A machine learning method applied to the evaluation of the condition in a fleet of similar vehicles

This paper presents a procedure for anomaly detection of temperatures in different key components of the power train in a fleet of similar vehicles. The anomaly detection is based on the characterization of the typical temperatures observed in the fleet of vehicles under all the working conditions that they develop. These typical temperatures are obtained by clustering methods and they are used as reference for identification of those vehicles where some abnormal behaviors, that can be symptoms of possible performance degradations or failures, are observed. The procedure uses data collected in real-time from the vehicle and they are used as inputs of a Self-Organized Map (SOM) able to discover the typical temperatures expected in their operation. The patterns obtained by the SOM cluster the vehicles according to similar behaviors concerning the temperatures observed at the different key points monitored. This offers a quick and effective view about the performance of each vehicle system respect to their reference temperatures obtained. Vehicles with untypical behaviors regarding the rest of vehicle fleet could suggest the existence of latent failures or degradations. Observing how each vehicle behavior shifts through the different neurons of the SOM, a prognosis can be made about the possible evolution of an anomaly detected. The paper includes some examples of application of the procedure used for the evaluation of the condition of the vehicle fleet.