Train an ADL activity recogniser from labelled windows

Learns a k-nearest-neighbours classifier over standardised adlFeatures(), mapping a window's features to an activity label. Dependency-free and model-agnostic: for a stronger engine, extract features here and classify with PhysioML (ROCKET/torch) instead – the ICF mapping (adlToICF()) is unchanged.

Usage

trainADL(features, labels, k = 5, window_sec = 2.56, overlap = 0.5)

# S3 method for class 'adl_model'
predict(object, newdata, ...)

Arguments

features

Training features: an adlFeatures() data.frame or a numeric feature matrix (one row per window).

labels

Activity label per window (length = number of windows).

k

Neighbours for the vote (default 5).

window_sec, overlap

Window settings to store so recognizeADL() can re-extract features consistently (defaults match adlFeatures()).

object

an adl_model.

newdata

features to classify (adlFeatures() data.frame or matrix).

...

unused.

Value

an adl_model object.

See also

recognizeADL(), adlFeatures(), adlToICF()

Examples

set.seed(1)
fs <- 20; t <- seq(0, 30, 1 / fs)
walk <- cbind(0.2 * sin(2 * pi * 1.8 * t), 0.2 * cos(2 * pi * 1.8 * t), 1)
sit  <- cbind(rnorm(length(t), 0, 0.01), rnorm(length(t), 0, 0.01), 1)
fw <- adlFeatures(walk, fs); fs2 <- adlFeatures(sit, fs)
feats <- rbind(fw, fs2)
labs <- c(rep("walking", nrow(fw)), rep("sitting", nrow(fs2)))
m <- trainADL(feats, labs)
m
#> <adl_model> kNN(k=5) | 44 training windows | 36 features
#>   activities: sitting, walking