Model-in-the-Loop Engineering · Capstone brief
ENG-02 · Predictive maintenance that earns its keep
On a public equipment-degradation dataset the teacher approves, can you beat a dumb time-based maintenance rule — and show whether the model’s false alarms cost more than the failures it prevents?
The question
On a public equipment-degradation dataset the teacher approves, can you beat a dumb time-based maintenance rule — and show whether the model’s false alarms cost more than the failures it prevents?
System / materials
Suitable public sets (confirm before assign): NASA C-MAPSS turbofan subsets, IMS bearing runs, or a teacher-exported CNC / pump log with no student PII. Fixed prediction horizon and cost table supplied by the teacher.
Expected failure modes
Optimizing accuracy instead of cost. Training on test units. Claiming “zero downtime” from a toy dataset.
Done looks like
Cost comparison table: time-based vs. model-triggered maintenance, with explicit false-alarm and missed-failure counts and a written recommendation.
Five C's
CT: the metric is dollars or downtime, not AUC alone. CR: defining a trigger threshold with a reason. CO: peer checks the cost arithmetic. CM: one-page memo to a maintenance lead. CZ: who pays for unnecessary stops.
Mentor role
Reliability or manufacturing engineer reviews metric plan at Checkpoint 1 and cost memo at Checkpoint 2. School-supervised.
Rubric calibration
R1: named horizon and cost assumptions. R2: splits documented. R3: time-based baseline present. R4: false-alarm cost shown. R5: memo actionable. R6: no production deployment claim.
Two ways this goes wrong
(a) Beautiful AUC, no dollar story. (b) Leakage from future labels into features.
Credit lane fit
Lane A; Lane B with Internship wrapper only if division supplies a real host. No verified credit claim.