070 – “Smart Data Cleanser for Just-In-Time Maintenance Risk Discovery”
Author: Gerry Falen
Company: Robins AFB/Warner Robins Air Logistics Complex
Phone: (478) 222-2912
• Roughly half of Air Force maintenance data is poorly coded, which degrades individual aircraft and fleet analysis. It is no small feat to correct that coding — after identifying rare subject matter experts, the corrections could take four years per platform, resulting in years to complete.
• Research shows that data collection for aircraft maintenance is pretty bad. Why?
• Unrestricted text input – Without pre-defined choices, this creates irregular, inconsistent and insufficient descriptions.
• Poorly coded – Work unit codes (WUC) should identify the needed maintenance. But inaccurate coding, for example a report using a WUC ending with a 0 or a 99 because of poor training or execution, foils maintenance plans.
• Working with the Air Force and Robins AFB in a Phase 2 SBIR and now STTP, Cybernet used artificial intelligence to develop a learning algorithm to correct the coding errors and accomplish in one day what previously took four years.
• SDC technology now touches every plane arriving at Robins AFB and is expected to save $3M annually in unscheduled maintenance and free hundreds of expert hours.
• Original case studies began with 3200 filters for C-130 developed by an Air Force SME over 4 years on and off – 7420 filters were developed by the SDC in 1 day.
• SDC is platform agnostic. With no knowledge of WUC definitions, it also scrubbed F-15 and E-8C data in 1 day.
• By using automation via auto-generated filters, there is an increase in useful and accurate maintenance information, as well as quicker follow through and completion, and even predictability for condition-based maintenance. In other words, garbage in – useful information out.
Smart Data Cleanser offers:
• Algorithms for any platform and coding.
• Advanced decision tree translation with keyword search ability in either discrepancy/corrective narrative.
• Supported keyword parentheses and variant generation.
• Promoted filters to execute at higher levels.
• Predictable filter quality.
• Developed pipelined, two-tiered filter execution.