Southwest Airlines is leveraging a new artificial intelligence tool to improve its maintenance operations. Startup company AIXI has been helping the airline more easily generate a database of the discrepancies and fixes for its 800-aircraft fleet in a way that dramatically improves engineer access to defect trends, most likely fixes and more to support better maintenance.
Southwest has long had a data file of discrepancies and corrective actions, manually entered for years and now entered digitally by technicians. However, hard-pressed technicians seldom get beyond two-digit ATA codes for either the problem or fix, and the correct code might be different for the problem versus the fix, explains Barry Lott, director of maintenance records and aircraft reliability at Southwest.
Therefore, a team had to review, correct and complete technician filings manually—a major and boring effort that burned out staff quickly, ate up labor-hours and yielded delayed results. Lott says the effort still fell far short of the completeness and accuracy required.
Southwest turned to AIXI, an artificial intelligence (AI) company that had developed a large language model, similar to those used for generative AI, but adapted for aviation. The tool can understand technician entries and turn them into complete, correct and detailed descriptions of the problem, the fix, the relevant 4-6 digit ATA codes and the required corrective actions.
For example, a technician might enter “NGS [nitrogen gas system] degraded” as the discrepancy and “R/R L/H pack flow” as the fix. The auto-coder would convert this shorthand to 4741 for the discrepancy ATA code and 2530 for the fix ATA code. It would name the object as the NGS System, classify the failure as degraded and classify the fix as removed/replaced. The resulting entry is then much clearer for later retrieval and analysis.
To train the new tool, Lott gave AIXI records representing three years of repairs filed by pilots and technicians. The training was a two-step process, according to AIXI CEO Cameron Byrd. The large language model adapted for aviation was created using tens of millions of records and took a month to train on a server. AIXI built a classification system on top of the language model, which consisted of two million records spanning five years. Building the classification system took approximately a month and a half.
AIXI’s system has been in development since 2019. It required millions of curated records and months of training on systems with multiple graphics processing units just to create the aviation-specific language model. “Further training and more curated records are required to create modules for each of the classification fields," Byrd explains.
Once the system had been thoroughly acquainted with Southwest’s repair data, Lott turned over another massive data set to AIXI to see how well the new system worked, seeking a minimum of 95% accuracy. “They pushed it back quick,” Lott says.
The system worked fine, and now Lott’s reliability team runs the AIXI auto-coder once a day, so defects entered one day are accurately available the next day. More than six years of Southwest’s maintenance history on hundreds of Boeing 737s has already been auto-coded.
The immediate benefit is that reliability engineers can easily extract data on trends, defects, repairs by ATA codes or type of component worked on, type of problem encountered and so forth to support advances in predictive maintenance and other improvements. Southwest uses Boeing’s Airplane Health Management and its own in-house version of aircraft health management software.
AIXI is now working with Southwest to automate the reports on a subset of faults that the airline must make regularly to the FAA under the Service Difficulty Reporting program.
Lott’s long-term vision is more ambitious. For example, suppose a future Southwest pilot sees a defect light pop up in the cockpit en-route and reports it. When the pilot lands the plane there should be technicians standing by with recommendations on the first- and second-most likely fixes for the defect and the parts needed. These recommendations would also reflect accurate, detailed records of the maintenance history of the individual aircraft, its age and the previous fleet-wide experience for handling this particular defect. Byrd calls this capability “prescriptive maintenance.” Lott just says it will mean faster and better fixes.
The principle behind this advance is simple: applying deep repair history to each new fault as it occurs. However, making this advance practical requires a volume, accuracy and level of detail in data that can only be supported by AI tools.
