FFT from CSV: free online FFT and PSD calculator

The peaks sit on the frequency steps nearest 50 Hz and 120 Hz, and they are the first two rows of the peak table next to the plot. A band around each tone gives that tone’s RMS, its amplitude divided by √2, give or take the noise in the band. The step at 2 seconds adds power below a few hertz: the rise at the left of the plot. The sample has no unit, so the PSD is per Hz; a signal in g gives a PSD in g²/Hz and a Grms in g.

The same in Python

With the sample file downloaded, these lines give SciPy’s answer for the same settings:

import numpy as np, pandas as pd
from scipy import signal

x = pd.read_csv("megarows-sample-signals.csv")["signal_a"].to_numpy()
f, p = signal.welch(x, fs=1000, window="hann", nperseg=1024, noverlap=512)
print(f"Grms {np.sqrt(np.trapezoid(p, f)):.4f}")
k, _ = signal.find_peaks(p)
for i in sorted(k[np.argsort(p[k])[-2:]]):
    print(f"peak {f[i]:.2f} Hz  PSD {p[i]:.3g}")

They print:

Grms 0.8786
peak 49.80 Hz  PSD 0.319
peak 120.12 Hz  PSD 0.0833

We ran them on 8 October 2026 with NumPy 2.1.3, SciPy 1.14.1, pandas 2.2.3 and Python 3.12. A test runs the same calculation through MegaRows on every change to the site and checks that its numbers are the ones on this page, and that they match SciPy’s to about 12 significant digits. Download the sample file (CSV, 185 KB).

FFT, PSD and Grms: what you get

FFT spectrum

The amplitude of every frequency in the whole signal, from 0 Hz up to half the sample rate. Choose peak amplitude (a 1 g sine shows as 1 g), RMS amplitude or power, and a window: rectangular, Hann, Hamming, Blackman or flat top. The mean or a straight-line trend can be removed first. By default every sample is used, which keeps amplitudes true; padding to the next power of 2, as some tools do, is there as an option.

Power spectral density (Welch)

The signal is cut into overlapping segments (by default 4,096 samples, fewer for a short file, with 50 % overlap), each is windowed and transformed, and the results are averaged. That smooths out the noise of a single spectrum, which is why a PSD is the usual way to look at vibration. The result is in unit²/Hz, or in dB.

Grms

The overall level of a random vibration: the square root of the area under the PSD. The result plot shows it for every frequency, and for any band you type in.

Peak table

Next to the plot, a table lists the spectrum’s highest peaks, highest first: the frequency of each and its value, and for a PSD the Grms of each peak, from the dip on its left to the dip on its right. It shows the top 5, or 10 or 20, and copies into a spreadsheet or downloads as CSV; the peaks are numbered on the plot, and a click on a row picks one out. A peak counts when it stands at least 3 dB above the dips either side, and two peaks are at least 3 frequency steps apart.

Sample rate and units

The sample rate comes from your time column when its steps are even, and the dialog says which column and what rate it found. A column of times in seconds, or of dates and times, both work. Without a time column, type the rate in. Units come from the channel, for example from a units row under the header: a signal in g gives a PSD in g²/Hz.

PSD for vibration: g, g²/Hz and Grms

Accelerometer data is usually in g. Give the column its unit in a units row under the header (g or [g]) and the PSD comes out in g²/Hz and the Grms in g, the units of a random-vibration test specification. The PSD uses every sample of the channel.

• The Grms of a band: type the band, such as 20 to 2,000 Hz, into Grms band in the dialog, or into the From and To boxes on the PSD plot. The Grms is the square root of the area under the PSD in that band.
• The highest frequency is half the sample rate. A file logged 1,000 times a second shows up to 500 Hz, so a band up to 2,000 Hz needs at least 4,000 readings a second.
• Frequency step and smoothness: the step is the sample rate divided by the segment length (1,000 ÷ 4,096 = 0.24 Hz). Longer segments give finer steps but fewer segments to average, so a noisier PSD; shorter segments give the opposite.
• Tones in random vibration: a sharp peak is a tone, such as a motor’s running speed. Its Grms over a narrow band is the tone’s RMS, as in the worked example above.

Checked against SciPy

• FFT spectrum and PSD: 55 checks across every window, scaling and option. The largest difference was 1.5 × 10⁻¹⁴ of the size of the result: the numbers agree to about 13 significant digits.
• Every operation: in 341 of the 526 checks the numbers are identical to the last digit. The largest difference anywhere was 1.2 × 10⁻¹¹ of the size of the result, far below anything you could see in a plot or a table.
• Tolerances: a check passes when the difference is at most 10⁻¹² of the result for spectra, sums and statistics (less for single functions such as log or sin), and 10⁻¹⁰ for the filters, which design their own coefficients.

A few results follow a different convention from the library default on purpose, such as rounding 2.5 up to 3 like Excel; the comparison uses the same convention on both sides. How we check our maths lists every check, with the reference call and the largest difference.

Private by design: nothing is uploaded

Test data is often confidential. This tool reads your file and calculates the spectrum in your browser, so the file and its name never leave your device; you can check it in the Network tab of your browser’s developer tools while the file opens. Read the privacy statement for details.

Need the signal itself rather than its spectrum? The CSV plotter opens the same files with zoomable charts and statistics, and our guide shows how to plot a CSV file with millions of rows.