one and compare both with actual execution. If only report format or indicator name changed, confirm that time boundaries stayed the same.

For several machines, coefficients should be separate. Even machines of the same class can differ in configuration, automation and the set of modelled events. Combine them in one statistic only after verifying similar error distributions.

It is useful to set a control date for the next review. Grounds for an unscheduled check are systematic error shift, software update, postprocessor change, a new material group or a noticeable change in the order mix. This makes calibration part of data management rather than a reaction to one unsuccessful program.

Separate bias from random spread

Two data sets can have the same average error but require different action. In the first, actual time is almost always a few minutes longer than forecast. This resembles a missing constant event or different cycle boundaries. In the second, the average difference is close to zero, but individual programs deviate strongly in both directions. A general coefficient will not help here: find the group where the spread changes.

For management review, it is enough to see:

• median absolute error in minutes;
• median relative error in percent;
• direction of systematic bias;
• the band containing most typical programs;
• a separate list of large deviations with reason codes.

The mean is useful but sensitive to a single long stop. The median better shows the typical case, but can hide an expensive tail of complex orders. Read the measures together rather than declaring one universally correct.

Form hypotheses from the error shape

The deviation distribution suggests where to begin checking, but is not proof of cause.

| Observation | Working hypothesis | What to compare | |---|---|---| | Nearly identical difference in minutes | An unaccounted fixed event | Start/end boundaries, automatic cycles | | Error grows with cut length | Non-uniform motion model or time database | Material groups, data versions, actual path | | Deviation grows with number of pierces | A separate cycle component | Event count in forecast and log | | Error only in one part group | Incorrect grouping or a local assumption | Geometry, material, revision, postprocessor | | Spread depends on shift | Different recording boundaries or auxiliary actions | Recording rule, not worker rating |

After forming a hypothesis, select control programs where it should appear. If changing a coefficient improves one group but worsens another, the correction is too broad. Separate classes or correction of the primary model are then needed, not another general percentage.

Check the new model in parallel with the old one

Before a correction is approved, it is useful to calculate two forecasts for a period: current and candidate. Both are compared with the same actual result, but only the permitted version goes to production planning. This prevents an experimental coefficient from silently changing deadlines or costing.

The verification record should include date, change author, program groups, old and new result, acceptance criterion and decision. If improvement does not repeat on a held-out control group, the new model should not be approved merely because it worked on the data used to tune it.

Align accuracy with its purpose

One forecast can be sufficient for a preliminary commercial estimate but insufficient for a tight hourly plan. Before calibration, therefore, state which decision the indicator supports: fast quotation, detailed costing, queue order or deviation analysis.

The shorter the planning window and the more costly delay is, the more important it is to see normal variation as well as average time. At the same time, there is no point requiring second-level accuracy from a model when the actual log is rounded to several minutes. Calculation accuracy cannot exceed input-data quality.

Before analysis, record measurement resolution: how time is rounded, who creates start and end events, and whether CAM, controller and production-log clocks are synchronized. Otherwise a small deviation can result from different timestamps rather than an error in the technological model.

Typical mistakes

Comparing CAM with full order time. Align boundaries first.

Calibrating on one material. The model may be accurate only for a narrow group.

Deleting “bad” actual results without a reason. An outlier can be excluded from a standard but must remain in the event log.

Updating a database without a version. Old and new programs then become incomparable.

Rating operators by deviation. This encourages hidden pauses; the data are for the process, not punishment.

Correcting technology with a coefficient. First separate data, organization and the real machine model.

Verification checklist

• Is it known which interval CAM predicts?
• Does actual time have the same start and end?
• Are nest, NC, CAM and postprocessor versions recorded?
• Does the sample represent the real product mix?
• Do stops have reason codes?
• Are absolute error and distribution analysed, rather than only the average?
• Are deviations grouped by material, geometry and machine?
• Do changes undergo control and versioning?
• Was a repeat check performed after the change?

Conclusion

Accurate CAM time is not a number adjusted once. It is a managed model regularly compared with homogeneous actual results. Clear cycle boundaries, data versioning, reason codes and analysis of groups where error repeats have the greatest effect.

When an enterprise separately sees the NC forecast, normal auxiliary operations and unplanned losses, costing and planning become much more reliable. The technologist also receives a concrete question to verify instead of a general demand to “make CAM more accurate.”

Need service advice?