As part of its nuclear fuel production activities, the customer used a boron deposition process which required cryopumps to maintain a vacuum in a chamber. These pumps were subject to progressive degradation which could trigger deficient cycles, destabilizing the vacuum generated and therefore the entire boron deposition process. The customer was looking to implement a predictive approach to anticipate failures in these production cycles. Data corresponding to 27 different sensors (flags, temperature, gas flow, etc.) were supplied to achieve this objective.
To anticipate, before the end of the previous cycle, the failure of a cryopump so as to avoid using it in the next production cycle, and switch instead to a second, healthy cryopump.
3 models were generated: pump 1 model, pump 2 model, 2-pump model
Metrics were provided for three types of cycles: definitely unhealthy cycles, possibly unhealthy cycles and definitely healthy cycles.
Satisfactory results were obtained with the models generated by DiagFit.
- Model 1: 99% of definitively unhealthy cycles detected, 61% of possibly unhealthy cycles detected, 70% of definitively healthy cycles detected, and 75% anticipation of 6 hours or more.
- Model 2: Detection of 90% of cycles definitely unhealthy, 73% of cycles possibly unhealthy, 58% of cycles definitely healthy, and 74% anticipation of 6 hours or more.
- Model 3: detection of 87% of permanently unhealthy cycles, 68% of possibly unhealthy cycles, 73% of permanently healthy cycles, and 40% anticipation of 6 hours or more.