An artificial neural network approach to fault detection in multiple stream binomial processes. A real-case study involving the monitoring of railway HVAC systems

Efficient regulation of the heating, ventilation and air conditioning (HVAC) sys tems has been playing a key role in the competition among railway transportation companies to ensure high comfort levels of passengers on board. Recent European regulations specify the requirements for thermal comfort of passenger rolling stock and railway companies have started to continuously monitor the HVAC systems in stalled on their fleets. Typically, a passenger train has more than one coach with a dedicated HVAC data acquisition system, and the simultaneous statistical process monitoring of the signals coming from each HVAC system can be regarded as a mul tiple stream process. In particular, a multiple stream binomial process (MSBP) is a process at a point in time that generates several output streams that can be modeled as binomial processes. Fault detection strategy for MSBPs based on artificial neural networks (NNs) is presented. In particular, a NN is trained to classify a sample as drawn from an in-control or out-of-control (OC) processes. The performance of the proposed strategy is evaluated through a wide Monte Carlo simulation and com pared with the Wludyka and Jacobs’s MSBP control charts in terms of OC average run length. The proposed strategy is also applied to real HVAC system operational data, made available by the rail transport company Hitachi Rail STS.