Waypoint: adaptive change-point detection with autoencoder and CUSUM

Implements the adaptive change-point detector described in the associated SBBD article: a non-supervised autoencoder learns a reference regime from temporal windows, the reconstruction error is standardized on a recent buffer, and a bilateral CUSUM supervisor validates persistent deviations. When a change is confirmed, the model is retrained on recent data and the supervisor is reset.

Usage

hcp_waypoint(
  input_size,
  encode_size,
  warmup = 500,
  retrain_size = 300,
  buffer_size = 100,
  k_factor = 0.35,
  h_low = 3.5,
  h_high = 6,
  prob_tau = 0.997,
  epochs_init = 40,
  epochs_retrain = 20,
  encoderclass = autoenc_base_ed,
  ...
)

Arguments

input_size

Integer. Window size used to build autoencoder samples.

encode_size

Integer. Latent size for the autoencoder.

warmup

Integer. Number of leading observations used to initialize the model.

retrain_size

Integer. Number of recent observations used when retraining after a confirmed change.

buffer_size

Integer. Number of previous residuals used to standardize the current reconstruction error.

k_factor

Numeric. CUSUM reference value (k).

h_low

Numeric. Warning threshold for the bilateral CUSUM supervisor.

h_high

Numeric. Confirmation threshold for the bilateral CUSUM supervisor.

prob_tau

Numeric. Quantile used to estimate the residual threshold from the warm-up regime.

epochs_init

Integer. Epochs for the initial autoencoder training.

epochs_retrain

Integer. Epochs for retraining after a confirmed change.

encoderclass

DALToolbox autoencoder constructor. Defaults to autoenc_base_ed.

...

Additional arguments forwarded to encoderclass; these are the autoencoder-specific parameters, not hcp_waypoint parameters.

Value

hcp_waypoint object.

Details

The method separates representation and decision:

• the autoencoder reconstructs windows and produces a scalar reconstruction error;

• the error is standardized with a rolling buffer;

• a bilateral CUSUM with lower and upper thresholds (h_low, h_high) acts as the statistical supervisor;

• after confirmation, the autoencoder is retrained on the most recent regime.

This detector is intended for regime-change monitoring rather than isolated anomaly marking.

References

• Ogasawara, E., Salles, R., Porto, F., Pacitti, E. (2025). Event Detection in Time Series.

• Salles, R. et al. (2020). Harbinger: Um framework para integração e análise de métodos de detecção de eventos em séries temporais. SBBD.

• Ygorra, B. et al. (2021). Monitoring loss of tropical forest cover from Sentinel-1 time-series: A CuSum-based approach.

• Ygorra, B. et al. (2024). A near-real-time tropical deforestation monitoring algorithm based on the CuSum change detection method.

• De Ryck, T. et al. (2021). Change Point Detection in Time Series Data Using Autoencoders with a Time-Invariant Representation.

• Cao, Z. et al. (2024). Change Point Detection in Multi-Channel Time Series via a Time-Invariant Representation.

• Corizzo, R. et al. (2022). CPDGA: Change point driven growing auto-encoder for lifelong anomaly detection.

Examples

library(daltoolbox)

data(examples_changepoints)
dataset <- examples_changepoints$simple

model <- hcp_waypoint(input_size = 12, encode_size = 4)
model <- fit(model, dataset$serie)
detection <- detect(model, dataset$serie)
#> Warning: number of items to replace is not a multiple of replacement length
print(detection[detection$event, ])
#> [1] idx   event type 
#> <0 rows> (or 0-length row.names)