Manual · Chapter 5

Measured signals & modal test

The other way to get modes: measure them. This chapter takes a recording from a file to a set of identified modal parameters, then gives the numbers behind the charts: resolution, windowing, averaging and dB references.

1 · Importing measurements

File ▸ Import Signal… (WAV/CSV/UFF), or Ctrl+I. Three formats come in:

Then the unit-assignment dialog opens. It shows the file name, the channel count, the sample rate and the duration, then one row per channel with three things to set:

• Quantity: Acceleration, Velocity, Displacement, Sound pressure, Force or Voltage.
• Unit: the list changes with the quantity. m/s², g, mm/s² for acceleration; Pa, kPa, mPa for pressure; N, kN for force. Everything is converted to SI on the way in and stored that way.
• Scale: the raw-to-unit factor, your sensitivity correction. Physical value = raw value × scale. A raw WAV in [−1, 1] with a scale of 2.0 becomes a signal reading ±2 in the unit you chose.

The dialog is mandatory. Unitless data cannot enter the project, and dismissing the dialog abandons the import.

Imported signals land under Signals in the project tree.

2 · Generating excitation instead

You do not need a measurement file to set up a virtual test. The Signals header in the tree carries a + button, Create signal (hammer, step, sweep…), which opens Create Signal. Pick a type and see it drawn before you commit:

• Hammer impact (half sine): the standard impact test excitation. Its usable bandwidth follows from the pulse width, and inWaves reports how far up the excitation stays flat when the signal is added.
• Step (load-release): load and let go.
• Sweep (linear): a shaker sweep across a band.
• White noise: broadband, with an explicit seed so the same signal can be regenerated.

Set the Quantity, one of Force, Acceleration, Velocity, Displacement or Pressure. A generated signal obeys the same rule as an imported one: no unitless data. Press Add to Project and it appears beside your measurements.

3 · Looking at a signal

Select a signal and the properties panel shows Sample rate, Duration, Channels and Source, then a table headed Channels (select a row → active channel) with each channel's Quantity and Unit. The row you click becomes the active channel, and every chart below acts on it.

Under Charts there are five views:

Frequency resolution is Δf = 1/T, where T is the length of data going into one transform. Two closely spaced resonances need a long T, which is what a transient does not have: a 4096-sample window over a hammer hit spreads the impact across the whole window. Spectrum for stationary signals, spectrogram for changing ones. If two peaks will not separate, record for longer. Section 7 gives the numbers.

Every line chart has the same toolbar: Zoom with a box, Undo the zoom, Reset to the full extent, and Save image for a PNG. Charts open as tabs. Pinning a tab keeps the current chart while the next run opens a new one, and Split view: chart and 3D scene side by side puts a chart next to the geometry.

Under Listen / download, ▶ Play plays the active channel, WAV and CSV export it. The panel prints RMS and peak in the channel's own unit; for a pressure channel it adds the sound pressure level in dB re 20 µPa. Playing anything that is not a pressure is a sonification, and the panel says so.

4 · Modal identification

This needs a multi-channel measurement: one channel is the excitation, the others are responses. A single-channel recording cannot give an FRF, and the button says so.

1. Choose the excitation. In the modal analysis section of the panel, pick the channel holding the hammer or shaker force. Everything else is treated as a response.

2. Compute the FRF. FRF + Coherence estimates H1 = S_xy/S_xx with Welch averaging and draws magnitude against Frequency [Hz], with Coherence [-] underneath.

3. Read the coherence before the peaks. Coherence is the fraction of the response linearly explained by the excitation, between 0 and 1. Near 1 across a band, the FRF there can be trusted. A dip means something else is contributing: noise, an unmeasured second input, a nonlinearity, a rattling fixture, or no energy there because the hammer tip was too soft. A peak sitting in a coherence dip is not a mode you have measured. Coherence also drops legitimately at antiresonances, where the response is genuinely near zero. That dip is narrow; a bad measurement's dip is broad.

4. Run the stabilisation. Stabilisation fits the FRF with a rational model at increasing orders and plots every pole it finds. A pole that reappears at the same frequency and damping as the order rises is physical; one that wanders is a numerical artefact of the fit. Physical poles are marked stable, and clicking them selects the mode frequencies. Select every stable column that lines up with an FRF peak.

5. Extract the shapes. The extract-shapes button carries the count of poles you have selected. It fits residues at those poles and produces an EMA mode set under Mode sets, beside the FEM ones and directly comparable with them.

The panel's Measurement Analysis checklist tracks the route: Import a signal and select it, Multi-channel measurement (excitation + response), Compute FRF + coherence and Select poles on the stabilisation chart. Each ticks as you go.

5 · Putting shapes on geometry

A mode shape from a test is a list of numbers per channel. To see it move, inWaves has to know which point on the geometry each channel measured.

1. Select the geometry with a signal active. The panel shows a Channel → node mapping editor listing every channel.

2. Pick a channel, press Assign Node, and click the point on the model where that accelerometer sat. The mapping stores the vertex and a direction; left as the surface normal, the direction is computed from the faces meeting at that vertex.

3. Repeat for every channel, then press Save. An unmapped channel blocks the animation and is named.

With the mapping saved, ODS… animates the operating deflection shape. You give a frequency, and inWaves takes the complex amplitude of every mapped channel at that frequency and spreads it over the mesh by inverse-distance weighting. An ODS is how the structure is actually moving at that frequency under whatever is exciting it: one mode, several, or a forced response with no resonance. The frequency is snapped to the nearest available bin and the displayed shape is normalised to its own peak, so read the colours as relative. The dialog also asks for a Window; without one, a strong neighbouring tone leaks into the line you asked for.

6 · Transmissibility between sets

Transmissibility… on a signal's panel asks the same question as an FRF, but with the response coming from a different signal set: this signal drives, another set responds. A base sweep against the accelerometers on the isolated side, for instance. Pick the excitation channel and the response set; all of that set's channels are used. The two sets must share a sample rate, and inWaves refuses the pair rather than resampling behind your back. The result is the same H1 magnitude and coherence chart, with the unit built from the two quantities: acceleration in and acceleration out gives a dimensionless ratio.

7 · Signal processing reference

Frequency resolution and record length

One rule underlies every spectrum: Δf = 1/T, where T is the duration of the data block being transformed. 0.5 Hz resolution means each spectrum consumes 2 s of data; 0.1 Hz means 10 s. Zero-padding makes the curve smoother, not the resolution finer: two resonances 0.3 Hz apart do not separate in a 1 s record no matter how many points the FFT has.

The averaged estimators use a fixed block of 4096 samples, so their Δf follows directly from your sample rate:

Sample at 51.2 kHz to catch a 20 kHz component, then try to separate two modes 8 Hz apart at the bottom of the band, and the block is too short by construction. Record the low-frequency test at a lower sample rate, or accept that the peaks merge.

Windowing

A transform assumes the block repeats forever. The signal never completes a whole number of cycles inside it, so the joint between end and start smears energy from a real peak across the whole spectrum. That is leakage, and a window tapers the block to zero at both ends to stop it.

inWaves applies a Hann window to every spectral estimate: amplitude spectrum, PSD, spectrogram, waterfall and the FRF. The PSD and spectrum chart titles name the window they used.

A window trades main-lobe width against side-lobe leakage, and Hann sits in the middle. Two consequences: it loses amplitude accuracy on a tone that falls between bins, and it tapers a hammer pulse like everything else, so a hit late in the block is attenuated. Hit near the start of the record.

Averaging (Welch)

A single spectrum of a random signal is a poor estimate: its scatter does not shrink as the record gets longer. Welch's method cuts the record into overlapping blocks, transforms each, and averages the results. Variance falls roughly in proportion to the number of independent blocks, so four averages halve the scatter and sixteen quarter it.

These are the defaults, and each chart lets you change them from the bar above it:

With 50 % overlap the blocks advance by half a block, so a record of N samples yields about 2N/4096 − 1 averages: a 10 s record at 8192 Hz gives roughly 39. The block length is capped at the record length, so a record shorter than 4096 samples is transformed whole and averaged once, and it will look noisy.

Changing the estimate from the chart

The spectrum, PSD, spectrogram and FRF tabs each carry a control bar above the plot. Window is offered on all four; Segment length and Overlap on the averaged estimates; FFT length on the spectrogram; and Scaling on the amplitude spectrum, where RMS and Peak differ by exactly √2 for a sine. Changing any of them re-runs the estimate on the same record and redraws the same tab.

The right-hand side of the bar reports what you got: Δf, and the number of averages the record could supply. Ask for a 16384-sample block on a short record and the segment count collapses; the bar shows it immediately.

Changing the estimator on an FRF discards the poles you had selected: they were identified from the previous estimate.

Sampling and aliasing

A sampled signal can only represent frequencies below half its sample rate. Anything above that Nyquist limit folds back into the band and appears as a peak at the wrong frequency, indistinguishable from a real one once recorded. There is no processing that undoes it.

Level and dB references

Levels follow ISO 1683, and the reference depends on the quantity, which is why the unit-assignment dialog is mandatory:

Acoustic results carry the same convention: a radiation result's Reference row reads dB re 20 µPa (ISO 1683), and so do the sound-pressure field maps. A quantity with no defined reference is given no dB value at all.

Next: sound radiation & rooms, turning a mode set, measured or computed, into sound.