Accurate anomaly detection in power equipment depends on the selection of the time-series input window, which defines the temporal context available to learning models. Conventional practices rely on fixed window sizes based on hardware constraints or operational cycles, often disregarding the autocorrelation structure of multivariate sensor data. This study derives candidate window sizes from variable-specific critical lag times based on the autocorrelation function (ACF) and evaluates them on a transformer sensor dataset using LSTM, 1D-CNN, and Transformer encoder models. Experimental results show that several ACF-based candidates achieved higher macro F1-score than fixed daily and weekly baselines. In this dataset, Case 2 Method 2 was selected as the final criterion across the basic and variable-removal experiments, and Case 3 Method 1 showed that excluding non-valid variables improved average-based window estimation. These results indicate that ACF-based window-size selection provides a data-driven alternative to fixed temporal windows for multivariate time-series anomaly detection in power equipment.
Ki et al. (Tue,) studied this question.