Putting the measurement station into the line cycle time

1. Four capabilities, one goal

The goal is simple: turn a measurement that used to be set up and judged by hand into a station on the line that runs at the push of a button, with traceable results.

2. Motor control

The biggest difference between a measurement station and an ordinary production mechanism is that "moving fast" is not the point; "stopping steadily, stopping accurately, and knowing where it stopped" is. Below are the decisions we go through on every project.

2.1 Servo or BLDC?

The difference is not just price. Choose wrong, and no amount of tuning later will make up for it.

The first thing to calculate in selection is the inertia ratio. The load inertia referred to the motor shaft, divided by the motor's own rotor inertia, is generally recommended to stay within 10:1; for measurement stages that need high response and short settling time, it is best kept to 5:1 or even 3:1. In a system with too high an inertia ratio, raising the gain causes oscillation and lowering it makes the axis soft, and nothing you tune is right. At that point what needs to change is the gear ratio or the motor, not more parameter tuning.

2.2 Three-layer control loops and tuning order

A servo system is not one loop but three nested loops. Without understanding this structure, tuning is just trial and error.

The standard tuning order is from the inside out:

1. Current (torque) loop: off-the-shelf drives usually come preset for the motor parameters; when developing firmware in-house, this is the first thing to get right, including current sampling timing, PWM dead-time compensation and commutation angle calibration.
2. Velocity loop: raise the proportional gain until high-frequency ringing appears, back off a little, then add integral action to remove steady-state error. If a mechanical resonance falls in this band, it has to be suppressed with a notch filter.
3. Position loop: its upper gain limit is tied to the velocity loop bandwidth, usually one fifth to one tenth of it. Feedforward (velocity and acceleration feedforward) reduces following error without touching the gain.

If halfway through you find that "the velocity loop oscillates however it is tuned", the answer is usually not in the parameters but in the mechanism: a coupling that is not stiff enough, a belt that is too long, an under-supported lead screw, or the inertia ratio mentioned above.

2.3 Angle feedback: choosing an encoder

However well the three loops above are tuned, they can only be as accurate as the feedback tells them. Common principles for angle feedback also include inductive encoders and resolvers; here we compare the three that are most often traded off against each other: optical, magnetic and capacitive. The differences between them are usually not in the resolution figures in the catalog, but in what each of them is vulnerable to, and that is often exactly what your site happens to have.

So once the accuracy requirement is settled, selection usually comes down to three things: how strong the nearby magnetic field is (with a direct-drive motor right next to it, a magnetic solution needs extra handling of stray fields, which often makes it uneconomical), the contamination and condensation conditions on site, and how much space and weight the mechanism has left. Once these three questions are answered, the options usually narrow down.

For the capacitive column, the brand we often use is Netzer Precision. The reasons for choosing it fall into two groups: one comes from its mechanical form, the other from the measuring principle and product design. Looking at them separately makes it clearer which are choices of this brand and which are properties shared by the capacitive principle.

From the mechanical form: the sensing ring is hollow and thin, and has no bearing of its own. The rotor is fixed directly to the measured shaft and the stator to the frame, with no contact between them. This kit-type, bearingless form is not unique to capacitive encoders; optical and magnetic encoders also come in similar forms. The following benefits come from this mechanical form itself:

• Cables can pass through the center of the hollow shaft instead of being routed around the outside of the mechanism. For an axis that swings through a large angle, the cable twists along the axis instead of being dragged around it, so stress and reaction torque are much smaller; when continuous rotation is needed, the hole is exactly where a slip ring should go
• One less bearing means less friction, less play, and less weight and inertia. Friction affects the low-speed smoothness of a direct-drive mechanism, and mass affects the reflected inertia, and the inertia ratio is exactly the first number to calculate in 2.1 above and in servo selection
• Mounting tolerances can be relaxed. According to the manufacturer's technical notes, the reading is taken from the electric field distribution around the full circle rather than from a single read head, so the effect of mounting eccentricity is averaged out; the same document also lists "relaxed mounting tolerances" as a key feature. This is very practical for thin-walled, lightweight mechanisms, because such structures cannot easily achieve high concentricity in the first place
• No contact means no wear. Removing a mechanical part that degrades over time makes a difference in reliability for long-running or hard-to-service equipment, not just a better-looking lifetime figure

From the measuring principle and product design: it measures an electric field, not a magnetic field and not light.

• Not affected by magnetic fields. The permanent magnets and drive currents of a direct-drive motor are only a few centimeters away; a magnetic solution needs extra handling of stray fields, while a capacitive one does not have this problem. The manufacturer also lists immunity to electromagnetic and radio-frequency interference as a design goal
• Absolute output. The angle is known at power-up, with no homing needed (optical and magnetic encoders also come in absolute versions; this is the manufacturer's product design, not something unique to the capacitive principle). For a production line that must resume immediately after a power cut, and for a gimbal that cannot easily perform a zeroing move in the air, this is a functional difference, not just a convenience. Note that this is single-turn absolute: if you need to record how many turns have been made, choose a multi-turn type; if turns are counted in the controller instead, any rotation of the shaft while power is off loses the count, so a backup power supply or re-homing is needed
• Wide operating temperature range. The manufacturer lists wide temperature range as a characteristic of its standard products, and the limits it covers let the same solution be used from inside a factory to outdoors and at altitude; the actual range varies by model, so check the individual datasheet when selecting

It is not a universal answer either. In environments with coolant spray or condensation, conductive contamination directly affects capacitive measurement, and the products themselves mostly have modest protection ratings that the mechanical housing has to make up for; for linear measurement needing extremely high resolution, an optical scale is still the more direct choice.

A UAV gimbal is a case where all of these properties are used at once: the hollow shaft lets video and power cables run through the center of the axis, the bearingless form reduces the overhung mass of two- or three-axis gimbals, absolute output lets the gimbal know where it is pointing as soon as it is powered up, and immunity to magnetic fields matters because the magnets of the BLDC direct-drive motor are only a few centimeters away. How the gimbal's counterweighting, inertia and pointing error go into the same budget is covered in the gimbal section of the UAV EO/IR payload page.

A final reminder about the easiest thing to get wrong: resolution is not accuracy. An encoder that can resolve 0.001° does not mean the mechanism stops to within 0.001°; the whole transmission chain and thermal drift lie in between. The full breakdown is in How to Choose a Servo Motor: Three Numbers That Decide Whether a Mechanism Moves Accurately.

2.4 Motion profile planning: the real cycle time bottleneck

Many people assume that shortening the cycle time means turning up the speed. For a measurement station, what really eats the time is often the part after arrival, "waiting for it to stop vibrating".

So when we plan cycle time, we do not measure "whether the command says it has arrived", but the time from the in-position signal to the mechanism actually coming to rest within tolerance. This time can be measured with an accelerometer, or by shooting a burst with the measurement camera itself and seeing when the image stops changing; the latter is closer to the real acceptance condition.

2.5 Synchronizing motion and image capture

When a measurement station has to shoot on the move (on-the-fly capture), the timing between motion and image capture becomes part of the accuracy.

There are two layers to the solution. The first is triggering by position rather than by time: the position compare output of a motion control card compares encoder counts directly and sends a pulse when the set position is reached, without going through the operating system, so jitter can be kept below a microsecond. The second is synchronizing the whole chain: EtherCAT distributed clocks give all nodes a common time base, typical control cycles are between 250 µs and 1 ms, and synchronization jitter between nodes can be kept within 1 µs; CANopen (with the CiA 402 drive profile) suits applications with few axes and less extreme real-time requirements.

If site conditions allow stop-and-shoot, stop and shoot: it is the simplest and most stable approach. On-the-fly capture is a measure for when the cycle time really cannot be brought down otherwise, and both of the points above have to be done along with it.

2.6 When you need your own drive firmware

In most cases an off-the-shelf drive is fine. Only in the following situations, where off-the-shelf products often do not fit, is in-house development worth considering:

• Direct-drive gimbals or scanning mechanisms: with no reduction gearing, low-speed smoothness depends mainly on commutation quality, the current loop and feedback resolution, and the parameter space of general-purpose drives is often not enough.
• Non-standard feedback: custom angle sensors, or interface and interpolation requirements that off-the-shelf drives do not support. Most mid- to high-end servo drives already support resolvers, sin/cos encoders and common optical scale interfaces directly, so check the drive model before deciding to do your own decoding.
• Tying the control loop to the measurement: for example correcting position from real-time image feedback, where the loop has to run on the same processor to keep up.
• Size, weight or power limits: payload-type equipment has no room for a standard drive enclosure.

Our usual split is: the FPGA handles encoder decoding, PWM generation and hardware-level synchronized triggering, the work that needs deterministic timing; the MCU runs the velocity and position loops, communication protocols and the state machine; the current loop goes in the FPGA or in the highest-priority interrupt of the MCU, depending on the bandwidth needed. The benefit of splitting it this way is that timing-sensitive parts are not affected by software scheduling jitter.

2.7 Common symptoms on site and their causes

This table is compiled from situations that really occur. The symptoms all look like "not accurate enough", but the causes lie at completely different levels.

3. How to configure a vision inspection station

"What needs to be seen" decides the waveband, "how large the defect is" decides the resolution, and "how fast the line runs" decides exposure and lighting. Do it in the reverse order, and the most common result is buying a camera that cannot see the defect.

3.1 Decide the waveband first

"Invisible defects" have completely different causes in different wavebands. First establish whether you are using reflection, transmission or the object's own emission; only then can you talk about cameras.

3.2 Resolution, depth of field and aperture: a worked example

Assume a 100 mm wide field of view and a 2448 × 2048 camera (about 5 megapixels) with a 3.45 µm pixel pitch:

• Each pixel corresponds to 100 mm ÷ 2448 = 40.8 µm
• Catching a 0.2 mm defect means 4.9 pixels, right at the practical lower limit of "the smallest defect should cover at least 3 to 5 pixels". Anything smaller calls for a narrower field of view or a different camera, not a different algorithm
• Magnification m = sensor width 8.45 mm ÷ field of view 100 mm = 0.0845

How to calculate the focal length itself (working back from field of view and working distance) is covered in How to Calculate Lens Focal Length from Field of View and Working Distance. Next, the aperture decides the depth of field, but the same aperture also decides the diffraction limit, and the two pull in opposite directions:

At F/11, the blur from diffraction (4.3 px) is already about as large as the defect to be detected (4.9 px): the depth of field gained by stopping down eats the resolution it was meant to protect. A sensible operating point for this station is around F/5.6: about 12 mm of depth of field, with diffraction blur of about two pixels. If the height variation of the part exceeds this range, what needs solving is the fixture (bringing the measured surfaces back to the same height) or focusing and capturing in several passes, not stopping down further. A telecentric lens keeps magnification constant with height, which matters for dimensional measurement, but its depth of field is still limited by aperture and diffraction.

3.3 Lighting often matters more than the camera

With the same camera and the same lens, changing the lighting method can turn "invisible" into "obvious at a glance". The four basic methods each correspond to a different type of defect:

• Backlight: for measuring outlines and hole diameters. The part becomes a silhouette with the highest edge contrast, the first choice for dimensional measurement
• Coaxial light: for mirror-like and polished surfaces. Light goes out along the lens axis and reflects straight back, so flat areas are bright and flaws are dark
• Dark field: low-angle side lighting. Only surface relief scatters light into the lens, so scratches and dents light up while flat areas stay dark
• Diffuse dome: uniform lighting for curved and reflective parts, used to remove highlights

With the wrong lighting method, no algorithm downstream can make up for it, because the contrast never made it into the image in the first place. The four lighting methods above apply to visible, NIR and SWIR (with light sources and filters changed to the matching band); thermal infrared (MWIR and LWIR) sees the radiation the object emits itself and needs no external lighting, and the radiation of objects at room temperature falls mainly in the LWIR.

3.4 Exposure time and motion blur

If the part is moving, it travels some distance during the exposure and the image blurs. The amount of blur is simply speed times exposure time. Using the 40.8 µm pixel size from above, to keep blur within one pixel:

• Conveyor at 100 mm/s → exposure must be shorter than 408 µs
• Conveyor at 200 mm/s → 204 µs
• Conveyor at 500 mm/s → 82 µs

An exposure of only two hundred microseconds means there must be enough light within two hundred microseconds, which usually means strobed lighting rather than constant lighting. The other option is to stop the mechanism before shooting, at a cost that shows up directly in the cycle time (see the next section). In addition, use a global shutter wherever things move: a rolling shutter reads out line by line, so moving objects are captured skewed; if only a rolling shutter is available, the strobe must fall within the window when all rows are exposing at the same time.

Error in the trigger timing itself also becomes position error; that part was already calculated in 2.5 Synchronizing motion and image capture.

4. How to budget the cycle time

The cycle time of a measurement station is the sum of five segments. Laid out side by side, they usually show that the camera is not what needs to change.

Taking the station above as the example: the stage moves 50 mm into position (with the S-curve profile from section 2.4), a 5-megapixel camera, GigE transfer.

This chart makes just one point: to cut cycle time, first look at where the time goes. On this station, replacing the camera with one twice as fast saves less than 5 % of the total. The truly effective measures, in order, are:

1. Improve the motion profile and mechanical stiffness: the 400 ms segment is by far the largest. Here we use the S-curve from section 2.4: on arrival at 0.40 s the residual vibration is only ±5 µm (about 0.1 pixel), so capture can start immediately, and these 400 ms are almost all move time. With a trapezoidal profile, the command arrives at 0.35 s, but the ±27 µm vibration still has to die down, and the 50 ms saved is often eaten back by settling time. To shorten this segment further, the mechanism's stiffness and resonant frequency must be raised first, so that steeper acceleration can be used without exciting vibration
2. Change the transfer interface: 10 GigE or CoaXPress (single-lane CXP-12, or dual-lane CXP-6 and above) can bring the 45 ms transfer time down below 5 ms; the actual lower limit also depends on the sensor's own readout speed
3. Move processing to an FPGA or GPU: the 30 ms segment usually has a lot of room for compression
4. Capture on the move: triggering by position instead of "stop, then shoot" removes the stopping and settling time, and lets transfer, processing and the next part's move overlap; the cycle time is then set by part pitch ÷ conveyor speed (for example a 50 mm pitch at 200 mm/s gives 250 ms). The cost is that exposure time is locked by motion blur to a few hundred microseconds (see 3.4; about 204 µs at 200 mm/s), so longer exposure can no longer make up for light, strobed lighting is usually needed, and trigger accuracy requirements rise by an order of magnitude

Every one of these has a cost. Whether it is worth it depends on how many parts the line makes a day, which is a question we always ask at the requirements stage.

5. Machine integration

Turning parts that can each move on their own into a machine that runs the whole process by itself.

This work covers interface definitions between motion mechanisms and measuring instruments, sensor and I/O configuration, timing of pneumatic and fixture actions, safety circuits and emergency stop logic, and state machine design for loading, unloading and process flow. In most integration projects the difficulty is not in any single component but in making "who waits for whom" explicit: image capture waits for the mechanism to come to rest, the mechanism waits for the fixture to clamp, the fixture waits for the part to arrive. If any step lacks a clear completion signal, the whole machine can only wait on fixed delays, and the cycle time will not come down.

At delivery we also provide interface documents and timing diagrams, so whoever maintains the machine later does not have to reverse-engineer the program to know what it is waiting for.

6. Automation GUI and data traceability

A machine that moves is only the first step. Whether the site can operate it on its own, and whether results can be traced afterwards, decides whether it has really gone into production.

The key to interface design is that operators see only what they need: select a recipe, load the part, press start, read the result. Engineering mode and parameter adjustment sit behind access rights. Measurement results are saved automatically with time, recipe, machine status and operator, and where needed can be connected to an existing MES or database, so a quality issue can be traced back to the batch, the machine and the parameter set.

Turning "the direction a customer describes into performance tests that can be measured" is already one of our services; at the automation stage, those tests become the pass/fail conditions and report fields in the GUI.

7. The control platforms we use

The choice depends on real-time requirements and the customer's maintenance conditions; no one platform is always better.

8. Project process and deliverables

Each stage has an output you can accept. You can commission just one stage, or hand us the whole run.

Stage 02 is deliberately placed before mechanical design. "Can this defect be seen in this waveband under this lighting?" is the highest-risk question in the whole project, and the time spent shooting samples is far cheaper than discovering after the mechanism is built that the defect cannot be seen.

9. Common questions and limits

Being clear about scope is more useful than promising everything.

How short can the cycle time get?

It depends on how the station is made up. For a "stop, then shoot" measurement station like the example on this page, the comfortable range is about 300 to 800 ms per part. Getting below 100 ms usually takes three things at once: switching to capture on the move (position triggering), changing the transfer to CoaXPress or 10 GigE, and moving processing to an FPGA. Going further down to the 20 ms class is mostly the territory of line-scan cameras with dedicated lighting, where the mechanical and optical design approach is entirely different.

What is the smallest defect that can be caught?

It is limited by the field of view and the sensor's pixel count, not by the algorithm. A defect must cover at least 3 to 5 pixels for a stable decision. With 5 megapixels and a 100 mm field of view (about 41 µm per pixel), the lower limit is about 0.12 to 0.2 mm. Going smaller means narrowing the field of view (shooting the same part at several positions, which lengthens the cycle time) or a camera with more pixels. This trade-off has to be decided at the requirements stage, not after the mechanism is built.

What about mirror-like and transparent parts?

That is a lighting problem, not a camera problem. Mirror-like surfaces are usually handled with coaxial light or a diffuse dome; transparent parts often rely on backlight to measure the outline, or switch to SWIR to see inside directly. For this kind of project we recommend shooting samples before discussing specifications: quoting a specification without having shot anything is not good for either side.

How accurately can thermal imaging measure temperature?

An LWIR camera measures radiation, and converting it to temperature requires knowing the surface emissivity. When emissivity is unknown or changes (oxidation, oil, changing viewing angle), absolute temperature error can range from a few °C to several tens of °C: on a painted surface, an emissivity error of 0.1 gives about 2 to 5 °C (object at 50–80 °C, background at 25 °C), and on bare or semi-bright metal parts it can reach several tens of °C. But as long as the surface condition is consistent within a batch, relative differences between parts in the same batch are usually still reliable, which is why thermographic inspection on production lines relies mostly on relative judgment: not "this bearing is at 68 °C", but "this one is 12 °C hotter than the others in the batch". When absolute temperature is needed, a reference body of known emissivity has to be placed in the field of view.

When would we advise against automation?

When volumes are low, every part is different, or the defect cannot be clearly defined. The value of automation comes from repetition: for an inspection done only thirty times a year, manual work plus a fixture is usually more economical. An unclear defect definition is a more fundamental problem: if three experienced inspectors disagree about the same part, a machine will not have the answer either, and the first step is to write down the judgment criteria.