Same Fault, Two Valves

I wanted to see, with real numbers, how much of a plant's trouble shows up in its recorded data. So I ran an experiment on a public benchmark: a simulated chemical plant that process-control researchers have worked with since 1993. It records 41 measurements and 11 controller outputs every three minutes.

I gave it the same fault in two places: a cooling-water valve that sticks, once on the reactor and once on the condenser, both from 08:00. A standard monitor raised an alarm on the first within 9 minutes. On the second, over the next forty hours, it flagged about as often as it does on a normal day.

01One plant, run twice

The data is a public dataset (Rieth and colleagues, 2017) of 21,000 simulated runs of the Tennessee Eastman process. I scored faults on its 48-hour test runs. In those, each faulty run has a twin: for the first eight hours it matches a normal run value for value. At 08:00 the fault switches on and the two days split. I checked this for all 400 faulty runs used here.

So each chart below shows the same day twice, with the fault and without it.

02A standard monitor, trained on normal running

I trained a standard monitor from the process-monitoring textbooks: principal component analysis (PCA) on all 52 recorded values, fitted to 250,000 samples of normal operation. It flags any sample that falls outside what normal looked like, at a 99% limit. Three flags in a row count as an alarm.

03Valve one: the reactor's cooling water

At 08:00 the reactor's cooling-water valve starts to stick, and the reactor temperature starts to swing with it. The first sample after 08:00 is already flagged. The alarm follows at 08:09.

• Sticking valve (run 1)
• Same day, no fault
• Field note (fiction)

Show the numbers, hour by hour

The monitor itself only says that something is off. To see which readings moved, I compared each run with its normal twin, which no real plant has. The three that moved most are the reactor cooling-water valve signal (13.9× its normal spread), reactor temperature (13.5×) and the reactor cooling-water outlet temperature (10.5×). Nothing else moved more than 3%.

04Valve two: the condenser's cooling water

Same fault, same 08:00, different valve. This time the index stays where a normal day keeps it.

• Sticking valve (run 6)
• Same day, no fault
• Field note (fiction)

Show the numbers, hour by hour

There is a trace, if you know where to look. Across the 20 runs, the condenser cooling-water outlet temperature is the value that widens most against its twin, by a median of 1.09 times.

• Sticking valve (run 6)
• Same day, no fault

Show the numbers, hour by hour

05Quiet is the right output, for the wrong reason

Researchers who sort these faults put the condenser valve among the controllable ones: the control system compensates and the process state barely moves. For faults like that, they write, a monitor's flag rate "should ideally be as low as" its false-alarm rate, so operators are not distracted (Sun and colleagues, 2020; they sort the faults from prior knowledge, on a different dataset from the same simulator).

So staying quiet is the right output here. PCA did not decide that, though. The same paper notes that PCA cannot tell whether a fault is controllable or not.

What this monitor lacks is concrete. The valve's real stem position is not one of the 52 recorded values, and the recorded signal for that valve, XMV(11), barely changes: its spread after 08:00 is 1.00 times its twin's. Methods built for faults like this do better. Du and Du (2018) report detecting it, and two other faults most methods miss, from the same kind of process data.

06What a person on rounds would see

A sticking valve has a look: the stem holds, then jumps, instead of travelling smoothly. Someone standing at it at the right moment can see it.

The simulator has no people in it, so I wrote these two field notes myself, the way a round could plausibly record them. They are fiction, and so are the gold marks on the charts above, which sit at the minute each note was written.

07What each source can answer at hour 48

Valve two, the run where the monitor stays quiet. Three questions, asked of this PCA monitor and then of the monitor plus one fictional field note.

For valve one, the monitor answers the first and third on its own: something is wrong, since 08:09. The note adds which valve, and that its stem is sticking.

08How this was done

• Data. Rieth, Amsel, Tran and Cook (2017), Harvard Dataverse, public domain. The three files used were checksum-verified at download: 500 normal runs to train, 500 more to measure false alarms, and test runs 1 to 20 of each of the 20 faults.
• Names. Recorded values, fault types and the short descriptions of faults 16 to 20 follow the model code in N. L. Ricker's Tennessee Eastman archive, which marks 16 to 20 as unknown. Where that code says the separator's cooling water, I say the condenser's, as Du and Du do.
• Monitor. PCA on the 52 standardised values, 31 components (90.1% of the variance), Hotelling T² and SPE limits at 99%. A sample is flagged when either crosses its limit. An alarm is three flags in a row, timed at the third.
• Measured. The flag rate on 480,000 normal samples, and the flag rate, alarm count and alarm delay for all 20 faults. Full table below.
• Checked. A second implementation, written separately, reproduced every measured number quoted in the text. Changing the tuning (80, 90 or 95% of the variance, or a 99.9% limit) does not change the picture.
• Tools. Python (numpy, pandas, scipy). The 837 MB R data file is read by a small streaming reader written for this, which matched a reference reader value for value on the two normal-running files. Charts are plain SVG. I wrote the pipeline with an AI coding assistant, to my specification; every number on the page comes from that code and its separate re-check.
• Code. The whole pipeline is public on GitHub, with the steps to download the data and re-run every number yourself.
• Not claimed. This is a simulation, not a plant, and it tests one kind of monitor. The field notes are fiction. Neither the dataset's authors nor Pacific Science & Engineering Group reviewed or endorse this page.

All 20 faults, measured

Fault	Type	What is disturbed	Flagged after 08:00	Runs that alarmed	Median alarm delay
1	Step	A/C ratio of stream 4, B composition constant	99.76%	20 of 20	15 min
2	Step	B composition of stream 4, A/C ratio constant	98.79%	20 of 20	42 min
3	Step	D feed (stream 2) temperature	2.02%	4 of 20	1491 min
4	Step	Reactor cooling water inlet temperature	99.94%	20 of 20	9 min
5	Step	Condenser cooling water inlet temperature	27.49%	20 of 20	9 min
6	Step	A feed loss (stream 1)	100.00%	20 of 20	9 min
7	Step	C header pressure loss (stream 4)	100.00%	20 of 20	9 min
8	Random	A/B/C composition of stream 4	96.96%	20 of 20	66 min
9	Random	D feed (stream 2) temperature	2.07%	5 of 20	1962 min
10	Random	C feed (stream 4) temperature	41.66%	20 of 20	136.5 min
11	Random	Reactor cooling water inlet temperature	77.19%	20 of 20	33 min
12	Random	Condenser cooling water inlet temperature	98.81%	20 of 20	25.5 min
13	Drift	Reaction kinetics	94.95%	20 of 20	126 min
14	Stiction	Reactor cooling water valve	99.98%	20 of 20	9 min
15	Stiction	Condenser cooling water valve	2.34%	8 of 20	1072.5 min
16	Random	(unknown) Heat transfer within the stripper heat exchanger	29.68%	20 of 20	118.5 min
17	Random	(unknown) Heat transfer within the reactor	92.45%	20 of 20	100.5 min
18	Random	(unknown) Heat transfer within the condenser	93.90%	20 of 20	160.5 min
19	Stiction	(unknown) Compressor recycle, separator and stripper underflow, stripper steam valves	17.80%	20 of 20	291 min
20	Random	(unknown)	55.95%	20 of 20	120 min

Tuning sweep

Variance kept	Limit	Components	Normal day	Valve one	Valve two
80%	99%	25	1.98%	99.98%	2.53%
90%	99%	31	1.95%	99.98%	2.34%
95%	99%	36	2.04%	99.98%	2.64%
90%	99.9%	31	0.19%	99.96%	0.28%

Which valve on your plant would you bet is sticking right now, with every screen reading normal?

Sources: Rieth, Amsel, Tran, Cook (2017), Additional Tennessee Eastman Process Simulation Data for Anomaly Detection Evaluation, Harvard Dataverse · Downs & Vogel (1993), A plant-wide industrial process control problem, Computers & Chemical Engineering 17(3) · N. L. Ricker, Tennessee Eastman Challenge Archive · Sun, Paiva, Xu, Sundaram, Braatz (2020), Fault Detection and Identification using Bayesian Recurrent Neural Networks · Du & Du (2018), ADCHEM