Dynamic condition monitoring method based on dimensionality reduction techniques for data-limited industrial environments

In recent years, the integration of intelligent condition monitoring systems has become crucial for the Operation and Maintenance in the industry. Condition-based maintenance strategies allow the identification of the current health status of industrial assets. Consequently, downtimes and maintenance costs are reduced by the efficient planning of corrective actions and the early detection of costly damages. However, the implementation of data-based monitoring techniques is often hampered by two main difficulties: Firstly, the lack of understanding of the monitored equipment; and secondly, the lack of data related to failures and their evolution, which is one of the major obstacles for the implementation. This article proposes a method based on the use of dimensionality reduction (DR) techniques to perform real-time monitoring of the state of the machine. The performance of the approach is studied with four different DR techniques: PCA, LDA, Relief, and Autoencoders. The algorithms are tested in a simulation of various asset lifetimes in which they determine when the assets are working under abnormal conditions. Next, the results of the simulation are evaluated under a specifically designed evaluation criterion based on three key performance indicators; cost, interpretability, and effectiveness. In this way, a common situation in industrial field applications is imitated, where the information is gathered in real-time and there is not enough prior knowledge of the machine nor data about its degradation or failures. (C) 2019 Elsevier B.V. All rights reserved.