Online Anomaly Detection in Time Series Using Moving Subsequences

The patent extracts only the necessary portion of the time series data by maintaining a sliding window of recent values rather than storing the entire time series. The algorithm keeps track of a limited number of most recent values and their subsequences, extracting only what is needed for online anomaly detection while discarding or not storing older data that is no longer relevant for current anomaly detection.

The patent applies local quality by focusing computational resources and memory storage on the most recent portion of the time series where anomalies are most likely to occur. Instead of uniformly processing or storing the entire time series, the algorithm concentrates on local subsequences within a sliding window, allocating memory and computation locally to where they are most needed for real-time detection.

The patent segments the anomaly detection task by dividing the time series into multiple subsequences of different lengths (e.g., length L, 2L, 3L, etc.). Instead of treating the entire series as one unit or finding a single top anomaly, the algorithm segments the detection into multiple categories based on subsequence length, allowing it to identify multiple distinct anomalies across different temporal scales without overwhelming computational complexity.

The patent applies partial action by detecting anomalies at multiple levels of granularity (different subsequence lengths) rather than performing a complete exhaustive search of all possible anomalies. The algorithm performs partial comparisons against benchmark subsequences for each length category, sufficient to detect significant anomalies without the excessive computational cost of analyzing every possible subsequence combination.