A multi-dimensional feedback engine for governed adaptation in human-in-the-loop predictive maintenance

When Predictive Maintenance (PdM) models must adapt to new operational priorities, we need to manage this process carefully. The main problem is that models can lose previously learned failure patterns, which reduces their usefulness. This paper proposes the Multi-Dimensional Feedback (MDF) Engine. This is an architectural pattern that structures human feedback for PdM systems into three levels: correcting specific predictions, validating model behavior patterns, and providing strategic guidance. The goal is to allow controlled model updates. We tested this approach using a simulation with the NASA C-MAPSS dataset. We created a scenario where the system's priority changes from saving costs to ensuring safety. The results show that our MDF method kept the model's performance utility stable. In contrast, the traditional method of retraining a new model from scratch caused a 67% drop in utility. Our method also used much less expert data - only 95 feedback samples, which is 210 times less data than full retraining. This work shows a structured way to use human feedback, helping PdM systems stay reliable when operational goals change.