Module preprocessing 

Sensor signal normalisation and denormalisation.

The raw sensor values captured by wearable devices span very different physical scales (e.g. heart rate ≈ 60–180 bpm vs. LF power ≈ 0–10 000). Before feeding data into the ViT encoder every channel is z-score normalised using population-level statistics.

Normalisation formula

z[t, c] = (x[t, c] - mean[c]) / std[c]

Denormalisation (used in caption generation)

x[t, c] = z[t, c] * std[c] + mean[c]

After denormalisation certain channels are clamped to be non-negative (e.g. steps, sleep_coefficient).

Structs

ChannelStats

Per-channel normalisation statistics resolved from NORM_PARAMS.

Functions

apply_mask

Apply a missingness mask, replacing imputed values with NaN.

average_downsample_ct

Average-pool a (C, T) array down to (C, target_t) time-steps.

channel_name

Return the human-readable name for a channel index.

channel_stats

Compute (mean, max, min, std) for every channel in a (T, C) array.

denormalize

Reverse a previous call to normalize.

denormalized

Denormalise without mutating: returns a new owned array.

f32_slice_to_normalised

Convert a flat Vec<f32> (row-major, shape [T, C]) into an ndarray after normalising.

normalize

Z-score normalise a (T, C) raw sensor array in-place.

normalized

Z-score normalise a raw sensor array and return a new array.