industrial//process monitoring

Process monitoring is the practice of watching the measured variables of a running plant to detect, as early as possible, that the process has left its normal behaviour, and it is how a refinery, a paper machine or a compressor station turns a historian full of readings into alarms an operator can act on. It sits between the SCADA alarm list and fault diagnosis: monitoring says that something is different, diagnosis says what.

Process monitoring is the practice of watching the measured variables of a running plant to detect, as early as possible, that the process has left its normal behaviour, and it is how a refinery, a paper machine or a compressor station turns a historian full of readings into alarms an operator can act on. It sits between the SCADA alarm list and fault diagnosis: monitoring says that something is different, diagnosis says what.

The oldest tools watch one variable at a time. A limit per variable (high temperature, low pressure) and the statistical process control charts of the SPC family, which plot a measured quantity against limits set at a few standard deviations of its normal spread, catch large, obvious faults and explain themselves: the operator sees which reading crossed which line. Slow drifts that never cross a limit are the territory of accumulating detectors such as CUSUM (change detection).

Some faults are combinations of normal values.

High flow with low pump current, or a temperature that should have risen with the load and did not: each reading is inside its own limits, and only the broken relation between them reveals the fault. Per-variable limits are blind to this by construction; multivariate monitoring exists for it.

Multivariate statistical process control learns the normal correlations of the plant with PCA on weeks of healthy operation and watches every new sample with two numbers: the T-squared statistic, large when the plant sits at an unusual but coherent point, and the Q statistic, large when the correlations themselves break. It costs one matrix-vector product per sample, fits in a PLC, and needs clean normal data rather than labelled failures.

The order of escalation follows the cost. Per-variable thresholds first, PCA with T-squared and Q when the faults hide in combinations, and an autoencoder (a nonlinear PCA) only when the relations between variables are strongly curved and data are plentiful, at the price of interpretability. This is the flyswatter rule applied to a plant.

Every monitor ends in a detection threshold that trades false alarms against missed faults, and an operator flooded with false alarms learns to ignore the screen; thresholds are set on held-out normal data and reviewed against what maintenance actually found. The general learning-based version of the problem is anomaly detection.