Predictive Maintenance

Predictive Maintenance technologies aim to detect, diagnose, and predict failures and degradation in machine components prior to criticality.  The ultimate goal is to prevent downtime, identify root causes for follow-up action, and enable efficient evidence-based maintenance planning and optimization. GE Research has a rich history in the development of Predictive Maintenance technologies, with deployed tools managing over one hundred thousand assets across the GE business units as well as countless more for GE's customers in the aerospace, power generation, transportation, oil exploration, and healthcare domains.

The primary research topics being pursued by GE fall into two categories: Early Warning and Prognostics.

Early Warning pertains to detecting anomalous behaviors in a system's operation at the earliest possible time, regardless of the operational mode or context that the system is operating in. The intent here is to provide the maximum lead time for any potential action, with additional technologies brought to bear once sufficient evidence exists to suggest potential problems. This technology is applicable across industrial verticals and is traditionally the first item in combined Prognostics Health Management (PHM) deployments. GE Research software solutions for Early Warning are built around unsupervised, semi-supervised, and fully-supervised data exploration, enabling a broad span of solution complexities based upon the availability of ground truth feedback regarding normal and abnormal asset operations. As part of this, research has included an emphasis on fusion algorithms to integrate alerts and mixed-type information from multiple models to improve prediction accuracy and lead detection time as well as reduce alert fatigue. Similarly, research activities have yielded a robust multivariate time series search pipeline to speed up human interaction when searching for signatures, features, and patterns in massive time series data (for root cause and diagnostic reporting).

Prognostics go a step further to provide long term predictions of behavior and life. Taking over once Early Warning has occurred, Prognostic algorithms aim to forecast remaining useful life, time to reduced capability, and emergent fleet segmentation for planning inspection, maintenance, repair, and spare parts inventory. GE Research technologies in this space are built upon a suite of hybrid modeling techniques that use embodied domain physics along with condition monitoring data from fielded systems and simulations. This also allows learning systemic behaviors from entire fleets through techniques such as transfer learning. Furthermore, auto-inspection technologies allow us to inspect and label component condition (and quantify performance of associated prognostic models) without human bias, adapting the fielded models in a continuous learning mode.

GE remains committed to advancing the state of the art in the area of Predictive Maintenance and is always seeking new frontiers in its associated technologies.