Towards Efficient Monitoring of Slowly Evolving Degradation Processes: Insights on Feedback Mechanisms in Future Neural Network

Abstract

Monitoring evolving degradation processes of systems, particularly in heavy industries such as mechanical and rotating machinery, is crucial for ensuring reliability and efficiency. However, accurate monitoring faces significant challenges due to unavailability of realistic data, often constrained by maintenance schedules, along with data drift and complexity. Data drift arises from dynamically changing working conditions and variations in health states due to factors such as damage propagation, aging, fatigue, and other evolving operational and environmental influences. Complexity, on the other hand, stems from the semi-random nature of recorded data under such conditions, making traditional analysis methods like signal processing and statistical analysis less effective in handling large, dynamic signals. This work primarily addresses these challenges, focusing on both data drift and complexity, while introducing an additional perspective on incorporating feedback mechanisms and error correction for improved adaptability. This is particularly relevant in the context of neural networks and deep learning, which are at the forefront of state-of-the-art monitoring techniques. Using a simulated synthetic dataset that emulates realistic degradation behaviors, we illustrate and analyze these key aspects. No preprocessing techniques were introduced to ensure that the effect of the feedback mechanism was assessed separately. The results demonstrate that incorporating feedback mechanisms leads to measurable improvements, with an increase in $\mathrm{R}^{\mathrm{2}}$ of approximately $\mathrm{1. 6 9 \%}$ for training and $\mathrm{1. 8 2 \%}$ for testing. This highlights the significance of feedback mechanisms in enhancing monitoring accuracy and underscores their inclusion in future research, whether in deep learning methods or small-scale machine learning systems.