Prototype Memory Control for Detecting Abnormal Operation Events in Industrial Robot Data Streams

Abstract

Industrial robot systems produce continuous event streams from joint controllers, torque sensors, motion planners, safety interlocks, tool changers, servo drives, and production task logs. These streams evolve with product switching, task reprogramming, tool wear, maintenance cycles, and operator intervention. Standard streaming clustering methods may adapt too quickly to repeated abnormal motion events, especially when early mechanical degradation appears as small but persistent deviations. This study develops a prototype memory control method for detecting abnormal operation events in industrial robot data streams. The method maintains dynamic prototypes for normal operation states and introduces a memory-control coefficient to preserve rare deviation patterns across long production cycles. A multi-event representation module encodes joint deviation, torque fluctuation, cycle-time variation, emergency-stop frequency, tool-change irregularity, and controller warning sequences into stream vectors. Experiments are conducted on a robotic manufacturing dataset containing 1,140 industrial robots, 26 production cells, 52 operation variables, and 10-second records collected over 118 days. The dataset includes 364 million streaming event records and 2,960 annotated abnormal segments, including joint overload, servo instability, tool misalignment, abnormal cycle delay, and repeated safety-interlock triggers. The proposed method shortens median detection delay from 22.8 minutes to 6.2 minutes compared with CluStream. False alerts decrease to 1.8 cases per robot-month. The online engine processes 73,000 event vectors per second and stores 9,300 active operation prototypes with 1.08 GB memory usage. Prototype memory control preserves 2,110 low-frequency degradation patterns that are lost by fast-update clustering baselines. These findings indicate that controlled cluster memory can improve anomaly detection in non-stationary industrial robot event streams.