Tensor SPC Tool

Tensor Contribution Decomposition

Residual localization shows where structured mismatch occurred. Contribution decomposition shows which sensors and time regions drove the Tensor Q statistic.

Purpose

Move from location to contribution.

A Tensor SPC alarm becomes more useful when the total residual energy is decomposed into ranked contributions. This helps the user understand which variables, time regions, and sensor-time patterns should be reviewed first.

Example finding: The review case is driven primarily by vibration and surface-finish behavior during the middle portion of the process cycle.

Core calculation

\[Q = \sum_s \sum_t \left(X_{s,t}-\widehat{X}_{s,t}\right)^2\]

The Tensor Q statistic is the total residual energy for one observation.

\[C_s = \frac{\sum_t E^2_{s,t}}{Q}\times100\%\]

Sensor contribution expresses each feature's share of the total residual energy.

\[C_t = \frac{\sum_s E^2_{s,t}}{Q}\times100\%\]

Time contribution identifies when the structured mismatch was strongest.

Theoretical background

What contribution decomposition adds.

Tensor Residual Localization answers where the model reconstruction error is concentrated. Tensor Contribution Decomposition ranks that same residual energy by sensor and time, turning a heatmap into a prioritized investigation path.

Residual Localization

Where did the structured mismatch occur across the sensor-time field?

Contribution Decomposition

Which sensors and time regions contributed most strongly to the Tensor Q result?

Manufacturing example

Normal behavior compared with a review condition.

The examples use the same structured manufacturing process introduced in the residual localization page. The normal example shows low, distributed residual energy. The review example shows concentrated contribution from vibration and surface finish during the middle of the observation.

Example 1: Normal structured behavior

Residual energy is low and broadly distributed. No single sensor or time region dominates the Tensor Q statistic.

Tensor Q4.2
Q Limit7.9
StatusNo Review Indicated

Normal contribution map

distributed residual energy
Feed Rate
Spindle Load
Tool Temperature
Vibration
Surface Finish
Coolant Flow
0102030405060708090100110
Time Step
low contribution moderate contribution

Top contributing sensors

normal example
Spindle Load
24%
Tool Temperature
21%
Feed Rate
18%
Vibration
15%

Time contribution profile

normal example
0102030405060708090100110

Interpretation

The normal example shows low residual energy with no dominant sensor-time contribution. The Tensor Q value remains below the review limit, and the contribution profile does not identify a concentrated structural mismatch.

Example 2: Structural deviation requiring review

Residual energy is concentrated in vibration and surface finish during the middle of the process cycle.

Tensor Q18.7
Q Limit7.9
StatusReview Needed

Review contribution map

concentrated residual energy
Feed Rate
Spindle Load
Tool Temperature
Vibration
Surface Finish
Coolant Flow
0102030405060708090100110
Time Step
low contribution elevated contribution high contribution

Higher contribution intensity indicates a larger share of the Tensor Q statistic from that sensor-time region.

Top contributing sensors

review example
Surface Finish
42%
Vibration
37%
Tool Temperature
11%
Coolant Flow
5%

Time contribution profile

review example
0102030405060708090100110

Interpretation

The review example shows a concentrated contribution pattern. Surface Finish and Vibration account for most of the Tensor Q statistic, and the time contribution profile peaks during the middle of the cycle. This suggests the review should begin with the coordinated vibration/surface-finish behavior rather than treating the result as a generic process alarm.

Practical meaning: Tensor Contribution Decomposition extends localization by ranking what drove the structural anomaly.

How to use the result

Contribution decomposition prioritizes investigation.

The output should not be treated as an automatic root cause. It is a structured diagnostic guide that helps focus review on the variables and time regions most responsible for the Tensor Q result.

OutputQuestion answeredUse
Tensor QIs structured behavior abnormal?Screen for a structural process change.
Residual localizationWhere is the mismatch concentrated?Locate the sensor-time region of concern.
Sensor contributionWhich features drove the result?Prioritize the investigation path.
Time contributionWhen did the mismatch dominate?Identify the process window requiring review.