Signal Processing — Fourier Transform, Filters & Sampling

A raw sensor trace often looks like noise: a jagged line that hides the one slow trend you actually care about. Signal processing is the set of moves that turns that mess into something readable, by asking which frequencies are present, which parts to keep, and how fast you must sample to store the signal digitally without corrupting it.

When to reach for this method

Use signal-processing tools whenever a measurement or wave carries information you cannot read directly off the time plot. That covers a changing voltage, a sound recording, a vibration, or light at a detector. The three workhorse ideas are the Fourier transform (which frequencies are present), filtering (keep some parts, suppress others), and sampling (turn a continuous signal into digital data, but only if the rate is high enough). If the raw waveform is too tangled to interpret by eye, this is the toolkit.

The procedure, step by step

1. Identify the signal. Start from the physical quantity changing over time, such as voltage, sound pressure, or displacement.

2. Look at both domains. In the time domain you see how the signal changes moment by moment, which matters when timing matters, such as when a pulse arrives. In the frequency domain you see the same signal as a mix of components. For a continuous signal $x(t)$, the Fourier transform is

You rarely compute this integral by hand; the point is that a complicated waveform is a combination of simpler oscillations, so a messy time plot can become a few clean peaks in frequency.

3. Filter with a purpose. Choose the filter by what should remain:

• a low-pass filter keeps lower frequencies and reduces higher ones,
• a high-pass filter keeps higher frequencies and reduces lower ones,
• a band-pass filter keeps a selected range and reduces the rest.

A filter only helps if the part you want to remove is separated in frequency (or some other property the filter uses) from the part you want to keep. It does not know what "noise" is on its own.

4. Check the sampling rate. If the highest frequency you must preserve is $f_{\max}$, the sampling rate must satisfy

to avoid aliasing in the ideal picture. That threshold $2f_{\max}$ is the Nyquist rate. Below it, high-frequency content can masquerade as a lower frequency and the data points to the wrong signal, so engineers sample above the minimum and often add an anti-alias filter before digitizing.

A full example, start to finish

A temperature sensor should change slowly, but its measured voltage carries a fast electrical ripple. Model it as

The constant $2$ is the baseline, the $1\ \mathrm{Hz}$ term is the slow change you care about, and the $60\ \mathrm{Hz}$ term is interference. Running the steps: the Fourier view places energy at $0\ \mathrm{Hz}$ (offset), $1\ \mathrm{Hz}$ (signal), and $60\ \mathrm{Hz}$ (interference). To keep the slow trend, a low-pass filter with a cutoff well above $1\ \mathrm{Hz}$ but well below $60\ \mathrm{Hz}$ reduces the ripple while leaving the trend intact. Now sampling: at $f_s = 20\ \mathrm{Hz}$ the Nyquist rate is $10\ \mathrm{Hz}$, so the $60\ \mathrm{Hz}$ component is above the limit and cannot be represented correctly unless removed first. In fact a $60\ \mathrm{Hz}$ sinusoid sampled at $20\ \mathrm{Hz}$ lands on repeating sample points and can vanish from the data, which is not harmless: the sampling has hidden it in a misleading way. That is precisely why pre-filtering matters before digitizing.

Where each step tends to break, and how to self-check

• Step 2: treating the frequency plot as a complete picture. Real measurements are finite, noisy, and time-varying; the transform is an interpretation, not a perfect fingerprint. Self-check: does the peak count match the components you expect?
• Step 3: filtering without stating what must be preserved. A low-pass filter also reduces any real high-frequency signal. Write down the goal first, then a filter is "good" relative to it.
• Step 4: using the Nyquist rate as a comfortable target. $f_s > 2f_{\max}$ is the bare minimum; working too close leaves no room for non-ideal filters or out-of-band content. Self-check: did you sample comfortably above $2f_{\max}$, not just barely?
• General: assuming sampling adds information. It only records what is already there; too low a rate cannot be undone later.

How the three ideas reinforce each other

It is tempting to learn the Fourier transform, filtering, and sampling as three separate tricks, but they form a chain, and most real mistakes come from breaking the chain. The Fourier view tells you where the wanted and unwanted content sit in frequency. That map is what lets you choose a filter intelligently, since a filter can only separate parts that are already separated in frequency. And the same map sets the sampling requirement, because you cannot pick a safe $f_s$ until you know the highest frequency present, including interference you may not care about but which still aliases. Skip the first step and the other two become guesswork: people filter blindly and hope, or sample at a round number and discover phantom low-frequency components later. Treating the three as one workflow, look, then filter, then sample, is what makes signal processing predictable rather than a bag of heuristics.

Where signal processing is used

It appears anywhere measurements or waves carry information: microphones, seismometers, radio receivers, ECG and EEG instruments, optical detectors, and control systems. It is most valuable when the raw signal is not directly readable, replacing a stare at a tangled waveform with transforms, filters, and sampling rules. To extend the idea, take $x(t) = \sin(2\pi \cdot 5\, t) + 0.4\sin(2\pi \cdot 40\, t)$ and ask which frequency a low-pass filter would keep and what sampling rate sits comfortably above its Nyquist limit, then compare with transfer functions to see how systems shape signals in frequency.