Day3_​S1_​ok

1 · Sensor fault

A dead transmitter writes a sentinel number, not a blank.

Step 1 · Read, flag, and blank the impossible values

1. Read the minute file with the CSV Reader, parse the stamp with String to Date&Time: 173,549 rows
2. One Expression node raises flag_range FIRST, then REPLACES each impossible cell with a missing cell: temp outside 20-120 C, negative flow or vibration
3. 12 rows carry flag_range = 1 and 0 rows are deleted - the other two tags in that row are still good data
4. The flag is what makes those 12 countable in the DQ report; blanking alone made them vanish

2 · The same minute written twice

The later write is the corrected value.

Step 2 · Keep the corrected write

1. Sorter by time, then Duplicate Row Filter keeps the LAST row: 173,549 rows become 173,449
2. One Expression marks the minute before the filter - measured on this machine, KNIME 5.11 writes no status column when it removes
3. 100 minutes carry had_dup = 1 all the way to the DQ report

3 · One time axis

Rows are not one minute apart, and the historian never writes an empty row.

Step 3 · Profile the gaps, then put every minute on the axis

1. Carry the previous stamp with the Expression node, measure it with Date&Time Difference, read the Histogram: 301 gaps of 2 minutes or more
2. 2 min x172, 3 min x64, 4 min x32, 5 min x30 - then nothing at all until 26, 71 and 181, the three real outages
3. Two cleanly separated populations, and that is what justifies cutting sparse from outage at 5 minutes
4. Date&Time Range Creator builds the 174,240 minute calendar, Joiner LEFT with the calendar on the left leaves 791 minutes unmatched, then Sorter - the Joiner returns ARBITRARY order

4 · A sensor that stopped moving

The value still arrives, it just never changes.

Step 4 · Catch the stuck tag on the time axis

1. Minute-to-minute change of all three tags with the Expression node
2. Sum and Count over a 10 row backward window with the Moving Aggregator
3. On the axis 10 rows is exactly 10 minutes, and a window over a hole has Count under 10 so it is skipped
4. Sum zero on a full window is stuck: 190 minutes, every one of them flow

5 · Count what was lost, and why

Every minute leaves this step with a label.

Step 5 · Label every minute and report

1. One Expression builds cov - how many of the three tags the minute really holds - and a 0/1 presence column
2. Two Moving Aggregator windows of 6 rows on that column, one backward and one forward: is there a real reading within 5 minutes on each side?
3. An empty minute with a real neighbour on BOTH sides is sparse, the logger simply wrote on change; with none it is outage
4. ok 173,214 - sparse 516 - outage 275 - frozen 190 - low_cov 45 = 174,240, and GroupBy on quality, had_dup AND flag_range writes 7 rows

6 · Hand over to Session 2

Only the hours you can defend leave this workflow.

Step 6 · Roll up to hours and write the file

1. Round to the hour with the Time Rounder, keep the 173,214 ok minutes with the Row Filter - nothing is imputed anywhere in this workflow
2. Group to 2,898 hours with GroupBy, rename for Session 2 with the Column Renamer
3. Read the Line Plot of minutes per hour: at 30 real minutes the hourly figure is off by at most 0.042 C, 0.29 m3/h and 0.033 mm/s against the pre-deadband truth, inside the per-minute noise of the tags
4. A second Row Filter at 30 on all three tag counts writes 2,895 rows - the thinnest tag count left in the file is exactly 30