Utilizing Historical Performance Data From Trajectory Logs and Service Reports to Establish a Proactive Maintenance Model for Minimizing MLC Downtime

PurposeThis study proposes a proactive maintenance model utilizing historical Multileaf collimator (MLC) performance data to predict potential MLC dysfunctions, promote preemptive maintenance and thereby reduce treatment disruptions. MethodsMLC failures were assumed to correlate with MLC performance quantitation from trajectory logs. A cohort of data from service reports and trajectory logs was used to establish a model for predicting MLC dysfunctions. Specifically, the service reports logged by our in-house engineers recorded failure status, including service date, service reason and actions taken, while trajectory logs recorded the ordered/actual leaf positions in 20-ms intervals. Leaf performance from trajectory logs was quantified, where an event was defined as detecting a leaf's position deviation a mm. Three a values, 0.05, 0.1, and 0.5mm, were used as candidates to determine the appropriate threshold for deviation event quantitation. Logged MLC failures from service reports were retrieved and classified into two categories based on the patterns of their deviation events calculated from trajectory logs: (a) failures with continuous deviations: deviation events lasted several days before failure, and (b) failures with a burst of deviations: deviation events only lasted 1 or 2days and then MLC failed suddenly. The proposed proactive model focused on the failures with continuous deviations since abnormal trends in their deviation events lasted couple of days, allowing preventive maintenance. The model was predefined with three parameters (x, y, z): if a leaf scored x deviation events per day in any y days within up to a z-day window, the leaf was marked as a potential failure. The distributions of the deviation events as functions of time (days or weeks) and leaves using the found a-value were then associated with logged failures to find model parameters. In a retrospective demonstration, a total of 28 logged failures with continuous deviations and 66397 trajectory logs from two TBs' 3-yr records were used to determine the model parameters (x, y, z). The established model was then applied to a third TB for validation. ResultsDeviation event threshold, a, was determined to be 0.1mm, and the resulting model parameters were (x=20, y=6, z=10). When validating the third TB's 3-yr record with 12 logged continuous deviation failures, the model predicted 16 failures: seven were confirmed from the records with a hit rate of 58.3%, while nine were not; further investigation of each unconfirmed failure convinced that some could be actual failures, but somehow not recorded. ConclusionThe model offers an addition to preemptive maintenance for reducing treatment disruptions.