Abstract

Objective: Study of the possibility of carrying out predictive analysis of the technical condition of locomotive equipment using neural network predictive models enabling to plan the scope of equipment maintenance for routine types of maintenance and repair. Methods: A comparative assessment of the accuracy of forecasts made using a feedforward neural network and a recurrent network with an LSTM layer (Long Short-Term Memory) has been carried out. For training and test-ing of predictive models, we used the results of monitoring the parameters of the lubrication sys-tem of the 2TE116 (2ТЭ116) diesel locomotive by means of on-board diagnostics. Results: The aver-age interval for preventive inspections (TO-3) of locomotives in the existing locomotive mainte-nance system is 25–30 days, and therefore it is this interval that determines the minimum duration of the lead-in period, which the predictive model should provide. We have established that a mod-el based on a feedforward neural network provides sufficient accuracy only for short-term fore-casts with a lead period of no more than 1–3 days. With a further increase in the lead-in period, the error of the model res¬ponse increases to 10–15 %, which prevents it from being effectively used for solving practical problems associated with planning the operation of service locomotive depots. At the same time, the ave¬rage response error of the predictive model based on a recurrent net-work with an LSTM layer does not exceed 3,5–5 % over a 30-day lead-in period, so it can be used to plan the scope and timing of locomotive maintenance procedures. Practical importance: The possi-bility of using time-series analysis methods for predictive analytics of the technical condition of units and systems of a locomotive is shown. Predictive models based on recurrent neural networks with LSTM layers provide prediction accuracy and lead-in period sufficient for solving practical prob-lems that are associated with planning the scope and timing of locomotive maintenance.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call