Streaming anomaly detection on edge devices must operate with fixed memory while adapt- ing to drift. We propose HDT-AD, a constant-memory detector that encodes sliding win- dows into real-valued hypervectors using a triangular-kernel Hyperdimensional Transform encoder with position binding, then scores them against a single EWMA-updated proto- type. To prevent adaptive collapse after prolonged anomalies, HDT-AD updates threshold statistics unconditionally while skipping prototype updates only on anomalous windows. We prove an exponential recovery bound for the statistics update rule and a concentration bound for the finite-dimensional kernel approximation. Across 59 benchmark data sets and additional controlled holdout experiments, HDT-AD preserves a 1.26 MB peak heap foot- print, recovers immediately in collapse stress tests, and remains competitive with constant- memory baselines under audited evaluation protocols, including a same-constraint-class comparison against EXPoSE. These results highlight the trade-off between collapse resis- tance and anomaly-segment recall while preserving the constant-memory guarantee. A channel-bound multivariate extension achieves the strongest F1 and AUC on 8-channel in- dustrial data, and a C implementation further demonstrates practical real-time speedups over the Python reference.
Keita Sode (Tue,) studied this question.