This is Post 5 of the six-part VHM statistical-methods series. The previous post (dependence measures) covered the pairwise question: how do you measure the relationship between two signals? This one generalizes to N signals: how do you monitor the joint behavior of an entire instrumented system?

The need is obvious in practice. A modern vehicle has hundreds of telematics channels. A modern battery pack has dozens of cell-voltage measurements. A modern motor controller logs current, voltage, temperature, position, and a dozen derived signals. Monitoring each one independently — with the methods of the previous four posts — misses the most diagnostic information, which is in the joint deviation. The classical multivariate toolkit (Hotelling’s T², PCA-based control charts, partial least squares) is the answer for moderate dimensions; the modern auto-encoder approach extends it to the high-dimensional case where N exceeds the sample count.

This is also the post where the false-alarm-rate problem becomes acute. If you have 100 sensors and you alarm whenever any one of them goes 3σ out, you’ll false-alarm constantly even when everything is fine. The multivariate methods are the principled answer — they alarm on the joint deviation, not the marginal deviation, and the false-alarm rate is controlled by design.

What it covers

About twenty-six minutes of careful reading.

Hotelling’s T². The multivariate generalization of the one-sample t-test. The chi-squared distribution under the null. The control-limit calculation. When this is the right reach (low-dimensional, Gaussian-ish data) and when it isn’t.

PCA-based control charts. The two-statistic monitoring approach. T² on the principal-component scores (changes within the model). SPE / Q-statistic on the residual (changes outside the model). Why both are needed and what each one catches.

Partial least squares. When you have both input variables and quality outputs. The structure that makes PLS the workhorse of pharmaceutical and chemical process monitoring. Adaptation to control-engineering applications.

Auto-encoder anomaly detection. The neural-network generalization. Train an auto-encoder to reconstruct healthy data; the reconstruction error is the modern SPE statistic. The bias-variance trade-off with model size. When this earns its keep over classical PCA.

False-alarm and detection-rate analysis. The ROC curve framework. The threshold-tuning discipline that’s the difference between a useful monitor and an annoying one.

A worked case study. A synthetic powertrain dataset — 24 telemetry channels, one slowly-developing fault. Run every method above; show the detection time, the false-alarm rate, the isolation strength. The decision matrix that picks the right method for the data shape.

Read it

Open the field guide →

Code & data

↓ multivariate-monitoring-kit.zip — the Python source for every figure plus the synthetic 24-channel powertrain dataset.

The VHM statistical-methods series

This is Post 5 of 6:

  1. Listening to the Machine — time-series and spectral analysis
  2. Reading the Residual — model-based fault detection and isolation
  3. Catching the Change — change and fault detection over time
  4. Reading the Couplings — dependence measures
  5. Reading the Whole Machine — (this post) — multivariate monitoring
  6. Reasoning Under Uncertainty — probabilistic methods

← Back to Autonomy