Fusion must time-align sensors whose rates and latencies disagree

What it is

ECG samples in milliseconds, respiration in tenths of a second, activity recognition emits labels by window, and a radio adds uncertain transport delay. Fusion that treats those numbers as simultaneous observations welds a time-staggered world together. Temporal alignment is a premise of fusion, not a finishing detail.

Why it happens

Each sensor has its own clock domain, buffer, and feature window. One PPG “beat” may correspond to an optical waveform that began 800 ms earlier; a phasic EDA response sits 1–3 s after a stimulus; an IMU window may be a two-second statistic. Without a shared time base (or at least an estimable relative delay), a Kalman update uses mis-timed observations and the filter treats delay as process dynamics. Alignment tools include a hardware sync pulse, interpolation onto a common grid, and treating slow channels as piecewise-constant priors. Typical symptoms of misalignment are “fusion shakier than either channel” or a phase error that walks (crystal drift).

Studying it

Use one physical event (a stomp, a deep breath) as a time landmark and estimate each channel’s delay distribution to that landmark. Factors: hardware sync present or not, buffer depth, feature-window length. Outcomes: mean and jitter of inter-channel delay, sensitivity of fused error to imposed misalignment. Treating log timestamps as ground truth ignores firmware buffering. Wireless retransmission after loss creates non-stationary delay that wired laboratory capture will not show.

Where it stops holding

For low-rate context labels (sit / walk), a few hundred milliseconds of misalignment is often harmless. For closed-loop control or capturing EMG lead, tens of milliseconds can eat the fusion advantage. Wall-clock sync across devices (watch + phone + chest strap) is at the mercy of OS scheduling; NTP cannot be assumed to have solved it. Daylight-saving shifts and users changing system time dirty long-run log alignment.

Applying it

• Record sample time and feature-window start/end per channel; fuse only aligned samples.
• When delay jitter exceeds a threshold, stop fusing and fall back to one channel rather than emit a mis-phased “high confidence.”
• For cross-device fusion, prefer an in-field calibration on a shared physical event over factory clocks.
• Verify by injecting a known channel delay and watching whether fused output biases with it; if the system still reports high confidence, alignment is not done.

Related

• Same group: C9.08.1 Noise and blind spots of a single sensor can be offset by fusing other sensors · 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: C9.11 Latency of Physiological Input · C2.10 Touch Latency and Directness
• Search: clock synchronization · sensor latency · multirate fusion