Certainty in uncertainty: Machine learning can improve engineering decisions
October 28, 2021
October 28, 2021
A look at how data can enhance engineering and science choices for transportation, flooding predictions, and water quality
Year after year, advances in technology revolutionize the world we live in. For engineers, these changes have made our jobs more efficient and improved our decision making. However, there remains an element of uncertainty in most engineering processes. For example, when designing a bridge engineers must account for its fatigue limit—the stress level the structure can endure before failing. To account for this limitation, we overdesign to apply a sufficient safety factor, meaning the fatigue limit will exceed the structure’s usable life.
While uncertainty will always be a factor in engineering and science, advances in machine learning can help us make sense of the unknowns to make more informed decisions.
Machine learning gives us the ability to find data patterns that we might miss. This, in turn, allows us to make better judgements to improve and expedite engineering processes. These data-driven machine-learning approaches provide insights on the uncertainty of the complex functional dynamics found in science and engineering. Simply put, machine learning lets us find certainty in uncertainty.
Uncertainty will always be present in engineering system because of the phenomenon of chaos. Chaos theory teaches that small changes to complex systems can give rise to seemingly unpredictable and random consequences. As such, we are unable to predict the future exactly. For example, systems such as climate have proven incredibly difficult to predict. To the human eye, a set of climate-system data may seem random and unrelated. However, with machine learning, we can find patterns in that data that we otherwise wouldn’t have seen, and by uncovering these unseen patterns, we may predict the uncertainty (i.e., risk) or outcomes with clarity and certainty.
This understanding of uncertainty that machine learning provides can help us make more informed engineering and policy decisions. Here are three examples:
One of the key metrics used in transportation and infrastructure decision making is traffic counts. Knowing how many vehicles are traveling along a particular road is vital to making decisions for road improvements and expansion. Traditionally, these counts were done by people standing at various places along the road making notes on the number and types of vehicles. Now, we can automate this data-collection process using traffic cameras as part of an intelligent data collection system.
In a recent project for the Tennessee Department of Transportation, our team collected more than 10,000 images of cars and trucks to create a model for traffic data collection. We trained the machine learning system to identify the types of vehicles that drive by cameras—cars, trucks, buses, and tractor-trailers—and then store this data in a central location.
This process greatly reduced the funds needed to get an accurate count. A typical manual traffic count averages about $100,000. With machine learning, the cost drops to roughly $14,500.
Machine learning gives us the ability to find data patterns that we might miss.
Traffic counts using machine learning also saves the human traffic counter from the stress and dangers of the job. It protects them from tiredness that may cause errors in the data. It improves safety since they are not physically present on roadways. Machine learning also improves the accuracy of these counts. For a recent study conducted on a traffic ramp in Tennessee, the machine-learning technology had a 99.3% accuracy rate and provided more information than a traffic counter. This type of technology can also collect data over a longer period, providing a larger data set.
Understanding the growth in traffic over time also allows us to predict future traffic increases. Being able to predict the increases in vehicle volume gives policy makers the insights needed to plan for building to projected growth. As a result, infrastructure projects meet needs for a longer period, saving time and effort on future construction.
As climate change continues to bring extreme weather, it is imperative that we identify and protect against the impacts. For example, when Hurricane Harvey hit Texas and Louisiana in 2017, most of the damage occurred to homes outside of the flood plain. Using machine learning, we rapidly can create predictions of potential flooding that allow for better decisions that save lives.
When Tennessee experienced deadly flash floods earlier this year, flooded roadways presented some of the biggest hazards. The floodplain modeling at the time did not show certain highways at risk of flooding and they remained open. When they did flood, it posed a serious risk to motorists. If the flood risk had been predicted accurately, highways could have been closed in advance.
With machine learning technology, we can combine flood map information from the Federal Emergency Management Association and key data on rainfall, land use, and topography to create a probabilistic model for flood events. This allows us to predict flood risk in a matter of minutes and make decisions that will protect communities.
In the wake of the tragic Tennessee floods in August 2021, our team created a machine learning model based on readily available terrain and land-use data. Based on the torrential rainfall amount, our model generated a prediction that mimicked the flooding brought on by the extreme rainfall. This demonstrates that putting extreme rainfall forecasts into the model could provide advance warnings that could save lives.
In addition to predicting flood risks, machine learning can also predict water quality and assist utilities. For example, giardia is a parasite found in soil that can infect both humans and animals. The risk of giardia in community water supplies is tied directly to the frequency of rainfall and runoff.
To assess this risk, we’re able to create a machine learning model that accounts for the relationship between giardia and runoff and derive the distribution from the runoff. The result is an accurate depiction of the giardia risk. Utilities can then use this information to prioritize water withdraws downstream from reservoirs with a lower risk.
Uncertainty is at the fabric of our universe—nature is uncertain. By embracing this truth, we can make the best use of the tools at our disposal to make more informed decisions in science and engineering.
Machine learning has demonstrated its ability to help us make sense of the chaos around us to provide the best, most efficient services and solutions for our communities.