An Unsupervised Pairwise Comparison Learning Approach With Adaptive Network Structure for Equipment Health Quantitative Assessment

Abstract

Accurate health assessment of large and complex equipment is essential to ensure their safety, availability, and affordability. Existing machine learning-based approaches for health assessment face the challenges of difficult access to high-quality labeled data, large variation in operation for different equipment under complex operating conditions, as well as data-intensive and time-consuming network training. In this paper, we propose an error minimized pairwise comparison learning approach (EM-PASCAL) that uses a lightweight network to achieve equipment health assessment under unsupervised conditions. Specifically, a lightweight network model with variable structure inspired by error minimized extreme learning machine is designed. Then, we propose EM-PASCAL, which makes full use of the non-increasing characteristic of equipment health state, to identify the appropriate network structure dynamically and calculate the corresponding network parameters efficiently. The proposed method is expected to achieve promising evaluation results with a low computational effort. Experimental studies using publicly available dataset show that our approach not only enables quantitative equipment health assessment under unsupervised conditions, but also offers advantages in terms of evaluation accuracy and computational effort compared to existing benchmarks of machine learning models.