Selecting the Best Units in a Fleet: Performance Prediction from Equipment Peers Anil Varma, Kareem S. Aggour, Piero P. Bonissone GE Global Research One Research Circle Niskayuna, NY 12309 {varma, aggour, bonissone}@research.ge.com Abstract. We focus on the problem of selecting the few vehicles in a fleet that are expected to last the longest without failure. The prediction of each vehicle’s remaining life is based on the aggregation of estimates from ‘peer’ units, i.e. units with similar design, maintenance, and utilization characteristics. Peers are analogous to neighbors in Case-Based Reasoning, except that the states of the peer units are constantly changing with time and usage. We use an evolutionary learning framework to update the similarity criteria for peer identification. Results indicate that learning from peers is a robust and promising approach for the usually data-poor domain of equipment prognostics. The results also highlight the need for model maintenance to keep such a reasoning system vital over time. 1. INTRODUCTION The problem of selecting the best units from a fleet of equipment occurs in many military and commercial applications. For example, given a specific mission profile, a commander may have to decide which five armored vehicles to deploy in order to minimize the chance of a breakdown. In the commercial world, rail operators often need to make decisions on which locomotives to use in a train traveling from coast to coast with time sensitive shipments. Asset selection for complex electromechanical equipment is often driven by heuristics and/or expert opinions. Some ‘obvious’ strategies include picking the newest, the most recently serviced, or the latest model equipment. Long-term data that allows reliability and MTBF (mean time between failure) computations at the fleet and individual unit level can also drive such decisions. However, this work was motivated by the special needs of military equipment on new platforms. In the case of a new aircraft, tank, or ship, there is simply no long-term data to assess reliability across the vast range of potential missions. Second, the usage pattern of military equipment can be described as a sequence of ‘pulses’—long periods of inactivity followed by relatively short periods of intense usage. Given the possibility of very sparse