Privacy-Preserving Active Learning for circular manufacturing supply chains for low-power autonomous deployments

The story begins with the author, frustrated on a Tuesday afternoon in their home lab, trying to train a supervised model on a sparse, imbalanced thermal imaging dataset from a pilot project tracking component degradation in refurbished electric vehicle batteries. The data was under strict privacy agreements, making it difficult to create a unified predictive maintenance model as each OEM partner wanted a shared model that could anticipate failures across the reverse supply chain.

Despite weeks of effort, the model achieved only a 68% F1-score on defect classification, highlighting the need for labeled examples. Labeling the data required domain experts to manually annotate thousands of thermal images, which was both slow and privacy-sensitive.

The author discovered a quantum computing paper about quantum-inspired optimization for active learning query strategies. This idea led to a three-month deep dive, fundamentally changing the author's approach to machine learning in privacy-critical environments. The article chronicles what the author discovered, built, and learned along the way.

Circular manufacturing poses unique challenges for machine learning due to data sovereignty, label scarcity, and energy constraints. The author identified three critical challenges: data sovereignty (OEMs won't share raw production data), label scarcity (defect labeling requires specialized knowledge), and energy constraints (edge devices often run on solar power or batteries).

The article explains that combining active learning, federated learning, and quantum-inspired optimization offers a compelling solution to these problems. Active learning minimizes labeling effort by strategically selecting which data points to label, maximizing model performance while minimizing labeling cost. Federated learning preserves privacy by allowing each participating node to train locally on its own data and share only model updates, rather than bringing data to the model.

Quantum-inspired techniques like simulated annealing and quantum annealing analogs can find near-optimal solutions efficiently for selecting optimal query sets in active learning, which is NP-hard in the general case. The author implemented a privacy-preserving active learning system with an architecture overview, but the source material does not provide specific details on its implementation.