Computational Science
Is the System Changing, or Is the Sensor Drifting?
2026-09-10
A slow sensor trend is not automatically a slow physical change. A computational review can ask whether independent channels and known relationships separate drift from system dynamics. Intuition may notice a mismatch first; the investigation must translate that impression into an observable contrast that can reject the preferred explanation.
The computational starting point
Begin with authorized archived measurements, timestamps, calibration history and a documented observation model. FilterPy provides a state-estimation implementation. Default filter settings are not a scientific model. Exclude personal surveillance and sensitive behavioral inference; the proposed scope concerns non-sensitive instrumentation or synthetic reference systems.
Sources: FilterPy Kalman filter.
Where intuition enters
The hypothesis might be that one channel is moving relative to the others while the underlying state remains stable. Specify the expected cross-channel residual pattern. The rival is a real state change that different sensors observe with different delays. A visual divergence alone does not distinguish these possibilities.
A test that can disagree
Construct separate drift and state-change models and test observability before interpreting fitted states. Use synthetic examples with known bias and dynamics to check whether the estimator can separate them at the available sampling rate. Numerical integration can generate reference trajectories under explicit equations rather than hidden simulator defaults.
Fit on earlier segments and predict withheld periods, preserving calibration events as distinct metadata. Check innovations, uncertainty coverage and sensitivity to assumed process noise. A filter can absorb almost any mismatch into a flexible hidden state; apparent smoothness is not validation. If drift and state remain equivalent, report non-identification.
Sources: SciPy initial-value solver.
An illustrative decision
Imagine a channel drifting while an independent reference and a conservation relation stay consistent. That supports a sensor-specific explanation within the model. If both channels share the same unrecognized calibration source, the apparent independence fails. The report should make such shared dependencies visible rather than count them as two confirmations.
What the research would deliver
A client would receive an observation-model audit, distinguishability test and a calibrated uncertainty statement. This can guide an analytics product's research roadmap without touching live equipment. It is not a maintenance order or a safety diagnosis. A preview can identify the discriminating record before a full monitoring implementation is commissioned.
Questions this raises
Can a Kalman filter automatically detect drift?
Only under an adequate, identifiable model. Adding a bias state does not guarantee the data distinguish bias from real dynamics.
Are two sensors always independent evidence?
No. Shared calibration, preprocessing or environmental sensitivity can create correlated errors that must be considered.
Sources and their limits
- FilterPy Kalman filter. State-estimation implementation; model adequacy and observability remain separate questions.
- SciPy initial-value solver. Numerical ODE integration; the research comparison is a proposed design.
Prepared with AI assistance. The linked sources support the specified technical points; they do not validate applied psionics as a whole or guarantee a result for a client.
Read the editorial and evidence standard.
Continue reading
Explore Scientific Oracle consultingfor a scoped review of an existing-data research decision. Start with a non-confidential outline of the question, available evidence and the decision it needs to inform.