Predicting Asset Degradation with Data-Driven Models

The Department of Defense was authorized $26.7 billion in FY2020 to construct, sustain, restore, and modernize its 585,000 facilities and infrastructure worldwide. While its annual expenditures are budgeted at 1.2 percent of the replacement value for these assets, an estimated $116 billion maintenance backlog remains. Inevitably, all federal agencies and private organizations with facility portfolios face degrading infrastructure. The nation’s infrastructure is currently rated a D+ by the American Society of Civil Engineers. Awareness of this concern appears to be gaining significant attention, with momentum toward an infrastructure bill looking likely. However, investment alone will not be a panacea over the long term.

Asset management databases are widely employed to help close the maintenance/sustainment gap and to more proactively identify potential risks. However, current data models often use population averages to make predictions, which can overlook individual asset performance over its lifespan. Recent research at the Air Force Institute of Technology investigated using stepwise asset condition forecast models to develop better predictions.

CATALOGUING CHALLENGES

Asset management techniques were first used to manage single infrastructure systems such as roads and pavements, railroads, bridges, airfields, and distribution pipelines. Over the last decade, industry has begun to harvest and catalog facility condition data in more comprehensive systems.

Decision-Making Efficiencies. The establishment of Enterprise Asset Management (EAM) and infrastructure inventorying that considers all portfolio assets enables policymakers and owners to more efficiently prioritize projects and plan long-term capital budgets. Decision-makers use the data to predict asset degradation and expected service life, reduce lifecycle costs, and achieve organizational objectives. Unlike original asset management techniques, EAM requires software, initial asset inventory, and ongoing condition inputs.

Condition Inspection Data. BUILDER SMS and BELCAM are two examples of EAM software. Although they have different technical approaches, they each use time-based condition inspection data to modify population service-life expectations. Unfortunately, this results in a simple scaling of a prediction curve instead of a tailored condition prediction based on assets with similar historical behavior. Current degradation models suggest that infrastructure assets age with time, yet several exogenous factors can cause degradation, including weather and maintenance. Because of this, asset condition is constantly changing. Databases must be updated on a routine basis to maintain accurate asset strategies and management decisions.

ENHANCING PREDICTIONS

Current degradation models suggest that infrastructure assets age with time, yet several exogenous factors cause degradation, including weather and maintenance. Because of this, asset condition is constantly changing.

Data-driven approaches to EAM are increasingly being adopted for use as a management tool. However, using data to power condition prediction models requires overcoming quantity and quality hurdles.

Current models use statistical functions like the Weibull distribution to capture population trends and make condition predictions as a function of age. Using lifecycle expectations of a population to make condition predictions of individual assets can result in large ranges, since individual assets can behave quite differently than a population average. So while viewing individual assets in terms of average population service-life ranges or by life cycles may be the industry standard, it results in seeing individual asset performance as stochastic. This creates large gaps in understanding an asset’s performance over its lifespan, which translates to weaker facility sustainment, restoration, and modernization management planning.

Data-driven forecasts can be developed to fill this gap.

STUDYING ROOFING DATA

Research conducted at the Air Force Institute of Technology suggests stepwise forecasts may outperform current models. Using data for five roof types from 61 unique Air Force locations, three models were developed and tested. Each was then compared with the state-of-the-art degradation model used by BUILDER SMS to determine how each approach improves degradation predictions. Roofing systems were selected over other assets because their shorter average expected lifecycle of 20 to 30 years is best covered by the data. The methods these models employ to convert data into predictions can be used for assets of all BUILDER system types.

There are several shared characteristics between the model types. Search space is a constraint that limits the population data that the model searches through to obtain input variables before applying mathematical computation, and it can be categorized by either age or condition. Different initial search spaces and mathematical computation are used to create the different model types. The stepwise computation is incrementally slope-based, which results in unique asset groupings being used to compute predictions at every time step. Discrete condition and age outcomes are then translated into a complete model by using stepwise computation and intelligent interpolation of predictions. The three new models are the slope, weighted slope, and condition-intelligent weighted slope.

Delta Condition Index is a validation metric that captures the difference between observed and forecast values. All model predictions are compared to observed conditions; a “win” is awarded to the model if its Delta Condition Index is lower than that of the BUILDER SMS prediction. The quantity of possible wins between the model and BUILDER SMS is equal to the service life. The weighted slope model outperformed the BUILDER SMS prediction an average of 92 percent of the time, while the condition-intelligent weighted slope model beat the BUILDER SMS prediction an average of 69 percent of the time. For built-up roofing, the weighted slope and condition-intelligent weighted slope accounted for over six times as much variation in outcomes as BUILDER SMS. These results show the stepwise models produced can outperform current predictions.

IMPROVING FORECASTING

While current service-life models use population averages to make predictions, there is a magnitude of improvement possible from stepwise, data-driven model predictions. The models developed as part of the Air Force research are better tools for translating BUILDER SMS data into enterprise-wide asset management plans. As EAM systems progress, taking time to align or create model types that target decision-maker priorities should be a focus.

Data-driven predictions are only as good as the data they employ. Maintaining the high quality of information entered into the EAM database is critical. As the Department of Defense devotes to EAM to better manage its resources, improving short- and long-term forecast skill is one fruit of that labor.

Kurt Lamm, RA, M.SAME, is Facility Engineer, 88th Civil Engineer Group, Wright-Patterson AFB, Ohio; kurt.lamm@us.af.mil.

Lt. Col. Justin Delorit, Ph.D., P.E., PMP, M.SAME, USAF, is Assistant Professor of Engineering Management, and Lt. Col. Steven Schuldt, Ph.D., P.E., M.SAME, USAF, is Assistant Professor of Engineering Management, Air Force Institute of Technology. They can be reached at justin.delorit@afit.edu; and steven.schuldt.1@us.af.mil.

[This article first published in the September-October 2021 issue of The Military Engineer.]