Enterprise AI for Aircraft Predictive Maintenance

Challenges

The U.S. Air Force faced challenges with its legacy predictive maintenance systems, as there was no government owned, consolidated, and authorized application that could scale to meet all use cases. Legacy predictive maintenance applications had critical gaps: (1) Lack of USAF ownership of their predictive maintenance applications, (2) Lack of Authorization to Operate (ATO) prevented establishment of automated data ingestion pipelines, (3) lack of data rights which hindered trust and adoption, (4) Traditional rules-based Sensor-Based Algorithms (SBA) models held less value compared to AI/ML approaches, and (5) SBA model development was not scalable without a unified data image or integrated data science environment.

The RSO already had access to data for multiple aircraft platforms, including the B-1B. However, each aircraft platform presented unique challenges with its telemetry sensor data, including requiring decoding, overcoming data transfer limitations, and dealing with different formats. Specifically for the B-1B, telemetry data was encoded and lacked a scalable, automated data pipeline for stakeholders positioned globally to access data quickly.

To address these issues, RSO sought a scalable solution to enhance the USAF’s predictive maintenance workflows. The desired solution needed to:

1. Utilize AI/ML techniques to extract deep insights from aircraft telemetry sensor data and predict system and component failures.
2. Handle a large volume of data from disparate sources, including co-mingling telemetry sensor data with maintenance, supply, and flight logs, in a unified model.
3. Allow easy deployment, management, and maintenance of SBA models using advanced ML techniques to enhance monitoring performance.
4. Enable streamlined investigative and case management workflows for stakeholders to coordinate predictive maintenance efforts.
5. Include robust data ingestion capabilities to facilitate rapid analysis by allowing direct uploads into the application or ingestion from cloud sources.

Approach

The project focused on the B-1B platform and began by ingesting and transforming historical telemetry sensor data from 5000 sorties into a unified data image. All SBA models utilized the virtual sensor approach, which predicts sensor values using surrounding sensors for a healthy system. By comparing actual sensor measurements to the predictions, the model determines if it resembles a healthy or degraded state. The virtual sensor model is trained solely on healthy system data, ensuring it only predicts healthy values.

SBA models have been developed at various levels: the system-level monitors whether the system is generating the desired symptomatic output, such as “Temperature Too Hot,” while the component-level identifies issues specific to components and assists in root cause analysis for fault isolation, for example, “Faulty Modulating Valve.” The Virtual Sensor Toolkit was built on the C3 AI Platform’s integrated data science environment, serving as the foundation for all future SBAs, including models developed for other aircraft platforms.

For the user interface, the CoE collaboratively designed and developed an application to manage and analyze alerts. It includes key data visualizations to provide System Program Office (SPO) engineers with confidence in the virtual sensor model’s output and ensure that alerts accurately detect failures.

About the U.S. Air Force

• 5,456 aircraft with an average age of 28 years
• 59 air bases in the U.S. and more than 100 airfields overseas
• 450,000 active personnel and 150,000 reservists
• 17,000 pilots, navigators, and air battle managers

Project Objectives

• Proactively detect signs of degradation prior to failure
• Reduce time between mission end and alert review
• Rapidly scale SBA models across any system or aircraft platform

Project Highlights

• Ingested 5000 B-1B sorties with 75 billion rows of data for use in SBA model training.
• Built an extensible data model to support USAF aircraft platforms.
• Created a machine learning model pipeline that includes autoencoders, hierarchical models, transformers, and post-processors to generate SBA models.
• Deployed C3 AI Readiness for Aircraft Predictive Maintenance supporting two USAF personas: System Program Office (SPO) Engineers and Major Command (MAJCOM) maintenance managers.
• Implemented flight line data uploader enabling ingestion of new sensor data as it’s offloaded from the aircraft and blending sensor data with other data sources already integrated into PANDA’s unified data image.
• Developed SBAs for 11 failure modes, spanning 29 models, to detect system and component degradation across the B-1B.

Built on the C3 AI Platform and extending the C3 AI Readiness application, PANDA delivers the following benefits to RSO and the USAF:

• Develop SBA models rapidly using the Virtual Sensor Toolkit in days, rather than weeks or months with traditional data science tools.
• Access the application from any internet-connected device with a CAC, allowing sensor data upload, alert analysis, and maintenance actioning directly from the flight line.
• Upload encoded sensor data immediately following a sortie, automatically decode and trigger alerts within PANDA, greatly reducing time to review alerts.
• Analyze the evidence package of each alert to confirm its validity as a true detection of degradation.
• Efficiently manage and address alerts at scale using case management capabilities.

About United States Air Force Rapid Sustainment Office (RSO)

Established in 2020, the RSO accelerates the delivery of critical operational solutions to the Department of the Air Force sustainment enterprise. https://www.aflcmc.af.mil/RSO