SSKD: self-supervised knowledge distillation for reliable remote surface defect detection under wireless transmission attenuation

Abstract Wireless sensor networks (WSNs) enable scalable, remote surface-defect inspection in smart manufacturing, but attenuation during wireless image transmission—manifesting as noise, compression artifacts, and packet loss—undermines diagnostic reliability. Existing solutions often assume high-fidelity inputs, rely on large paired datasets of pristine and degraded images, or require costly hardware upgrades, and they degrade sharply under adverse channel conditions. We propose a Self-Supervised Knowledge Distillation (SSKD) framework that reconstructs attenuated observations on the fly without historical pairings or rigid data correlations, avoiding hardware changes. SSKD couples a self-supervised learning architecture with cross-domain knowledge transfer to autonomously restore defect-salient cues and stabilize downstream detection. Across diverse attenuation scenarios and domains (including Ball Screw Drive monitoring), experiments show that SSKD consistently preserves diagnostic fidelity and robustness relative to strong baselines under severe degradation, enabling reliable remote inspection with commodity WSN deployments.