How machine learning can help the military make more accurate predictions of intermittent demand
Military service branches rely on modeling, simulation and training systems to enhance warfighter readiness. Like all systems, they are composed of hardware and software components that experience wear and tear and require upgrades, repairs, or replacement over time.
Analyzing military systems to find the most efficient way to sustain system components in the supply chain is a research area of interest for Dr. Corey Hendricks, vice president and chief engineer at Leidos. We caught up with Hendricks after he presented a technical paper titled “Comparison of Intermittent Demand Forecasting Methods in Predicting the Repair of Simulators based upon System State” at I/ITSEC. Answers have been edited for clarity and length.
What is intermittent demand and why is it a challenge to the Department of Defense?
Intermittent demand occurs when spare parts are needed only occasionally but historically has been a challenging element of the supply chain to forecast, leading to costly overages of excess spare parts. For instance, in a four-year period examined by GAO, Army data reflected an excess of $3.6 billion in spare parts.
What makes spare parts problematic to forecast is threefold. First, the sheer size and complexity of the military’s supply chain can be daunting, with tens of thousands of unique parts for different vehicles, aircraft, and equipment to oversee. Next, predicting which part will fail, especially if using traditional methods with historical averages, is difficult. And finally, integrating advances in technology to address spare part forecasting challenges – like machine learning – into a massive system can be slow and complex.
What did your research in demand forecasting compare, and why is that important?
My research compared demand forecast methodologies using simple exponential smoothing (SES) and a machine learning technique called support vector regression (SVR). SES works well for stable and predictable demand but fails when demand is irregular or intermittent. In contrast, SVR can analyze complex patterns and adapt to unusual demand trends. Comparing these methods ensures that the chosen approach fits the specific needs of the system being forecasted.
How does SVR with machine learning predict intermittent demand better than historical averages would?
Predicting intermittent demand is challenging because there isn’t a consistent pattern to analyze. For example, a rarely used aircraft part might sit in storage for years before suddenly being needed due to an unexpected failure. Traditional forecasting models struggle with these irregularities because they rely on historical averages that don’t capture rare spikes in demand.
Imagine a military truck that only breaks down under extreme weather conditions, like snowstorms. If snowstorms happen infrequently, it’s hard to predict when the truck will need repairs or replacement parts. Using advanced forecasting methods – like machine learning – can help identify subtle patterns, such as seasonal changes, that might influence demand for these rarely needed parts.
Is machine learning as effective at predicting intermittent demand when there are changes in the asset’s operational status?
Machine learning is very effective at predicting intermittent demand even when an asset’s operational status changes, but the model’s accuracy is dependent on how it was trained. Machine learning methods like SVR or Random Forest are great at recognizing patterns in large amounts of data, including changes in how changes in military operational tempo might impact part consumption. For example, a tank that is fully mission capable located in the continental U.S. might need routine maintenance parts, like oil filters, to stay operationally ready. On the other hand, a tank in a deployed location supporting tactical operations that is non-mission capable might need major components, like a new transmission, to become mission ready again. A good machine learning model can analyze the data available from dimensions, (e.g., location, weather, past failures, etc.) to figure out how these changes affect part demands.
Did your research find that machine learning was as accurate as traditional statistical analysis?
Yes, and that’s a big deal because machine learning methods often have the potential to be more flexible and scalable than traditional statistical methods. If a machine learning model achieves similar accuracy, it could eventually outperform the statistical approach as more data is added. Additionally, machine learning models can adapt to complex and changing patterns, which is especially valuable in dynamic environments like military operations.
Machine learning models can adapt to complex and changing patterns, which is especially valuable in dynamic environments like military operations.
Dr. Corey Hendricks
Vice President, Chief Engineer, Commercial & International Sector, Leidos
For example, consider how we can approach forecasting the demand for spare parts for a fleet of drones. A traditional statistical method might only consider basic trends, like monthly usage. A machine learning method, however, could analyze additional factors, such as weather, mission types, and terrain. Achieving similar accuracy initially sets the stage for machine learning to provide more advanced predictions in the long term.
Based on this research, can SVR with machine learning help make military planning more efficient?
SVR can help the military plan more efficiently by providing highly accurate predictions for intermittent demand. SVR excels at finding patterns in complex datasets, even when demand is irregular or hard to predict. For example, if a specific helicopter part is only needed during certain high-stress missions, SVR can identify those conditions and forecast demand accordingly.
In a scenario where an armored vehicle’s cooling system tends to fail during desert deployments. SVR can analyze past maintenance data, temperature records, and mission types to predict when and where the cooling system might fail next. This insight allows logisticians to pre-position parts, reducing downtime and ensuring mission readiness.
Why is this research important not just for the military but for anyone managing supplies?
This research is valuable for anyone managing supplies because the challenges of intermittent demand apply to many industries. For example, hospitals need to predict demand for rarely used medical equipment, like ventilators, while retail businesses must stock seasonal items, like snow shovels. Accurate forecasting helps these organizations avoid stockouts and overstock, saving money and ensuring customer satisfaction.
Consider a company that repairs industrial machinery. If it uses advanced forecasting methods like SVR, it can better predict which replacement parts will be needed during peak seasons, reducing downtime for its clients. Just as the military benefits from improved logistics, so do businesses that rely on efficient and accurate supply chain management.
