Free sensor entropy audit
Every column in your file is profiled for entropy across the record, and the ones that changed most are flagged. Takes seconds, needs no account, and runs entirely in your browser — the file is never uploaded anywhere. Open your network tab and check.
| Channel | n | Entropy first third |
Entropy last third | Movement | Status |
|---|
What a flag means, and what it does not. A flag means one thing only: That is a fact about your data against a threshold you can change below.
It is not a diagnosis that the instrument is faulty. Entropy shifts for real reasons too — a process change, a different feedstock, a recalibration, a season. Cross-check a flagged channel against your own maintenance and calibration records first. If the dates line up, that is worth a conversation. If they do not, the flag is telling you something changed in the process, not the sensor.
Defaults: 0.10 absolute movement, minimum 60 points per third. Absolute rather than percentage on purpose — a channel whose entropy goes from 0.005 to 0.025 has "changed 400%" and means nothing. Both are arbitrary starting points, not thresholds derived from your plant. Move them and watch which channels come and go — if a channel flips between 0.09 and 0.11, it was never a strong flag.
This tool screens. It does not diagnose. If the flagged dates line up with something real in your records, send me what you found and I will go deeper on those channels.
Nothing is sent from this page. The button opens your own email with the summary — channel names and counts only, never your readings.
Each numeric column is converted to fractional changes between consecutive readings. Those changes are binned into five buckets at ±0.0005 and ±0.002, and Shannon entropy is computed over the bin distribution, normalised to 0–1. Low entropy means the channel is moving in a structured way; high entropy means it is behaving like noise. The record is split in thirds and the first third is compared against the last. Columns that are monotonic counters — timestamps, row indices, sample numbers — are detected and excluded, since their "entropy" describes the clock, not an instrument.
Those bin thresholds are not arbitrary — they were tested against 10× tighter, 100× tighter and 10× wider alternatives on public benchmark data, and performed best. Columns that are constant, non-numeric, or too short are skipped and reported as such rather than silently dropped. A channel marked "flatlined" is worth looking at — a sensor reporting one unchanging value across the whole record is either out of service or stuck, and neither shows up as an alarm.