A model for preemptive maintenance of medical linear accelerators—predictive maintenance

BACKGROUND: Unscheduled accelerator downtime can negatively impact the quality of life of patients during their struggle against cancer. Currently digital data accumulated in the accelerator system is not being exploited in a systematic manner to assist in more efficient deployment of service engineering resources. The purpose of this study is to develop an effective process for detecting unexpected deviations in accelerator system operating parameters and/or performance that predicts component failure or system dysfunction and allows maintenance to be performed prior to the actuation of interlocks. METHODS: The proposed predictive maintenance (PdM) model is as follows: 1) deliver a daily quality assurance (QA) treatment; 2) automatically transfer and interrogate the resulting log files; 3) once baselines are established, subject daily operating and performance values to statistical process control (SPC) analysis; 4) determine if any alarms have been triggered; and 5) alert facility and system service engineers. A robust volumetric modulated arc QA treatment is delivered to establish mean operating values and perform continuous sampling and monitoring using SPC methodology. Chart limits are calculated using a hybrid technique that includes the use of the standard SPC 3σ limits and an empirical factor based on the parameter/system specification. RESULTS: There are 7 accelerators currently under active surveillance. Currently 45 parameters plus each MLC leaf (120) are analyzed using Individual and Moving Range (I/MR) charts. The initial warning and alarm rule is as follows: warning (2 out of 3 consecutive values ≥ 2σ (hybrid)) and alarm (2 out of 3 consecutive values or 3 out of 5 consecutive values ≥ 3σ (hybrid)). A customized graphical user interface provides a means to review the SPC charts for each parameter and a visual color code to alert the reviewer of parameter status. Forty-five synthetic errors/changes were introduced to test the effectiveness of our initial chart limits. Forty-three of the forty-five errors (95.6 %) were detected in either the I or MR chart for each of the subsystems monitored. CONCLUSION: Our PdM model shows promise in providing a means for reducing unscheduled downtime. Long term monitoring will be required to establish the effectiveness of the model.