News & Events

Big data and automation


May 3, 2019

Moth hoving above flower

To apply machine learning to real-world problems, researchers need to understand how to work with nonlinear and dynamical systems, in chaotic environments, and with ever-changing factors. Some in ME are looking to nature for solutions. Image by Schwoaze from Pixabay.

As society’s challenges grow more complex, engineers are using machine learning (ML) — a process of building models to manage and describe big data sets and automate their analysis — as an increasingly important research tool. Yet to effectively apply ML to real-world problems, researchers need to understand how to work with nonlinear and dynamical systems, in chaotic environments, and with ever-changing factors.

These are areas in which mechanical engineers have decades of experience. Because of their diverse background and deep knowledge of systems, sensors, controls and fluids, mechanical engineers are playing a key role in shaping the future of ML.

Here we highlight ME faculty-led projects demonstrating how ML is being used across the department, from medical device innovation to alternative energy research to improving and advancing manufacturing processes.

ML for improved disease detection, diagnosis and precision treatments

An example of imaging in which prostate glands have been automatically segmented through an ML-based algorithm.

An example of imaging in which prostate glands have been automatically segmented through an ML-based algorithm. Courtesy of Jonathan Liu.

The Molecular Biophotonics Laboratory, directed by ME associate professor Jonathan Liu, is designing advanced microscopy tools to revolutionize the field of pathology. In particular, the lab is developing miniature hand-held microscopes for noninvasive disease detection and surgical guidance, as well as methods for nondestructive slide-free 3D pathology of biopsies and surgical specimens. The 3D pathology technology (being commercialized by LightSpeed Microscopy) generates massive amounts of data, which should enable more accurate diagnostic determinations. However, this information-rich imaging data is time-consuming for pathologists to interpret.

Therefore, Liu’s team is working with collaborators to develop ML algorithms to automate the process of segmenting and quantifying key features from the 3D pathology datasets in order to classify tissues in terms of disease aggressiveness (prognostication) and the likelihood to respond to specific treatments (precision medicine). One of the team’s current clinical studies aims to demonstrate that 3D pathology can help health care professionals better identify and predict the level of risk in prostate cancer patients to guide their treatment. The goal is to prevent the over-treatment of men with low-risk disease while correctly identifying patients who would benefit from treatments such as surgery and radiation therapy, which have the potential to cause serious side effects.

Insect flight and the future of advanced manufacturing and control

Optimal sensor and actuator placement is an important unsolved problem in systems and control theory. Since it’s not feasible to determine the best placement by trying all possible configurations, ME associate professor Steve Brunton and his research collaborators are looking to nature for solutions.

Biological systems provide models for future interpretations and applications of machine learning. For example, moths don’t need to understand physics to know how to fly; they rely on the bio-sensors in their wings and their neural control systems to inform flight actions. In a biological neural system like a moth’s, there may be many complex biosensor patterns, but only a few dominate. Working with researchers in biology, applied mathematics and physics, Brunton is developing ML algorithms based on the moth’s sparse sensing. Translated to inform aircraft design, these algorithms can determine optimal sensor locations and configurations, helping to make aircraft manufacturing more cost-efficient and performant.

A moth hoving above flowers

An airplane flying in the air

Steve Brunton and collaborators have developed an ML algorithm based on the sparse sensing of a hawk moth (left) to inform sensor placement in aircraft design and manufacturing. Images by Gayulo (left) and Korneel Luth (right) from Pixabay.

The researchers have demonstrated this framework on several problems, including to understand insect flight and for control design. They have also worked with The Boeing Company to deploy this technology for advanced aerospace manufacturing.

Monitoring marine life, maximizing turbine power output

Led by ME associate professor Brian Polagye, the Pacific Marine Energy Center is applying ML to two projects. One is to identify marine life in sonar imagery collected by the lab’s Adaptable Monitoring Packages (AMPs), underwater recording devices used to understand how marine life interacts with marine energy converters. AMPs generate so much data that it’s nearly impossible to sort through it all to detect specific moments of marine life interaction. The team has been using ML algorithms for real-time identification of underwater creatures and sea birds. This way, researchers can record and study only those moments, drastically reducing the amount of low-value data to collect and curate.

AMP's field of view and a diving bird

AMP's field of view and a diving seal

Brian Polagye's team has been using ML algorithms for real-time identification of marine life, such as this diving bird (left) and seal (right), as they move through an AMP’s field of view. Courtesy of Emma Cotter.

The team is also using ML to maximize the power output from cross-flow turbines operating in water currents or wind. They use a different type of ML running in parallel with experiments to seek out optimal strategies for accelerating and decelerating a single turbine within each rotation and to identify operational strategies for dense arrays of turbines. Like the AMP project, without ML, progress in this area would require months, possibly years, of experimentation.

Improving supply chain and aircraft inspection processes

ME and ISE assistant professor Ashis Banerjee is applying ML to projects for manufacturing and production process improvement. In supply chain operations, transactions among suppliers and buyers can be inefficient and unreliable due to limited information exchange. For effective production scheduling and inventory management, suppliers and buyers need access to information about part availability and delivery estimates. However, such predictions can be challenging because of lead times, hierarchical dependencies among the suppliers, time-varying production capabilities and capacities, and unexpected changes in materials procurement.

Examples of automatically detected tow boundaries

Examples of automatically detected tow boundaries

Examples of automatically detected tow boundaries — which form the edges of the individual layers of composite materials — during in-process inspection of composite parts. Courtesy of Wei Guo and Agnes Blom-Schieber.

Banerjee is addressing some of these challenges by employing supervised ML on historical supplier-buyer transactional data. Results on actual datasets show substantially lower prediction errors than the current supplier-based or heuristic-based estimates. His research team is also developing an easy-to-use visualization tool to enable real-time use of the learning models.

Banerjee is also working at the Boeing Advanced Research Center to develop deep learning-based methods for visual inspections related to aircraft assembly. Composite materials are increasingly popular in aircraft manufacturing; however, processes to inspect composite parts require manual review and are time-consuming. Since quality control is a necessary part of assembly, Banerjee is developing ML methods to replace manual inspections, improve accuracy and automate the overall process.


Learn more about key ME research areas.