Learning center

Anomaly scores: baseline scope, candidate flags and complete exports

Separate a chosen statistical baseline from healthy truth, read every Z-score row and compare complete native/server exports with current-source limits.

A written guide to one measured-record Z-score workflow. Engineering Signal Processing anomaly_detection produces method-dependent candidates. The example here has complete saved numeric exports and an earlier independent arithmetic check. It does not have per-sample fault truth. In the 9 November 2026 source review, the native app, base window and viewer matched the accepted payload; the developer and historical server identities were recorded separately. Source descriptions in this example refer to that reviewed snapshot.

1. Preserve the input and identify the scored column

The saved example uses CWRU record 106 from the official 12 kHz drive-end bearing fault collection. The reviewed MATLAB input explicitly supplies data, sample_rate_hz and channel_names. Its 121,991 × 3 matrix preserves all 365,973 source waveform values, ordered as X106_DE_time, X106_FE_time and X106_BA_time.

The supplied sampling rate is 12,000 Hz, documented in the preparation receipt. The original MAT file has no sample-rate field; the accompanying RPM value is not a sampling frequency. The saved workflow used the explicit waveform representation rather than silently treating the scalar RPM as a fourth signal. Preserve both original and prepared identities. This preparation record does not establish that every original MATLAB structure is accepted by current intake.

Load the measured matrix and retain all three channels through the launcher’s channel selection. Set Primary column (1-based) to 1. The Z-score route operates on this primary waveform; the presence of three selected feature columns does not make it a three-channel Z-score aggregate. In this input, column 1 is the drive-end waveform. The saved input-QC table uses generic column labels, so keep the preparation receipt’s original channel mapping.

Use Time column (0=none) = 0 and Sampling Hz = 12000 for this representation. Time is derived from sample index divided by sampling frequency. Preserve the amplitude label source units; the export does not establish a calibrated physical amplitude unit. The standardized score itself is dimensionless.

2. Declare the statistical reference and decision scope

SettingSaved value and consequence
Selected routeZ-Score / zscore. One of twenty choices; other methods are outside this example.
Feature representationuploaded_columns. Keep the uploaded measurement columns; Z-score still uses the primary column.
Phase-I baseline fraction0.25. First floor(121,991 × 0.25) = 30,497 samples.
Detection scopepost_baseline. Only the following 91,494 samples can receive candidate flags.
Z / modified-Z limitsZ limit 3.0. Flags require scores strictly greater than 3, within the evaluation scope.
Known-label column0 / not supplied. No precision, recall or fault-detection accuracy follows from this run.

Choose only Z-Score for this workflow. The first quarter is a chosen statistical reference, not a verified healthy operating period: the source recording belongs to a fault-labelled collection. The current method estimates a mean and sample standard deviation from that reference. It needs a varying signal and a nonzero baseline spread; selected NaN or infinite values are refused rather than silently filled or removed.

The interface also exposes Expected anomaly fraction, significance and other method-specific controls. The saved Input Contract records an expected fraction of 0.01, but the Z-score decision uses the explicit threshold of 3. It does not select exactly one percent of rows. Changing a requested fraction does not create ground-truth labels or establish a fault rate.

3. Read the score and the rounded summary separately

For this route, the score is abs(x − baseline mean) / baseline sample SD. The sample SD uses the n − 1 convention. A score of 3 means a departure of three baseline sample standard deviations under this calculation; it is not a probability that the bearing is faulty.

The saved independent reference mean is 0.005928293683152107 and sample SD is 0.29154642636783734, in source amplitude units. The summary table shows 0.005928294 and 0.2915464 because its formatting uses seven significant digits. Those strings are presentation summaries. Use the complete Numeric Data table for exported score precision; do not recompute scores from rounded summary strings and expect the same low-order digits.

The earlier independent check compared all 121,991 scores against scalar arithmetic without the product’s numerical helper. Its maximum absolute difference was about 8.88 × 10−16, below the declared 10−12 absolute tolerance. This guide reopens the saved exports and their decisions; it does not refit the baseline or calculate new scores from the waveform. Numerical agreement for one method is separate from detection performance.

4. Use the full six-column decision ledger

ColumnHow to read it
source_sample_indexZero-based row in the loaded record, from 0 through 121,990
source_time_coordinateSeconds, equal to index / 12,000 for this input
scoreAbsolute standardized departure; dimensionless
decision_threshold3 for every saved row in this method
evaluation_allowed0 in baseline rows 0–30,496; 1 from row 30,497 onward
candidate_flag1 only when evaluation is allowed and the score exceeds the threshold

The full ledger includes scores for baseline rows even though those rows cannot be flagged under post_baseline. A baseline flag of zero therefore does not demonstrate that the sample is normal. The first eligible sample is index 30,497 at approximately 2.541416667 seconds. The last saved sample is index 121,990 at approximately 10.165833333 seconds. The time of the last sample is not the same quantity as sample count divided by sampling rate.

There are 1,575 candidates among 91,494 eligible rows, approximately 1.7214%. The denominator is the selected evaluation population, not all 121,991 rows. The result also states that known labels were not supplied. Do not turn this fraction into a failure rate, diagnostic accuracy or percentage of confirmed faults.

5. Save every row, then verify the package boundary

  1. After the selected route completes, use Open Viewer and inspect the Input Contract, Z-Score Detection summary and Z-Score Detection Numeric Data table. The viewer’s visible table is limited to the first 20,000 rows. This preview ends before the evaluation region begins in this example.
  2. Use Export Package to retain the full saved tables. The reviewed native ZIP has nine members: manifest, summary text, one PNG, and JSON/CSV pairs for three tables. Its Numeric Data CSV and JSON each contain all 121,991 rows × 6 columns. The standalone saved CSV is byte-identical to the package member.
  3. Use the manifest’s titles, row counts and hashes to locate and verify each member. The example numeric paths are tables/003.csv and tables/003.json; table order can differ for another selection. Read rows beyond 20,000 and the baseline/evaluation boundary rather than judging completeness from the visible grid.
  4. The separately saved server ZIP has eight members and represents the complete numeric table as six JSON column arrays. It does not supply the native CSV as a server export. All 731,946 numeric cells agree across native CSV, native JSON and server JSON in the saved artifacts. This is a cross-format check of existing outputs, not a fresh server execution.
  5. Keep the original input, preparation receipt and settings with the snapshot. The native manifest says reloadable_session: false, numeric_plot_arrays_included: false and scientific_validation_performed: false. Complete selected score tables do not turn this into a restorable analysis project or validated diagnosis.

The current score-export guard budgets six values per row plus overhead within a 1,000,000-value bound. Oversized output is refused without silently truncating rows. For this six-column route, the arithmetic ceiling is 166,664 score rows; other input and method limits still apply, and that ceiling is not a benchmark or guarantee for a larger dataset.

The package also declares that source-identifier scrubbing was not performed and content review is required. Inspect retained content before sharing it. Exporting a local package does not publish anything.

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

← Back to the learning center