The article looks at optimization of maintenance and repair services on the base of analysis of operational data as an implementation of real movement away from mileage/calendar maintenance towards the risk based condition based system, when servicing is prescribed not on the schedule, but on basis of actual operational behavior in actual service environment, proving, that data streams (telematic streams, OBD-II/CAN parameters, DTC codes, service history, repeat service records) build an “digital trace” of operation, which value is revealed when this trace is transformed into repetitive service operations (scheduling, further diagnosis, prevention maintenance or acceptable deferral with risk control) within a service loop (the data started to “work”). From a methodological viewpoint, the paper confirms an end-to-end proposal that mixes disparate data integration, time-series-based degradation-features engineering, anomaly detection, estimating the probability of a repair event over a certain horizon and an operations-and-economics model that determines a threshold for intervention, taking into account both downtime costs and costs of missed failure, capacity and logistics of service resources. In our opinion, it’s precisely the ‘price of error’ that gives the forecasting a practical meaning, defining when it’s useful to carry out early works and when the ‘cost-effective’ choice is to wait with an additional investigation during a short validation period or collect more features, thereby avoiding unnecessary jobs and not compromising confidence in suggestions. Scientific actuality would be the formalization of the movement from fragmented operational data toward a governed service process. Meanwhile we adopt standard vocabulary such as “symptom – diagnosis – work – result” and a feedback loop after repairing, because we need to update the system and account for changes in the data, like in the case of different software versions. Practical usefulness looks like being on determining rules on fleet segments, trigger levels, etc.
Roman Mokrynskyi (Thu,) studied this question.