Convert a MAT file to CSV

How to convert a .mat file to CSV

1. 1Open your .mat file. Drop the file onto the tool at the top of this page, or pick it with “Choose a file”. Every MAT-file version from v4 to v7.3 works. Your signals are found and grouped into tables, one per time base, and the first one is plotted straight away.
2. 2Pick what to convert. For the whole table, skip this step. For a few signals, tick them in the Channels list (up to eight per chart), switch to another table from the menu in the top bar, and zoom into a range if you only need part of the recording.
3. 3Download the CSV. Open “Do more”. “Download all columns (CSV)” saves every column and row of the table, time first; for timeseries and Simulink Datasets, “Download as wide CSV” with “All signals” saves every signal side by side on one time axis. “Download chart data (CSV)” saves just the ticked signals, for the range in view.

No .mat file to hand? Try a sample MATLAB file in the tool: a made-up Simulink-style drive log with two tables, ready to export. Or download a sample MATLAB MAT file: a Simulink simulation of made-up UK weather.

What the CSV looks like

One header row with the signal names, then one row per sample, with the time in the first column. These are the first rows from a timetable saved in MATLAB, with two signals ticked:

Time [s],Leeds_Humidity,Bristol_Pressure
0,80,1010
10,79.9383350853889,1010.1986693308
20,79.7541008538945,1010.38941834231
30,79.4495694631474,1010.56464247339

Numbers have 15 significant digits, the same precision as MATLAB’s own CSV export (writetable) and Excel; tick “Full precision (17 digits)” under “Do more” for every digit of each number. The table on screen rounds to ten. A duration’s seconds name their unit in the header (Time [s]), a single-precision signal is written in its own shortest form, and a signal you renamed on the chart keeps its new name in the header. “Download chart data (CSV)” covers the range in view: zoom out fully for the whole recording, or zoom in for a single event; “Download all columns (CSV)” always takes every row.

Signals logged at different rates

Simulink logs and timeseries often hold signals at different rates. On one chart each keeps its own time axis, and the CSV keeps every sample: one row for every time of any plotted signal, with a signal’s cell left empty where it has no sample.

Time,Lone_Signal,Leeds_Humidity
0,0,80
0.01,0.187381314585725,
0.02,0.368124552684678,

To get one value per row for every signal, choose Download as wide CSV under “Do more”. Pick the signals on the chart or all of them (the whole file), the time axis (every time of any signal, where it starts when the signals have different times, so that no sample is lost; the times of the first signal; or a fixed interval such as every 0.1 s) and how values between samples are filled: the nearest sample or linear interpolation. The dialog says how many rows the CSV will have, and names the signals a time axis would leave samples out of. Outside a signal’s own time range its cells stay empty.

How your variables become CSV columns

The common complaint about converting a .mat file is that a struct ends up as one cell instead of one column per field. Here every numeric variable and field becomes a column of its own, and the time comes first:

• Several variables. Vectors of the same length line up as columns of one table, next to the time vector that matches them (t, time, tout or a name like t_rain), which becomes the first column. Variables of another length get a table of their own: pick it from the menu in the top bar.
• Structs, nested ones too. Every numeric field becomes a column named by its path inside the struct, so weather.leeds.humidity becomes the column leeds.humidity. Fields of the same length share a table, like separate variables; text fields are listed in the notes instead.
• Matrices. One column per matrix column, named from the list of channel names saved with it, where there is one.
• Timeseries, on their own, in structs, in a tscollection or in a logged Simulink Dataset: each becomes a column named after the signal, with its own time; a timeseries with several columns becomes name(1), name(2). Signals that share a time vector are also offered as one table.
• Simulink simulation output saved as it is (out from sim), and Simulation Data Inspector exports: logsout, yout, xout, tout and every To Workspace variable become the same tables as when you save them on their own; each Inspector run gets tables of its own.
• Tables and timetables. One column per variable, with its unit; a timetable’s row times become the Time column.
• Simulink To Workspace output (Structure with time, Structure, Array): each block’s signals become a table, with their labels as column names.

To convert a whole table, choose Download all columns (CSV) under “Do more”: every column, every row. Timeseries, including those in a Dataset, all go into one wide CSV (“All signals”), up to 200 at once.

Which .mat files it reads

MATLAB v4, v6, v7 and v7.3, including v7.3 files of several GB, and files written by Python. Signals are found in Simulink simulation output saved as it is, Simulation Data Inspector exports, Simulink To Workspace output (Structure with time, Structure, Array and Timeseries formats), logged Datasets (v7 and v7.3), tscollections, MATLAB tables and timetables, dSPACE-style measurement files, structs, matrices with channel names and data-logger exports with a time matrix and a value matrix. Dates and times (in the time zone saved with them), durations (as seconds in the CSV), strings and categories are read too. Objects the viewer can’t read yet, such as containers.Map and your own classes, are listed in the notes and the rest of the file opens. The MAT file viewer page has the full list.

Good to know

• A chart holds up to eight signals, so for more use “Download all columns (CSV)” for a table, or the wide CSV for timeseries (up to 200 signals).
• Between two signals logged at different rates, the wide CSV fills a value only within each signal’s own time range: before its first sample and after its last, its cells stay empty.
• Excel opens at most 1,048,576 rows. For a longer recording, zoom into the part you need first, or pick a coarser interval in the wide CSV.
• Your .mat file is never changed: the tool only reads it.

MAT to Excel

Excel can’t open a .mat file, so convert it to CSV first. To look at the signals without Excel, plot them in the MAT file viewer.

Open it in Excel

MegaRows writes CSV, not .xlsx: there is no direct .xlsx download, and Excel opens the CSV directly. To check the separator first, use Data › From Text/CSV in Excel.

• One table per CSV. A header row with the signal names, then one row per sample, the time first. Signals of another length or time base are a table of their own: pick it from the menu in the top bar and download it too.
• Signals at different rates. “Download chart data (CSV)” keeps each signal’s own samples and leaves a cell empty where a signal has none. “Download as wide CSV” puts timeseries and Simulink logs on one time axis, with the nearest sample or linear interpolation in between.
• Excel’s row limit. A worksheet holds 1,048,576 rows, the header row included. For a longer recording, zoom to the part you need before “Download chart data (CSV)”, split the CSV into Excel-sized files with the CSV splitter, or keep the whole recording in the MAT file viewer.
• Saving as .xlsx. If you need a workbook, open the CSV in Excel, then use File › Save As and choose Excel Workbook (.xlsx).

Other ways to convert .mat to CSV

In MATLAB

With MATLAB at hand, writetable and writematrix do the job in a line:

% A table or timetable, with its column names as the header
writetable(timetable2table(TT), "weather.csv")

% Vectors with a time vector, as columns
writematrix([t(:), temperature(:), humidity(:)], "weather.csv")

In Python

SciPy reads v4 to v7 files, and pandas writes the CSV:

import pandas as pd
from scipy.io import loadmat

data = loadmat("weather.mat", simplify_cells=True)   # v4 to v7 only
df = pd.DataFrame({"time": data["t"], "temperature": data["temperature"]})
df.to_csv("weather.csv", index=False)

On a v7.3 file scipy.io.loadmat stops with NotImplementedError (“Please use HDF reader for matlab v7.3 files”), which is why code written with loadmat fails on files saved with -v7.3. The mat73 package reads them (h5py does too, with the arrays transposed):

import mat73                 # pip install mat73: reads v7.3 files
import pandas as pd

data = mat73.loadmat("weather_v73.mat")
df = pd.DataFrame({"time": data["t"], "temperature": data["temperature"]})
df.to_csv("weather.csv", index=False)

In GNU Octave

Octave’s load reads v7 and older files, and csvwrite writes the numbers (without a header row):

s = load("weather.mat");                          % v7 and older
csvwrite("weather.csv", [s.t(:), s.temperature(:)])

Our guide on opening a .mat file without MATLAB covers these in more detail, with tested code for every MAT-file version.

Going the other way? Use the CSV to MAT converter, or see how to convert CSV to a .mat file in MATLAB, Python or Octave.

Private by design: nothing is uploaded

Your browser reads the .mat file straight from your disk and writes the CSV on your device: the file and its name never leave it, which you can check in the Network tab of your browser’s developer tools. Read the privacy statement for details.