Privacy-Preserving Active Learning for Circular Manufacturing Supply Chains for Low-Power Autonomous Deployments

说明与用途

一种隐私保护的主动学习系统结合联邦学习、差分隐私和量子启发式优化，在分布式循环制造供应链中训练缺陷预测模型，保护原始数据并减少标注工作量。

问题

OEM partners could not share raw sensor data due to privacy and regulatory constraints, yet a unified predictive maintenance model was needed across a distributed circular manufacturing supply chain. Label scarcity and energy constraints on edge devices further limited traditional supervised learning.

结果

The integrated system reduced labeling effort by selecting only the most informative samples, preserved raw data privacy through differential privacy and federated learning, and improved defect classification performance beyond the initial 68 % F1‑score. It enabled a unified predictive maintenance model across a distributed, privacy‑sensitive supply chain using low‑power edge devices.

工作流

• 1. Deploy low‑power edge devices (Raspberry Pi Zero, ESP32) to collect sensor data and run local inference.
• 2. Edge devices run a local privacy‑aware query selector that uses uncertainty sampling with differential‑privacy noise to choose informative unlabeled samples.
• 3. Selected samples are summarized into privacy‑preserving query summaries and sent to a regional federator.
• 4. Federated learning rounds train local models on each node’s data and aggregate only model updates, protecting raw data from reconstruction attacks.
• 5. The central orchestration tier manages the active learning loop, distributes the global model, and coordinates federated rounds.
• 6. Quantum‑inspired genetic algorithms optimize the query set selection, balancing uncertainty, diversity, and privacy cost.
• 7. Experts label the summarized queries locally, and the new labels are fed back into the local models for continual improvement.

证据

I had spent weeks trying to train a supervised model on this fragmented data, and the results were abysmal—a paltry 68% F1-score on defect classification.

来源

My research revealed a crucial insight: naive federated learning is vulnerable to gradient inversion attacks.

来源

I discovered that using a quantum‑inspired genetic algorithm to optimize query selection—balancing uncertainty, diversity, and privacy cost—significantly outperformed greedy approaches.

来源