Tensor SPC Method

How Tensor SPC works

Tensor SPC extends traditional process monitoring by preserving the natural multiway structure of process data. Instead of treating a run as one long row of numbers, Tensor SPC keeps the relationship between observations, sensors, and time visible to the model.

Core idea

Tensor SPC monitors structure, not just individual values.

Many process signals are meaningful because of how they move together. A single variable may look acceptable while the coordinated process behavior changes. Tensor SPC is designed to detect those structured changes.

Traditional SPC

Excellent for individual variables, shifts, spikes, and instability in one signal at a time.

Flattened multivariate monitoring

Useful for multivariate behavior, but the sensor-time layout is converted into a long vector.

Tensor SPC

Preserves observation × sensor × time structure so the model can learn coordinated behavior.

Plain-language summary: Tensor SPC asks whether the process is still behaving coherently across variables and time.

Data representation

Flattening versus preserving structure

The same data can be represented two different ways. The first approach flattens sensor-time data into one long vector. The tensor approach keeps each mode separate, which makes the relationships easier to model and interpret.

Flattened representationTensor representationSensor × time structure is collapsed into one long row.Observation, sensor, and time remain separate modes.one long feature vectorsensor/time relationships must be inferred indirectlysensor × time structure is preservedmodes remain interpretableVector:x ∈ ℝS·TTensor:𝓧 ∈ ℝN×S×T
Why this matters: flattening can still be useful, but it makes sensor and time relationships less explicit. Tensor SPC keeps those relationships available for modeling, monitoring, and diagnosis.

Modeling step

The tensor model learns normal structured behavior.

A low-rank tensor model represents the dominant patterns in the data. In practical terms, it learns common sensor relationships, time patterns, and observation-to-observation variation.

Low-rank tensor approximation

𝓧 ≈ 𝓖 ×₁ A ×₂ B ×₃ C

𝓧 is the structured data tensor. 𝓖 is the core tensor. The factor matrices represent major patterns by mode: observations, sensors, and time.

Mode interpretation

ModeMeaning
AObservation/run behavior
BSensor relationship patterns
CTime evolution patterns
1
Collect structured data

Keep observation × sensor × time layout.

2
Fit normal model

Learn dominant low-rank process structure.

3
Reconstruct behavior

Estimate what normal structured behavior should look like.

4
Monitor residuals

Flag behavior the model cannot reconstruct well.

Monitoring statistics

Tensor SPC uses both latent scores and residual energy.

Two common monitoring ideas are retained: a score-space statistic for unusual movement within the learned structure, and a residual statistic for behavior outside that structure.

Score-space monitoring

T² = tᵀ Σ⁻¹ t

This monitors whether a new observation is unusual within the learned latent tensor structure.

Residual monitoring

Q = ‖𝓧 − 𝓧̂‖²

This monitors how much structured behavior remains unexplained after reconstruction. A high Q value indicates behavior the normal tensor model could not reproduce.

Important distinction: a Tensor SPC alarm does not have to mean one sensor exceeded a limit. It can mean the coordinated sensor-time pattern no longer matches the learned process structure.

Visual explanation

Residual maps help show where the structure changed.

Because the tensor keeps sensor and time modes separate, residual energy can be displayed as a sensor-time map. This makes the result more diagnostic than a single alarm number.

Observed structureTensor reconstructionResidual energyWhat happenedWhat the model expectedWhere the mismatch occurredlocalized mismatchThe residual map is not the raw data. It shows where the tensor model failed to reconstruct expected structure.

Where it fits

Tensor SPC is an extension, not a replacement.

Traditional SPC should remain the first layer for individual process stability. Tensor SPC adds a second layer for structure-aware monitoring when processes involve meaningful relationships across variables and time.

Monitoring layerBest at detectingTypical question answered
Traditional SPCSingle-variable shifts, spikes, instability, and special causesDid this signal move outside expected behavior?
PCA/MSPCOverall multivariate variation after flatteningIs this observation unusual in multivariate space?
Tensor SPCSensor-time relationship changes and structured residual patternsDid the process stop behaving coherently?
Best use case: Tensor SPC is most valuable when the process contains meaningful multiway structure, such as repeated runs, multiple sensors or features, and time-evolving behavior.