A Two-Stage Parallel Inception 1D-CNN for Event-Driven Condition Monitoring

Digital twins (DTs) of collaborative robots need continuous synchronization, but streaming high-rate kinematic telemetry to a central DT can create bandwidth bottlenecks. Protocol overhead exceeds 70% of the packet volume for raw sensor data in a projected 1000Hz scenario. This work proposes an Edge-Enabled Digital Twin Networked Control System (E-DTNCS) in which an edge node classifies IMU data windows and sends a low-rate heartbeat in normal operation and a raw-data burst when an anomaly is detected. A compact two-stage Parallel Inception 1D-CNN (PINC-noBN, 37,154 parameters) is evaluated as a classifier on the public CASPER dataset (20Hz UR3e IMU data) with a chronological split and three seeds. With the given labels, it reaches a test accuracy of 0.86±0.02, slightly below a CASPER-style 1D-CNN with kernel size 50 (0.88±0.01, 169,794 parameters). Errors concentrate near class transitions, where labels are offset from the signal by several minutes; excluding these zones, accuracies are 0.97 and 0.99. An analytical projection including false-positive bursts indicates bandwidth reductions of 85–92% (labels as given) and 90–99.8% (transition zones excluded) for anomaly prevalences between 10% and 0.1%. In float32, PINC-noBN needs at least 2 MB of activation memory.