Design a Train Wheel Damage Detection and Failure Prediction System

Design a system that detects damage to train wheels. It must collect measurement data about wheels as trains run, analyze that data to find damaged wheels, and predict which wheels are likely to need attention before they fail. The report names only these goals: the sensors, scale and response times are open, so settle them with the interviewer before designing. ```hint Follow one wheel Trace a single measurement from a sensor to a maintenance decision, and ask how you know which physical wheel, on which car, the reading belongs to. ``` ```hint Two time scales Separate what must be caught within minutes, while the train is still running, from wear trends that only show up across many passes over weeks. ``` ### Clarifying Questions - Where does the data come from: sensors installed beside the track that measure each wheel as a train passes, sensors mounted on the cars, inspection images, or a mix? - How many sensor sites, trains and cars are involved, and how often does a given wheel pass a sensor? - What must happen when a severe defect is found, and how quickly: alert a dispatcher to stop or slow the train, or schedule repair at the next yard? - How are cars identified at a sensor site, and how is each reading mapped to a car, an axle and a side? - Are there maintenance records, such as wheels removed and the reason, to train and check predictions against? - How are the sensor sites connected: always-on links, or intermittent ones? ### What a Strong Answer Covers - Requirements and scale estimates agreed first, separating urgent alerts from long-term trend analysis - Processing at the sensor sites, with store-and-forward for unreliable links - Associating every reading with the right car, axle and wheel, including when identification fails - Ingestion, a per-wheel time-series data model, and separate storage for raw and derived data - Immediate rule-based alerts plus trend- or model-based prediction, evaluated against maintenance records - Alert routing and the human workflow, the cost of false alarms, failure handling, and monitoring of the sensors themselves ### Follow-up Questions - Readings from one sensor site slowly drift upward for every wheel. How do you tell sensor drift from real wear? - A car's wheelset is replaced. How does the wheel's history stay correct? - A missed defect can cause a derailment, but every false alarm stops a train. How do you set alert thresholds? - Actual wheel failures are rare. How would you validate a failure-prediction model?