MovingAverageFilter

MovingAverageFilter smooths a noisy signal by averaging it over time, tuned by a sample count rather than by a frequency. It is the filter to use when you think in samples: “average this over a hundred readings”. Its memory cost is the same at any sample count, so a very long average is as cheap as a short one.

It averages by weighted decay, not over a fixed window: older samples fade out rather than dropping out. So output approaches a step gradually and does not finish after numberOfSamples cycles — it reaches 63% at that point and 90% at a little over twice it.

flowchart LR
    i1(["input — signal to smooth"]) --> B["MovingAverageFilter"]
    i2(["reset"]) --> B
    i3(["disable"]) --> B
    B --> o1(["output — smoothed signal"])
    B --> o2(["isEnabled"])

Time constant $\tau = (N+1) \times$ task period [s], with $N$ = numberOfSamples. At the default 100 on a 1 ms task that is 0.101 s, a cut-off of 1.58 Hz: 63% of a step in 0.1 s, 90% in 0.23 s. Phase lag at the cut-off is 45°, and that lag is the whole trade. Note the time constant scales with the task period, so the same sample count is a different filter in seconds on a different task rate.

For the first numberOfSamples cycles after a start or a reset the filter converges faster than the numbers above, because it is averaging everything it has seen rather than decaying. That is deliberate and useful.

Signals

Inputs

Outputs

Parameters

Both parameters are persistent and survive a controller restart. The inputs and outputs do not. No parameters exist below this block.

Setup

1. Link the source signal: …/sensor/rawPressure → …/pressureFilter/input. input follows the source on a trace.

2. Set enable false. output equals input sample for sample and isEnabled reads false.

3. Set numberOfSamples to 10. Any value is accepted, so there is nothing to read back and check.

4. Set enable true. isEnabled reads true and output tracks input closely, with the noise still visible.

5. Raise numberOfSamples in steps of roughly double, watching the noise and the loop consuming output. Stop one step below where that loop feels soft or hunts.

   Step 5 changes what the controller sees. Each doubling of the sample count doubles the delay. Raise it with the axis at rest before trying it in motion — the added delay can destabilise a tightly tuned loop.

6. Step the input and time how long output takes to reach 63% of the step. It should be about numberOfSamples plus one task periods.

7. Trace every element of input and output. An element that reads zero while the machine moves is an unwired channel, not a filtered one.

Tuning

1. Read the ripple you want gone off a trace of input, and count its period in task cycles. Set numberOfSamples to at least a few times that count.
2. Measure the delay you have bought: numberOfSamples plus one task periods is the time constant, and the delay below the cut-off equals it.
3. Keep that delay below a tenth of the response time of the loop consuming output. Above that, the loop pays for the smoothing in stability.
4. Double numberOfSamples to halve the noise. It is a cheap knob — there is no memory or computation penalty for a large value — so the only cost is delay.
5. Check the loop in motion, not at rest. Delay costs phase margin only while the axis is moving fast enough to need it.
6. Re-check the sample count after any task-rate change. The count stays the same in samples but the delay in seconds moves with the task period.
7. If you need a filter with a hard cut at a known frequency — to kill a mains harmonic, say — this block cannot do it. A finite-impulse-response filter with equal taps can.

Read the time constant off any curve as the moment it crosses 0.63.

A conservative starting point for a pressure signal on a 1 ms task, feeding a loop with a 100 ms response time:

numberOfSamples = 50
enable          = true

This is a starting point, not a final tuning. Work step 1 with a trace of your own signal.

Limits and errors

The block raises no errors or warnings and logs nothing. Every failure above shows as a value on a trace, not as a message.

Verified against motorcortex-control3 3.30.0 (bc348fd).