Learning center

CWRU spectra: measured inputs and complete numerical exports

Explore a bearing-vibration recording, interpret its frequency spectrum and understand the values saved with the result. The example covers selected measurements, not fault diagnosis or prognosis.

Explore a bearing-vibration recording, interpret its frequency spectrum and understand the values saved with the result. The example covers selected measurements, not fault diagnosis or prognosis.

1. Preserve the measured input and sampling rate

The source is Case Western Reserve University Bearing Data Center record 106 / IR007_1: a 0.007-inch inner-race defect, 1 HP load and recorded 1772 rpm. The original MAT contains 121,991 samples in each of three measured channels: drive end, fan end and base. The prepared NPZ retains all 365,973 values in their original order, with an explicit 12,000 Hz sampling rate. The recorded scope is about 10.166 seconds, with no launcher resampling or normalization.

Load the import-ready NPZ, select all three channels and the complete recording, then inspect the input contract before running Frequency-Domain Analysis. Choose channel 1, the drive-end measurement, as primary. Keep the original MAT, conversion provenance, file hashes and sampling metadata alongside your results. A filename alone does not establish sample identity. The website does not redistribute this raw recording or assert a new redistribution license.

2. Reproduce the recorded settings

Set Window: hann, Detrend: constant, Segment samples: 1024, Overlap: 50% and Amplitude unit: source units. The input contract should report 512 overlap samples, 121,991 rows and uniform sampling at 12,000 Hz. The earlier narrated tutorial may show different segment settings; use this continuous workflow’s configuration when comparing its exported values.

Mean removal applies once to the complete primary channel. Welch then uses periodic Hann windows with no additional per-segment detrending. There are 1 + floor((121991 − 1024) / 512) = 237 complete segments. An incomplete tail does not become another Welch segment. The full-record FFT and the segmented Welch estimator use different grids and normalizations; selecting Hann does not turn the full-record amplitude table into a Welch table.

3. Read amplitude and density on their own scales

ResultRecorded numerical contractInterpretation
FFT Numeric Spectrum60,996 rows; frequency spacing 12000 / 121991 ≈ 0.098368 Hz; one-sided peak amplitude in source units.The complete mean-removed primary channel is transformed. Positive-frequency amplitudes are doubled; DC is not. This odd-length record has no exact Nyquist bin. For even lengths the Nyquist amplitude is not doubled.
Welch Numeric Spectrum513 rows; spacing 12000 / 1024 = 11.71875 Hz; density in source units²/Hz, including 0 and 6000 Hz.Average the complete periodic-Hann segments. Density is power per unit frequency, not peak amplitude. A taller density bin is not directly comparable to an FFT amplitude.

The Welch plot’s nominal pointwise interval uses 2K = 474 degrees of freedom. It does not correct for correlation between overlapping segments. Do not read it as a calibrated diagnostic interval. The recorded summary gives integrated PSD about 0.0857626 and record mean square about 0.0857522; windowing, averaging and numerical integration need not yield exact equality. Use complete numerical tables for precision; descriptive summary values are rounded.

The seven recorded routes also produce spectral concentration, cepstrum, harmonic, Parseval and spectrogram views. These help inspect measured structure, but the independent receipt covers selected FFT/Welch calculations and energy, not every route. Neither the largest peak nor a visually repeating pattern establishes a bearing-fault diagnosis.

4. Read the complete tables beyond the viewer limit

The recorded viewer displays the first 20,000 FFT rows. Use the native ZIP export to obtain all 60,996 bins; copying visible rows or saving a screenshot is incomplete. In this saved package, tables/002.csv and tables/002.json contain FFT frequency and peak amplitude; tables/004.csv and tables/004.json contain all 513 Welch frequency/density rows. Consult the manifest’s table titles for other runs instead of assuming their numbering is fixed.

Check row counts excluding CSV headers, inspect the final frequency row, and compare CSV values with the corresponding JSON rows. The guide check reopens this existing archive and verifies every numeric row in both spectra. It does not rerun the scientific application or extend the scope of the recorded 25 reference checks.

5. Save a result snapshot with its limits

  1. Review the input contract, settings, figures and numeric tables before exporting. Preserve the selected channel and unit labels.
  2. Save the figures, summary and native ZIP package. The recorded package has seven figures and ten tables, each table supplied as CSV and JSON. Its internal manifest.json identifies table row counts, member hashes and snapshot limits.
  3. Reopen the ZIP and verify the hashes of its figures, tables and summary against the internal manifest. The separate artifact_manifest.json from this walkthrough verifies nine outer files: seven PNGs, the summary and ZIP. That outer checksum file was produced by the walkthrough verifier, not by a promised native feature.
  4. Keep the original input and settings. The native manifest explicitly says reloadable_session: false and numeric_plot_arrays_included: false. Complete table exports are available, but the ZIP is a result snapshot, not a reloadable analysis session or a replacement for all plot source arrays.
  5. Open a PNG and the numerical tables outside the product to confirm files are usable. Review identifiers before sharing; the manifest does not claim automatic identifier scrubbing or publication.

Worked examples describe particular inputs, methods and recorded results. Check current availability and input requirements before planning an evaluation.

← Back to the learning center