Privacy-Preserving Federated Semi-Supervised Learning for Industrial Machine Vision: A Real-World Deployment Study With FedSSL-ADIM

Abstract

This study presents FedSSL-ADIM, a privacy-preserving federated semi-supervised framework for industrial machine vision that addresses data-centralization constraints, high annotation costs, and cross-site heterogeneity. Deployed across three textile factories with 80 robotic systems and edge devices rated at 0.1–1.0 TFLOPS, the framework combines a novel Adaptive Data-Driven Aggregation (ADDA) mechanism, online incremental updates, and differential privacy. ADDA weights client updates using dataset size, labeled-data fraction, and pseudo-label entropy to mitigate non-IID effects. The system achieved a mean average precision (mAP) of 0.984 for Empty Pocket Detection (EPD) and an accuracy of 0.994 for Bobbin Head-Tail Identification (BHTI). At Factory C, the most heterogeneous site (61.4% contamination; 0.1 TFLOPS), ADDA improved mAP by 7.4 percentage points (8.2% relative) over the FL+SSL baseline (0.901 to 0.975). Annotation costs fell by 90%, and the differential privacy mechanism provided formal guarantees with only 0.2–0.5% mAP degradation. In a three-month operational evaluation, production efficiency increased by 8.1–12.3%, false positives decreased by 12.9–18.2%, and downtime decreased by 18.7–22.1%. The economic analysis showed a <inline-formula> <tex-math notation="LaTeX">$4.0\times $ </tex-math></inline-formula> return on investment and a 2.3-month payback period. The results support privacy-compliant, annotation-efficient federated vision as a replicable approach for heterogeneous industrial settings. Note to Practitioners—Textile factories can use FedSSL-ADIM to improve visual inspection across sites without sharing raw production images. This is useful when privacy regulations (e.g., PIPL and GDPR), limited bandwidth, or sparse labels make centralized training difficult. Field-tested across three factories operating 80 robotic systems, the framework runs on industrial PCs with 0.1–1.0 TFLOPS and supports Empty Pocket Detection (EPD) and Bobbin Head-Tail Identification (BHTI). It achieved 0.984 mAP and 0.994 accuracy, respectively. The ADDA mechanism weights each site’s model updates according to dataset size, labeled-data fraction, and pseudo-label entropy, reducing the influence of unreliable data and addressing site heterogeneity. It reduced manual labeling effort by 90% and improved production efficiency by 8.1–12.3%. At the most heterogeneous factory, ADDA raised mAP by 7.4 percentage points (8.2% relative) over the FL+SSL baseline. Differential privacy (<inline-formula> <tex-math notation="LaTeX">$\epsilon = 5.0$ </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">$\delta = 3\times 10^{-3}$ </tex-math></inline-formula>) caused 0.2–0.5% mAP degradation. During a three-month evaluation, false positives decreased by 12.9–18.2%, downtime by 18.7–22.1%, and net savings totaled <inline-formula> <tex-math notation="LaTeX">${\$}62.2$ </tex-math></inline-formula>K, corresponding to a <inline-formula> <tex-math notation="LaTeX">$4.0\times $ </tex-math></inline-formula> return on investment and a 2.3-month payback. The architecture is designed to work with existing industrial protocols. The reported system meets a 400-ms production cycle, with inference latencies of 42–187 ms for EPD and 68–312 ms for BHTI. Further work should reduce latency for faster lines and evaluate multimodal inputs such as vibration and acoustic signals. A three-slide multimedia deck illustrating deployment is provided as supplementary material.