Uncertainty Aware Functional Behavior Prediction and Material Fatigue Assessment for Circular Factory

Researchers present a machine learning framework that predicts functional performance and material fatigue for remanufactured tools in circular economy settings, using LSTM neural networks combined with finite-element stress analysis to assess whether returned products can safely re-enter production.

This research addresses a critical operational challenge in circular manufacturing: determining whether returned products are safe and functional for reuse. Traditional inspection methods provide only a snapshot of current state, failing to predict how degraded components will perform under future operational stress. The proposed framework bridges this gap by combining two previously disconnected domains—functional prognosis and material fatigue analysis—into a unified reliability assessment workflow.

The technical approach leverages deep learning to extract patterns from sensor data (spindle forces and shaft torque) and predict nine functional variables simultaneously while quantifying prediction uncertainty through Gaussian distributions. Parallel finite-element analysis translates the same loading history into damage accumulation forecasts using established fatigue models. This dual-branch methodology provides manufacturers with comprehensive reliability trajectories rather than isolated predictions.

The practical implications for circular factories are significant: accurate reuse decisions reduce waste, lower production costs, and minimize field failures that damage brand reputation and warranty economics. The achieved 96.52% accuracy on held-out tests demonstrates production readiness. Thermal predictions near perfect accuracy contrast with motor current and load speed predictions requiring continued refinement, suggesting the framework handles some failure modes better than others.

Industry adoption hinges on integration into existing remanufacturing workflows and validation across diverse product categories beyond angle grinders. The finding that conventional LSTM outperforms newer architectures in short-history settings has implications for edge deployment and real-time assessment systems with limited historical data.

• →Framework combines LSTM-based functional prediction with finite-element fatigue analysis to assess remanufactured product viability in circular manufacturing.
• →Achieves 96.52% mean tolerance accuracy across nine predicted functional variables with uncertainty quantification via Gaussian distributions.
• →Thermal variables predict with near-perfect accuracy while motor current and load speed remain challenging, indicating model strengths and limitations.
• →Enables data-driven reuse decisions that reduce waste and improve circular economy economics by replacing inspection-only assessment methods.
• →Conventional LSTM outperforms modern architectures in scenarios with limited historical data, suggesting practical deployment advantages.