Plant data assessment — TEP-Plant-10days-AI

Tennessee Eastman benchmark process · period analysed 2026-04-01 00:00 to 2026-04-10 23:57 · 41 tags · prepared by Yahya Nazer, CTRL Designer LLC · 11 September 2026

In short. 3 cause(s) are worth acting on out of 7 the review assembled from 11 statistical findings. The highest priority is xmeas_9 - Reactor temperature - deg C is cycling (7 min) (2026-04-02 01:45 to 2026-04-08 00:00). 4 cause(s) did not survive review and are listed openly in section 2. 7 question(s) need the plant to answer before any of this is certain.

1. What we found

xmeas_9 - Reactor temperature - deg C is cycling (7 min)supported · 75%

2026-04-02 01:45 to 2026-04-08 00:00 · severity 9.1 · event, oscillation

What the data shows. Oscillation index 0.10 against 0.02 in the reference window (5.1x). Variability 5.4x the reference.

Why we believe it. cycling 5.1x stronger than in the reference window - this is new behaviour; a fast, regular cycle is the signature of a control-loop problem - a sticking valve or aggressive tuning.

What to do. Operations review - what changed in the plant over this window; Loop check - stroke the valve, review the tuning, try the loop in manual. Pull this tag at 1-minute or faster to pin the period down.; Check the Step A availability map over this window; re-run Step C without those tags if the overlap is large.; Confirm with the operators whether anything was done to the plant in this window.

the tags behind this

Tag name	AI/AO	Description	Engineering unit
xmeas_9	AI	Reactor temperature	deg C
xmeas_10	AI	Purge rate (stream 9)	kscmh
xmeas_16	AI	Stripper pressure	kPa gauge
xmeas_2	AI	D feed (stream 2)	kg/h
xmeas_21	AI	Reactor cooling water outlet temperature	deg C
xmeas_27	AI	Reactor feed composition E	mol%
xmeas_29	AI	Purge gas composition A (stream 9)	mol%
xmeas_35	AI	Purge gas composition G	mol%

The plant left its normal envelope for 38.2 hsupported · 90%

2026-04-09 09:48 to 2026-04-11 00:00 · severity 8.9 · event, shared period

What the data shows. SPE reached 908 and T2 1337 against limits from the reference window. Most of the deviation sits in: xmeas_38, xmeas_16, xmeas_33, xmeas_1, xmeas_20.

Why we believe it. the plant went 91x past the reference limit - far outside anything seen in the normal window; it held for 38.2 h - far too long to be noise or a single bad sample; 5 tags moved together, so it reads as a plant event rather than one instrument.

What to keep in mind. 1 of the 5 named tags were repaired by Step A (held (not interpolated)): xmeas_38.

What to do. Operations review - what changed in the plant over this window; Trace the common cause - a shared utility (steam, cooling water, feed) usually carries it. Compare raw against corrected for those tags in the Step A fixed-tags report before acting.; Compare raw against corrected for those tags in the Step A fixed-tags report before acting; Look at the reference window before concluding anything: it may not be a quiet period; Look upstream: a shared utility (steam, cooling water, feed) is the usual carrier.

the tags behind this

Tag name	AI/AO	Description	Engineering unit
xmeas_1	AI	A feed (stream 1)	kscmh
xmeas_16	AI	Stripper pressure	kPa gauge
xmeas_20	AI	Compressor work	kW
xmeas_33	AI	Purge gas composition E	mol%
xmeas_38	AI	Product composition E	mol%
xmeas_10	AI	Purge rate (stream 9)	kscmh
xmeas_11	AI	Product separator temperature	deg C
xmeas_13	AI	Product separator pressure	kPa gauge
xmeas_22	AI	Separator cooling water outlet temperature	deg C
xmeas_7	AI	Reactor pressure	kPa gauge

xmeas_21 - Reactor cooling water outlet temperature - deg C is cycling (7 min)supported · 75%

whole period · severity 8.5 · oscillation

What the data shows. Oscillation index 0.10 against 0.02 in the reference window (4.5x). Variability 4.7x the reference.

Why we believe it. cycling 4.5x stronger than in the reference window - this is new behaviour; a fast, regular cycle is the signature of a control-loop problem - a sticking valve or aggressive tuning.

What to do. Loop check - stroke the valve, review the tuning, try the loop in manual. Pull this tag at 1-minute or faster to pin the period down.

the tags behind this

Tag name	AI/AO	Description	Engineering unit
xmeas_21	AI	Reactor cooling water outlet temperature	deg C

2. What we did not conclude

These came out of the analysis but did not survive review. They are listed so that nothing is hidden, and so nobody re-runs the same study expecting a different answer.

verdict	what it claimed	why we set it aside
likely artefact · 30%	The plant left its normal envelope for 0.1 h 2026-04-06 00:51 to 2026-04-06 00:57	1 of the 5 named tags were repaired by Step A (held (not interpolated)): xmeas_38; only 1.2x past the limit - close enough to the boundary that the limit itself could explain it; only 6 min - short enough to be one upset or a data glitch
weak · 38%	The plant left its normal envelope for 0.2 h 2026-04-01 18:42 to 2026-04-01 18:54	only 1.2x past the limit - close enough to the boundary that the limit itself could explain it; only 12 min - short enough to be one upset or a data glitch
weak · 38%	The plant left its normal envelope for 0.1 h 2026-04-02 18:36 to 2026-04-02 18:42	only 1.2x past the limit - close enough to the boundary that the limit itself could explain it; only 6 min - short enough to be one upset or a data glitch
weak · 38%	The plant left its normal envelope for 0.1 h 2026-04-06 18:36 to 2026-04-06 18:42	only 1.2x past the limit - close enough to the boundary that the limit itself could explain it; only 6 min - short enough to be one upset or a data glitch

3. What to do

priority	#	cause	action	tags
High	1	xmeas_9 - Reactor temperature - deg C is cycling (7 min) 2026-04-02 01:45 to 2026-04-08 00:00	Operations review - what changed in the plant over this window; Loop check - stroke the valve, review the tuning, try the loop in manual	xmeas_9, xmeas_10, xmeas_16, xmeas_2, xmeas_21, xmeas_27, xmeas_29, xmeas_35
High	2	The plant left its normal envelope for 38.2 h 2026-04-09 09:48 to 2026-04-11 00:00	Operations review - what changed in the plant over this window; Trace the common cause - a shared utility (steam, cooling water, feed) usually carries it	xmeas_1, xmeas_16, xmeas_20, xmeas_33, xmeas_38, xmeas_10, xmeas_11, xmeas_13, xmeas_22, xmeas_7
High	3	xmeas_21 - Reactor cooling water outlet temperature - deg C is cycling (7 min) whole period	Loop check - stroke the valve, review the tuning, try the loop in manual	xmeas_21
Low	4	The plant left its normal envelope for 0.2 h 2026-04-01 18:42 to 2026-04-01 18:54	Operations review - what changed in the plant over this window	xmeas_11, xmeas_25, xmeas_31, xmeas_33, xmeas_35
Low	5	The plant left its normal envelope for 0.1 h 2026-04-02 18:36 to 2026-04-02 18:42	Operations review - what changed in the plant over this window	xmeas_1, xmeas_25, xmeas_27, xmeas_33, xmeas_35
Low	6	The plant left its normal envelope for 0.1 h 2026-04-06 18:36 to 2026-04-06 18:42	Operations review - what changed in the plant over this window	xmeas_1, xmeas_25, xmeas_27, xmeas_33, xmeas_35

4. What we need from the plant

Nothing above is certain until these are answered. They are the fastest way to turn a statistical result into a decision.

The reference window is 2026-04-03 12:00:00 to 2026-04-05 11:57:00 (47.95 h, 23.4 samples per tag). Was the plant genuinely running normally then? Everything called abnormal is measured against it.
xmeas_9 - Reactor temperature - deg C is cycling (7 min) (2026-04-02 01:45 to 2026-04-08 00:00) — what happened here? The tags carrying it are xmeas_9 - Reactor temperature - deg C; xmeas_10 - Purge rate (stream 9) - kscmh; xmeas_16 - Stripper pressure - kPa gauge. Was there maintenance, a grade change, a setpoint move or a trip?
The plant left its normal envelope for 38.2 h (2026-04-09 09:48 to 2026-04-11 00:00) — what happened here? The tags carrying it are xmeas_1 - A feed (stream 1) - kscmh; xmeas_16 - Stripper pressure - kPa gauge; xmeas_20 - Compressor work - kW. Was there maintenance, a grade change, a setpoint move or a trip?
xmeas_21 - Reactor cooling water outlet temperature - deg C is cycling (7 min) (whole period) — what happened here? The tags carrying it are xmeas_21 - Reactor cooling water outlet temperature - deg C. Was there maintenance, a grade change, a setpoint move or a trip?
2 loop(s) cycle with a short period. When were they last tuned, and has the valve been stroked or serviced recently? Can one be put in manual briefly to confirm?
Step A repaired 7 things in the data (gaps<=5 interpolate; slow tags held; frozen unknown; spikes flag). Are the frozen or missing stretches known instrument problems, or was the plant simply steady?
What matters to you — throughput, energy, quality or reliability? The ranking above is by statistical size, not by value; say which and the next pass can rank by money.

5. What happens next

A study either closes here or returns to Step B with a new scope. For this one:

Re-run without the repaired tags. 3 finding(s) rest partly on tags Step A had to fix. Excluding them in Step B and re-running C and D shows whether the finding survives on clean tags alone.
Drop 1 cause(s) the review rejected from the next scope so they stop consuming attention: "The plant left its normal envelope for 0.1 h"

6. How this was produced

Five steps, each writing its own report, so any number here can be traced back to the raw export.

	step	what it does	program
A	Prepare	Raw historian exports become one clean, regular Data table. Merge files, normalise timestamps, replace bad-value strings, put on a regular grid, fill short gaps, apply the repair rules, and record every change.	step_a_prepare.py + A-Prepare-GUI-v01.py
B	Explore	Trend and cluster the clean data. The engineer chooses which tags, which period, and which window counts as normal, and writes it down as the scope.	step_b_explore.py + B-Explore-GUI-v01.py
C	Analyze	Operability (PCA, T2/SPE, events), oscillation and slowness per tag, control-loop metrics, batch phases - all driven by the scope from B.	step_c_analyze.py + C-Analyze-GUI-v01.py
D	AI review	An AI reads the results and the scope and drafts findings, anomalies and questions for the engineer.	step_d_review.py + D-Review-GUI-v01.py
E	Report	The engineer edits, decides, and either closes the study or returns to B with a new scope.	to build

The data behind it

Raw export	/Users/yahyanazer/Dropbox/__C_2026_Work/CTRL-Designer/B-CDL-Data-Analysis/A-Raw-Data/TEP/tep_plant_10days.csv
Glossary	tep_schema.csv
Period	2026-04-01 00:00 to 2026-04-10 23:57
Sampling	3 minutes
Tags	41 analysed of 41 available
Reference window	2026-04-03 12:00 to 2026-04-05 11:57 (47.95 h, 23.4 samples per tag)
Repairs made	7 (gaps<=5 interpolate; slow tags held; frozen unknown; spikes flag)
Model	29 components holding 91% of the variation
Programs	step_a_prepare.py 2026-09-10 10:49:28 · step_b_explore.py 2026-09-10 19:50:26 · step_c_analyze.py 2026-09-10 19:35:16 · step_d_review.py 2026-09-10 10:49:28 · step_e_report.py 2026-09-10 10:49:28

Data quality recorded in Step A

issue	count
spike	114
missing	41
slow update	6
bad values	1
frozen	1

7. Where the detail lives

Step A, data preparation and every repair: A-Data/TEP-Plant-10days-AI/overall_report.html
Step B, clusters and the reference window: B-Scope/TEP-Plant-10days-AI/explore_report.html
Step C, the analysis and all findings: C-Analysis/TEP-Plant-10days-AI/analysis_report.html
Step D, every finding challenged: D-Review/TEP-Plant-10days-AI/review_report.html
This report, its charts and the action list: /Volumes/YN-4T/WIP/CTRL-Designer/B-CDL-Data-Analysis/E-Report/TEP-Plant-10days-AI

Charts in this document are live: drag to zoom, double-click to reset, click a name in the legend to hide it. They are copied into charts/ beside this file, so the whole folder can be zipped and sent as one piece.