Signal processing is the art of getting a defensible answer from a changing measurement, without turning missing information into a confident story. Whether the observation comes from a microphone, medical sensor, seismometer or radar, the same discipline applies: understand the source, acquire it properly, transform it deliberately, and inspect the limits.
This reader-first guide begins with one central puzzle: how can two correct-looking graphs tell different stories about the same event? It then builds from first principles to sampling, filters, worked calculations and trusted specialist reading.
Five ways into the guide
1. Understand the measurement · Chapters 1–4 · Start here
2. Sample without inventing patterns · Chapters 5–8 · Start here
3. Choose mathematical tools · Chapters 9–12 · Start here
4. Follow worked examples and applications · Chapters 13–16 · Start here
5. Diagnose, navigate and teach · Chapters 17–20 · Start here
Chapter contents · Related guides · The Gold Standard master
Full chapter index
Understand the measurement · Chapters 1–4
- 1. A signal is an observation, not an explanation
- 2. Start with the measurement chain
- 3. Continuous time and sampled time are different objects
- 4. Choose the question before the transform
Sample without inventing patterns · Chapters 5–8
- 5. Sampling and aliasing: why a tidy curve can lie
- 6. Quantisation, dynamic range and clipping
- 7. Filtering: what you remove and what you delay
- 8. Frequency views, spectra and Fourier ideas
Choose mathematical tools · Chapters 9–12
- 9. Windowing and spectral leakage
- 10. Convolution and impulse response
- 11. Noise, background and the signal-to-noise ratio
- 12. Record length, resolution and latency
Follow worked examples and applications · Chapters 13–16
- 13. Worked example 1: a 12 Hz vibration becomes 8 Hz
- 14. Worked example 2: averaging helps noise but hides an alert
- 15. Worked example 3: a decibel figure has a denominator
- 16. Applications across sound, health, radar and infrastructure
Diagnose, navigate and teach · Chapters 17–20
- 17. Diagnose the wrong move before changing the software
- 18. Specialist routes across the eduKate ecosystem
- 19. Questions readers ask
- 20. A practical next step and a source trail
| Processing stage | Question to ask | Mistake it helps prevent |
|---|---|---|
| Acquisition | What physical quantity, unit and calibration? | Treating a voltage as the quantity itself |
| Sampling | What rate, bandwidth and analogue filter? | Aliasing a fast event into a slow one |
| Transformation | What was removed, delayed or amplified? | Confusing a smoothed record with the raw event |
| Interpretation | What independent evidence supports the claim? | Treating a chart feature as proof of a cause |
1. A signal is an observation, not an explanation
A microphone measures a changing electrical output; a seismometer records motion; a radar receives echoes. None of those outputs is automatically an explanation of its source. Signal processing is the set of mathematical and engineering methods used to recover useful structure from observations while keeping uncertainty visible.
The practical question is: What did we actually measure, what operation did we apply, and which conclusion does the result support? A polished graph is not the answer until that chain is clear.
↑ Full contents · Next chapter →
2. Start with the measurement chain
Trace the physical quantity, sensor, analogue conditioning, conversion to digital samples, processing, display and final decision. Record the unit at every handoff. A sensor reading of 2 volts is not yet 2 metres per second, 2 pascals or 2 degrees.
A calibration curve may translate voltage to pressure; timestamp correction may align sensors; missing readings may require rejection rather than imaginative interpolation. If the original quantity is unclear, adding an algorithm only makes the confusion faster.
← Previous chapter · Next chapter →
3. Continuous time and sampled time are different objects
An analogue signal can vary between observations. A sampled signal is a sequence of measurements made at specified instants. A digital record also has finite numerical precision. Those two changes—time discretisation and value quantisation—have different consequences.
Write the sample rate in samples per second (hertz), the record duration in seconds, and the number of samples separately. A file with 1,000 samples at 100 samples per second spans 10 seconds of observation, not 1,000 seconds.
← Previous chapter · Next chapter →
4. Choose the question before the transform
Some questions require timing: when did the impact arrive? Others require frequency: is a machine vibrating mainly at 25 Hz? Still others require a trend or a classification. A method that answers one question may erase the evidence needed for another.
For a safety alarm, preservation of short peaks may matter more than a very smooth average. For long-term environmental patterns, noise reduction may be useful. Define the decision, permitted error and response time before selecting a filter.
← Previous chapter · Next chapter →
5. Sampling and aliasing: why a tidy curve can lie
Sampling too slowly allows a fast oscillation to masquerade as a slower one. The ideal Nyquist condition says a strictly band-limited continuous-time signal needs a sampling frequency greater than twice its highest frequency for unambiguous reconstruction under the usual assumptions. Real measurements need suitable analogue anti-alias filtering and practical margin.
An aliased frequency cannot be recovered merely by drawing a smooth curve through the stored points. The missing information was lost before plotting.
← Previous chapter · Next chapter →
6. Quantisation, dynamic range and clipping
Once values are digital, each sample sits on a finite set of representable levels. Small changes may fall below the quantisation step; large ones may hit the measurement limit and clip. More bits do not correct a badly placed sensor, incorrect gain or a saturated amplifier.
Always inspect the raw waveform for flat-topped peaks, a drifting baseline and abrupt gaps. Clipping is a different problem from ordinary noise: information beyond the limit has not been recorded.
← Previous chapter · Next chapter →
7. Filtering: what you remove and what you delay
A low-pass filter suppresses rapid variation; a high-pass filter emphasises changes; a band-pass filter selects a range. A moving average may reduce random fluctuations but can also delay events and reduce sharp peaks. Filters have gain and phase behaviour, not only a tidy appearance.
Document a filter’s cutoff or window length, the expected signal bandwidth and whether real-time delay is acceptable. Compare the unfiltered and filtered observations when judging whether an event occurred.
← Previous chapter · Next chapter →
8. Frequency views, spectra and Fourier ideas
A time plot answers when; a spectrum helps answer how much signal is associated with different oscillation rates. Fourier analysis represents a signal through components with different frequencies. A spectral peak is evidence of oscillation in the measurement, not automatic proof of the physical source.
For example, a 50 Hz feature might reflect electrical interference rather than the phenomenon the researcher hoped to measure. To distinguish them, compare sensor placement, reference channels and plausible mechanisms.
← Previous chapter · Next chapter →
9. Windowing and spectral leakage
A finite recording cannot reveal an infinite-duration oscillation. If its endpoints do not join smoothly, a discrete transform can spread energy into neighbouring frequency bins. A window can reduce some leakage, but it changes amplitudes and the width of spectral features.
Record which window was used, how results were scaled and whether the amplitude represents a peak, an RMS value or a spectral density. Changing the window changes the measurement interpretation.
← Previous chapter · Next chapter →
10. Convolution and impulse response
For a linear time-invariant model, convolution describes how the system’s impulse response shapes the input. A short impulse sent through a room becomes a longer recorded response because of reflections; a sensor may also impose its own time response.
If the system is nonlinear or changes with time, the simple LTI picture may fail. Test these assumptions instead of applying a mathematically convenient model by habit.
← Previous chapter · Next chapter →
11. Noise, background and the signal-to-noise ratio
Noise can come from sensors, electronics, the environment or the process itself. Signal-to-noise ratio describes signal relative to noise only after the quantities and measurement bandwidth are defined. When comparing powers, SNR in decibels is 10 log10(Psignal/Pnoise); for compatible amplitudes under comparable impedances, 20 log10 of the amplitude ratio is used.
Subtracting a background measurement is not guaranteed to isolate a source. If source and background vary independently, differ in time, or interact, subtraction can create an unstable estimate.
← Previous chapter · Next chapter →
12. Record length, resolution and latency
For a conventional N-point transform of N samples recorded at sample rate fs, adjacent displayed frequency bins are separated by fs/N. That spacing is not the same thing as guaranteed physical resolving power. A longer observation generally provides more frequency discrimination but may react too slowly to a changing event.
Check time-window length, frequency-bin spacing, actual spectral resolution and update delay as four related but distinct design decisions.
← Previous chapter · Next chapter →
13. Worked example 1: a 12 Hz vibration becomes 8 Hz
Teaching case with an ideal pure sinusoid: A machine has a real 12 Hz oscillation but the logger samples at 20 Hz. The sampling Nyquist frequency is 10 Hz. In this simplified case the 12 Hz sinusoid can produce samples indistinguishable from an 8 Hz component (20 − 12 = 8 Hz), subject to phase/sign conventions.
The wrong move: identify an 8 Hz peak and announce that the machine vibrates at 8 Hz. The repaired method: examine analogue bandwidth, anti-alias filtering, sample rate and independent physical evidence. Increasing the plotting resolution of the old samples does not restore the lost frequency.
← Previous chapter · Next chapter →
14. Worked example 2: averaging helps noise but hides an alert
Invented readings, unit arbitrary: five consecutive samples are 2, 2, 2, 12, 2. Their arithmetic mean is 20/5 = 4. A five-sample moving average replaces a short, clearly visible peak of 12 with a value of 4 in that window.
If the task is estimating a slowly varying baseline, averaging could help. If the task is detecting brief excursions above 10, that same average could hide the event. The calculation is correct; the decision is wrong if the processing objective was not stated.
← Previous chapter · Next chapter →
15. Worked example 3: a decibel figure has a denominator
Invented powers measured over the same bandwidth: signal power is 100 units and noise power is 1, so SNR = 10 log10(100) = 20 dB. If noise rises to 4 while signal stays 100, the ratio becomes 25 and SNR is about 13.98 dB.
Calling both recordings “clear” obscures a 6 dB deterioration. Conversely, comparing power SNR across different bandwidths or incompatible measurement conditions is not a fair comparison. The measurement conditions belong beside the headline number.
← Previous chapter · Next chapter →
16. Applications across sound, health, radar and infrastructure
Audio processing supports speech intelligibility and room measurement. Radar processing estimates echo timing and motion subject to ambiguity. Medical imaging relies on acquisition physics, calibration and interpretation. Vibration monitoring can support maintenance only when sensor geometry and operating state are controlled.
These applications use related mathematical tools but different truth tests. A good algorithm is not portable just because its code runs. Ask which physical assumptions transfer and which do not.
← Previous chapter · Next chapter →
17. Diagnose the wrong move before changing the software
Common failures include confusing sampling rate with record length, treating every spectral peak as a source, averaging away a short critical event, comparing decibels with incompatible references, and removing a trend that was part of the signal.
Recovery is practical: (1) return to raw data; (2) label units and timestamps; (3) state the intended decision; (4) change one processing step; (5) compare against an independent reference or controlled test. Keep a log of what each change improves and sacrifices.
← Previous chapter · Next chapter →
18. Specialist routes across the eduKate ecosystem
Signal processing is not one giant isolated chapter. Its foundations and real-world applications are already developed across the linked guides below. Use the foundation route to understand waves, then the instrumentation route to follow actual measurements, and finally the application route to test the claim made from them.
Linked specialist article directory
Foundations
- The Gold Standard of Waves
- The Gold Standard of Sound
- The Gold Standard of Acoustics
- The Gold Standard of Radar
- The Gold Standard of Bioacoustics
Measurements
- The Gold Standard of Acoustics | Background Subtraction and Unstable Source Estimates
- The Gold Standard of Palaeoclimatology | Sampling Intervals, Aliasing and False Climate Periods
- The Gold Standard of Industrial Machinery Vibration Analysis Services
Applications
- The Gold Standard of Medical Imaging Engineering
- The Gold Standard of Biosensor Engineering
- The Gold Standard of Structural Health Monitoring Engineering
- The Gold Standard of Microwave Engineering
- The Gold Standard of Photonics Engineering
← Previous chapter · Next chapter →
19. Questions readers ask
Does filtering always make data more accurate? No. It may reduce one form of variation while removing useful events or shifting timing. Accuracy depends on the task and an independent reference.
Is a high sampling rate enough? No. Sensor response, analogue filtering, saturation, calibration and timestamps matter too.
Does an FFT show exactly what caused a sound? No. It shows mathematical structure of the recorded data. Source identification needs physical context and corroboration.
Can AI replace the measurement plan? No. Pattern recognition inherits the recording’s gaps, sampling biases and labeling decisions.
← Previous chapter · Next chapter →
20. A practical next step and a source trail
Choose one real signal—a voice recording, an accelerometer trace or a safe classroom sensor. Record its units, sampling rate and duration. Save the raw data. State one intended question; then apply one transformation and write what evidence it retains, removes or cannot establish.
Educators can begin with a synthetic dataset and ask learners to predict the effect of averaging before seeing the processed graph. In research and engineering, document the collection setup, calibration, code/version, assumptions and limits so another reader can repeat the reasoning.
Authoritative references
- MIT OpenCourseWare — Lecture 16: Sampling
- MIT OpenCourseWare — Sampling, anti-aliasing and reconstruction
- MIT OpenCourseWare — Signals, systems and information syllabus
← Previous chapter · Return to reading routes
Return to the Gold Standard master library. For independent reading, begin with Waves and Acoustics, then test your ideas with the linked instrument and application cases.