Distinguishing Process Change, Instrument Fault and Data-Path Corruption in Streaming Telemetry: A Reference Implementation and Benchmark

Abstract

A reading departs from what a monitoring system expects for one of three reasons: the physical process changed, the instrument is failing, or the value was corrupted between the instrument and the software. The three call for different responses and look alike in a single sample. We state the three-cause problem, prove which pairs of causes one channel cannot separate, and propose a bank of online detectors - a Kalman tracker with CUSUM and EWMA charts on its innovation, a Hampel median with generalized-ESD confirmation, a robust noise statistic with ADWIN, exact-value checks and an optional two-channel complementary filter - combined by a deterministic, inspectable attribution rule. The detectors were merged into a public algorithm library. On 2,942 injected events of thirteen fault types the method detects 96.3% (95% interval 95.7-97.0) at 0.22 false alarms per hour and attributes 91.2% (90.6-91.8) of detected events correctly from one channel and 99.0% (98.6-99.4) with a second, rate channel; from one channel only 2 of 207 detected sensor drifts are attributed to the instrument, as the problem statement predicts. Injected into recorded industrial signals, the same catalogue is detected at 80.8% and attributed correctly at 84.5%, at 3.6 false alarms per hour. On the Skoltech Anomaly Benchmark the method detects 32 of 33 experiments with a median delay of 15 s, at 46 false alarms per hour against 11 for a rolling z-score.