MSPC

What Is MSPC? Catching Faults Univariate Charts Miss

By J. de Vries · · 7 min read

Time into batch Hotelling's T² 95% control limit Fault detected — 40 min before failure

Hotelling's T² chart flagging a developing fault (illustrative data).

What is MSPC?

Multivariate Statistical Process Control (MSPC) is classical SPC upgraded for correlated data. Instead of one control chart per sensor, it builds a model of how your process behaves when everything is normal — including how variables move together — and then watches for departures from that joint behavior.

The univariate blind spot

Picture a reactor where temperature and cooling-water flow always rise and fall together. One day the temperature rises while the flow stays flat. Both values are individually inside their control limits — every univariate chart on the board is green — yet the relationship between them has broken, and that's precisely what an early-stage cooling fault looks like. A chart per tag can't see it; a model of the correlation structure can.

Two numbers instead of forty charts: T² and Q

MSPC condenses all monitored tags into two statistics. Hotelling's T² measures how far the process has moved within its normal operating pattern — unusually hot and fast, but in a familiar way. The Q-statistic (squared prediction error) measures behavior the model has never seen — the correlation structure itself breaking, as in the cooling example. Each gets a control limit, and together they replace a wall of individual charts with two lines an operator can actually watch.

From alarm to cause: contribution plots

An alarm is only useful if it says where to look. When T² or Q crosses its limit, a contribution plot breaks the excursion down by variable — showing, say, that cooling-water tags account for most of the deviation. The operator gets a shortlist of suspects in one click instead of a fault-finding meeting.

What it takes to set up

You need a set of historical batches you'd call good — typically a few dozen — to train the normal-operation model. From there, ProcessLens scores new data against it, either batch-by-batch or as a live stream from your historian, and raises T²/Q alarms with contribution plots attached. No thresholds to hand-tune per sensor tag.