Matlab Prepare Signal Data

MathWorksOfficialCreated by MathWorks

Skill
Tool integration
Data and AI
BSD-2-Clause
No issues found

From the author

Use this skill when conditioning, loading, preparing, or labeling signal data for analysis or ML training. Covers: cleaning a single signal (fill gaps, remove drift, deoutlier, denoise, resample/align a time base) BEFORE analysis; building a `signalDatastore` pipeline; creating a `labeledSignalSet` for Signal Labeler; deriving labels (filename, folder, in-file, ROI, time-frequency ROI); stratified train/val/test splits; framing long signals; parallel processing; and shaping datastore output for `trainnet`. Triggers include "clean up this signal", "remove drift / detrend", "fill gaps", "remove spikes / outliers", "denoise", "resample to a uniform rate", "align channels", "labels from filenames", "stratified split", "prepare for Signal Labeler", and function names like `fillgaps`, `fillmissing`, `detrend`, `filloutliers`, `smoothdata`, `resample`, `synchronize`, `signalDatastore`, `labeledSignalSet`, `filenames2labels`, `folders2labels`, `splitlabels`, `framesig`, `framelbl`, `createDatastores`.

— the description by MathWorks

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

This skill provides a structured approach to signal data preparation, guiding users through the lifecycle of cleaning raw signals, managing datastores, applying labels, and partitioning data for machine learning. It emphasizes using high-level Signal Processing Toolbox functions to ensure readability and maintainability, while offering clear heuristics for ordering conditioning steps like detrending, smoothing, and outlier removal. The skill is designed to help users move from raw data to a ready-to-train state for models like trainnet.

What it does

Cleans signals by filling gaps, removing drift, denoising, and handling outliers.
Synchronizes and resamples multi-channel data onto a uniform time base.
Builds data pipelines using signalDatastore for efficient loading.
Creates labeledSignalSet objects for structured labeling and Signal Labeler interoperability.
Performs stratified splits and framing for long signal sequences.
Shapes and prepares datastore outputs for machine learning workflows.

When to use it

When you need to clean a single signal before analysis or training.
When preparing a dataset of signals for machine learning with specific labeling requirements.
When you need to align multiple signal channels or resample them to a uniform rate.
When you need to partition signal data into training, validation, and test sets.
When you need to frame long signals for windowed analysis or supervision.

Advantages

Promotes readable, maintainable code by preferring high-level toolbox functions.
Provides a logical, verified order for conditioning steps to prevent data corruption.
Reduces manual effort by automating common tasks like label derivation from filenames.
Ensures compatibility with Signal Labeler and deep learning training functions.
Offers clear guidance on when to escalate to lower-level methods.

Usage examples

“How do I remove a linear trend from my signal?”: The agent will recommend using detrend(x, 1) and explain why explicit degree specification is preferred for code clarity.
“I have a folder of signal files; how do I create a datastore and split them for training?”: The agent will guide you through using signalDatastore to load the files, followed by splitlabels to perform a stratified split.
“My signal has gaps; how should I fix them?”: The agent will suggest using fillmissing for short gaps or fillgaps for longer gaps in oscillatory signals, depending on the signal characteristics.
“How do I align two signals with different sampling rates?”: The agent will recommend using the synchronize function to place both signals onto a shared time grid.

Install

Download the skill as a ZIP:
https://autogineer.com/agent-skills/matlab/matlab-prepare-signal-data/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip matlab-prepare-signal-data.zip -d .agents/skills

Documentation by MathWorks

Prepare Signal Data

Look in Signal Processing Toolbox first. The conditioning, labeling, splitting, framing, and partitioning helpers here live in Signal Processing Toolbox — not in Stats & ML Toolbox or generic-MATLAB string utilities.

The arc: condition a raw signal (clean it) -> load a folder into a datastore -> label -> split / frame -> hand off to trainnet. Each stage is a workflow file; this page routes you to the right one.

When to Use

Cleaning a single signal before analysis: fill gaps, remove drift, deoutlier, denoise, put it on a uniform time base, align multiple channels.
Loading / preparing signal data for ML training: datastores, labels from filenames or folders, stratified splits, framing, parallel processing.
Structured labeling: labeledSignalSet for Signal Labeler, all label types.

When NOT to Use

Raw .wav audio classification with Audio Toolbox available. audioDatastore is the canonical path (this skill's custom-ReadFcn workflow handles .wav only when Audio Toolbox is absent — references/wf-custom-readfcn.md).
Frequency-selective filter DESIGN (band isolation, notch, custom FIR/IIR) — see the matlab-design-digital-filter skill. This skill's conditioning is about cleaning, not designing filters.
Computing per-frame features (RMS, crest factor, spectral / bandwidth, time-frequency features) from an already-conditioned signal — see the matlab-extract-signal-features skill. This skill's framesig / framelbl are for manual per-window labeling / supervision, not for deriving a feature table; the signal*FeatureExtractor objects window internally and emit the table.

Best practices

Deliverable is a runnable .m script the user can save, version, and re-run — not workspace state.
Prefer the highest-level function that does the job. detrend / smoothdata / fillmissing / resample read cleanly and are easy for a non-expert to follow. Drop to a lower-level / more-configurable path (designfilt + filtfilt, a hand-built AR model, a named primitive) only when you need control the high-level call cannot give, or when the user asks. Readability first; escalate to low-level for necessity, not by default.
- The high-level call usually exposes the control you think you need. In particular smoothdata(x, "sgolay", fl) takes the frame length fl as an argument — it does NOT hide it — so prefer it over calling sgolayfilt directly. Reach for sgolayfilt only for what the dispatcher genuinely lacks (derivative output via dn, or an unusual polynomial order).

0. Common reflexes

If your first instinct is one of these, the canonical replacement is one row away.

Reflex	Canonical	Detail
Hand-design a highpass/`designfilt` to remove a smooth drift	`detrend(x, n)` — escalate `n` = 1 -> 2 -> 3 before reaching for a filter; polynomial detrend has unity passband gain	references/fn-detrend.md
Invent a gap-filler (`regularizeNaNs`, `inpaintn` — not real)	`fillmissing` (interp) for short gaps; `fillgaps` (SPT, AR) for long gaps in oscillatory signals	references/wf-repair-missing.md
Hand-roll `retime` + shift + `retime` + concat to align channels	`synchronize(A, B, ...)` — one call to a shared grid	references/wf-align-channels.md
Custom `ReadFcn` for a `.csv`	`signalDatastore` default reader + `SignalVariableNames`	references/fn-signaldatastore.md
`cvpartition` for a datastore split	`splitlabels` + `subset(ds, idx{k})`	references/fn-splitlabels.md
`regexp` / `extractBefore` / `fileparts` for labels from filenames	`filenames2labels(sds, Extract=...)`	references/fn-filenames2labels.md
`regexp` / nested `fileparts` for labels from subfolders	`folders2labels(sds.Files)`	references/fn-folders2labels.md
Manual framing loop with `(i-1)*hop+1`	`framesig(x, fl, OverlapLength=...)`	references/wf-frame-and-label.md
Manual ROI-to-frame vote with `containers.Map`	`framelbl(rois, ...)`	references/wf-frame-and-label.md
`for` loop `load(file)` to read in-file label variables	`signalDatastore(folder, SignalVariableNames=["x","label"])`	references/fn-signaldatastore.md
`signalMask` when you need Signal Labeler interop	`labeledSignalSet` with ROI labels (signalMask can't import)	references/fn-labeledsignalset.md
`signalLabeler(lss)` (pass the set as an arg)	Launch bare `signalLabeler` (zero args), then Import -> From Workspace or From File	references/wf-label-and-export.md

SPT-specialized functions exist — reach for them, don't reinvent. fillgaps (AR gap fill), medfilt1 / hampel (impulse handling), sgolayfilt / smoothdata(...,"sgolay") (feature-preserving smoothing) are in Signal Processing Toolbox.

1. Workflows

Each workflow file is the entry point and lists the functions it uses. Start here.

Workflow	Use when	Reference
Repair missing samples	NaN gaps / dropouts to fill.	references/wf-repair-missing.md
Detrend, smooth, deoutlier	Drift, spikes, and/or broadband noise on one signal (smoothing/denoising lives here).	references/wf-detrend-smooth-deoutlier.md
Align multi-rate / offset channels	Several channels onto a shared time base.	references/wf-align-channels.md
Put one channel on a uniform rate	One channel -> uniform grid at a chosen rate: jittery timestamps to regularize, OR already uniform but the wrong rate to `resample`.	references/wf-uniform-rate.md
Wavelet denoising (escalation)	Non-stationary/multi-scale noise a tuned `sgolayfilt` can't remove; `wdenoise` (Wavelet TB).	references/wf-denoise.md
Envelope extraction	Amplitude outline (AM demod, peak hull) — not cleaning.	references/wf-envelope.md
Load + label + split	Folder of files -> datastore for training.	references/wf-load-and-split.md
Frame long signals + per-frame labels	Long signals, per-window supervision.	references/wf-frame-and-label.md
Label + export (all label types)	Structured labels (attribute/ROI/point/TF-ROI), export to Signal Labeler / DL.	references/wf-label-and-export.md
Parallel processing across a parpool	Per-signal work across workers.	references/wf-parallel-process.md
Custom ReadFcn (only when needed)	Format isn't `.mat` / `.csv`, or has a metadata prelude.	references/wf-custom-readfcn.md
Hand-off to `trainnet`	Datastore ready; shape for `trainnet` / `combine`.	references/wf-handoff-to-dl.md

Each workflow file names the fn- reference pages for the functions it uses; there is no separate function index — enter through the workflow that matches your task, or the reflex table above.

2. Ordering when a signal needs several conditioning steps

The governing principle (this is the real rule): order the steps so an earlier operation does not corrupt the input to a later one. Spikes bias least-squares fits and get smeared by filters/resamplers; an un-removed trend gets averaged into the signal by a smoother; most operations choke on NaN. Reason from that for the signal in front of you — do not follow a fixed chain blindly.

Default heuristic (a good starting order, not a universal law):

outliers -> detrend -> smooth, with fill and align placed by the principle above.

outliers -> detrend -> smooth is the verified core: remove spikes before a polynomial detrend (a spike biases the fit) and before a smoother (a smoother spreads the spike across its window); detrend before smooth so the smoother isn't averaging across a trend.
Fill NaN before any step that can't handle missing data (detrend, filters, most smoothers).
Align / resample: putting a signal on a new grid (retime/synchronize) creates NaN at non-overlapping times, so fill after aligning. BUT if the signal has spikes, deoutlier before resampling — resample's anti-alias filter will smear an un-removed spike. So align-vs-outliers order depends on the signal; the principle decides, not a fixed sequence.

Not every signal needs every step — identify which apply, order them by the principle, and each workflow file has an off-ramp if your problem is actually a different family.

Files

ATTRIBUTION.mdAdded by AutoGineer1.0 KiB
LICENSELicence file added by AutoGineer1.4 KiB
SKILL.md9.5 KiB
manifest.yaml650 B
references/fn-countlabels.md2.1 KiB
references/fn-detrend.md4.6 KiB
references/fn-filenames2labels.md4.8 KiB
references/fn-fillgaps.md3.3 KiB
references/fn-fillmissing.md4.2 KiB
references/fn-filloutliers.md5.0 KiB
references/fn-folders2labels.md3.4 KiB
references/fn-framelbl.md7.1 KiB
references/fn-framesig.md4.1 KiB
references/fn-hampel.md3.3 KiB
references/fn-labeledsignalset.md9.1 KiB
references/fn-labelspectrogramoptions.md4.6 KiB
references/fn-medfilt1.md3.3 KiB
references/fn-partition.md4.6 KiB
references/fn-resample.md4.6 KiB
references/fn-retime.md4.7 KiB
references/fn-sgolayfilt.md3.5 KiB
references/fn-signaldatastore.md11.9 KiB
references/fn-signallabeldefinition.md5.9 KiB
references/fn-signalmask-getmask.md8.6 KiB
references/fn-smoothdata.md4.8 KiB
references/fn-splitlabels.md5.3 KiB
references/fn-subset.md3.9 KiB
references/fn-synchronize.md4.7 KiB
references/fn-wdenoise.md5.2 KiB
references/wf-align-channels.md6.6 KiB
references/wf-custom-readfcn.md5.0 KiB
references/wf-denoise.md2.6 KiB
references/wf-detrend-smooth-deoutlier.md6.6 KiB
references/wf-envelope.md5.3 KiB
references/wf-frame-and-label.md4.7 KiB
references/wf-handoff-to-dl.md9.1 KiB
references/wf-label-and-export.md8.7 KiB
references/wf-load-and-split.md4.6 KiB
references/wf-parallel-process.md4.8 KiB
references/wf-repair-missing.md6.7 KiB
references/wf-uniform-rate.md4.2 KiB

references/wf-detrend-smooth-deoutlier.md

Workflow: detrend + smooth + de-outlier a single channel

Functions used: filloutliers, hampel, medfilt1, detrend, smoothdata, sgolayfilt

Condition one noisy channel before analysis: remove spikes, remove drift, then smooth what remains. Three steps, and the order matters - do them in the wrong order and each step corrupts the next.

Off-ramp. If the "noise" you want gone has a frequency story you can name in Hz (60 Hz hum, a noise floor above 100 Hz), this is a filter job, not a data-cleaning job - use highpass / lowpass / bandpass directly. This workflow is for the case where the input has a stationary character (sparse spikes, smooth polynomial drift, broadband wiggle) and no clean cutoff frequency.

When to use this

Reach for this recipe when the channel shows some combination of:

Sparse spikes / dropouts - a few samples far off the local level.
Slow baseline drift - a smooth wander over the record (thermal, sensor drift, gravitational).
Residual high-frequency wiggle - broadband noise on top of the signal.

The core lesson: when the input looks like this, data-cleaning primitives beat DSP filters. detrend(x, n) removes polynomial drift exactly with unity passband gain; filloutliers fixes spikes without touching clean samples. Filters have transition bands that touch the signal of interest. Reach for the primitives first; reach for a filter only when you can name a frequency.

Why the order is outliers -> detrend -> smooth

Each step assumes the previous one ran:

Outliers first. A single spike wrecks a polynomial fit - detrend minimizes squared error, so one outlier tilts the whole trend line. Remove spikes before you fit anything.
Detrend second. Smoothing a signal that still has drift just produces a smooth version of the drift. Get the baseline flat before you smooth.
Smooth last. Once spikes and drift are gone, whatever wiggle remains is the broadband noise you actually want to reduce.

Do not smooth first: smoothing spreads a spike across its whole window (turning one bad sample into win mildly-bad samples) and bakes the drift into the result.

Recipe

Step 1 - Remove outliers

Pick the outlier tool by spike density and whether you need an audit trail:

Situation	Reach for	Detail
Sparse spikes; want a smooth fill across the spike	`filloutliers(x, "linear" \| "pchip", "movmedian", win)`	fn-filloutliers.md
Sparse spikes; do not care about the fill rule	`filloutliers(x, "clip", "movmedian", win)`	fn-filloutliers.md
Sparse spikes; want to know which samples were flagged	`hampel(x, k, nsigma)` (returns mask + local median + local sigma)	fn-hampel.md
Baseline drifts, so a global threshold misleads	`filloutliers(x, fill, "movmedian", win)` (or `hampel`)	fn-filloutliers.md
Dense impulse noise (>10% of samples corrupted)	`medfilt1(x, n)`	fn-medfilt1.md

Anti-pattern. Do not use medfilt1 for sparse spikes. It is a denoiser that replaces every sample with a local median, distorting clean samples too. It earns its place only when the burst is so dense that windowed detection has no "normal" majority to compare against.

% Sparse spikes, drifting baseline -> movmedian detection, smooth fill:
xc = filloutliers(x, "pchip", "movmedian", 25);

Step 2 - Remove the trend

Pick the detrend tool by whether you can describe the drift's shape:

Situation	Reach for	Detail
Drift is smooth and slow, looks polynomial	`detrend(x, n)` for `n` in {1, 2, 3}	fn-detrend.md
Drift jumps at known points (sensor reset, calibration)	`detrend(x, n, bp, Continuous=false)`	fn-detrend.md
Drift is oscillatory or non-stationary; no polynomial fits	`x - smoothdata(x, "lowess"\|"loess", win)`	lowess = local linear (steadier vs noise), loess = local quadratic (follows curvier trends); fn-smoothdata.md
Drift band is well below signal band, shape unknown	`highpass(x, fc, fs)`	base SPT filter; pick `fc` below signal band
Remove only the mean	`detrend(x, 0)`	fn-detrend.md

Escalate the polynomial degree before switching to a filter. If a linear detrend leaves a smooth low-frequency bend, escalate n (2, then 3) before reaching for highpass - a smooth drift is often a low-order polynomial that detrend removes exactly with unity passband gain. Switch to highpass or smoothdata("lowess") only when the drift is genuinely oscillatory or non-stationary. Full rationale, the escalation example, and when the polynomial is genuinely wrong: fn-detrend.md.

Step 3 - Smooth the residual noise

smoothdata is the dispatcher and the default first reach. sgolayfilt is the escape hatch for the two cases smoothdata cannot cover cleanly.

Situation	Reach for	Detail
General smoothing, no special story	`smoothdata(xd)` (default `"movmean"`) or `smoothdata(xd, "gaussian")`	fn-smoothdata.md
Outliers might remain (belt-and-braces)	`smoothdata(xd, "movmedian", win)`	fn-smoothdata.md
Peak shapes / derivatives matter (spectra, chromatograms, transients), degree 2 fine	`smoothdata(xd, "sgolay", fl)`	fn-smoothdata.md
Peak shapes matter and you need degree 3-4, or the FIR/weighted form	`sgolayfilt(xd, m, fl)`	fn-sgolayfilt.md
Named cutoff frequency in Hz	`lowpass(xd, fc, fs)`	base SPT filter; use when a Hz cutoff separates noise from signal
Non-stationary, multi-scale noise a tuned `sgolayfilt` cannot remove without blurring features	`wdenoise(xd, ...)` (Wavelet Toolbox)	escalation only; wf-denoise.md

sgolayfilt has no defaults - both m (order) and fl (odd frame length, >= 2*m+1) are required. If you find yourself wanting the defaults, you want smoothdata.

y = smoothdata(xd, "gaussian");        % default first reach
% peaks matter and degree 2 is not enough:
y = sgolayfilt(xd, 3, 11);             % order 3 over an 11-sample frame

Worked end-to-end example

% Sparse spikes + smooth cubic drift + broadband wiggle, no named cutoff.
xc = filloutliers(x, "pchip", "movmedian", 25);   % 1. spikes gone, clean fill
xd = detrend(xc, 3);                              % 2. cubic drift removed exactly
y  = smoothdata(xd, "gaussian");                  % 3. residual noise smoothed

Next in the chain

Fill NaN gaps left after outlier removal, before detrending, if any step introduced or exposed missing samples -> fn-fillmissing.md / fn-fillgaps.md.
Hand-off to trainnet / dlarray once the channel is conditioned -> wf-handoff-to-dl.md.

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Security

No issues found by our static scan· No issues found

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by

MathWorks

Publisher

MathWorks (Official)

Source

matlab/matlab-agentic-toolkit · skills-catalog/signal-processing/matlab-prepare-signal-data · commit 2a7f215

Licence

BSD 2-Clause License (redistribution allowed)

LICENSE (added by AutoGineer from LICENSE.md)
ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-09

AutoGineer is not affiliated with or endorsed by MathWorks or MathWorks.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download matlab-prepare-signal-data.zip