Tag Analytics

Can I use it?

Holding a role in Instrument Graphs does not give you this. It is a separate module on purpose: looking at a chart and building a predictive model are different jobs.

If you have no role at all you will see: "You do not hold a Tag Analytics role in any company yet. An administrator grants it under Administration ▸ User Access."

What it actually does

Your instrument tags have been recording values every few seconds for months - speeds, temperatures, throughputs. Tag Analytics looks at that history, finds a pattern, and then - this is the part that matters - checks whether the pattern actually predicts anything on days it was not allowed to look at.

That last step is the whole point. It is easy to find a pattern in the past. Most of them are coincidences. This tool's main job is telling you which is which, and it will happily say "this model is useless" rather than dress up a coincidence as an insight.

The four pages

Menu ▸ Tag Analytics

Step 1: check the tags are usable (Model Health)

Before building anything, an administrator should open Model Health, pick a date range and press Measure. You get a row per tag, and three columns decide whether that tag can be used.

Four tiles at the top count: tags in total, tags fit to model, tags below the coverage floor, and tags whose clock is unknown. There is also an Only show problems switch.

Coverage - the bit that confuses everyone

You see two numbers: a big one, and a smaller one labelled reported underneath.

• reported = how many five-minute slots the sensor actually wrote something into.
• The big number = how many slots we know the value for.

These are different, and the gap is normal, not a fault. Most sensors only write when the value changes. A room temperature that sits at 21.4 degrees for an hour writes once, not twelve times - but we still know what it was for that whole hour. So the big number is much larger, and the big number is the one that counts.

Coverage only drops when a tag goes genuinely quiet for a long stretch - the machine was off, the sensor died, the network dropped. That is real missing data, and the tool refuses to invent it.

If a tag you expected shows 0%, it reported nothing at all in that period. Either it is broken or it has never been commissioned. That is a tag problem, not a Tag Analytics one - see the Instrument Tag setup guide.

Time basis - why some tags say "Unknown"

Different equipment sometimes records timestamps in local time and sometimes in UTC, which are two hours apart. Mix a local-time tag and a UTC tag in one model and the tool would "discover" a confident two-hour delay that does not exist.

So ThreadWare works out each tag's clock from its daily rhythm, and refuses any model that mixes tags on different or unknown clocks. An administrator can settle it by hand with the Declare SAST / Declare UTC buttons on the row.

Step 2: build a model (Model Studio)

On the left is your list of saved models with a + button to add one (Editors and Admins only). On the right are three tabs: Relate, Results and Forecast.

Start on Relate. There are five decisions.

Decision 1 - Model kind

Decision 2 - Target tag

The measurement you want to explain or predict. Pick an instrument tag, or a derived tag if your administrator has set one up.

Decision 3 - Target form

• Change over the horizon - recommended, and the default. You are asking "how much will this move in the next hour?"
• Level - "what number will it be?" ThreadWare shows an orange warning if you pick this, because a slow-moving value is trivially easy to predict as a level and the result looks far better than it is.

Decision 4 - Drivers

The measurements you think move the target. Add each one, and for each you can set:

• Lag - how long the effect takes. Leave it on Auto unless you know.
• Form - level or change.
• An Enabled switch, so you can turn one off without deleting it.

Do not know what to add? Press Suggest drivers. ThreadWare screens every eligible tag and lists the promising ones. Two things to understand about that list:

• It is always labelled Exploratory. It is a starting point for your thinking, never a result to quote.
• ThreadWare runs the same screen on deliberately scrambled data as a control. If the scrambled run also "finds" things, the whole screen is thrown away rather than shown to you.

Decision 5 - Window and horizons

Then Save, then Fit. Fitting happens on the server, so you can leave the page and come back.

Step 3: read the results honestly

Start at the top. Always.

The first thing on the Results tab is a big green or red box. Read it before anything else.

Under the box are four small numbers:

"drivers add" - read this one twice

A model always knows the target's own recent history, because it needs that to predict a change from where it is now. On a slow-moving process, that alone earns a little apparent skill - even if your drivers are pure noise.

So ThreadWare also fits a version using only the target's own past, and reports the difference. That difference is drivers add, and it is the honest answer to "are my chosen drivers telling me anything the target's own history did not already?"

• 0.05 or more - yes, they are earning their place.
• About 0.00 - no. The verdict box says so bluntly: "NO SKILL FROM THE DRIVERS... everything it appears to know, the target's own past already said."

The four tiles

Method chosen - which technique won. If a simple one beat the machine learning, that is a good result, not a disappointment.

Rows against Effective n - this catches people out. Rows might say 25,000 and Effective n might say 400. That is not an error. Readings five minutes apart are nearly identical, so 25,000 of them are not 25,000 independent facts. Effective n is the real sample size, and it decides whether the numbers mean anything. If it is small, treat everything with caution.

The chart

Actual against predicted, over the most recent slice of history the model was never allowed to see while learning. Two extra lines show the two "dumb guess" benchmarks.

If the prediction tracks the actual better than the benchmarks do, the model is doing real work. If you cannot see a difference by eye, believe your eyes over the numbers.

The coefficients table

One row per driver, saying how much the target moves when that driver moves. Look at the last column first - "Reading":

The three reasons you will see:

• "the 95% interval spans zero" - the effect might be zero. The data cannot tell.
• "VIF 15.2 - largely explained by the others" - two of your drivers move together so closely that the maths cannot tell which deserves the credit, so it splits it arbitrarily. The pair together still predicts fine; you just cannot read either one on its own.
• "Sign agreed in only 40% of folds" - ThreadWare sliced the history into chunks and the effect flipped direction between them. That is not a real effect.

This is deliberate. The tool shows you the number and refuses to let you read it, rather than quietly hiding it.

The lag profile chart

How strongly each driver relates to the target at different delays. A negative delay means the driver moves first.

There are two lines per driver. The grey raw line is usually big and impressive. The green prewhitened line is the honest one - it is the raw relationship with "both things happen to drift together" removed.

Trust the green line. If green is much smaller than grey, most of what you were seeing was just two signals drifting through the day together. If the green peak sits at, say, minus 40 minutes, that is your answer to "how long does it take?".

Forecasts

Pick a model and see what the measurement is likely to do next.

• Live scoring switch - keeps the forecast updating. Editors and Admins only.
• Horizon - how far ahead.
• The chart shows the forecast with a shaded band around it. That band is measured, not assumed: ThreadWare also tells you what fraction of past values actually landed inside it.
• You get the probability of crossing a threshold you care about.
• A severity figure flags when recent behaviour is drifting outside the normal band. It is deliberately called severity and not a probability, because it is not one.

A Stale badge on a model means reality has drifted away from what the model was trained on. Re-fit it.

If a forecast is empty, the page tells you which of four reasons applies: scoring is switched off, the model was never fitted, it was fitted before the scoring parts existed, or it simply has not run yet.

Fault Risk

Lists your Event models only, with a live scoring switch and a risk readout.

No models yet? The page says: "No fault-prediction models yet. In Model Studio, create a model of kind Event whose target is the fault tag and whose drivers are the measurements meant to see it coming."

When it refuses - and what to do

ThreadWare refuses a fit rather than producing a bad one, and every refusal says why in plain words. A refusal is the tool working. A model fitted on 40% coverage would still produce confident-looking numbers - they would just be made up.

Three traps worth knowing about

1. Temperature and humidity will always look strongly related, and it is mostly physics. Relative humidity is defined partly by temperature. Cool the air and it rises without a drop of moisture moving. You get a big relationship that tells you nothing about your process. If you are chasing moisture, ask an administrator to set up a derived tag such as dewpoint, absolute humidity or VPD. Those are what materials actually respond to.

2. Two drivers that move together cannot be told apart. Two sensors on the same shaft rise and fall as one. The model predicts fine; it just cannot tell you which of the two deserves the credit. Watch for the VIF flag in the Reading column.

3. Do not ask a model to justify a change you have already decided on. Try enough combinations and something will eventually look significant by luck. Write down what you expect before you fit, and treat a surprising result as a reason to go and look at the machine, not as proof.

Quick reference

Related guides

• Instrument Tag setup - where tags and their sample rates are configured.
• Instrument charts - looking at the raw readings.
• All guides