Inline Measurement for Machined Parts: Gauging at the Machine

A machined part is finished the moment the tool leaves the cut, and that is exactly when the measurement becomes urgent: if the diameter has drifted out of tolerance, the next operation adds value to scrap. Bench inspection answers the question an hour later, from a sample, in a queue. An inline measuring head answers it at the machine, on every part, while the part is still in the fixture or on the transfer path - which is a different job from bench metrology, with different optics, different failure modes and a different way of going wrong.

This page is about the inline case specifically: a POMEAS HM series head mounted over a machine tool exit, a conveyor or a robot pick position, gauging turned and milled features in a production cycle. It sets out which features a silhouette measurement can hold and which it cannot, how to size the head against a part family, how to do the tolerance arithmetic before a capability figure is promised, and what coolant, chips and thermal drift do to a number that otherwise looks perfectly stable. The bench and one-shot options are covered by their own pages, and the last sections say plainly when they are the better instrument.

What changes when the gauge moves to the machine

Three things change, and each of them has a cost. The first is sampling: a bench system inspects a few parts per shift, so a drifting process is found by a trend, while an inline head inspects every part and finds the drift at the part that caused it. That is the reason to accept the engineering cost - not throughput, but the ability to stop a bad batch at part ten instead of part four hundred.

The second is that the part is not presented to the optics; the optics are presented to the part. On a bench system the operator places the part on a stage, the software finds the datum, and the measurement is made in a controlled pose. On a machine tool exit the part arrives in a fixture or on a pallet, at whatever orientation the process leaves it, with a locating error that belongs to the fixture rather than to the head. That locating error has to be inside the error budget, because no optical correction removes it. A head that is accurate to a micron is still measuring a part whose datum moved by five microns.

The third is environment. Coolant mist settles on a window, chips land on a transfer path, a spindle grows as it warms, and the shop floor temperature swings between the start of the shift and the afternoon. None of these are optical problems, and all of them appear in the measurement as though they were. A layout that ignores them produces a head that reads beautifully on the day it is commissioned and drifts out of credibility in a month. The sections below treat those three costs as part of the specification rather than as installation details.

What a silhouette head can and cannot hold on a machined feature

The HM family measures a shadow: a collimated beam projects the part onto a receiver, and the software evaluates dimensions from that outline in a single exposure. Everything that is measurable follows from that one sentence, and so does everything that is not. A turned part is a good subject for this kind of instrument because most of its drawing is a set of axial and diametral dimensions that live on the silhouette. A milled part is a mixed subject: its outline dimensions are measurable, its depth dimensions are not.

Feature on the drawing	Inline silhouette	What governs the result
Outer diameter of a turned shaft or boss	Yes	Edge definition and head position accuracy; measure at a defined axial section, not "the maximum"
Overall length, shoulder-to-shoulder length, step position	Yes	Both end faces visible in one view; datum choice must match the drawing
Groove or slot width (axial)	Yes	Groove flanks must be clear of the shadow of adjacent features
Chamfer, fillet start, corner break	Yes, with a caveat	Chamfer is measurable as a projected length; a small radius is edge-limited rather than resolution-limited
Groove depth, step height, bore depth	No	Depth is along the optical axis; a silhouette cannot separate it from a width
Thread pitch diameter	No	The thread projects as a serrated band; major and minor envelope are visible, the pitch diameter is not
Cross-drilled hole, internal bore, keyway floor	Only if it lies on the silhouette edge	The same part can measure differently cycle to cycle if it rotates; internal features are absent from the data
Roundness, cylindricity, taper in three dimensions	No, as a form error	One projection gives one section; form needs rotation, multiple views or a dedicated form instrument
Surface roughness, subsurface porosity, hardness	No	Not an optical outline quantity at any magnification
Coplanarity of two faces, flatness	No	Out of plane; needs a height sensor or a different measuring principle

The table is worth reading twice before a project scope is written, because the failures it prevents are commercial rather than technical. A supplier who is told "measure the part" will quote a silhouette system, and the customer who expected groove depth will discover the gap at acceptance. The honest framing is that an inline head replaces the caliper, the micrometer, the snap gauge and the bench projector for outline dimensions, and it does not replace the depth gauge, the thread gauge, the roundness tester or the CMM for the quantities those instruments exist to measure.

Two subtleties decide whether a "yes" in the table survives contact with production. First, a diameter is not a single number: a turned shaft is measured at a section, and the drawing usually says where. If the recipe evaluates "maximum diameter over the visible length", it will report the largest section in view, which is not the same as the diameter at the datum section, and the difference on a tapered or barrel-shaped part can be several microns. Second, edge definition is a lighting question before it is a resolution question. A deburred, chamfered or plated edge produces a gradient that the sub-pixel edge algorithm interprets somewhere inside the transition, and the same physical edge can move by a fraction of a pixel depending on where the light comes from. That is why the lighting and edge settings belong in the recipe, under revision control, next to the tolerance.

Matching the head to the part family, not to the sample

The published family spans five fields of view, and the trade-off between field and accuracy is the whole selection argument: the same receiver resolves a smaller field more finely, so the smallest head holds the tightest numbers while the largest head covers the most part. Selecting a head is therefore a question of the largest feature the head must see at once and the tightest tolerance that feature must hold - not of buying the best number on the sheet.

Head	Field of view	Repeatability	Position measurement accuracy	Typical machined-part fit
HM-1006	φ6 mm	±0.05 µm	±0.2 µm	Small turned pins, dowels, insert seats, drill-bit shanks, watch and connector-scale parts
HM-1015	φ15 mm	±0.05 µm	±0.3 µm	Shaft sections up to 15 mm, circlip grooves, small sleeves, bushings
HM-1040	φ40 mm	±0.08 µm	±1.0 µm	Motor shafts, spindles, flanges, gear blank OD, multi-step turned parts
HM-1065	φ65 mm	±0.1 µm	±1.2 µm	Hub and housing diameters, discs, longer shafts measured in one view
HM-1120	φ120 mm	±0.2 µm	±2.5 µm	Whole-shaft span, plate and bracket envelopes, parts with several features whose spacing is the requirement

Read the table as a decision, not a ranking. If the real requirement is a 4 mm pin diameter to ±2 µm, the φ6 head is the only sensible choice, and putting a φ120 head on that job buys coverage nobody asked for at twenty times the position uncertainty. If the requirement is the spacing between two shoulders 80 mm apart, no small head can see both at once, and the answer is either the φ120 head or a fixture that indexes the part twice - and indexing twice means the fixture error enters the budget twice.

Two rules keep a family from being designed around a single sample. Size the field against the largest part in the family with margin, because a field that just contains the nominal part has nowhere to put a reference feature and nothing to spare when a variant arrives 3 mm longer. And size the accuracy against the smallest tolerance in the family, not the loose one: a part family whose critical callout is a 3 mm groove will be sampled at the resolution that groove needs, whatever the 60 mm overall length suggests. A head that satisfies both rules for the whole family is cheaper than two heads that each satisfy half of it, and it removes the changeover mistake where a recipe is run on the wrong station.

The station-level starting point is the HM telecentric measurement system for in-line inspection, which packages head, collimated illumination, controller and software; the head records above are the optical front ends for that system. Head-by-head selection logic, including the ranges and the model naming, is set out in the HM series selection guide.

The tolerance arithmetic that has to happen before a Cpk is promised

The published table carries two different quantities and conflating them is the most expensive mistake available in an inline project. Repeatability is short-term scatter on the same feature: it says how tightly the instrument repeats, not whether the value is right. Position measurement accuracy describes how close the reported dimension is to the true dimension anywhere in the field, including calibration and field distortion, and it is the figure that governs a tolerance decision. A head can repeat to ±0.05 µm and still read a micron high, and the repeatability will make that error look perfectly stable.

The arithmetic a project should do is a worst-case combination of the two, compared against the drawing band. For the φ40 head, ±0.08 µm repeatability plus ±1.0 µm position accuracy gives about ±1.08 µm of instrument uncertainty (derived: 0.08 + 1.0). On a ±10 µm drawing band that is 10.8% of the band before the fixture, the part temperature or the operator contribute anything (derived: 1.08 ÷ 10). The same head on a ±20 µm band spends 5.4% (derived: 1.08 ÷ 20). Run the same numbers across the family and the shape of the decision appears: the φ6 head sits at ±0.25 µm and 2.5% of a ±10 µm band (derived: 0.05 + 0.2 = 0.25; 0.25 ÷ 10), the φ65 head at ±1.3 µm and 13% (derived: 0.1 + 1.2 = 1.3), and the φ120 head at ±2.7 µm and 27% (derived: 0.2 + 2.5 = 2.7; 2.7 ÷ 10).

Those percentages are not a capability study and they are not a promise. They are the share of the tolerance that the instrument consumes, and they leave the rest for everything else. A common rule of thumb in gauge selection is that the measurement system should consume no more than 10% to 20% of the tolerance band; a head that consumes 27% before the fixture is counted will struggle to demonstrate capability no matter how good the process is, while a head at 2.5% leaves room for the real error sources. Stated the other way round: if the drawing tolerance is ±5 µm on a 60 mm part, an inline silhouette head is the wrong instrument for that callout, and no amount of software or magnification fixes it.

Two honest caveats belong with every figure on this page. The published repeatability and accuracy values are specification-sheet values; the test conditions - environment, sample count, artefact, integration time - are not stated in the material we hold, so they must be read as instrument capability under defined conditions rather than as a production tolerance. And the combined uncertainty above excludes fixture repeatability, part temperature, edge definition on a burred or chamfered edge, and evaluation settings. The only number that can enter a control plan is a gauge study on your own artefact, at your own line temperature, with your own fixture. The method for building that budget term by term is set out in building an error budget for inline diameter measurement, which is worth reading before the purchase order rather than after acceptance.

Cycle time: the exposure is never the constraint

The published exposure for the HM family is 35 µs. At that exposure a part moving through the field is frozen, so the shadow is captured without motion blur and the transfer path does not have to stop for the measurement. It is also the number most often misquoted in project meetings as if it were the measurement time.

The interval that matters is the one between two parts arriving at the station. At 600 parts per hour that interval is 6,000 ms, at 1,200 it is 3,000 ms, and at 3,600 it is 1,000 ms. A 35 µs exposure occupies 0.0035% of a 1,000 ms interval (derived: 0.035 ms ÷ 1,000 ms), so the optical capture is effectively free and the cycle is decided by everything around it: transfer and settle, trigger, image readout, recipe evaluation, result handshake with the PLC, and the reject actuation downstream. The published statement that a single capture can output up to 100 items is a statement about how many dimensions the software can evaluate from one silhouette, not a promise that 100 callouts keep the same accuracy - and each additional evaluation adds processing time and a place for the recipe to be wrong.

Two design consequences follow. First, put the head where the part is already stationary or predictably moving: a robot pick position, a transfer index station, or the exit of the machine before the part is released. Mounting over a free-running conveyor is possible because of the short exposure, but it demands a trigger that is repeatable to the part's position, otherwise the datum shifts from part to part and the repeatability figure becomes meaningless. Second, count the evaluations rather than the parts. A recipe with forty dimensions on every part is a different computational load from a recipe with eight, and the difference shows up as cycle time on the slowest station, not on the datasheet.

Coolant mist, chips and the window: the maintenance that decides accuracy

A machining cell is an aerosol of coolant, fine chips and, on castings, graphite or sand dust. The measuring window sits in that aerosol by definition, and a film of mist on the window does not produce an error message; it produces a slightly softer edge, then a slightly shifted edge, then a drift that looks like a process problem. This is the most common reason an inline installation loses credibility in the first months, and it is entirely preventable with layout and a maintenance routine.

Four measures carry most of the benefit. Stand the head off the splash zone and use an air purge across the window rather than a jet at the part, because a jet directed at the part puts energy into the part and moves the silhouette during the exposure. Give the window a replaceable protective pane so that cleaning is a wipe and not a re-calibration. Put the window check on the shift checklist with a master artefact, and record the reading rather than a pass or fail, because a drift trend is what tells you the purge is losing. And position the head so that chips fall away from the optical path rather than through it; on a vertical spindle the difference between mounting above and below the transfer plane is the difference between a weekly wipe and a daily one.

There is an honest limit here as well: a silhouette head cannot tell you that the window is dirty. The measurement stays plausible, the repeatability stays good, and only the absolute value moves. A master artefact measured at the start of each shift, with its value logged against the commissioning value, is the cheapest instrumentation for that failure mode, and it turns an invisible drift into a trend line with a threshold.

Part temperature and machine thermal growth

Machined parts are measured warm. A turned steel shaft leaves the cut at a temperature above the shop floor, and it cools while it travels. Steel expands by roughly 11.5 µm per metre per kelvin (typical handbook value for structural steel, not a POMEAS specification), so a 100 mm steel part that is 5 K above the gauge temperature reads about 5.75 µm larger than it will be at 20 °C (derived: 100 mm × 11.5 × 10⁻⁶ / K × 5 K). Aluminium at roughly 23 µm per metre per kelvin doubles that to about 11.5 µm on the same part (derived: 100 mm × 23 × 10⁻⁶ / K × 5 K). On a ±10 µm band, the difference between measuring at the machine and measuring in a temperature-controlled room is the same order as the whole tolerance.

This is not an argument against inline measurement; it is an argument for defining what the number means. Three practices resolve it. Decide whether the drawing tolerance applies at 20 °C or at the process temperature, and write that decision into the control plan, because a part that is "in tolerance hot" can be out of tolerance cold. Measure at a consistent point in the cooling curve - the same station, the same delay after the cut - so that the systematic part of the thermal error is constant and can be compensated. And use the master artefact to separate instrument drift from part temperature: if the master moves, the instrument or the window moved; if the master is stable while the parts trend, the process or the part temperature moved.

The machine contributes as well. A spindle, a ballscrew and a fixture grow as the cell warms up, so the first parts of a shift are measured against a different geometry from the parts after two hours. Inline data makes this visible for the first time on many lines, and the usual response is a warm-up cycle plus a first-article check, with the drift absorbed by the offset rather than by a tolerance change. Both effects are slow and both are trend-shaped, which is exactly what a 100% inline measurement is good at detecting - provided someone is looking at the trend rather than at a single pass or fail.

Where a bench flash system, a telecentric lens or a zoom lens is the better answer

The inline head is one instrument in a family, and the boundary between them is a matter of presentation and scope rather than of quality. A one-shot flash measuring system takes a single exposure of a whole part placed on a stage, evaluates the full drawing in that pose, and is the right choice when the part can be carried to a bench: it sees the whole part, it is easy to re-programme, and it does not have to survive coolant. The comparison between a flash system and a video measuring machine, including where each one stops, is set out in what a flash measuring machine is and when it beats a VMM, and the machine-by-machine boundary in instant measurement versus CMM. Flat, stamped and machined parts that are inspected off-line at a bench are covered specifically in flash measurement for CNC machined parts.

When the requirement is a good image rather than a measurement, the choice moves to the lens. A fixed telecentric lens gives a stable magnification and no perspective error for a single field of view, which is why it is the standard front end for a static gauging station; the machined-part case is worked through in telecentric lens for machined parts. When one camera must serve several part sizes or several features at different magnifications, a zoom lens replaces the turret, as described in zoom lens for machined parts. The inline head differs from both in that the measurement is computed in the head's own controller, in the production cycle, without a PC-based vision application built around it - which is a software and validation argument as much as an optical one.

The overlap deserves a plain statement. If a line runs one part number at one magnification and the part can be presented on a stage, a telecentric lens with a camera and software is usually the cheaper and more flexible route. The inline head earns its place when the part cannot leave the process, when every part must be measured, or when the measurement has to close a loop with the machine tool. Buying an inline head for a job that could be done at a bench pays for integration, purging and drift management that the application does not need.

Acceptance: what to test on your own parts before the line is released

An inline installation should be accepted on artefacts and on parts, in that order, and the results should be written down. The sequence that catches the expensive problems is short. Measure the master artefact ten times at the production trigger rate and compare the scatter with the published repeatability; a large discrepancy points at the trigger, the fixture or vibration rather than at the optics. Measure thirty parts that span the tolerance band, three times each, and compare each part against a reference instrument whose uncertainty is at least four times better than the head's contribution; this is where a wrong datum, a wrong section or an inverted sign appears. Log the master artefact every shift for a week and plot it; a monotonic trend is the window, a step is a fixture, and a sawtooth that follows the shift pattern is thermal.

Then run the line and compare the inline distribution with the bench sample for the same period. Two outcomes are informative. If the inline distribution is wider than the bench distribution, the difference is usually presentation - datum, orientation, edge setting - rather than optics. If the inline mean differs from the bench mean by a constant, the difference is a calibration or a temperature convention, and it should be resolved as an offset with a documented definition rather than by editing the tolerance. Neither outcome is a failure of the instrument; both are the normal content of a commissioning report, and both are much cheaper to find in week one than in month six.

What to send for a head and station recommendation

A recommendation that can be quoted needs six pieces of information, and missing any one of them turns the answer into a range. The part family, with the largest and smallest envelope and the variant that is expected next year. A drawing with the callouts that must be measured marked, and the ones that are currently measured by other means, so that the scope is explicit. The tolerance on each of those callouts, and whether it applies at 20 °C. The required rate in parts per hour, and whether the station is in-cycle or after the machine. The presentation: fixture, pallet, robot pick or free conveyor, with the locating scheme. And the environment: coolant type, chip load, ambient temperature range, and whether the head can be enclosed.

With those six, the head follows from the field of view and the accuracy arithmetic, the station follows from the presentation, and the maintenance interval follows from the environment. Without them, any answer is a catalogue page. The published range and per-head figures used throughout this page are the ones in the inline dimensional measurement system overview; the battery-line and automotive-powertrain applications of the same family are worked through on their own pages, because the feature mix and the cycle structure differ from a general machining cell (battery cell and tab inspection, automotive powertrain parts).

Related reading

• HM series online image measuring heads: five models, ranges and selection
• Inline dimensional measurement system: series overview and published specifications
• Building an error budget for inline diameter measurement
• Flash measurement for CNC machined parts: OD, grooves and limits
• Telecentric lens for machined parts
• Zoom lens for machined parts: one camera, several magnifications
• Instant measurement versus CMM: when one-shot vision is enough
• Inline measurement for automotive powertrain parts

Product starting points

• HM telecentric measurement system for in-line inspection - head, illumination, controller and software as a station
• HM-1006, φ6 mm field - small pins, dowels and insert seats
• HM-1015, φ15 mm field - shaft sections, circlip grooves, bushings
• HM-1040, φ40 mm field - motor shafts, spindles, flanges
• HM-1065, φ65 mm field - hubs, housings and discs
• HM-1120, φ120 mm field - whole-shaft span and plate envelopes