2026 International Conference on Networks Computers and Communications (ICNCCom 2026)

Predictive maintenance for belt-driven conveyor
systems often relies on motor-current signature analysis to detect
degradation. In practice, however, changes in transported item
mass can produce current variations that look much like wearinduced
signatures, which leads to false alarms and premature
maintenance actions. This paper presents VPD-Net (Vision-Physics
Decomposition Network), a method that uses a depth camera
mounted above the conveyor to estimate the mass of each passing
item from its bounding volume and then passes that estimate to
a physics-informed neural network that separates the measured
motor current into a load-explained component and a wear
residual. A gated recurrent unit tracks the wear residual over
time to predict remaining useful belt life. The depth estimate
provides something the model otherwise does not have: a per-item
load reading that helps it distinguish normal operating variation
from actual degradation without depending on manifest data
that may be incomplete or delayed. We evaluate VPD-Net on an
anonymized industrial conveyor testbed dataset whose degradation
characteristics are benchmarked against published motor-current
fault studies. Across five random seeds, VPD-Net achieves a
remaining-useful-life prediction RMSE of 41.3 ± 3.2 cycles,
compared with 78.6±5.1 for current-only baselines, 62.4±4.3 for
manifest-informed models, and 45.8±3.6 for a physics-constrained
manifest baseline that isolates the contribution of vision-based
load estimation (p < 0.05, one-sided Wilcoxon signed-rank). Falsepositive
maintenance alerts fall by 54% relative to the current-only
approach, and the vision-assisted mass estimator reaches 94.1%
accuracy within ±10% of true mass. These results suggest that
combining direct load observation with current-signal analysis
can make conveyor predictive maintenance substantially more
reliable.

predictive maintenance, motor current signature analysis, conveyor systems, depth estimation, physics-informed neural networks, remaining useful life, vision-based inspection