Noise and blind spots of a single sensor can be offset by fusing other sensors

What it is

Optical heart rate fears arm swing; a chest-strap ECG fears a lifted electrode. EDA fears heat; a respiratory belt fears speech. Sensor fusion stacks measurements of different physics on one judgment, so one channel’s noise lands where another is still clean. What is complementary is the blind spot and the noise mechanism, not sampling the same signal more often.

Why it happens

Each sensor has its own failure-generating process: PPG motion artifact tracks gait frequency; ECG muscle artifact tracks nearby muscles; they do not fully covary, and a joint likelihood can push down “heartbeat-like motion peaks.” Blind spots likewise split by physics: cameras go dark, radar dies behind metal, a pressure mat dies when the person leaves the chair. Fusion gain comes from conditionally independent residuals. If two streams share a disturbance (PPG and wrist acceleration in the same watch, both from the same arm swing), joining them does not automatically clean them. Filters such as Kalman assume process and measurement noise that can be modeled; physiological state is often nonlinear and non-Gaussian. Whether complementarity holds is a question about residual correlation, not sensor count.

Studying it

Put a single channel in its blind spot on purpose (PPG while walking, camera in a dark room, pressure with the person off the chair) and see whether fused error falls and whether residual correlation stays high. Factors: sensor combination, whether the physics is truly heterogeneous. Outcomes: per-channel versus fused mean-squared error, misses in known blind-spot windows. Reports should include “is this channel still alive in the other’s blind spot.” Fusion gain measured only in a lab where every sensor is comfortable sells complementarity as averaging.

Where it stops holding

More sensors add latency, power, and calibration failure points. When two streams are strongly correlated, fusion mainly inflates confidence. If a safety feature writes complementarity as “therefore a high-consequence action may execute on one remaining channel,” it has overspent the gain. Complementarity is statistical; both channels can fail the same day (sweat plus motion).

Applying it

• For each judgment, list known single-channel blind spots and add a channel only when the failure mechanism differs.
• Estimate residual correlation before fusing; if it is high, do not claim complementarity.
• State in product copy “chest strap while walking, watch while seated,” making complementarity visible.
• Verify by creating each channel’s typical failure and confirming that fused judgment in that window is carried by the stream that is still alive.

Related

• Same group: C9.08.2 Fusion must time-align sensors whose rates and latencies disagree · C9.08.3 Fusion raises confidence while widening the kinds of data collected and the privacy scope · C9.08.4 When one sensor fails, a fused system should degrade rather than fail as a whole
• Adjacent: C9.04 Respiration and Posture Sensing · C9.06 Cost of Sensor False Positives
• Search: sensor fusion · complementary sensing · common-mode artifact